Karsten Röpke Clemens Gühmann (Ed.) and 88 Co-Autors Automotive Data Analytics, Methods, DoE Proceedings of the International Calibration Conference Automotive Data Analytics, Methods, DoE Proceedings of the International Calibration Conference Dr.-Ing. Karsten Röpke, Prof. Dr.-Ing. Clemens Gühmann (Ed.) and 88 Co-Autors With 267 Figures and 33 Tables Bei der Erstellung des Buches wurde mit großer Sorgfalt vorgegangen; trotzdem lassen sich Fehler nie vollständig ausschließen. Verlag und Autoren können für fehlerhafte Angaben und deren Folgen weder eine juristische Verantwortung noch irgendeine Haftung übernehmen. Für Verbesserungsvorschläge und Hinweise auf Fehler sind Verlag und Autoren dankbar. © 2017 by expert verlag, Wankelstr. 13, D -71272 Renningen Tel.: + 49 (0) 71 59 - 92 65 - 0, Fax: + 49 (0) 71 59 - 92 65 - 20 E-Mail: expert@expertverlag.de, Internet: www.expertverlag.de Alle Rechte vorbehalten Printed in Germany Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlags unzulässig und strafbar. Dies gilt insbesondere für Vervielfältigungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. ISBN 978-3-8169-3381-6 Bibliografische Information Der Deutschen Bibliothek Die Deutsche Bibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet über http: / / www.dnb.de abrufbar. Bibliographic Information published by Die Deutsche Bibliothek Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available on the internet at http: / / www.dnb.de Preface The International Calibration Conference will expand on the topics discussed at the conference entitled “DoE in Powertrain Development” with the related areas of “machine learning” and “big data”. Now it its ninth outing, the conference will thus be a forum on which to critically engage with the future challenges of the digital revolution. Real driving emissions (RDE), worldwide harmonized light-duty test procedures (WLTP) and the next round of CO 2 guidelines all demand ongoing technical refinement of the drive train. The combination of changed environmental requirements, stricter limit values and new measurement techniques additionally require changes to existing processes and the development of new methods. To reduce costs, many OEMs are scaling down the size of their engine ranges. A small number of standard engines are then installed in numerous vehicle models with minor hardware modifications. The result is an increased focus on the use of derivatives and the systematic validation of an application In this book, the lectures of the International Calibration Conference - Automotive Data Analytics, Methods, DoE held on May 11 and 12, 2017 in Berlin are contained. I would like to thank all authors for their contributions to this conference. Dr. Karsten Röpke, IAV Contents Preface 1 Big Data I ......................................................................................1 1.1 Machine learning and artificial intelligence for engine calibration .......... 1 Mirko Knaak, Michael Hegmann, Daniel Reppel, Felix Springer 1.2 Big Data and Machine Learning Made Easy ............................................. 11 Dmytro Martynenko, Arvind Hosagrahara, Seth de Land 2 Calibration ..................................................................................18 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions ............................................................................................ 18 Kassem Wehbi, Andreas Wurm, Dieter Bestle, Jörg Beilharz 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions ............................................. 34 Tom Berghmans, Aymeric Rateau, Kotaro Maeda, Thiebault Paquet 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes ................................................................... 52 Stefan Angermaier, Mukunda Gopalakrishnan 3 Design of Experiments I ............................................................79 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine ............................................................................... 79 Yooshin Cho, Donghee Han 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration .................................................................... 89 Jan-Christoph Goos, Matthias Brumm, Sebastian Weber, Frank Kirschbaum, Yagiz Dursun, Thomas Koch 3.3 Dynamic Route-Based Design of Experiments (R-DoE) ........................ 109 Thomas Mayer, Mohamed Ayeb, Ludwig Brabetz Contents 4 Poster session .........................................................................121 4.1 System for Real-time Evaluation of Real Drive Emission (RDE) Data ......................................................... 121 Rajesh Reddy, Sven Meyer 4.2 System optimization for automated calibration of ECU functions ....... 129 André Sell, Frank Gutmann, Tobias Gutmann 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization ........................................................................ 136 Kento Fukuhara, Daniel Rimmelspacher, Wolf Baumann, Karsten Röpke, Yutaka Murata, Yui Nishio, Masato Kikuchi, Yukihisa Yamaya 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU ........................................................ 163 Simon Wunderlin, Antoine Hurstel, Ernst Kloppenburg 5 Dynamic....................................................................................173 5.1 Dynamic Modelling for Gasoline Direct Injection Engines ................... 173 Taro Shishido, Jing He, Masataka Kaihatsu, Carsten Haukap, Thomas Dreher, Michael Hegmann 5.2 Excitation Signal Design for Nonlinear Dynamic Systems ................... 191 Tim Oliver Heinz, Mark Schillinger, Benjamin Hartmann, Oliver Nelles 6 Design of Experiments II .........................................................209 6.1 Application of a DoE based robust design process chain for system simulation of engine systems .............................................. 209 Matthias Hekrenz, Daniel Reppel, Michael Hegmann, Jochen Broz, Christoph Brands 6.2 Application of Emulator Models in Hybrid Vehicle Development ......... 222 Justin Seabrook, Pascal Revereault, Mike Preston, Markus Grahn, Johan Bringhed, Björn Lundberg 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines ................................. 233 Owen Parry, Julian Dizy, Vivian Page, Amit Bhave, David Ooi Contents 7 Big Data II .................................................................................256 7.1 The Connected Car and its new possibilities in ECU calibration ......... 256 Lars Hagen, Andreas Walter, Michael Bargende 7.2 Processing vehicle-related measurement data ...................................... 270 Leon Evgeni Kusnezow, Ortwin Escher 8 Data Analytics ..........................................................................283 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system ....................... 283 Matthias Kahl, Andreas Kroll, Robert Kästner, Manfried Sofsky 8.2 Data Plausibility at the Engine Test Bench - How important is the Human Factor in the Process? ............................ 300 Hanno Ihme-Schramm, Alexandra Schramm 9 Design of Experiments III ........................................................309 9.1 Non-Convex Hulls for Engineering Applications ................................... 309 Nico Didcock, Nikolaus Keuth, Andreas Rainer 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning ....................... 321 Mark Schillinger, Kadir Mourat, Benjamin Hartmann, Carola Eckstein, Martin Jacob, Ernst Kloppenburg, Oliver Nelles 9.3 Model-based iterative DoE in highly constrained spaces ..................... 336 Stefan Scheidel, Marie-Sophie Vogels 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds ......... 349 Yagiz Dursun, Sebastian Weber, Richard Jakobi, Frank Kirschbaum, Stephan Rinderknecht The Authors.........................................................................................366 1 Big Data I 1.1 Machine learning and artificial intelligence for engine calibration Mirko Knaak, Michael Hegmann, Daniel Reppel, Felix Springer Abstract Model-based calibration (MBC) has been revolutionizing the way for calibrating combustion engines for nearly two decades. In particular, the utilization of DoE has contributed to the dramatic reduction of time per calibration item that made the complex engines of today possible. To reach the goal, a high number of sophisticated statistical methods have been invented or adapted - resulting in a decent statistical knowledge that is present in virtually all OEM, suppliers and engineering companies. Having said that, the knowledge has been applied to regression mainly in the past. By increasing computational power and an abundance of cheap memory, there is no need to keep this limitation anymore. On the contrary, using the knowledge for regression, clustering and classification could spark the next level of development: Could statistical learning be the “next DoE”? At the same time, neural networks have achieved great success in solving many problems of machine learning. After decades of stagnation, successful applications under the name of deep learning have been published frequently in the last months and years. Methods of artificial intelligence such knowledge graphs are seeing a renaissance, too. What does this mean for engine calibration? Can they change the entire calibration process - as evangelists of AI foresee for fields like health care? After giving the general introduction, the key note will show some specific examples that make the direction visible. Special emphasis will be placed on the challenges that emerge by the introduction of the RDE test procedure. A possible solution to this issue is the usage of synthetic RDE trips in combination with methods of machine learning. 1 Introduction The availability of affordable large memory has the potential to change the calibration process drastically. We are seeing a steep increase in the data volume generated and processed in the calibration process. In particular, the new RDE (real driving emissions) legislation leads to a significant amount of data as many driving situation need to be tested. If a typical RDE drive consists of two hours driving, and one measures 1000 channels ten times per se- 1 1.1 Machine learning and artificial intelligence for engine calibration cond, roughly 72 million data points are generated. Many RDE trips will be driven to ensure the compliance, resulting in megabytes and terabytes of data. But not only RDE drives have the potential of generating a large amount of data as can be seen in Fig. 1. Durability runs and long test trips are another good source for large data files. As an example, OBD validation needs many hours of testing that equals to a large data volume. Other data sources are chassis and engine test benches, as well as simulation results generated by numerical models. Even though these sources are not that large in volume, they can increase the variations in the data. Simulation results may surprise here, however thanks to the increase of computational power models they can be executed in a much faster time. The increased data volume - big data in calibration - is both a boon and dame at once. At the first glance big data causes new issues. The data need to be stored and processed in larger data bases consuming infrastructure investments and dedication of people to topics such as parallel processing and data storage. Terms such as Hadoop or spark are becoming household names. Moreover, when compiling data from several sources into one data lake, precautions in the calibration process are needed to avoid inconsistency in the data. Some examples are: How to deal with different software versions? How to deal with different naming conventions? This becomes particularly important if data from several projects are being utilized. How to deal with changing reference values? For some of the issues, smart algorithms occasionally using A.I. can solve these inconsistencies. Some of the methods described later in detail are being used in IAV to increase the quality of the data lake in IAV. Generally, some even small changes in the calibration guidelines can contribute significantly to better data. As a rule of thumb on can state: Collect and store as all (or as many as possible) channels even if you do not know the reason in advance. There will be one! Figure 1: Big data based calibration process Big Data Machine Learning Roller dynamometer Durability run and test trips Engine test bench Virtual Calibration RDE-test Calibration optimization Worst case cycle reproduction Guided Processing of critical driving situations Manage and process Search and evaluate 10110011110001100 11001111011010110 01101010110101111 10100100110010111 00101010110111000 00001111110011100 01101001111111100 10111111001011010 Remote Calibration + Learn from data Recommendation Knowledge base 2 1.1 Machine learning and artificial intelligence for engine calibration After having seen the “price tag”, the benefits of big data are also enormous. The past conferences of this series were focused on gaining knowledge from minimal amount of data. Data were generated from highly expensive tests on engine and chassis dynamometers and the community put a strong effort into methods such as DoE to gain sufficiently accurate models with the least amount of measurements. While this remains valid in many development steps where data generation remains expensive, we are seeing a new field where the opposite is true: We can make models from many more data than coefficients to be estimated. These benefits can be even amplified when taking data from several sources or even projects into account. Virtually speaking: Combining the data sumps of each division into a single data lake. Why not use experiments made for transmission calibration for emission calibration of the engine? While this task is cumbersome for humans, machine learning algorithms can easily utilize additional data. The following section (Sec. 2) shows briefly the calibration process using large amounts of data and machine learning. Sec. 3 provides some back ground about machine learning and results in specific examples for RDE calibration. Sec. 4 shows the application of the method for the validation of RDE compliance. 2 Big data based calibration process After gaining data from various sources on left hand side of Fig. 1, they need to be stored into one system. As joint folders in Windows are not the matter of choice for big data, IAV is using a server cluster AMEDA to manage and process measurement files. Figure 2: Calibration steps for variant calibration Black: Mainly human and test facilities Light grey: Mainly algorithms o.k. Experimental evaluation Measurements Assessment of data Validation measurements Counter measure n.g. Initial investigation o.k. Event search/ visualization Machine learning Artificial intelligence Assessment of data o.k. Validation measurement Counter measure n.g. o.k. Proposal of reasons Event search/ visualization Experimental evaluation o.k. Validation Measurements with counter measure o.k. Proposal of counter measure Experimental evaluation Classical process 3 1.1 Machine learning and artificial intelligence for engine calibration The initial assessment is done with the IAV product Mara that utilized algorithms similar to business intelligence to evaluate the measurements. Mara is rule based and has already implemented many typical evaluations in a calibration process. The results are visualized in the typical charts used in calibration as developed by calibration engineers. The next step is using machine learning to evaluate measurements without any given rule. The machine learning finds structure in the data and relationships between channels as well as between channels and calibration issues. This will result in a selection of potential root causes for a calibration issue. Together with employing a calibration knowledge base, the tool is even capable of recommending calibration changes. Finally, the calibration changes need to be verified by repeating the worst case situation for the calibration issues. The machine learning can also suggest worst situations. The benefits of the described processes are illustrated in Fig. 2 with the following example. A calibration task for the calibration of an engine variant may have the following steps: After conducting the measurements, the calibration engineer evaluates them by searching for potential issues. If some points are not good (n.g.), he/ she searches for some engine parameter to be calibrated differently to avoid the calibration issues. Usually, additional expensive measurement are necessary. After the calibration changes, validation measurements ensure the positive effect, the lack of negative side effects and document the results. In the first step, Mara helps to find the calibration issues much easier. Then, the potential root causes are identified on base of the existing data by applying various machine learning algorithms. The engineer can now focus on the search for counter measures. Fig. 3 shows the monetary benefits of this procedure. When applying existing engineering knowledge from unstructured data such as calibration guidelines, results of earlier projects and other, a recommendation of calibration changes is possible. The main role of the calibration engineer in such a calibration step is the final selection of the calibration changes. Figure 3: Top: Typical calibration process. Bottom: Calibration with machine learning Assessment of data Initial investigation Validation and test measurements Counter measure (new calibration) Release measurement Iterations as needed Resources/ Cost Steps/ time Preprocessing by Mara New calibration by ML tool and engineer Release measurement Resources/ Cost Steps/ time 4 1.1 Machine learning and artificial intelligence for engine calibration Fig. 4 shows some examples of using machine learning for engine calibration and development. For OBD and engine calibration can utilize the process depicted above. In addition of OBD development, machine learning can also contribute to improve diagnosis algorithms. One example field is the configuration data management as the number calibration variants is increasing. For one engine, we differentiate between markets, production dates, software version, variants and many more. If we vary parameters for these categories, they creating a huge space of variation when combining into function, control units and even sub-systems of the vehicle. Figure 4: Examples for machine learning in engine development Predictive diagnosis and maintenance is another huge field of application. Here, machine learning is used to develop advanced self-diagnosis functions that allow for warnings even before the actual fault is happening. Field investigations often have two goals. One is finding the root causes for frequent problems, the other is identifying all vehicles that need to be updated. Machine learning can contribute as the data source is not only huge but also heterogeneous. We need to combine information about garage inspections, vehicle configuration data, environmental data, vehicle history, fault protocols, fault environmental data as well driving patterns or load profiles among many others into the data lake. 3 Machine learning in engine calibration 3.1 Overview about machine learning Machine learning is used in the paper even though similar concepts exist under the name data mining or knowledge discovery in data bases, often collectively referred to as data science. The reason for this choice is the strong root in statistical theory (machine learning is often called as statistical learning). In fact, many algorithms used here have a strong foundation in statistics as most of the regression algorithms do. The second main field is pattern recognition that contributed some of the most valuable algorithms such as support vector machine and other kernel based methods. The last field that constitutes machine learning are neural networks whereby deep recurrent networks have gained a decent popularity recently. Continuous improvement Configuration data x 1 x 2 State of fault Issue of warning Predictive diagnostics/ maintenance OBD development Engine variant calibration Field investigations 5 1.1 Machine learning and artificial intelligence for engine calibration Algorithms for machine learning can be characterized by way of learning, illustrated for the calibration processes described in Fig. 5. The first category is supervised learning. For supervised learning a target function is known. Depending on the type of target function, continuous or categorical, the applied methods are regression or classification, respectively. Regression is very familiar for engine calibration as the statistical modelling used for DoE models is inherently a regression task. In contrast to the earlier usage of regression, the relevant input channels are not known in advance and potentially huge. Model selection is not only choosing the right non-linear kernel for a set of input parameters, but automatically the right input parameters. Methods for sub-set selection are used here. Figure 5: Categories of machine learning for engine calibration Classification is used for finding the relationships to categorical events such as jerks, misfire und others. When using classification, a feature vector is generated by nonlinear transformations of the input channels, done manually with expert knowledge. This state-of-art concept tries to avoid this step by defining the feature vector automatically. In contrast to supervised learning, unsupervised learning does not require a target value. The data are being structured by finding the largest similarity of certain groups of measurement, clusters. Those clusters are often used for selecting driving situations of high similarity that it is are likely to have the same properties. These clusters can later be used for identifying worst case driving maneuvers, as shown in Sec. 4. The remaining categories for learning are reinforcement learning and active learning. Reinforcement learning is often used during the “run-time” to increase the performance of the algorithm. If the algorithm is suggesting a calibration chance that is rejected by the calibration engineer, it will use this information to adjust its configuration. For active learning the algorithm estimates the accuracy of its recommendation and requests additional information if the estimated accuracy is low. Feature vector Signal processing • Regression analysis • Classification • Best subset selection • Deep neural networks Root causes/ Calibration changes • Clustering • Association rule learning • Deep neural networks • Genetic algorithms Significant pattern Unsupervised learning Supervised learning Event information, Target value Discussion with engine expert Reinforcement learning 6 1.1 Machine learning and artificial intelligence for engine calibration 3.2 Example for machine learning for engine calibration The methods for machine learning can be illustrated with the following use case. In OBD calibration, the self-diagnosis of a sensor is supposed to show the defined level when no violation is visible. In the use case, the sensor had shown a distribution as depicted in Fig. 7. The input value was 100 and the show tolerance band was allowed for the sensor. In a large number of test situation under normal calibration conditions the diagram clearly shows that the sensor is not working properly in many situations. Figure 6: Self-diagnosis of an OBD sensor If we are going back to the calibration process in Fig. 2, an engineer would search for main influences by executing designated experiments. That means he/ she would identify potential input parameters in which the sensor shows a malfunction. Then he/ she would modify theses parameters, e.g. environmental conditions, to see whether the frequency of malfunction increased. In a consecutive execution of this procedure, one can find the root cause. When applying the machine learning tool, huge amount of previous measurements can be utilized. As seen in Fig. 5, the first step is the formation of a feature vector. In classical pattern recognition, this was a manual task using a deep level of expert knowledge. The expert transformed some measurement signals with signal processing into features by using gradients, time-frequency transformation, bandpass filtering and many others. Both the selection of input signal and the transformation require knowledge about the system and signal processing. In our use case, this would mean the engineer should already guess the potential root cause of the malfunction. In the approach suggested here, all measurement data are taken into account. That means that we start with a vector length of 600-800, equaling the number of labels recorded in an INCA experiment. The first step is an automatic reduction of redundancy. This step is needed as exactly the same information is occasionally written in different ECU labels. In addition, some other channels may contain the same information only differing by a linear transformation. An example is the value of exhaust flow as the raw sensor value in V, the processed value in mg/ s. Other values may contain almost the same information. The scientific concept used here is the reduction of joint information in the signals. tolerance band Self-diagnosis values of a sensor Frequency 7 1.1 Machine learning and artificial intelligence for engine calibration After reducing the feature vector by redundancy reduction, it is being enlarged by the most common signal processing approaches being already mentioned. Following this step, a redundancy reduction is executed again as the signal processing (including nonlinear parts) may induce new redundancies. Following the process in Fig. 5, the next step is classification since we now have categorical target signal. The target signal is in the category “on” when the sensor output is outside the tolerance band in Fig. 6. For the actual classification, several classifiers where trained, including random forests, support vectors machines, and others. As usual a part of data was taken as a validation data. Compared to other situations of modelling, this step is not critical in the advent of big data. There are enough measurements. We could reach cross validation rates in the high 90% rates. However, the cross validation rate is of limited importance as our final goal is not the prediction of a malfunction. Instead, the structure of the classifier delivers the desired information: Which components of the feature vector has it selected? In which combinations of these components the classifier are important for predicting a malfunction? As in earlier discussions of model-based calibration, a similar concept appears: One is using a model to avoid new dedicated measurements. The main difference is that the availability of big data allow an automatic identification of the model structure. Concluding the use case, by using the procedure explained above we could define an importance measure for all 800 input signals. Fig. 7 clearly indicates that the sensor has a high sensitivity to ambient pressure. Figure 7: Categories of machine learning for engine calibration 4 RDE complying calibration using synthetic cycles The RDE test procedure adds a strong element of stochasticity to conventional repeatable test bench based measurements. Even for the same route, vehicle and driver, a larger spread of emissions appears. A single worst case cycle or route may therefore not be enough to cover the large variety of valid speed traces and ambient conditions. A possible solution to this issue is the usage of synthetic RDE trips in combination with a model based approach. 8 1.1 Machine learning and artificial intelligence for engine calibration In a first step, velocity profiles from real road measurements are analyzed by methods of machine learning. The first step is a clustering to dissect the road measurements into their constituent parts (maneuvers) and stored into a data base. Thereby, each maneuver is classified as acceleration, constant or deceleration and marked by its initial and terminal velocity. In addition, information on driver, vehicle, ambient conditions, engine state and emissions are stored. In a second step, a transition matrix is created that describes the probability of changing from a current velocity and acceleration state to another state. This transition matrix reflects the routes, traffic situation and driving style of the underlying data. Based on the transition matrix, the basic driving maneuvers are then statistically reassembled assuming an underlying Markov chain model for the velocity profile. The underlying transition probabilities can be modified to reflect different driving styles and ambient conditions. In a last step, critical cycles are selected from all synthetic trips compliant with RDE requirements. Possible criterions for the selection of critical trips may be the catalyst temperature, raw emissions or the lambda value before catalyst. These quantities can be taken directly from the measurements or predicted by means of a simple data driven models. The selected critical synthetic cycles can be afterwards used as an input for a model-based calibration or as test sequences for a chassis dynamometer. Fig. 8 illustrates the process chain. Figure 8: Categories of machine learning for engine calibration 9 1.1 Machine learning and artificial intelligence for engine calibration Literature [1] Hastie, Tibshirani, Friedman, The elements of statistical learning, Springer, 2nd ed., 2009 [2] Röpke, K. (ed.), DoE in Engine Development, Expert Verlag (Proceedings 2001-183) [3] Meintrup, D., Schäffler, S., Stochastik - Theorie und Anwendungen, Springer Verlag, Berlin, Heidelberg 2005. [4] Rasmussen, C. E., Williams, C. K. I. Gaussian Processes for Machine Learning, The MIT Press, Cambridge, Massachusetts; London, England, 2006. [5] Bernhard Schölkopf, Alex Smola: Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond (Adaptive Computation and Machine Learning), MIT Press, Cambridge, MA, 2002 [6] Oliver Nelles: Nonlinear System identification, Springer, Berlin, 2001 [7] Lecture material of MDT: “Technical diagnosis and pattern recognition“ [8] Mitchell, Tom M.: Machine Learning. The Mc-Graw-Hill Companies, Inc., 1997 [9] Finn V. Jensen: Bayesian Networks and Decision Graphs. Springer-Verlag, New York 2001 10 1.2 Big Data and Machine Learning Made Easy Dmytro Martynenko, Arvind Hosagrahara, Seth de Land Abstract MATLAB is designed for engineers, by engineers. MathWorks is committed to making MATLAB the easiest and most productive tool for working with big data and with data analytics applications, because making sense of large data sets is becoming an important part of solving complex design challenges. In this paper you will learn how automotive companies use MATLAB to turn large volumes of complex data into actionable information for vehicle design and decisionmaking processes. We look at recent technology for big data and machine learning, and see how it can be applied to traditional problems faced by automotive engineers. 1 Introduction For many years, engineers across the automotive industry have been analyzing data in MATLAB to improve performance and efficiency. Automotive engineers have learned to apply MATLAB to analyze data sets collected from the engine, transmission, chassis, and virtually all parts of the vehicle. But what happens when the data gathered mounts into the terabytes or larger, as it often does for vehicle data at the fleet level? Can MATLAB scale up and meet increasingly demanding fleet data analytics requirements? Is it possible to build big engineering data analytics algorithms and deploy them into a production environment for data scientists, product engineers, and business analysts? The short answer to both questions is “Yes.” This paper explains how MATLAB can be used to extract key insights from big engineering data for improving engine and vehicle design. Specifically, it describes a realworld system constructed to demonstrate big engineering data analytics with MATLAB. We built this system to collect our own data, so that we could further streamline and refine techniques for applying MATLAB analytics on big data. This system includes infrastructure for gathering actual fleet data, as well as a web-based fleet data analytics application built and deployed with MATLAB. The paper concludes with examples of valuable insights that can be gained from such a setup. 11 1.2 Big Data and Machine Learning Made Easy 2 What Is Big Engineering Data and Why Is It Challenging to Derive Insights from It? Across industries, big data is commonly described in terms of the three Vs: volume, variety, and velocity. (A fourth V, veracity, is frequently included and will be covered later in this paper.) In engineering applications, we use the term “big engineering data” to characterize data that exhibits these qualities while presenting a subtly different set of problems. The data is typically better structured and more varied, and it has the potential to grow much faster than big data in other industries. For example, when dealing with fleet data: • Volume refers to the scale of the data. For automotive OEMs, it is not uncommon to work with data sets of up to 20 TB at once and collect more than 30 GB of data per car per day. At these rates, the data can easily grow to sizes in the order of petabytes and sizes that cannot be analyzed at once in memory. Even the relatively modest demonstration system we constructed was capable of collecting 25 MB of sensor data per day, per car, and it produced almost 2 GB of data over several months with just a few drivers. These figures do not include video data, commonly used in advanced driver assistance systems, which can amass at a rate of several gigabytes per hour. • Variety reflects the recognition among analysts that including data from many sources can lead to more valuable and unexpected insights. Today’s vehicles are equipped with dozens of sensors - and instrumented fleet vehicles have many more - all generating a variety of signals including speed, fuel consumption, and temperature. Different vehicles produce different types of data. For example, hybrid and electric vehicles may report battery charge information rather than fuel flow. In our system, data from the vehicle is combined and timealigned with data from other sensors, including GPS devices. Lastly, all of this time-series data can be complemented by simulation, audio, video, and CAN log data, among other types. • Velocity, in big data terms, describes the speed at which data is accumulated. The MathWorks demonstration system collected data from several sensors every second. In practice, automotive sensors with sampling rates of 10 milliseconds generate a hundred data points per second. Data streaming in from multiple sensors across numerous vehicles quickly mounts into the gigabyte and terabyte range. Organizations seeking to extract value from big engineering data must have systems in place to handle the sheer volume, variety, and velocity of that data. These systems must include flexible, scalable data analytics algorithms deployed in a production environment so data scientists, product engineers, and business analysts can derive insights from the data. 12 1.2 Big Data and Machine Learning Made Easy 3 Building a Big Engineering Data Analytics System Using MATLAB The MathWorks Consulting Services team, who have worked with many customers to build data analytics applications, wanted a framework for gathering and analyzing big engineering data that we could experiment with, explore, study, and share. There are many architecture options for MATLAB data analytics applications. The system we built (Figure 1) begins with an adapter that plugs into a vehicle’s ODB2 port and transmits automotive data wirelessly via Bluetooth to a smartphone. The phone relays the data to a cloud application running on a Linux-Apache-MongoDB- Rails (LAMR) stack on Amazon EC2. In this system, the vehicles act as edge nodes in an Internet of Things (IoT) solution, generating data that is gathered and sent to a data aggregator. The system uses the Java-based Hadoop software framework to simplify the distributed storage and processing of big engineering data across computer clusters. Figure 1: Hardware and software infrastructure for collecting fleet data and performing data analytics with MATLAB and Hadoop. With this data collection infrastructure in place, we employed a MATLAB based workflow that starts with standard data analysis techniques on a single computer and culminates with the deployment of a production web application. We used MATLAB on the desktop to explore the data we gathered and to develop, test, and visualize ideas for processing it. Building upon our work on the desktop, we developed and deployed the Fleet Data Analysis web application, which combines packaged MATLAB analytics with Hadoop and other web technologies to enable the visualization, optimization, and analysis of vehicle fleet performance characteristics (Figure 2). 13 1.2 Big Data and Machine Learning Made Easy Figure 2: The MATLAB based Fleet Data Analytics web application. 3.1 MapReduce with MATLAB and Hadoop Among the challenges of processing big engineering data sets is that they are often too large to fit into available memory and take too long to analyze on a single processor. The MapReduce programming technique addresses these challenges by processing data in small chunks that individually fit into memory to produce intermediate results, and then aggregating these intermediate results to produce a final result. Hadoop MapReduce is a popular implementation that works with the Hadoop Distributed File System (HDFS), and MATLAB provides its own implementation of the MapReduce technique. Analysts can interactively develop and debug algorithms with MapReduce in MATLAB on the desktop. They then have two options for running their algorithms - without any modifications to the code - using Hadoop. They can use MATLAB Distributed Computing Server™ to execute MATLAB MapReduce based algorithms within Hadoop MapReduce for data that is stored and managed on Hadoop. Alternatively, they can use MATLAB Compiler™ to create applications based upon MATLAB MapReduce for deployment within production instances of Hadoop. Traditional data analysis techniques typically involve moving data into a computational environment to be analyzed. This approach - bringing the data to the analytics - may not be feasible for big engineering data. Support for Hadoop in MATLAB enables teams to bring analytics to the data, and leave big data where it is stored. 14 1.2 Big Data and Machine Learning Made Easy 4 Gaining Insights from Big Engineering Data Analytics Organizations seeking to use big data to drive real improvements in engine and vehicle design should enable anyone in the organization to extract insights from the data that has been accumulated and stored. This section describes three examples of insights gained through data analytics with MATLAB. 4.1 Understanding Real-World Brake Specific Fuel Consumption Automotive engineers know that even a well-designed engine cannot operate at maximum efficiency when attached to a transmission that is not tuned to operate with that engine. Brake-specific fuel consumption (BSFC) measures, which reflect not only energy efficiency but also how well an engine is tuned, are used to optimize gearboxes, shift schedules, and other transmission parameters. Although BSFC is often addressed as a simulation-optimization problem, a data-driven approach provides a different perspective, giving a view of fuel economy across real-world driving patterns. We used MATLAB data analytics to study BSFC for the approximately 25 vehicles in the sample fleet, operated by real drivers going about their daily commute and other trips. In MATLAB, we plotted torque as a function of engine speed and added hyperbolas of constant power to create the BSFC map shown in Figure 3. Figure 3: BSFC map created in MATLAB based on real-world fleet data. The real-world data points used to create the BSFC map carry much more information than just engine speed and torque. Extra data may include latitude, longitude, and altitude, so we can determine where any vehicle in the fleet is while its engine is operating in its efficiency sweet spot. Additional data may also include noise and vibration metrics, which can yield insights into how measures to improve fuel efficiency may run up against noise, vibration, and harshness (NVH) constraints. 15 1.2 Big Data and Machine Learning Made Easy 4.2 Assessing Infrastructure Changes Traffic engineers can use MATLAB data analytics with the same fleet data to identify problematic traffic patterns and evaluate ways to resolve them. For example, we developed analytics to identify areas in which fleet vehicles were consuming the most fuel. We integrated the MATLAB analytics with third-party business intelligence analytics software to rapidly construct an interactive web dashboard (Figure 4). User interactions with the dashboard trigger calls to MATLAB code to recompute results, which are then used to update the dashboard’s visual components. Figure 4: Interactive web dashboard created and powered by MATLAB and business intelligence analytics software. This dashboard highlighted several intersections at which lengthy idle times led to particularly inefficient fuel consumption. Exploring the data further in MATLAB, we computed how much fuel was consumed by a vehicle waiting for nearly 90 seconds at one of the most problematic intersections. Based on the result - 0.035 gallons per vehicle - we estimated that replacing the intersection with a roundabout would save more than 120 gallons of fuel per day for local commuters and reduce CO 2 emissions 16 1.2 Big Data and Machine Learning Made Easy by about 4.5 million pounds per year. For an automotive OEM, such an analysis could be used by engineers to optimize car features, such as a start-stop system, using real-life traffic patterns. 4.3 Understanding Veracity Volume, variety, and velocity are central to any discussion of big data. Organizations have recognized that understanding the veracity of their data is vital to avoid drawing the wrong conclusions from data analytics. Consider, for example, an automotive engineering team that is creating prognostics models by training them with fleet data. If the data used to train the models is erroneous, then business decisions based on those models will be unfounded. Similarly, a team training networks for machine learning applications would face a comparable challenge, because a classifier model that is trained on faulty data will not perform as anticipated. 5 Conclusion Across a wide range of automotive use cases, MATLAB provides an open and extensible data analytics platform that supports a proven workflow for the rapid development, refinement, and deployment of data analytics applications. MathWorks is continuing to invest in big data analytics and build upon existing MATLAB support for datastores, MapReduce, Hadoop, Tall Arrays, Spark integration and related technologies. For those who seek a quick start and expert guidance, MathWorks Consulting Services is available to help organizations improve their ability to extract valuable insights from big engineering analytics. Literature and Resources [1] Bain & Company - Big Data revolutioniert die Automobilindustrie [2] Jonny Andersson, Scania - Model-Based Approach to Resource-Efficient Object Fusion for an Autonomous Braking System (MathWorks Automotive Conference 2015) https: / / www.mathworks.com/ company/ events/ conferences/ automotiveconference-stuttgart/ 2015.html [3] MathWorks Fleet Analysis Projekt http: / / ec2-54-187-132-60.us-west- 2.compute.amazonaws.com [4] Dmytro Martynenko, MathWorks - Big Engineering Data Analytics with MATLAB (MathWorks Automotive Conference 2015) https: / / www.mathworks.com/ company/ events/ conferences/ automotiveconference-stuttgart/ 2015.html © 2017 The MathWorks, Inc. MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See mathworks.com/ trademarks for a list of additional trademarks. Other product or brand names may be trademarks or registered trademarks of their respective holders 17 2 Calibration 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Kassem Wehbi, Andreas Wurm, Dieter Bestle, Jörg Beilharz Abstract For several years, the automotive industry has been aiming to reduce cost and effort in vehicle calibration by automating the transmission calibration process. However, the quality of results may be deteriorated by the presence of environmental, operating or manufacturing uncertainties. Already in the design process of automatic transmissions, therefore, robustness should be considered as important design goal to generally assure high quality and reliability of gear shifting and launch control. Besides the implementation of adaptive algorithms for clutch control, this may also be achieved by using robust design strategies. By such a procedure, mean values for best shift or launch quality, and variances for highest robustness against change of system parameters are minimized simultaneously. In this paper, such an approach is developed for transmission calibration of vehicle launch control, where the aim is to achieve a fast and smooth clutch engagement at the same time. Optimization results are discussed in different criterion spaces and the achieved improvements confirm the validity of the proposed procedure. Kurzfassung Der Trend in der Automobilindustrie geht seit mehreren Jahren hin zur Automatisierung der Getriebekalibration, um die Kosten und den Aufwand zu reduzieren. Die Qualität der Ergebnisse kann jedoch durch Umwelt-, Betriebs- oder Fertigungsunsicherheiten negativ beeinflusst werden. Bereits im Entwicklungsprozess von automatischen Getrieben sollte daher Robustheit als wichtiges Entwurfsziel Berücksichtigung finden, um eine hohe Qualität und Zuverlässigkeit des Schaltablaufs und des Anfahrvorgangs unter allen Umständen zu gewährleisten. Neben der Implementierung von Adaptionsalgorithmen für die Kupplungssteuerung kann dies auch durch die Verwendung von robusten Optimierungsstrategien erreicht werden. Bei einer solchen Strategie werden gleichzeitig Mittelwerte für die beste Qualität und Varianzen für die höchste Robustheit gegen Änderung der Systemparameter minimiert. In dieser Arbeit wird ein solcher Ansatz für die Kalibrierung der Kupplungssteuerung entwickelt, mit dem sowohl ein schnelles als auch komfortables Anfahrverhalten erreicht werden kann. Die Optimierungsergebnisse werden in verschiedenen Kriterienräumen diskutiert, wobei die erzielten Verbesserungen die Eignung des vorgeschlagenen Verfahrens bestätigen. 18 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions 1 Introduction The demand for vehicles with automatic transmissions is continuously increasing. Together with the growing complexity and diversity of modern powertrains, the calibration effort is increasing dramatically, which is why human search for optimal parameter sets is not sufficient anymore and automation of the calibration process for automatic transmissions becomes increasingly important [1]. This applies especially to automated engagement of friction clutches during gear shift or vehicle launch. It has been shown that such calibration automation can be achieved by applying a multi-objective optimization strategy in combination with simulation models, software-inthe-loop or hardware-in-the-loop test rigs, or even in the car, e.g. [1], [2], [3]. In terms of assessing shift quality, the optimization strategy is based on two design objectives assessing comfort and sportiness by shift time and longitudinal acceleration of the vehicle [4]. Correlation studies in [5] revealed that the typically used launch criteria in industry may also be reduced to just these two kind of objectives assessing the launch quality regarding sportiness and comfort. High sportiness of a launch event may be represented by low launch time measured from the beginning till clutch synchronization as ! : ( ) ( ) ( ) 0 sync sync e sync c sync t t t t ω ω ω Δ = − = (1) where e ω and c ω are the engine and clutch speed, respectively. For discomfort evaluation of shift events, deviations of the actual acceleration ( ) a t from an ideal acceleration transition ( ) ideal a t have been proposed [4]: ( ) ( ) 2 0 1 1 sync t ideal sync sync D a t a t dt t t = − ∫ . (2) In [6] it was shown that it also applies to launch events where, however, the ideal acceleration transition ( ) ideal a t has to be defined differently. It is an optimal solution with minimum jerk and jerk derivative, but maximum travel distance after synchronization time sync t . Simultaneous minimization of both objectives (1) and (2) typically results in optimal tradeoffs between these conflicting launch quality criteria. This was demonstrated in [6] for well-defined operation conditions and model parameters. In reality, however, the calibration of clutch control parameters is faced with uncertain transmission parameters like uncertain friction coefficients and engine torque as well as uncertain vehicle load due to changing number of passengers, slope of the road, etc. Thus, the real launch process will usually differ from the deterministically optimized clutch control and will not behave optimally at all. The key to tackle this issue is to consider robustness as inherent design goal of calibration already in the early development phase of automatic transmissions, similar to [2], in order to assure high launch quality and reliability also in the presence of uncertainties. In this paper, a robust design strategy for launch control will be developed and applied to a proper driveline model, which will be discussed in the next section. By using the described objectification and simulation, effects of uncertainties on the launch behavior will be discussed in Section 3. Then special effort will be made to focus on uncertain quantities with highest influence on the launch behavior, where samples 19 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions are created according to empirical assumptions. Based on these, an optimization problem will be formulated in Section 5 and solved in Section 6 with the multiobjective genetic optimization algorithm NSGA-II. The resulting performance benefit will be discussed in detail, where final optimization results show improved launch control performance compared to a given reference design. 2 Driveline Modeling for Automatic Calibration The intended design strategy is partly based on a driveline model created in Velodyn , which is a toolbox based on Simulink® and combines the advantages of signal-floworiented with object-oriented modeling. The tool allows modularization of models by providing standardized interfaces using a signal and control data bus as illustrated in Figure 1. Over the years an extensive library with electrical, mechanical and hydraulic components was developed, which allows to build up complex models rather quickly. A graphical user interface enables an intelligent signal management and advanced handling of sub-models and their parameter sets. Such models may be used for concept studies of electrified and conventional drivetrains and layout of components, and further as plant models for software development, for software testing or for initial calibration. By compilation of the Velodyn models and parallelization of the simulation and evaluation, the simulation time can be reduced significantly and thus allows the application of effective genetic algorithms for design optimization. Due to all these benefits, Velodyn models are used in the entire vehicle development process and frontloading strategy at the IAV company. Figure 1: Velodyn key elements Figure 2 gives an overview of the most important drive train components for driveability investigations. Depending on the kind of application, various parametric and nonparametric approaches with different degree-of-fidelity are available. Conceptually, there is a distinction of physical models with regard to dynamics as rigid, dynamic or highly dynamic model approaches where an increase in dynamics is accompanied by an increase in the depth of detail. For the launch control investigations in this paper, 20 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions model approaches marked by arrows are used, where the general request for the application here is to keep them “as simple as possible and as detailed as necessary”. Figure 2: Component models for driveability investigations of the drive train (components marked by arrows are used for launch control investigations) 3 Effect of Uncertainties on Launch Behavior The calibration of clutch control parameters has to account for uncertain factors which may influence the clutch engagement during operation. According to [7], deviations of engine and clutch torque from their nominal values may become noticeable during clutch engagement. Subsequently, not only launch criteria (1), (2) may get worse, but also critical engagement maneuvers at vehicle takeoff such as engine flare or engine stalling may occur. Further, uncertain environmental conditions result in major challenges for robust system behavior. For instance, when launching uphill with high slope, synchronization time will be extended associated with longer slipping and thus more dissipation energy. In this section, different uncertainties and their effect on the driveline model will be discussed in more detail. 3.1 Friction clutch and hydraulics Friction clutches play a major role in power transmission from the engine to the wheel, especially when launching a vehicle. During the engagement phase, clutch control with the clutch capacity cap T as control input should be performed accurately in order to meet requirements such as high launch comfort [8]. For the slipping clutch, the torque capacity cap N m T F r z μ = (3) results from Coulomb’s friction law, where m r is the mean friction radius, μ is the dynamic friction coefficient and z the number of contact surfaces. The normal force N F is realized by a hydraulic actuator and depends on the hydraulic system. Investigations performed in [2] have shown that friction behavior represented by cap m r z τ μ = , the stiffness s k of the return spring acting on the clutch piston as well as oil delay oil t Engine TCU Transmission Vehicle stationary characteristics transfer function oscillations engine process strategy clutch control point masses transfer function tire model chassis kinematics Gear set Converter Clutch Shaft Lots Hydraulic rigid coupling power distributed kinematics elastokinematics stationary characteristics Voigt-Kelvin transfer function flow model invariable structure variable variable/ stabilized friction model rigid coupling transfer function elastic/ phys. joint shaft rigid coupling transfer function elastic/ phys. joint shaft stationary characteristics transfer function incompressible compressible dynamic highly dynamic rigid/ stationary physical approach: mathematical / empirical approach: analytics / heuristics Depth of detail 21 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions of the hydraulics are important sources of uncertainties to be considered during optimization. In the following these uncertainties are added to the nominal values nom cap τ , nom s k , nom oil t as ( ) 1 (1 ) , , cap s nom cap cap nom s s k nom oil oil oil k k t t t τ τ τ ε ε = + = + = + Δ (4) where , cap s k τ ε ε and oil t Δ represent stochastic deviations of the actual values , s cap k τ and oil t from their nominal values, respectively. 3.2 Engine map The internal-combustion engine generates mechanical power through complex chemical and mechanical processes. In the control of such engines, a torque map is stored in the engine control unit (ECU) as a function of engine speed e ω and accelerator position α . This determines the fuel-quantity and the air mass to provide a requested engine torque. While such a system may operate well for a specific set of engine operating conditions, variations in engine wear, fuel characteristics and injection dynamics may interfere with the proper functioning of the system [9]. In addition, the mass flow rate of inlet air is significantly degraded at high altitude due to reduction of air pressure, which has a significant effect on the engine torque [10]. Sensing of the actual engine torque usually requires additional sensor hardware, such as rotary transformers, which is not readily adaptable to existing engine or transmission calibrations [9]. Since the engine torque, however, has a major impact on the drive shaft torque [7], discrepancies in the engine torque have to be considered in robust design optimization. The modeling of engine discrepancies due to unknown uncertain parameters may be reflected by random deviations e T ε from the nominal engine torque nom e T . These uncertainties, however, may highly depend on the engine speed e ω as discussed in [10]. Here we restrict them to specific uncertainties ( ) e T i ε Ω at some pre-defined break points: {100,150, 200, 250, 300, 350, 400} e i rad s ω = Ω ∈ (5) and we use a linear interpolation in between for each subinterval 1 [ , ] i i i I + = Ω Ω resulting in actual engine torque ( ) ( ) , ( ) 1 ( ) e nom e e e e T e T T ω ω ε α ω = + (6) where 1 1 2 ( ) ( ) ( ) ( ) ( ) e e e T e T i T i G x G x ε ω ε ε + = Ω + Ω , 1 e i i i x ω + − Ω = Ω − Ω for e i I ω ∈ (7) and 1 ( ) 1 G x x = − , 2 ( ) G x x = are linear Hermite basis functions, see Figure 3. 22 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions 100 150 200 250 300 350 400 50 100 150 200 [ / ] e rad s ω [ ] e T Nm nom e T ( ) e T e ε ω ± Figure 3: Variation range of engine torque for throttle value 0.6 α = 3.3 Drive shaft In the used driveline model, the drive shaft between gearbox and vehicle axle is assumed to have no inertia, but some elasticity and damping, where this also accounts for compliance of engine mounting and vehicle suspension. Due to the fact that strong non-linearity exists in the spring and damper components, uncertainties in the associated linear models may reflect this which then may effect the vibration level of the powertrain. Therefore, parametric uncertainties in the stiffness d k and damping coefficient d d may be taken into account [11] during launch resulting in actual values (1 ) d nom d d k k k ε = + and (1 ) d nom d d d d d ε = + (8) deviating from their nominal values nom d k and nom d d by uncertain percentages d k ε and d d ε , respectively. 3.4 Gravitational force According to [12], vehicle load has significant influence on vehicle operation and impacts perceived driveability as well. This may be represented by the down-hill force sin sl F mg θ = acting on a vehicle with mass m on a slope with angle θ . The calibration of the clutch control parameters is typically independent of the road slope, which may be related to sensor costs [13]. For this reason, it has to be handled as an uncertain parameter during optimization, where uncertainty is due to slope percentage tan pl λ θ = . (9) 3.5 Sensitivity investigation Detailed statistical analysis of all effects mentioned above would require a lot of measurements. In practice, however, only a limited number of sample observations is carried out with specially designed vehicles to map individual effects of the above mentioned uncertainties. In this paper, therefore, only crude assumptions are made about uncertainties and their distributions based on the experience of some engineers employed by the industrial partner. Especially, uniform distributions are assumed between the following pre-defined bounds: 23 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions 0 0.05 10% 10% 10% 10% 30% 30% 30% 30% 10% ( ) 10% 1(1)7 15% 15%. , , , , , , , cap s d d e oil k k d T i pl s t s i τ ε ε ε ε ε λ Δ − − − − − Ω = − ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ ≤ (10) In order to see the effects of uncertainties mentioned above, the launch process is simulated for a random sample of size 50 N = being generated by optimal Latinhypercube sampling (oLHS). Figure 4 shows the simulation results as bundles of trajectories (gray) resulting from parameter variations, where only specific groups of parameters are varied according to Eq. (10). In Figure 4a we observe large variations in the acceleration and in both engine and clutch speed due to , cap s k τ ε ε and oil t Δ with obvious impact on the synchronization time. On the other hand, variations in the engine torque (Figure 4b) have only a visual effect on the engine speed during slip, whereas the whole powertrain is influenced only after synchronization. Variations in the elasticity and damping components obviously have only minor effect in Figure 4c, where the synchronization time is almost not affected. Finally, both acceleration and clutch speed behavior are extremely influenced by variations in the road slope during and after the synchronization, Figure 4d. Figure 4: Variation of launch events (gray lines) compared to nominal design (black lines) resulting from uncertain parameters for a) clutch friction and hydraulics, b) engine torque, c) drive shaft and d) road slope 24 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions 4 Sensitivity Analysis for Reduction of Uncertain Quantities Taking into account all uncertainties from the previous section (11) during robust multi-criterion optimization would result in a high number of simulations being required for each design change combined with each uncertain value. Therefore, the number of uncertain parameters has to be reduced by restricting them to those with highest influence on the launch behavior assessed by criteria (1) and (2). Figure 4 has revealed that there is some potential due to highly different impact. In order to enable a proper selection, a non-linear sensitivity technique based on regression trees may be applied. According to [14], regression trees are robust, flexible, and can deal with highly nonlinear relationships between design parameters and model output. The main idea behind this technique is that differences between data points of a given model and predictions of a specific regression model obtained from a subset of design parameters can be interpreted as capability of this subset to describe the major behavior of the underlying model. Regarding our application, if all uncertain parameters with the strongest influence on the launch behavior are used and the regression model provides a sufficient level of flexibility, then the regression errors will be small. On the other hand, if some important uncertainties are missed, the regression errors will become large. Therefore, the goal here is to find a subset of uncertain parameters with least regression errors of the resulting regression model, which then reproduces the behavior of the original model sufficiently well. In order to build-up regression trees, firstly an uniform oLHS sampling with 2000 N = design evaluations is performed for a specific throttle position ( 0.6 α = ) of the driveline model. These samples are generated by varying all uncertain parameters (11) within the bounds given in Eq. (10). Subsequently, each design evaluation k results in an output ( ) k y corresponding either to sync t or D . The high number of design evaluations is needed to detect non-linear perturbation effects of these uncertainties from their nominal values in this 13-dimensional space. In order to assess the quality of the regression, the Pearson coefficient of correlation ˆ ρ is used, see [14]. A value close to one indicates a very high correlation between original data and regression outputs, and thus a reliable regression model. The sorting procedure of uncertainties with decreasing influence is shown in Figure 5 which is applied two times, i.e., independently to both criteria sync t and D . It starts with the complete index set J referring to all uncertain parameters (11) and an empty index set I of already sorted uncertainties. In the inner loop, each uncertain parameter with index j J ∈ is combined with the other elements found in I creating a new uncertain subset x . Then, a regression tree ˆ( ) y x is built-up from the sample points ( , ) 1... , , k k y k N = x and the correlation coefficient ˆ j ρ is determined. At the end of this loop, the uncertain parameter * j , which due to its presence enables the highest coefficient of correlation ˆ j ρ , is added to I and removed from J . Because the number of variables in I is increased in every iteration, the quality of the regression model is improved and thus the correlation coefficient is increased. This procedure is continued until all parameters are removed from J . Subsequently, the set I includes all 25 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Figure 5: Sorting procedure of uncertainties with decreasing influence indices referring to uncertainties (11) in a sorted manner with decreasing influence on the launch behavior. The results of this sorting with respect to both launch criteria are shown in Figure 6 as correlation coefficients ˆ sync t ρ and ˆ D ρ , respectively. The uncertain parameters with the highest influence on the launch behavior are located at low i -values. As can be seen, sequentially adding the next important uncertain parameter always increases the correlation coefficient of the regression model, and simultaneously increases the level of variability in the analysis output that can be explained. In addition, removing parameters from the right with low additional change in ˆ ρ does not effect the correlation coefficient and thus the regression model too much. Obviously, the uncertain parameters with strongest influence on both objectives are variations in the slope value pl λ , changes cap τ ε in the clutch friction and variations in the engine map for 150 200 , e ω = and 250 / rad s . Variations d k ε in the stiffness coefficient may impact the discomfort criterion, but not so much the sportiness. All other uncertainties have only minor influence. Figure 6: Progress of the regression quality with increasing number i of uncertain parameters for sportiness sync t and discomfort D , {1,..., } I J M = ∅ = : j J ∀ ∈ J = ∅ yes 1 2 ! [ , ...], { } ˆ ˆ ( ) : ( ) ˆ ˆ ˆ ( ) regression correlation i i n k j k u u i I j y y y y ρ ρ = = ∈ ∪ → → = x x x * ˆ arg max j j j ρ = * : { }, I I j = ∪ no * : \ { } J J j = index set of with decreasing influence i I u uncertanties with decreasing influence 0.6 1 ˆ sync t ρ ˆ D ρ pl λ cap τ ε e T ε (200) d k ε e T ε (150) e T ε (250) i ˆ ρ 26 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Consequently, the first five most important uncertain parameters from each criterion study are selected and summarized in the reduced uncertainty vector where * {150, 200, 250} i rad s Ω ∈ . (12) Only these selected uncertain parameters will be considered in the following optimization problem, which are less than half of the original uncertainties and thus reduce the evaluation effort. The eliminated uncertain parameters with only weak influence on the launch behavior are kept constant at their nominal values. 5 Formulation of a Robust Design Problem For design evaluation by objectives (1) and (2), simulations of the launch behavior are performed by the driveline model introduced in Section 2. Minimization is achieved by varying some design parameters summarized in design vector where the control law of the clutch is parametrized similarly to real TCU’s as clutch torque rates depending on actual clutch torque, see [6]. However, here the criterion values (1) and (2) not only depend on the deterministic design variables p , but also on the uncertainties (12) resulting in probabilistic criterion values. In order to improve launch quality in the presence of uncertainties, an adapted optimization problem has to be formulated and solved. The idea of robust design is to simultaneously improve the mean design quality and reduce the sensitivity of designs on random perturbations of the uncertainties (12), and thus to reduce the randomness of the launch criteria. Typically the mean launch quality is estimated from evaluations of each launch criterion for a given sample ( ) i u , 1... , , i N = by the arithmetic mean ( ) 1 1 ( ) N i y i y N μ = = ∑ p (13) where ( ) i y represents either ( ) ( ; ) i sync t p u or ( ) ( ; ) i D p u . However, as can be seen from the sensitivity analysis, the objective values are highly influenced by variations in the road slope, and therefore large values would dominate low values. In order to tackle this issue and to see improvements more clearly, we firstly simulate the launch process with reference design values found from manual calibration for different slope values [ 15%,15%] pl λ ∈ − resulting in reference objective values , pl ref sync t λ and pl ref D λ . Then we use ratios instead of absolute values for each design evaluation as where ( ) i pl λ represents the slope value being the last component of the current perturbation vector ( ) i u according to Eq. (12). Applying this to Eq. (13) results in average relative improvements of the synchronization time and discomfort as ( ) ( ) ( ) , ( ; ) ( ; ) i pl i sync i sync ref sync t t t λ = p u p u % and ( ) ( ) ( ) ( ; ) ( ; ) i pl i i ref D D D λ = p u p u % (14) 27 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions ( ) 1 1 ( ; ) ( ) sync N i t sync i t N μ = = ∑ p p u % % and ( ) 1 1 ( ) ( ; ) N i D i D N μ = = ∑ p p u % % , (15) both to be minimized for high launch quality. Robustness of a design is assessed by the variances of criteria (14). High deviations of synchronization time and discomfort from corresponding mean values (15) due to uncertainties (12) have to be kept as low as possible. This can be achieved by minimizing the variances 2 sync t σ and 2 D σ for sportiness and discomfort, respectively, which may be estimated from simulations with the perturbations ( ) i u as ( ) 2 2 ( ) 1 ( ) 1 ( ; ) ) 1 ( sync sync N i t sync t i t N σ μ = = − − ∑ p p u p % % % , ( ) 2 2 ( ) 1 1 ( ( ; ( ) ) 1 ) N i D i D D N σ μ = = − − ∑ p p u p % % % . (16) The search for optimal results has to account for reasonable constraints. In practice, synchronization should occur while the first gear is still engaged. This means that long synchronization times must be avoided by restricting them with an upper bound u sync t . Moreover, very short synchronization time at low engine speed may deteriorate the sportiness behavior of the vehicle during takeoff and the engine has the risk of stall [5]. Therefore, lower bounds l sync t and l e ω may be introduced: 1 2 ( ) ( ) ( ) 3 ( ) ( ; ) ( ; ) ( , ; ) ( ) : ; . l l i sync sync i i e e i u sync sync t t t t t h h h ω ω ⎡ ⎤ ⎡ ⎤ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ = = ≤ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎣ ⎦ ⎣ − − ⎦ − p u h p u p u 0 p u (17) Summarizing statistical estimates (15), (16) in a vector 2 2 ( ) [ , , , ] sync sync T t D t D μ μ σ σ = f p % % % % and adding constraints (17) to be fulfilled for all perturbed system parameters results in a constrained multi-objective optimization problem (18) Since the multi-objective genetic optimization algorithm NSGA- II proposed by Deb [15] shall be used for solving the problem, it needs to be transformed into an unconstrained, but bounded design problem where ( ) ( )[10,10] ( ) [ , ] for regular , else. T T γ γ ⎧ + = ⎨ ∞ ∞ ⎩ f p p p f p (19) The penalty function 3 0 2 ( ) ( ) 1 1 ( ) ( , ) ( , ) N i i j j i j h h γ = = ⎛ ⎞ = + ⎜ ⎟ ⎝ ⎠ ∑∑ p u p u p (20) 28 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions combines step functions and quadratic penalties in case of constraint violations expressed by the Föppl notation 0 0 0. for , for n n χ χ χ χ ⎧⎪ = ⎨ > ≤ ⎪⎩ (21) Irregular designs p are defined as parameter sets where simulation of any launch event fails for at least one sample point ( ) i u . Such a case cannot be handled by the penalty function which is why for such parameter sets both criteria are penalized with very poor criterion values (infinity stands for a sufficiently high number). 6 Optimization Results The solution of optimization problem (19) is obtained within about 106 hours using an Intel i5 core@2.9 GHz computer, where a population of 80 individuals is evaluated by NSGA-II over 80 generations. For every design evaluation, 200 N = simulations have to be performed for uncertain parameters (12), where the ( ) i u are generated only once in the initialization step and used identically later on. At each of these support points, objectives are evaluated according to criteria definition (1), (2) and related to references according to Eq. (14). Inequality constraints (17) are checked for all ( ) i u separately and statistical analysis is performed according to Eqns. (15) and (16). Figure 7 shows the Pareto-optimal designs obtained from this procedure in the fourdimensional criterion space. The two axes are associated with the mean performances normalized with respect to a reference setup R manually calibrated by an application expert, i.e., 1 2 , sync sync t D R R t D f f μ μ μ μ = = % % % % % % . (22) Figure 7: Pareto-optimal solutions in the four-dimensional criterion space 1 min f ← % 2 min f ← % min ← 3 : inner circle f % 4 : outer circle f % R T C S 29 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions The grey shades of the inner and outer circles indicate increasing normalized robustness from white to black according to 2 2 3 4 2, 2, , sync sync t D R R D t f f σ σ σ σ = = % % % % % % . (23) The spread of Pareto-optimal solutions in the criterion space demonstrates that the chosen strategy has enough design freedom. Obviously, a lot of designs dominate the reference design R in terms of mean performance (22) and robustness (23). Furthermore, the most robust designs (indicated by dark colors) with respect to both discomfort and sportiness are located at the lower right of the figure, which reveals a clear conflict with the mean sportiness performance 1 f % . Let us have a look on three optimal robust designs with different characteristics and compare them with the reference setup R. We choose design C with optimal mean comfort, design S with high mean sportiness, and an optimal tradeoff T in between. Besides, all selected designs have very good robustness characteristics with respect to comfort and sportiness. The corresponding normalized design criterion values are given in Table 1. Table 1: Normalized criteria for selected designs shown in Figure 7 design 1 f % 2 f % 3 f % 4 f % R 1 1 1 1 C 1.15 0.49 0.33 0.16 S 0.77 1.09 0.33 0.62 T 0.97 0.66 0.32 0.30 Figure 8 shows the relative frequency-of-occurrence plots determined for an uniformly distributed sample (oLHS) of size 2000 N = . Obviously, all selected designs dominate the reference setup R in terms of robustness (23). Especially the spread of the probability distribution of the reference setup with respect to D % is wider than for all other designs indicating worse robustness. The plots located at the lower right of sync t H in Figure 8a are related to unsuccessful speed synchronizations, which is why poor values are assigned to the sportiness criterion. As expected, design C shows the best mean performance of D % (or 2 f % ) and the spread of the probability distribution is also fine, Figure 8b. Design S shows the best mean sportiness sync t % (or 1 f % ) and a rather good robustness regarding synchronization times as shown in Figure 8c. On the other hand, this design clearly shows less comfort robustness compared to T and C, but much better than the reference setup. Only the tradeoff design T dominates the reference design regarding all criteria, i.e. (22) and (23), which becomes visible in Figure 8d. 30 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Figure 8: Probability distribution of a) reference setup R compared to b) design C , c) design S and d) design T In order to see the advantages of the proposed robust design approach more clearly, let us also have a look on the time behavior of these selected designs. Mean comfort improvements of design C are associated with the fact, that the uncertainty band shows much lower acceleration oscillations in comparison to R, Figure 9c. Moreover, the band of uncertainties in the acceleration as well as in the engine speed behavior of this optimal design is much more narrow indicating higher robustness against parameter changes. The same phenomena can be observed when regarding design S in Figure 9d. In this figure, design S also reveals shorter synchronization times than the reference setup at higher acceleration values, indicating higher mean sportiness. On the other hand, design R shows some samples with very short synchronization times (Figure 9a) which are, however, associated with the risk of engine stall worsening sportiness and excessive acceleration oscillations worsening comfort. This confirms how important it was to take constraint 1 h in Eq. (17) into account during optimization. Finally, the unsuccessful speed synchronizations already seen in Figure 8a for the reference setup can also be observed in the time domain (see 3 samples in Figure 9b). They are avoided in the selected optimal designs, since this is considered as constraint 3 h in Eq. (17) for each design evaluation during optimization. sync t H sync t H min [ ] sync t s ← % 3 min [ / ] D m s ← % 0 2.3 1 0.1 2.5 1 0 2.3 1 0.1 2.5 1 0 2.3 1 0.1 2.5 1 0 2.3 1 0.1 2.5 1 3 min [ / ] D m s ← % 3 min [ / ] D m s ← % 3 min [ / ] D m s ← % min [ ] sync t s ← % min [ ] sync t s ← % min [ ] sync t s ← % 1 R f % 1 R f % 1 R f % 1 R f % 2 R f % 2 R f % 2 R f % 2 R f % 2 C f % 2 C f % 2 S f % 2 S f % 2 T f % 2 T f % sync t H sync t H a) b) c) d) 31 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Figure 9: Comparison of a,b) reference setup R with c) design C and d) design S in the time domain 7 Conclusions The paper proposes a new robust design strategy for automatic calibration of automatic transmissions during launch, which besides finding optimal tradeoffs between sportiness and comfort also accounts for robustness w.r.t. manufacturing and operational uncertainties. It is based on a clear definition of such uncertain parameters, where those with highest influence on the launch behavior are identified by means of a non-linear sensitivity investigation based on regression trees. The solution of the design problem is performed with a multi-objective genetic algorithm, where the deterministic problem formulation is transformed into a robust four-criterion problem; mean values and variances of both launch criteria are minimized simultaneously for achieving best mean quality and highest robustness against change of uncertain parameters at the same time. The found Pareto-optimal designs allow a tradeoff between good mean performance and high robustness. A comparison with a reference setup found by calibration experts shows improved performance and validates the eligibility of the proposed design strategy. The use of multi-objective optimization for finding multiple optimal solutions allows to finally select an appropriate launch control characteristic according to individual customer needs and manufacturer philosophies. 32 2.1 Automated Calibration Using Simulation and Robust Design Optimization Improving Shift and Launch Quality of Automatic Transmissions Literature [1] S. Kahlbau, Mehrkriterielle Optimierung des Schaltablaufs von Automatikgetrieben , Dissertation, Brandenburg University of Technology. Aachen: Shaker, 2013. [2] A. Wurm and D. Bestle, "Robust Design Optimization for Improving Automotive Shift Quality", Optimization and Engineering, vol. 17, pp. 421-436, 2016. [3] M. Bachinger, B. Knauder and M. Stolz, "Automotive Vehicle Launch Optimization Based on Differential Evolution Approach for Increased Driveability", Proc. of 3rd International Conference on Engineering Optimization, Rio de Janeiro , 2012. [4] S. Kahlbau and D. Bestle, "Optimal Shift Control for Automatic Transmission", J. Mechanics Based Design of Structures and Machines, vol. 41, pp. 259-273, 2013. [5] K. Wehbi, D. Bestle and S. Kahlbau, "Multi-Objective Launch Control Optimization for Automatic Clutch Engagement", Proc. of 6th IC-EPSMSO, Athen, pp. 272-279, 2015. [6] K. Wehbi, D. Bestle and J. Beilharz, "Automatic Calibration Process for Optimal Control of Clutch Engagement During Launch", J. of Mechanics Based Design of Structures and Machines, vol. 1 , pp. 1-16, 2016. [7] K. Van Berkel, T. Hofman, A. Serrarens and M. Steinbuch, "Fast and Smooth Clutch Engagement Control for Dual-Clutch Transmissions", Control Engineering Practice, vol. 22, pp. 57-68, 2014. [8] Z. Sun and G. Zhu, Design and Control of Automotive Propulsion Systems , Boca Raton: CRC Press Inc., 2014. [9] I. Kay and . R. Lehrach, Torque Sensor for Internal Combustion Engine. US Patent 4.691.288 , 1987. [10] M. Ghazikhani, M. Ebrahim Feyz and A. Sabazade, "Effects of Altitude on the Soot Emission and Fuel Consumption of a Light-Duty Diesel Engine", Transport, vol. 28, pp. 130-139, 2013. [11] N. Singh, H. Chhabra and K. Bhangal, "Robust Control of Vehicle Active Suspension System", International Journal of Control and Automation, vol. 9, pp. 149-160, 2016. [12] A. Chunodkar, T. Proctor, V. Sujan, K. Follen, G. Salemme, P. Vajapeyazula and A. Wu, Vehicle Controls Including Dynamic Vehicle Mass and Road Grade Estimation during Vehicle Operation . US Patent 14/ 840.861 , 2016. [13] Y. Sebsadji, S. Glaser, S. Mammar and J. Dakhlallah, "Road Slope and Vehicle Dynamics Estimation", Proc. of American Control Conference, Seattle, 2008. [14] M. Lockan, P. Amtsfeld, D. Bestle and M. Meyer, "Non-Linear Sensitivity Analysis Based on Regression Trees with Application to 3D Aerodynamic Optimization of Gas Turbine Blades", Proc. of International Conference on Engineering and Applied Sciences Optimization , Kos Island, 2014. [15] K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, "A Fast Elitist Multi-objective Genetic Algorithm: NSGA-II", J. of IEEE Transactions on Evolutionary Computation, vol. 6, pp. 182-197, 2002. 33 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Tom Berghmans, Aymeric Rateau, Kotaro Maeda, Thiebault Paquet . Abstract Due to the change of regulation and the strong focus on reducing emission in realworld-driving, simulating the exhaust emissions is a must during diesel engine development. The requirements for such simulation strongly increased. The simulation should be fast (focussing on real time), accurate and valid for a wide area of driving conditions and ambient conditions like temperature and altitude. Toyota developed a base diesel engine simulation platform as part of the V-cycle development. In predevelopment, this platform is used to optimize the hardware and calibration of the engine. It combines a driver model, 1D engine model, vehicle model, a predictive combustion model and ECU model. For simulation of RDE (Real Driving Emissions), this platform was adapted with the goal to increase simulation speed, keep high accuracy and ability to simulate both engine-out and tailpipe emissions and the related RDE conformity factor. In order to do so, a software tool was made to create realistic driving speed & shift profiles for any road, which can be defined by selecting a route from a geolocalisation webservice. This road will be “driven” offline by a driver model which has been trained for the specific vehicle which is the simulation target. The result is a realistic vehicle speed and shifting profile for the route which is the input for the simulation platform and the online driver model. In order to improve the simulation speed, the engine-out emissions are now predicted by a statistical model in the form of dynamic data-based (Gaussian process) model. By optimising the model training, very accurate emission reproduction of RDE cycles is possible at very high calculation speed. Prerequisite here is to have an engine or vehicle available with a reasonable maturity of hardware and software. In order to keep modularity, the 1D engine simulation was kept and converted to a mean-value model to increase the simulation speed and provide physical inputs to the mathematical emission model. Finally the exhaust gas condition predicted by the engine and emission model is connected to an aftertreatment model, including DOC, DPF and SCR catalysts, with integrated logic for Urea dosing. This toolchain gives the possibility to validate the engine-out and tailpipe emissions for any route which can be selected from a map, and develop optimisation for the engine hardware, software and aftertreatment control system, reducing the evaluation duration significantly. 34 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions 1 Introduction The upcoming RDE (Real Driving Emission) is a further milestone regarding the emission legislation for passenger cars, ensuring the vehicles are clean in a wide variation of engine operation conditions, environmental conditions and vehicle setups. For vehicle development, it is very time-intensive to optimize and validate the emission performance in these wide conditions, and especially their different possible combinations, through measuring the emissions with a real vehicle. In this background, the capability to develop models representing accurately real driving emissions, which can predict the emissions in a very wide range and combinations of conditions, is a must. Based on the original Toyota Virtual Diesel Engine (VDE) platform [1], modifications were implemented to ensure fast and accurate simulation of typical real driving cycles of 2 hours. A Gaussian Process statistical dynamic databased model is implemented to predict the engine-out emissions. In this paper the focus will be mainly on the NOx emissions. This toolchain is used by development engineers to validate and optimise the emission performance for real driving. 2 Toyota VDE (Virtual Diesel Engine) concept The typical V-process based on Model Based Design has been introduced as the key approach throughout the vehicle and engine development. The former style of separating the design and evaluation process caused many iterative actions and a relatively long development duration. In the V-process, the design and evaluation process interact through simulation and give the possibility to early detect design failures and develop improvements. For predevelopment, Toyota developed the Virtual Diesel Engine as displayed in Fig.1, which has the capability to simulate the engine’s emissions on a single PC, by combining a 1D GT-Power® engine model (for calculation of gas flow), a 0D cycle predictive in-house combustion model (UniDES-D) [2] , an ECU SIL (Software-In-the- Loop) model and a driver and vehicle model. For the validation of real driving emissions later in the development process, further improvements regarding simulation speed and emission prediction in highly dynamic engine operation are implemented and described in this paper. 35 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Figure 1: Toyota Virtual Diesel Engine (VDE) Toolchain 3 VDE application for real-world driving validation 3.1 Targets & Requirements The main targets of the RDE toolchain are to validate and optimise the engine-out and tailpipe emissions in the wide boundary conditions of RDE, once an engine prototype and software with reasonable maturity is available: In this way the number of actual on-road validation tests can be reduced. The targets are as follows: • Validate RDE emission in a wide range of driving patterns & environmental conditions • Optimise the aftertreatment control system to reach tailpipe emissions target • Identify improvement points in the engine-out emission calibration • Identify worst case conditions for focussing the development In order to reach these requirements, following capabilities are necessary: • Ability to simulate any road conditions in Europe • Ability to simulate the effect of environmental boundaries (ambient temperature, altitude, slopes) within the extended RDE boundaries • Ability to simulate both engine-out and tailpipe emissions • Fast calculation speed, target = real time simulation • Potential to evaluate air path hardware & control modifications • Fast model training data collection in case of significant hardware or calibration updates 36 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions 3.2 RDE toolchain description Based on Toyota’s original VDE toolchain used in predevelopment, a number of modifications were implemented in order to develop an RDE development tool to quickly and accurately simulate real world driving emissions. The toolchain is composed of various connected submodels in order to predict the tailpipe emissions for any selected road, as shown in Fig. 2 Figure 2: RDE toolchain from road to tailpipe 3.2.1 Driving road preparation with off-line driver Target is here to prepare a realistic target vehicle speed and gear shifting profile for RDE driving with different levels of driver aggressiveness as option. In this offline preparation process, the user can select a driving route from a geolocalisation database by selecting a series of GPS waypoints; A trained offline driver will drive the road, in order to obtain a realistic speed & shift profile based on the road attributes extracted from the geolocalisation database. This offline driver model uses machine learning techniques to imitate human driving behaviour according to the driving context. For training this driver, actual on-road data is used. The methodology will be further described in Chapter 6 of this paper. 3.2.2 Driver (online) The function of the developed online driver module is to follow the driving profile (target vehicle speed, shift) that was prepared by the offline driver. Feedback is provided by the vehicle & engine model. 3.2.3 Engine & vehicle model The 1D airpath GT-Power® model is kept in the engine model to keep flexibility and give the possibility to evaluate the impact of airpath modifications on the real driving 37 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions emission simulation. This model is calibrated to accurately provide the gas dynamics and content to the emission model. A mean-value cylinder model is used based on maps of the IMEP, volumetric efficiency and exhaust temperature, which use simple correction maps for environmental conditions. Special effort was dedicated to good matching of the exhaust temperature, since this is critical input for the turbocharger operation and finally for the aftertreatment system activation and efficiency. 3.2.4 Engine control unit model (SIL) The full engine control for the airpath and injection control is implemented and connected to the driver, engine/ vehicle and catalyst models. The inputs of the ECU model are the same as in the actual vehicle, and are now provided by the 1D engine model and driver model. The result is the setpoint (including feedback) for the different engine actuators and the injection system. In order to improve simulation speed, the calculation time step was optimised without impact on dynamic behaviour or accuracy. 3.2.5 Engine-out emission model The predictive UniDES-D diesel combustion & emission model was replaced by a data-based emission model in order to improve the simulation speed and keep high accuracy for simulating engine-out conditions in a wide area of driving conditions and environmental conditions. It is coupled with the 1D engine model, which has good potential to reach high accuracy [3]. The dynamic data-based model can be trained quickly and has very accurate behaviour especially in highly dynamic driving conditions, which is a must for RDE simulation. In order to train the dynamic data-based model, actual measured data is necessary. Although it is a purely mathematical model, the goal was to increase it’s understanding of physical phenomena. Chapter 4 will focus on the training and accuracy of the emission model. 3.2.6 Aftertreatment model To simulate the tailpipe emissions, a 1D chemical aftertreatment model was connected to the dynamic data-based engine-out emission model and the 1D engine model (for exhaust massflow and temperature). Included in the model are oxidation catalyst, diesel particulate filter and selective catalytic reduction (SCR) catalysts. The aftertreatment model is calibrated and validated trough synthetic gas bench and engine tests in a wide temperature and massflow range, required for RDE simulation. The dynamic temperature and emission conversion/ storage is calculated along each catalyst of the aftertreatment system. It is connected to the aftertreatment control logic, which controls the quantity of urea used for converting NOx by the SCR catalysts. 3.3 Data-based emission model selection For the emission model, both steady-state data-based models and dynamic databased models were considered. In this case, the focus is on the RDE validation of the powertrain, rather than the base calibration itself. Dynamic-data based models [4] 38 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions have the advantage that the training data collection is a matter of hours, while measuring the training data of a DoE based steady model typically would take weeks. This means that although dynamic models have less prediction possibility for calibration changes than typical steady DoE based models (considering it is difficult to measure a dynamic DoE for calibration parameters), it is possible to quickly retrain the dynamic model and have updates in the tool chain. This training is possible both on an engine bench and roller bench. Table 1 shows a comparison of both approaches. Table 1: Comparison of steady (DoE) and dynamic data-based models Steady data-based Dynamic data-based Training type Steady DoE Dynamic cycle/ Dynamic DoE Training data collection Weeks Hours Facility Engine bench Engine bench & Roller bench Issues Engine & facility drift Facility responsiveness Model inputs Typically 8-12: Engine operation Airpath & Injection system Environmental params. Typically 2-6: Engine operation Airpath Environmental. params. Running speed Far below real time Far below real time Main usage Base calibration Validation & dynamic calibration 4 Dynamic data-based models for RDE 4.1 Training experience on engine bench Since the dynamic data-based model is merely a “black-box” for end-users, investigations at the engine bench were performed in order to understand which training methodology would give the best results in accuracy and efficiency. 4.1.1 Selection of number and type of model inputs The number and type of inputs of the dynamic data-based model have to be fixed in the initial stage and contain at least the engine operational parameters (speed and load). Since the 1D airpath model predicts relatively accurately & dynamically the intake air related inputs (EGR rate, airmassflow, pressure & temperature of the inlet gas at intake manifold), it was possible to enhance the physical understanding of the model by adding them as further inputs. This brings a clear improvement in NOx emission prediction accuracy, as showed in Fig.3. In this case, the intake gas physical inputs are a result of the different speed-load combinations in a dynamic bench training with speed and load the actively changed parameters, while the further calibration parameters are resulting from a given engine calibration. In this case, the model was trained by a single 2h RDE cycle and is predicting a WLTC. 39 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions In order to get a better idea of the model prediction quality for changes in EGR calibration strategy (without actively changing EGR calibration during the training), several WLTC driving cycles with different EGR calibrations were run at the engine bench and various dynamic models were trained by combining the different cycles, in order to check the dynamic model extrapolation and interpolation potential for EGR calibration changes. As can be seen on Fig. 4, the dynamic model has good interpolation capabilities when trained only with the extreme EGR rate cycles (WLTC with base EGR calibration and with 30% reduced EGR) and predicting intermediate EGR cycles (WLTC with 10% and 20% reduced EGR). In case the model would be used to predict the effect of different EGR calibrations, training with at least the extreme EGR calibrations might be a reasonable approach. On the other side, the capability for predicting EGR reductions that were not part of the training is limited. This demonstrates that there is limited possibility to predict or extrapolate outside of the trained domain. Figure 3: Dynamic data-based model NOx R² accuracy for various inputs (WLTC) Figure 4: Dynamic data-based model predictivity for EGR calibration changes . 4.1.2 Type of training plan & amount of data The dynamic model can be trained by existing transient cycle data or by a dedicated dynamic DoE test plan, which in this case needs to cover the absolute values and 40 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions gradients of the operational parameters (speed & load) for various RDE type cycles. In this case, an existing 2h “harsh” RDE cycle, which covers a very wide engine operation area as can be seen in Fig. 5, was used to train the model. The selected model inputs are engine speed, fuel mass, and the intake manifold physical condition (pressure, temperature, EGR rate) Figure 5: Engine operation area of the training & validation cycles Contrary to the experience with DoE based steady models, the dynamic data based models showed immediately a good potential in predicting highly transient cycles, but limitations when simulating more steady cycles like the NEDC. A typical highly dynamic (harsh) RDE-cycle based model training showed difficulties in predicting steady driving emissions. Adding a short smooth cycle to the training data improves the NOx accuracy for cycles with less dynamism, as shown in Fig. 6. Figure 6: Dynamic data-based R² NOx fitting and prediction accuracy 4.1.3 Accuracy results By using the experience in training the model and optimising the training data frequency and settings, a very good accuracy of NOx & CO2 was achieved, both in dynamic behaviour (Fig. 7) and as cumulative engine-out emissions. The R² accuracy for CO and THC is slightly lower, which could be due to emission analyser stability and / or physical long-term phenomena which are not trained well in a short test plan. 41 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Figure 7: Emission validation (R²) for different driving cycles 4.2 Training & validation for RDE environmental conditions For training the dynamic emission model for the full RDE boundaries also the effect of ambient temperature, slopes and altitude need to be covered, including the various combinations of these environmental parameters. A matrix style of measurements to cover the different altitude-temperature-slope combinations would give strong interpolation capabilities to the model but would result in a relatively long measurement campaign, in order to cover all these combinations. The available facility for altitude emission training was a mobile climatised roller bench as shown in Fig. 8.[5]. On a vehicle roller bench it is difficult to create a dedicated mathematically optimised testing plan since there is no direct dynamic control of the speed/ load of the engine like on an engine bench. Therefore using the experience gathered at the engine bench, a dedicated model training driving cycle was developed with the target to train the emission behaviour in full RDE environmental circumstances in the shortest possible timeframe. Figure 8: Mobile & climatised roller bench 4.2.1 Training methodology with dynamic environmental inputs Apart from the dynamic engine load/ speed behaviour including their gradients, the most important environmental aspects which have an influence of the engine calibration and emission formation are the ambient temperature, the ambient air pressure (altitude) and the slope gradient. The target is here to train the dynamic data-based emission model to cover the extended RDE requirements. To improve training efficiency, as much as possible the environmental parameters where dynamically 42 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions changed during the test on the roller bench. For the following parameters, this was actually possible: • target vehicle speed & shift pattern • slope • test cell temperature The test cell pressure could not be controlled and was equivalent to the outside pressure. Therefore the training cycle was repeated at various altitudes to extend the training domain to higher altitudes. 4.2.2 Optimised dynamic test plan for roller bench Based on the experience gathered on the engine bench, a dynamic cycle was developed for the roller bench with an optimal variation of the environmental parameters in order to efficiently train the model. Due to the emission measurement equipment limitation, the cycle duration was limited to 2 hours. The optimisation of the cycle was done as follows: • The engine load/ speed and gradient of load/ speed was varied by the vehicle speed/ shift profile in combination with the variation of slope gradient, which was varied in a sinusoidal pattern [6]. At the same time, these profiles were optimised in relation to the cooling capacity of the test cell, since cooling down is difficult when engine load keeps persistently high. The slope variation was optimised to cover a wide part of the engine map with different gradients. • The test cell temperature target was optimised to train the widest engine operation range at the various temperatures, taking into account the cooling performance limitation of the test cell. Due to limitations in test cell cooling capacity at high load driving, the test was divided in a low-temperature range cycle and high-temperature range cycle with minimised overlap. The final roller bench test cycle vehicle speed, slope, air temperature and altitude is visualised in Fig. 9. In total 8 cycles of 2 hours (2 temp. ranges x 4 altitudes) were recorded to train the emission model, covering a large environmental domain. Since the training duration at lowest temperatures (below 0degC) was limited, the trained speed-load domain of a low-temperature range training cycle was broken down for different temperature ranges on Fig. 10. As can be seen, also below 0degC the trained domain covers the whole engine map reducing the need for extrapolation through the model. In this way the boundaries of the engine map could be sufficiently trained at the various ambient temperatures. 43 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Figure 9: Developed dynamic roller bench training cycle Figure 10: Training cover area for different engine intake air temperature ranges, for a low-temperature range cycle Full cycle Temp. <0°C Temp. >6°C 0<Temp.<6°C 44 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions 4.2.3 Validation in WLTC and RDE of dynamic-data emission model The dynamic data-based emission model was trained using all roller bench data as described in 4.2.2. The model inputs were the physical engine operation and intake state related parameters (Fig. 11), the ambient temperature/ pressure itself is not known for the model. Figure 11: Validation methodology of emission submodel In this validation step, the trained emission model was used to predict the emissions of actual measured on-road RDE cycles and WLTC cycles on the roller bench, with a variation of location, slope gradients, altitude, ambient temperatures and driver aggressiveness. To check the prediction accuracy of the emission submodel for real driving, the model inputs are measured and are not yet coming from the engine & vehicle model. Fig. 12 shows the cumulative NOx mass prediction accuracy of the dynamic emission model for different RDE cycles (measured on-road) and WLTC cycles (measured on roller bench), showing an accuracy for cumulative engine-out NOx emissions within +/ - 5%. The R² values for NOx were on average 0.95. This means that the dynamic data-based emission model is very accurate if the provided inputs are correct. This also means that the training program on the roller bench provided wide model training data in order to predict various driving and environmental conditions. The trained domain (cover range) of the model inputs vs. the RDE and WTLC validation tests is plotted on Fig. 13. The trained domain is wide enough to predict the emissions while avoiding extrapolation outside of the trained domain for each input parameter. Figure 12: Dynamic model cumulative NOx mass prediction accuracy 45 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Figure 13: Model input range of several validation tests vs. trained domain 5 Full toolchain validation results in RDE After all the toolchain submodels were validated individually, they were connected and the toolchain was validated for engine-out emissions in RDE and WLTC, based on existing speed-shift-slope profiles, as described in Fig. 14. In this set-up, the inputs for the dynamic emission model are predicted by the 1D GT-Power model, connected with the ECU SIL model and driver model. Figure 14: Engine-out emissions toolchain validation for known routes A roller-bench measured hot WLTC at higher altitude was initially simulated to make a back-to-back comparison of the toolchain with the measurement data possible. (On the roller bench, the vehicle roadload is fully known and validated). As can be seen in Fig. 15, a fuel consumption gap of 0.3% and a cumulative NOx gap of 2.1% was achieved. The instantaneous NOx behaviour is reproduced well, both in transient and steady parts of the cycle, providing accurate information for the catalyst submodel. Additionally the engine-out toolchain was validated with on-road measured RDE cycles. In such a case, the vehicle roadload and resistance is related to the road surface, wind and the slope gradient, while their instantaneous impact on the vehicle resistance is not fully known or measurable on the road. Also, it is difficult to get very precise slope gradient information from geolocalisation databases. Therefore the validation with real on-road RDE data is mainly indicative. The difference of cumulative NOx emissions with measured RDE data is within +/ -6% for 2 measured cycles as shown on Fig. 16. 46 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Figure 15: Toolchain engine-out gas validation in WLTC at higher altitude Figure 16: Toolchain engine-out NOx validation in WLTC & RDE 6 Virtual route profile To run the toolchain emission prediction function, a speed/ shift/ environmental conditions profile is required as the general input. This can be retrieved from measured onroad driving data or official cycle profiles. Another possibility is to retrieve the route attributes from existing databases (for example: maximum allowed speed) and develop functions to realistically accelerate to this speed. The main issue with such method to prepare a speed profile is that actual driving is more dynamic due to various events on the road, which have a large effect on the vehicle emission. In this background, an artificial intelligence longitudinal driver was developed, which learns real driver behaviour (acceleration/ brake, clutch and gearshift) on measured routes and reproduces it on any routes selected from a map trough GPS waypoints. Machine learning was selected as the best approach to develop the offline driver model. An innovative algorithm, inspired by t-distributed Stochastic Neighbour Embedding (t- SNE) was applied. 47 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions 6.1 Training of the offline model The target of the offline model is to learn the relation between driver behaviour (accelerate, brake, shift) and the driving context (traffic signs, slopes, curves etc.), in order to be able to predict the driver behaviour for new routes. In order to train the model, both GPS and vehicle ECU data are required. By a Map Attributes Extractor, the driving context is defined by extracting road attributes from a geolocalisation database using a GPS trace, and is combined with engine ECU data (vehicle operational parameters) for training. In this way, the model is able to learn the interaction of the driver with the driving environment. 6.2 Algorithm A novel algorithm named ‘Bi-Directional- Map’ (BiMap) [7,8] was developed, as an extension of the dimensionality reduction algorithm t-SNE. The original t-SNE algorithm [9] has a performant self-clustering capability and can use twoor threedimensional space maps to visualise links between the multi-dimensional data. The benefit of the BiMap algorithm is the possibility to extrapolate in a t-SNE cluster/ map new driving contexts without retraining the whole model, which makes it possible to predict new contexts, therefore new route profiles. A combination of geometric road attributes (like slope of segment, length of segment, angle between current and previous segment etc.) and visual driver input (like speed limits, distance to traffic light, junctions, proximity to stop signs etc.) are used input for the training. Examples of BiMap visualisations are shown in Fig. 17. In the left case, the colour variations represent the different speed limits in the training data, retrieved from the geolocalisation database. In this case, a wide variation (combination of the different attributes and inputs) is available in the training data, which gives good prediction ability. In the right case of Fig.17, the Bimap for the parameter “distance to traffic light” is shown. Since in this case the training data was recorded in a small town with very few traffic lights, there is little colour variation. In such a case, it will be difficult to predict the driver behaviour related to the distance to the traffic light in other conditions (for example other speed limits) than available in the training data. This shows the importance of “rich” training data to have a high-quality training of the off-line driver model. Figure 17: Example of Bimap visualisation of road attributes in driver model training Distance to traffic light Speed limit 48 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions 6.3 First training & validation results The offline driver model was trained by using available on-road RDE data and then used to virtually predict the driving profile of an RDE route that was not part of the training data of the driver model. This means that based on the original training, the driving profile (speed, shift etc.) was fully predicted after selecting the known road on the map. The first results are shown in Fig. 18 and are compared to a single on-road measurement of this route. The vehicle speed and shifting profiles are matching well when comparing to a measured driving event on this route. Further statistical analysis including more driving events and drivers is required to make final conclusions, but the first results show that fully virtual RDE simulation (from selecting a road on a map to the resulting emissions) is feasible. Further work will also focus on the possibility to set different levels of driving aggressiveness. Fig.18: RDE rural predicted driving profile vs. on-road measurement which was not part of the driver model training 7 Practical usage of RDE toolchain The validation of the 1D engine & vehicle model, online driver-model, SIL and catalyst model were finalised with good results but are not reported here. The full toolchain makes it possible to fully virtually evaluate the engine-out and tailpipe emissions. The different steps and output is displayed in Fig. 19, showing an actual example from an RDE route in France. The usage could be further expanded to engine calibration if the calibration parameters are included in the emission submodel and the training plan through a dynamic DOE of calibration parameters. 49 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Fig. 19: Workflow and usage cases in development of the developed RDE toolchain 8 Conclusions A complete toolchain to validate and optimize RDE emission both engine-out and tailpipe was created and validated. The combination of a 1D engine model with a dynamic data-based emission model demonstrated a very good accuracy for predicting RDE cycles in various environmental conditions. Valuable experience was accumulated to train the dynamic data-based emission model in a relatively short amount of time, both on an engine bench and a roller bench. Alltough the dynamic data-based emission model currently lacks the predictivity towards engine calibration changes, it is possible to update it with a very short amount of time for major engine calibration modifications, and therefore can be a regular standard procedure in the development process. Additionally, a tool was created to virtually create driving profiles for any road selected on a map. Validation in rural conditions showed a good match with measurement data, making fully virtual RDE validation and optimization possible. Acknowledgments The author would like to thank the co-authors and colleagues from TMC/ Japan, including Takayuki Hamada, for their strong support in this development and especially thank Daniela De Lima Moradell for her big contribution in developing the toolchain. 50 2.2 Development of a Simulation Platform for Validation and Optimisation of Real-World Emissions Literature [1] N. Oikawa, T. Fukuma, Y. Hamamura, T. Yamamoto, H. Kaneko, G. Kishimoto, T. Toda: The New Toyota Inline Four-Cylinder 2.8L ESTEC GD Diesel Engine, 36 th International Vienna Motor Symposium, 2015 [2] Y. Takasu, S. Kaneko, H. Tominaga, H. Namura et al: Universal Diesel Engine Simulator (UniDES) 2 nd Report: Prediction of Engine Performance in Transient Driving Cycle Using One Dimensional Engine Model”, SAE Technical Paper 2013-01-0881, 2013 [3] G. Cornetti, T. Huber, T. Kruse: Simulation of Diesel Engine Emissions by Coupling 1-D with Data-based Models, FKFS Stuttgarter Symposium, 2014 [4] T. Huber, M. Hanselmann, T. Kruse: Use of data based models to predict any RDE cycles - Challenges, Experiences and Results, 9. Emission Control Conference Dresden, 2016 [5] N. Yedikardachian, J-C Yvorel, J. Roselier, N. Bion: How to control the impact of external factors to ensure the reproducibility of RDE test runs, SIA Powertrain Conference Rouen, The Clean Compression Engine of the Future, 2016 [6] N. Sakushima, W. Baumann, K. Röpke, M. Knaak: Transient Modeling of Diesel Engine Emissions, JSAE Annual Congress, 2010 [7] Seiji Ogura, Hirotaka Kaneko, Kenji Ota, Isao Kobayashi, Naohiko Oikawa: Diesel Engine System Development Method Based on MBD (Second Report), JSAE Annual Congress,Spring 2016 [8] Patent publication number WO2017/ 012677 [9] L.J.P. van der Maaten, G.E. Hinton: Visualizing High-Dimensional Data Using t- SNE, Journal of Machine Learning Research 9, 2008 51 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Stefan Angermaier, Mukunda Gopalakrishnan Abstract Increasing complexity of internal combustion engines demands the design and implementation of powerful control structures in the engine control units in order to fully exploit the potential in terms of fuel consumption, exhaust emissions and driving behavior. The Advanced Modeling Unit (AMU) developed by Bosch allows the calculation of pure data-based (structure-free) models in real time on the ECU. This technology is used to represent virtual sensors as well as complex pre-control structures, e.g. the basic ignition timing of a gasoline engine. A hybrid structure consisting of data-driven modules and structurally supported, physically motivated parts can also be used for the modeling. We present here an overview of different application types of databased models using the AMU, together with the respective potential to increase accuracy based on selected applications. As an example we present the function representing the basic ignition timing. By implementing a data-based model, a significant improvement has been realized, which will contribute to meet future fuel consumption requirements. The function development process will extend to the area of data generation and the formation of the data model, since this has a significant influence on the performance of the function and is the most important precursor to the function itself. Knowledge of the specific needs and limitations of data-based models is necessary, furthermore expertise on mathematical modeling methods is helpful. We will explain the handling of these models and their basic conditions by means of the steps we performed while developing the function representing the basic ignition timing. In addition, it is also explained how to deal with the handling of a time-critical function module when using the AMU. The calibration process is, even more than for conventional approaches, an integral part of the development, since the ability to efficiently generate corresponding measurement data is a prerequisite. At the same time, a higher demand is placed on the quality assurance of the resulting multi-dimensional models. We show the corresponding calibration steps for the above mentioned function. Differences in the development processes in the context of the above-mentioned method compared to the conventional approach are explained and the potential for agility and sustainability is pointed out. In this context, we also deal with the handling of function variants for different engine systems. 52 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Kurzfassung Die steigende Komplexität der Motorsysteme erfordert leistungsfähige Kontrollstrukturen in den Motorsteuerungssystemen, um Potentiale in Bezug auf Kraftstoffverbrauch, Abgasemissionen und Fahrverhalten vollumfänglich ausschöpfen zu können. Die von Bosch entwickelte Advanced Modeling Unit (AMU) erlaubt die Berechnung rein datenbasierter (strukturfreier) Modelle in Echtzeit auf dem Motorsteuergerät. Diese Technologie wird eingesetzt zur Darstellung virtueller Sensoren oder komplexer Vorsteuermodelle, beispielsweise des Basiszündwinkels bei einem Ottomotorsystem. Auch hybride Strukturen aus datenbasierten Modulen und strukturgestützten, physikalisch motivierten Anteilen sind möglich. Wir präsentieren einen Überblick über unterschiedliche Anwendungsklassen datenbasierter Modelle auf der AMU und zeigen anhand ausgewählter Anwendungen das Potential zur Steigerung der Genauigkeit auf. Für die Funktion zur Ausgabe des Basiszündwinkels, welche wir im Besonderen vorstellen, wurde durch die Implementierung eines datenbasierten Models eine deutliche Verbesserung erzielt, welche dazu beitragen wird zukünftige Kraftstoffverbrauchsziele zu erreichen. Der Funktionsentwicklungsprozess weitet sich auf den Bereich der Datenerzeugung und Bildung des Datenmodells aus, da dieser erheblichen Einfluss auf die Leistungsfähigkeit der Funktion hat bzw. die Funktion überhaupt erst ermöglicht. Wissen über mathematische Modellierungsverfahren, mindestens aber Kenntnis über deren spezifische Erfordernisse und Grenzen, sind erforderlich. Die Handhabung dieser Modelle und ihrer Rahmenbedingungen erläutern wir anhand der Entwicklung der Funktion zur Ausgabe des Basiszündwinkels. Darüber hinaus behandeln wir an diesem Anwendungsfall die Handhabung eines zeitkritischen Funktionsmoduls im Umgang mit der Ressource AMU. Der Applikationsprozess ist mehr denn zuvor integraler Bestandteil der Entwicklung, da die Fähigkeit effizient entsprechende Messdaten zu generieren eine notwendige Bedingung darstellt. Gleichzeitig geht ein höherer Anspruch an die Qualitätssicherung der resultierenden, hochdimensionalen Modelle einher. Wir zeigen den korrespondierenden Applikationsablauf für die oben genannte Anwendung. Unterschiede im Entwicklungsprozess im Zusammenhang mit der genannten Methode im Vergleich zur konventionellen Vorgehensweise werden erläutert und daran ihr Potential im Hinblick auf Agilität und Nachhaltigkeit aufgezeigt. Wir behandeln dazu auch die Handhabung von Funktions-Varianten für die unterschiedlichen Motorsysteme. 53 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes 1 Introduction Besides the constant need to meet future emission standards, combustion engine development has the objectives to improve driving characteristics, reduce fuel consumption, and achieve ideal torque performance. To enable these improvements a steadily increasing number of actuators are being used. More precisely, gasoline engine actuators are used to reduce losses influencing charge exchange such as camshaft phasing or variable valve lift as well as other characteristics with the aim of optimizing combustion behavior such as swirl and turbulence control, external EGR flow or water injection. Moreover, in the nearby future, it is expected that variable valve train as well as variable compression ratio will be introduced resulting in even more complex cause-effect relationships. Together with engine speed, engine load and lambda (air/ fuel ratio), the aforementioned actuator positions affect the basic ignition timing (basic spark advance) that is needed to achieve the engine operating condition with the best possible efficiency. The multitude of parameters and their non-linear dependencies on basic ignition timing leads to a major challenge during function development when defining a plant model capable of accurately representing the desired ignition timing. Today’s stateof-the-art solutions are based on control value maps which are at most three dimensional, severely limiting their applicability in high-dimensional scenarios. The resulting inaccuracies in the function output can lead to additional fuel consumption, drivability deficiencies, and in the worst case to engine damage by knocking. With the new generation of engine control units (MDG1, from device 3 upwards), Gaussian Process Regression (GPR) models can be evaluated in real time on the control unit by use of the Advanced Modeling Unit (AMU). Such models are capable of mapping multi-dimensional non-linear relationships. With this modeling approach, the limitations imposed by the conventional function structures described above can be overcome. The term ASC@ECU includes the workflow implementing the GPR models on the engine control unit and the required tool chain covering the process from model training to post-application [1]. 2 Application categories with potential for using the AMU Data-based structure-free models are generally the preferred solution if multidimensional nonlinear system behavior is to be represented for which no physical description is available or if the physical model does not provide sufficient accuracy. Possible applications can be divided into the following categories: Virtual sensing, pre-control-/ set point-structure and hybrid structure. The categories differ from each other in the way the models are created and structured. In the following, we describe the categories, show an application and the potential in terms of the achievable accuracy. Figure 1 gives an overview on the application categories. 54 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 1: Overview on the application categories with potential for using the AMU 2.1 Virtual sensing Physical sensors are costly, sometimes unreliable, and need to be maintained [2]. In addition, their operating range is in general restricted: the lambda sensor for example can only be activated after reaching a minimum temperature threshold. Virtual sensors are used to replace a temporarily installed sensor or to provide a continuous signal from discrete sensor measurements (e.g. stationary measurements) and may have predictive capabilities. They can also be used to provide a redundant signal to a physical sensor, thus enabling the diagnosis of faults and serving as a substitute. [3] x i y A) Virtual sensing mapping of I/ O-relation of a measured value y to input vector x C) Hybrid structure - combination of conventional structural elements and data-based model(s), where the intermediate value is not accessible through measurement B) Pre-control or set-point structure - mapping of I/ O-relation of an output y* that is a result of an optimization procedure to input vector x x i y* x i y LUT/ Filter/ … s LUT/ Filter/ … t +-*/ +-*/ I 1 I 2 I 3 55 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes The data-based model establishes the relationship between the selected input parameters and a value y, where the value y is either determined empirically by means of a dedicated measurement or is the result of a (physically motivated) formula which is too complicated to calculate in the ECU such as a numerical solution of differential equations. As an example for this category we present the 50%-Mass-Fraction-Burned (MFB50) signal. This signal is derived from in-cylinder pressure measurements using known state-of-the-art thermodynamic equations. The signal is defined as the crank angle where 50% of the fuel mass is burned. This signal can be represented by a databased model. To this end for a warm gasoline engine the input parameters engine speed, cylinder air charge, lambda and ignition timing as well as the parameter values of the actuators that influence the charge exchange and the turbulence are to be used. Figure 2 shows the accuracy of a MFB50-model for a complex gasoline engine in homogeneous single injection mode. The engine is complex because it has, in addition to continuously variable inletand outletcamshafts, another continuously adjustable parameter which affects charge exchange. In below figure we depict the training error and the error on the test data that has not been part of the model training. The same is done for the other examples that follow. It is seen that a highly accurate model covering the entire input parameter space could be created, even for a complex engine system. Figure 2: Accuracy of the virtual sensor 50%-Mass-Fraction-Burned (MFB50) [°CA] on training data as well as on test data. 2.2 Pre-control or set-point structure Set-point structures differ from virtual sensing application in that the output to be modelled results from a preceding optimization process. Such an optimization process can, in turn, be based on a model or originate from a control process. As an example for this category we present the set-point function for the basic ignition timing. This function calculates the ignition timing that leads to optimal combustion. The ignition timing to be represented is determined by an optimization process, using a data-based (engine) model which describes the combustion of the engine by Data-based model approach compatible with the AMU 56 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes means of all relevant input parameters including the ignition timing. The ignition timing is thus optimized for all input parameters according to the optimization goals in a space-filling manner. By use of the values thus determined, a model with output ignition timing set-point is created, that can further be evaluated directly in the ECU by using the AMU. Figure 3 shows the accuracy of the data-based model approach compared to the accuracy of a function based on conventional structural elements such as 2-D/ 3-D lookup tables (LUT). It is seen, that the accuracy can be significantly increased. The application basic ignition timing set-point is discussed in the following chapters in detail. Figure 3: Accuracy of the ignition timing set-point [°CA]. Comparison of the databased model approach A) and the a. using conventional structural elements B). 2.3 Hybrid structure Hybrid structures use purely data-based models in combination with conventional structural elements such as 2D-/ 3D lookup tables, filters and arithmetic operators. Such applications differ from the previously described types when the parameters of the structure are to be optimized simultaneously, not individually. This becomes necessary when the intermediate values are unknown, like I 1 to I 3 in Figure 1 C). B) Approach using conventional structural elements A) Data-based model approach compatible with the AMU 57 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Conventional structural elements are used for various reasons, for example if (physical or empirical) prior knowledge needs to be introduced into the structure or when there is a necessity to realize an invertible function. As an example for this category we show the torque model of a gasoline engine. The torque must be represented by the same input parameters as listed in Chapter 2.1. This application requires, in particular, that the calculation can also be carried out inversely. This is necessary to determine the desired air charge and ignition timing that generates the desired torque. Here, a part of the efficiency calculation with regard to the ignition timing has been implemented by using a RBF (Radial Basis Function) net. This modeling algorithm has the same evaluation formula as the GPR and therefore is suitable to be calculated on the AMU as well. The RBF net has been determined by fitting the parameters of the entire hybrid structure with ETAS MOCA. The result is shown in Figure 4. The accuracy can be increased from 3.4 Nm to 2.2 Nm (RMSE). Figure 4: Accuracy of torque model [Nm]. Comparison of the data-based model approach A) and the approach using conventional structural elements B). B) Approach using conventional structural elements A) Data-based model approach compatible with the AMU 58 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes 3 Function development and Calibration 3.1 Impact on function design In the previous chapter we showed applications in which data-based models can be used. In the following we describe how a meaningful functional design is developed for a potential application and what has to be considered. In the following, we explain the procedure for defining the data-based structure-free part of the function described by a flow chart in Figure 5. Figure 5: Recommended procedure for defining the data-based part of the function First, a reasonable system boundary of the subsystem must be defined. It is helpful to visualize the system boundary considered, in particular, to identify possible influencing parameters. Figure 6 shows the subsystem of combustion in a gasoline engine. In order to determine reasonable input parameters, it is necessary to examine the system boundary in detail. To exemplify this we will discuss the system boundary with regard to the injection system. The influence of the fuel quantity on the combustion can be represented in three ways. If the behavior of the injector is unknown, this subsystem has to be additionally represented by a model that additionally requires the input parameters high pressure, back pressure and injection duration, see A) in Figure 6. If the injection quantity is Define a meaningful output variable and further measures to ensure reasonable extrapolation behavior Define system boundary and input parameter candidates Select reasonable input parameters as a subset of all the candidates Acquisition of space filling training data is feasible? No Yes Done 59 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes already determined by a control unit function, use of this value reduces the input parameter space by 2 dimensions in comparison to A). This consideration is of great importance when taking into account that the databased structure-free modeling requires space-filling training data. Each additional dimension thus significantly increases the measurement effort. Variant B) has the further advantage that the fuel quantity is corrected by means of the lambda control, thus compensating system tolerances which would have occurred in the other case as faulty input signals. Variant C) describes the influence of the fuel quantity on the combustion by means of the air/ fuel ratio lambda. Since the air mass is also chosen as the input variable of the model, using lambda instead of fuel quantity is essentially just a transformation, but with a helpful effect. Since the amount of fuel correlates strongly with the air mass, training data will be represented as a narrow distribution around a straight line, while lambda forms a rectangular parameter space referable to the air charge (Figure 10). In general, an approximatively rectangular parameter space is desired. This reduces the challenge of establishing reasonable extrapolation behavior. We will explain this further below. Furthermore, strongly correlated signals such as intake manifold pressure and air mass have to be avoided, thus ensuring that dependencies will be distinguished properly. Figure 6: System boundary of the subsystem combustion of a gasoline engine A) Injector behaviour is unknown B) Fuel amount is det. by ECU function C) Lambda represents fuel amount A) B) C) Detail 60 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes After these considerations we derived the following input variables for our subsystem: • Engine speed • Cylinder (air) charge • Actuators, influencing charge exchange and turbulence • Ignition timing • Lambda (air/ fuel ratio) • Injection timing and rail pressure • Fuel properties (further discussed below) • Engine and intake manifold temperature (further discussed below) We consider the three last-mentioned parameters to further support our hypothesis that in addition to physical meaningfulness the selection of the input parameters essentially depends on the ability to acquire space-filling training data and on the other hand, to be able to manage the extrapolation behavior. In particular, the representation of the engine in cold condition poses a great challenge to data-based modeling. The generation of space-filling training data is only possible to a limited extent and even then it is extremely complex. The parameters engine temperature and intake temperature are thus in general not suitable as input parameters of a data-based model. Hence approximations using conventional structural elements must be developed. Before we describe the solution w.r.t. the representation of cold condition of the engine that was developed for the basic ignition set-point function (see chapter 2.2), we will give further explanations on this application. The essential criteria for determining the optimum ignition timing is the achievement of the optimal combustion w.r.t. thermodynamic considerations and on the other side avoiding knocking combustions, which limits ignition timing in the early direction. The influence of the engine temperature and the intake temperature on the ignition timing which leads to the thermodynamic optimum can be approximated by a simple correction curve for each of these two parameters. The knock limit also shifts but in a different manner, due to a different underlying phenomenon, see Figure 7 A). We have thus separated the two phenomena functionally, which enables the approximation of the temperature dependency with a sufficient accuracy by means of a simple correction function on the respective paths. 61 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 7: Influence of the intake manifold temperature distinguished between ignition timing w.r.t. the thermodynamic optimum and the knock limit Figure 7 B) shows the resulting curve of the desired function output when the dependence of the intake temperature is approximated as a simple characteristic curve, as implied in Figure 7 A). One can easily imagine that a global correction function would not provide a sufficiently accurate function since this would mean a compromise for the two behaviors depicted in Figure 7 A). Figure 8 shows the implementation of the above described distinction between the ignition set-point w.r.t. the thermodynamic optimum and the knock limit. The correction values are added to each path individually, thus the influences are approximated as close as possible to the physical behavior. In this way, we introduced physical prior knowledge into the function structure. 20°C Intake Temperature -30°C 70°C Cylinder (Air) Charge Ignition Timing desired output @70°C -30°C 20°C 70°C desired output @20°C thermodyn. optimal ignition earlier earlier influence on knock limit influence on thermodyn. optimal ignition A) B) Ignition Timing 62 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 8: Implementation of the separation of the ignition timing set-point w.r.t. the thermodynamic optimum and the knock limit After considering a parameter candidate for which data acquisition is almost impossible, we will discuss the others that are manageable, but involve a high effort. Expenditure for training data acquisition increases in general with each additional input dimension, hence each input parameter is to be weighed with regard to the ratio of potential accuracy gain to effort for data acquisition. As an example, we discuss the implementation of the ethanol content. The ethanol content has a significant influence on the optimal ignition timing, in particular on the knock tendency. The parameter can, however, only be varied with high effort in a space-filling manner. A change in the parameter value requires the emptying and refilling of the fuel system with the desired fuel mixture, so that the supply of different mixtures also leads to high logistical effort. In Figure 9, three implementation approaches are shown. In A), the influence of the ethanol content is expressed by means of a lookup table (defined over the maximum parameter range), where intermediate values are interpolated by means of a characteristic curve. This variant represents the greatest simplification in terms of accuracy, but also the least effort for calibration. Variant B) uses another data-based model considering the maximum ethanol content with identical input parameters as for the base model where intermediate values are interpolated between the two models. The effort for calibration is doubled, but it is still manageable. Variant C) includes the parameter ethanol content in the data-based model which leads to almost unmanageable effort as described MDL_TERMODYNOPT eng_speed cyl_charge inlet_cam outlet_cam lambda knocklim' ign. timing Δ ign. 1 + MN CURVE 1 temperature MDL_KNOCKLIM eng_speed cyl_charge inlet_cam outlet_cam lambda knocklim' Δ ign. 2 + CURVE 2 temperature 63 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes above and is therefore not recommended. All variants are software-programmable, the choice depends on the requirements for the accuracy. Figure 9: Implementation approaches for considering the ethanol content Reasonable extrapolation behavior is the other major challenge in the application of data-based models in addition to ensuring space-filling data acquisition. In Figure 5 this is mentioned as the last step of the process since the previous steps define the parameter space for which extrapolation has to be considered. The modeling algorithm substantially determines the extrapolation behavior. A GPR model for example, approaches far away from the data to its mean function. The extrapolation behavior of this modeling approach further depends on the smoothness of the model described by the hyper-parameters. In particular the hyper-parameter length-scale influences the way in which the model transitions from the trustworthy training-data range to the mean function. The mean function is in general the average value of the data, but can also be a linear regression model etc. Furthermore, it is important to be aware of the fact that the model provides a function value for all input parameter values from minus infinity to plus infinity. When using such a model on the ECU, it must be ensured that a meaningful function output results for all input parameter values. With the above description of the extrapolation behavior, it becomes clear that measures are required to ensure a reasonable extrapolation behavior but the definition of reasonable extrapolation behavior in the multi-dimensional space is anything but trivial. We suggest to follow the measures listed below to attain a functional design which yields reasonable extrapolation behavior. A) Basic approach B) Complex Approach C) High complex approach MDL_KNOCKLIM eng_speed cyl_charge inlet_cam outlet_cam lambda ethanol knocklim INTERPOLATION ethanol x DELTA_MAP eng_speed cyl_charge Δ knocklim MDL_KNOCKLIM_E100 eng_speed cyl_charge inlet_cam outlet_cam lambda - INTERPOLATION ethanol x knocklim_E0 Δ knocklim 64 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes • Reasonable function design (1) - The input parameters of the data-based part of the function have to be defined such that the needed data can be acquired space-fillingly in an axis-aligned parameter space. Applying the Bounding Box , which is a feature of the AMU driver, the parameter space without training data available can then be reduced to a minimum. To this end, the Bounding Box limits the input parameter range according to specified values, thus restores the known behavior of a 2-D/ 3-D lookup table. Figure 10 illustrates this for the two quantities air mass and fuel mass, where in both of the depictions the same information content and data is presented but in a different parameter space. The choice of the input parameters in the lower graph is better because of the smaller area that results for which no data is available. Figure 10: Reasonable choice of input parameters and the Bounding Box (dash-dotted-line) to reduce the parameter space where no data is available. The input data range marked in red shows the area where extrapolation behavior of the model has to be investigated. • Reasonable function design (2) - By understanding the extrapolation behavior of a GP model, it becomes obvious that functions are favorable that only vary around a mean value. Thus, the extrapolation value is defined and the transition area becomes smaller. To this end the function to be modelled has to be formulated so as to come as close as possible to this ideal scenario. Use of physical and empirical prior knowledge as core of the function structure could serve this purpose. To illustrate this, Figure 11 A) depicts the extrapolation behavior of the model that represents the sum of two functions f 1 (x) and f 2 (x). In the extrapolated range (<0 and >1) the model doesn’t follow the true function (black). On the 0,6 0,7 0,8 0,9 1,0 1,1 1,2 0 20 40 60 80 100 120 140 160 Lambda [-] Cylinder (air) charge [%] 0 10 20 30 40 50 60 70 0 100 200 300 400 500 600 700 Fuel quantity [kg/ h] Air mass [kg/ h] 65 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes other hand, the sine function f 1 (x) as a part of the whole true function varies around the mean value 0, thus the extrapolation behavior is good to handle, it is meaningful still far in the range without available data, see B). If prior knowledge about the function f 2 is applied, the extrapolation behavior can be markedly improved, see C). An analogous example from practice is the ideal gas equation in context of modeling the cylinder (air) charge out of intake manifold pressure. Figure 11: Reasonable choice of output parameters to achieve more meaningful extrapolation behavior. • Creating artificial training data - In order to force the model to certain values in the parameter space where no data is available, artificial data can be created to be included in the training data set. This may also be simulation data. Be aware that creating data always requires an idea on how to extrapolate. If the idea is the result of prior knowledge, we recommend to rather include this prior knowledge in the function structure itself (see second measure). = 0.1 sin 2 = 0.5 training data available no training data available no training data available A) The function f 1 (x)+f 2 (x) is modelled B) The function f 1 (x) is modelled C) The function f 1 (x) is modelled and added to the true function f 2 (x) true function true function 66 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes 3.2 Impact on calibration processes In the previous chapter, we explained that functional design is strongly interlinked with the ability to efficiently generate space-filling data. Function development therefore requires the expertise of the test engineers on experimental design and measurement execution. In addition, the calibration engineers bring in their system knowledge, for example, to identify the parameters affecting the (sub-) system. On the other side, the requirements for calibration increases as the function behavior is now almost solely determined by the generated data. Sophisticated measurement data acquisition and clean processing of the acquired data are of great importance for the quality of the models generated out of them and consequently affecting the function behavior. The function development is thus closely linked to the calibration process up to series production since validation of the function is only possible with the availability of the data, which in turn is project-specific. Creating space-filling measurement data which allows to characterize the entire system or function behavior is, in many cases, more complex than acquiring dedicated data for a specific function structure. This occurs when the functional structure restricts the degree of freedom regarding the systems multi-dimensional non-linear interdependencies and the data acquisition is strictly reduced to the given degrees of freedom. On the other hand, global measurement procedures that are needed to serve for such an approach can open up synergies between individual fields of work, so for example the applications mentioned in chapter 2 can be determined out of the same measurement campaign. This finding has led Bosch to develop the method ECIMMM+ (Engine Character Identification by Advanced Measurement, Modeling and Mapping Methods). ECIMMM+ achieves the greatest possible synergy for processing all the fields of work of gasoline engine base calibration, which are calibrated at the engine test bench. Substantially to this end space-filling measurement data for the parameters engine speed, cylinder (air) charge, lambda, all the chargeand turbulence influencing actuators and the ignition timing is generated. Figure 12 shows an overview on the applications that can be processed with such a measurement campaign that comprehensively characterizes the engine system. With this approach, however, the demands for the ability to automate the measurements on the test bench continue to increase, in particular because the reproducibility of the measurement must be guaranteed over a very long period of time. This places high demands on the test system as well as the test equipment. 67 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 12: Fields of work that can be processed with ECIMMM+ In the following sections we will show how the ignition set-point w.r.t. the thermodynamic optimum and the knock limit is determined out of such a measurement. Figure 13 shows the workflow. Ignition timing sweeps are performed in full factorial manner for all the rest of the parameters, which are themselves varied in space-filling design. If the knock limit is achieved in an ignition timing sweep this gives one training point for the data-based model of the knock limit. All the information about the knock limit determined in this way serves for creating the model that will be implemented in the ECU. For determining the ignition timing w.r.t. the thermodynamic optimum all the data is used to generate a model that gives torque as output based on all the input parameters as listed above. The model is used to generate information about the ignition timing leading to the thermodynamic optimum in a space filling manner by maximizing the torque, other specified boundary conditions have to be respected as well. This information is then used to create the model for the ignition set-point w.r.t. the thermodynamic optimum that will be implemented in the ECU. Measurement + Mapping + Charge Determination Torque Model Paramater Optimization Ignition Setpoint Virtual Sensor (e.g. MFB50%) ECIMMM+ Combustion Limit 68 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 13: Calibration workflow: from measurement data to AMU model(s) For a complex engine as described above ca. 3000 parameter permutations, i.e. ignition sweeps, are measured and evaluated, which easily leads to more than 30000 points. The use of data-based models, implemented on the ECU and the way measurement data is acquired as described above further reduces recursions in the calibration process. Recursions occur when system parameters change, which are only implicitly considered in the respective function structure or when the function focuses on specific system states, e.g. defined operating modes. With the multidimensional representation of the system by means of the data-based model, all changes are covered. Over and above that, better accuracy throughout the calibration process reduces the necessity for recursion, as no inadequate compromises are made. As a last point we want to mention, that the way calibration engineers work is affected not only by the subject of data generation and modeling but also in the post processing of such a model. The two most important points are: • Introducing of local (short-term) adjustments, e.g. on a test trip • Reviewing of the multi-dimensional data-based models, e.g. as part of the quality assurance process Methods and tools are under continuous development and validation to manage these subjects. 0 40 80 120 160 -40 -30 -20 -10 0 10 20 30 40 Torque, Knock Frequency Crank Angle [°CA] Ignition timing knock limited Ignition timing w.r.t. the thermodynamic optimum Knock limit Design and execution of experiment Modeling -10 0 10 20 30 40 -10 0 10 20 30 40 50 60 70 Torque, Knock Frequency Crank Angle [°CA] Ignition timing w.r.t. thermdyn. opt. Preparation of measurement data ETAS ASCMO ETAS ASCMO Modelling of optimal ignition Modeling of Knock limit Script based evaluation of measurement […] Modeling and Optimization of Torque (i.a.) ETAS ASCMO 69 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes 3.3 Functional advantages We have shown in Chapter 2 that a more precise function output can be achieved by means of data-based models, in particular in complex, multi-dimensional settings. Figure 3 in Chapter 2 presents the accuracy of the overall function output of the basic ignition timing set-point which includes both the resulting accuracies of the ignition timing w.r.t. the thermodynamic optimum as well as the knock limit. It should be remarked that the results of the knock limit model tend to have a larger test error (approx. 1.5°CA RMSE) than the model of the ignition timing w.r.t. the thermodynamic optimum (approx. 0.5°CA RMSE). This can be explained by the lower degree of reproducibility of the knock limit. This result coincides value of repeatability usually seen in measurements. Next we present the effects of the improved function on the engine behavior. The exact output of the ignition timing is needed to achieve the optimum efficiency of the engine in the respective operating point without provoking knock events and as a result, best possible fuel consumption with respect to the ignition timing. The dependence on the engine efficiency and thus the fuel consumption of the deficit at the ignition timing is shown in Figure 14. Figure 14: Dependency on the fuel consumption of the deficit at the ignition timing. The values represent the efficiency loss [relative %] from a 3°CA retardation in relation to the optimum ignition timing As basis for comparison we use a conventional function that has an average deviation of 3°CA. The potential in non-knock-limited operating range then is around 0.5% - 1%. In knock limited area it may be up to 5%, while these values do not yet take into account the further effect of reduced enrichment due to component protection. 70 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Such large deviations are common when the actuator positions and lambda do not coincide with the target values for which the ignition timing is optimized under stationary (warm) conditions. This scenario occurs in the following cases: • Dynamic driving situations, due to the limited adjustment speed of the actuators. Higher dynamics tend to have higher values of deviation. • Lambda component protection (de-) activation, resulting in different values for lambda. • Advanced charge control system using the charge exchange influencing parameters in addition to throttle valve and the turbocharger actuator to control the cylinder (air) charge and inert gas ratio. • In cold start and the warm-up phase of the engine. To summarize, we improved the function due to comprehensive consideration of all varied physical parameters’ influences and their interdependencies. The parameters for which data could be (efficiently) acquired space-fillingly became input to the databased model while others like intake manifold temperature can now be extrapolated in a correct manner. This has been realized by the separation of the ignition set-point w.r.t. the thermodynamic optimum and the knock limit. Furthermore, the separation of the knock limit enables an advanced knock limit adaptation functionality, which becomes part of future investigations. 4 Software Implementation The use of data-based models that can be flashed on the engine ECU directly not only increases the accuracy of non-linear models by offering more modeling capabilities than existing 2-D/ 3-D lookup tables but also has a large impact on the software development as a whole. The following section aims to compare and elucidate the changes in the automotive software development practices by using ASC@ECU method as compared to conventional software development processes. The following section is further divided into impact of ASC@ECU on the software development cycle, the handling of function variants using data-based models and scheduling approaches. 4.1 Impact of ASC@ECU on Software Development Life Cycle To compare the impact ASC@ECU has on the software development lifecycle, the existing state-of-the-art life cycle that is recommended by the ASPICE consortium is depicted in Figure 15. As an example to explain the changes in this life cycle, the following real life scenario that arises in engine development is considered. Consider the scenario that a new actuator is added to an existing engine platform. An engine, that has inlet and outlet continuously variable camshafts, now needs to be fit with a charge motion flap that can be continuously varied in the inlet manifold. The charge motion flap at low to medium loads can introduce a swirl effect on the air entering the combustion chamber which has an effect on both the ignition angle w.r.t. thermodynamic optimum and the knock limit. 71 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 15: Software development V cycle according to the ASPICE approach highlighting the steps affected by ASC@ECU The current software structure for an engine where state-of-the-art 2-D/ 3-D lookup tables are used to describe the pre-control model of the basic ignition timing would look like as shown in Figure 16. For every new actuator, such as the charge motion flap the set of 3-D lookup tables based on engine speed and load typically would have to be doubled and the resultant values interpolated. The section marked in blue in Figure 16 is what is needed to be added for depicting the influence of the charge motion flap described in the aforementioned example. This current design of lookup tables then has to be calibrated and validated to check if all the interdependencies have been represented correctly. If during the validation phase it is found that there arises a new interdependency that has not been known before a new software structure has to be deduced for representing this interdependency in the form of more lookup tables. An investigation of such an unknown dependency needs a lot of effort and time on the dynamometer and results again in software development which must pass through the complete V cycle. For actuators causing influences that have not been studied in detail such as low pressure EGR or water injection the development phase from Software Requirements Analysis to Software Integration and Test becomes even more critical and time consuming specifically because a large time and effort has to be spent in analyzing measurement data from the engine dynamometer and then arriving at plausible interdependencies which are transferable in the form of 2-D/ 3-D lookup tables. For already existing complex engines, analysis of the existing Bosch MDG1 software have shown that some engines use up to 1024 lookup tables just to represent the software for the basic ignition timing. Steps affected by ASC@ECU method System Architectural Design System Requirements Analysis Requirements Elicitation SW Unit Test SW Integration and Test Software Qualification Test ECU System Integration ECU System Test Software Architectural Design Software Requirements Analysis Software Design and Construction System Engineering Software Engineering 72 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Figure 16: Implementing an additional parameter comparing the approach based on 3-D lookup tables and the data-based model approach With the advent of ASC@ECU the steps between Software Requirements Analysis to Software Integration and Test can be also largely combined with the calibration workflow of the engine development. ETAS ASCMO not only offers a representation of the nonlinear interdependencies of the various input parameters involved but also solves the problems of having to analyze the measurement data and arrive at software structures that are easily maintainable, understandable and easy to implement without causing a huge increase of runtime. The model that is formed in the calibration data analysis phase of the engine development can be directly flashed on the ECU without any major changes. The precursors for deciding how the model should be built have been described in section 3.1 As seen clearly from Figure 16 the software is not only easier to be presented structurally, but also the changes required in terms of lines of code and implementation complexity of the software is also drastically reduced. The block representing the data-based model is actually implemented as a functional call where the AMU is called eng_speed cyl_charge inlet_cam outlet cam MAP 3 eng_speed cyl_charge MAP 4 Hi Lo fac 1 Δ eng_speed cyl_charge MAP 1 eng_speed cyl_charge MAP 2 Hi Lo fac 1 Δ Hi Lo fac 1 Δ eng_speed cyl_charge inlet_cam outlet_cam MAP 7 eng_speed cyl_charge MAP 8 Hi Lo fac 1 Δ eng_speed cyl_charge MAP 5 eng_speed cyl_charge MAP 6 Hi Lo fac 1 Δ Hi Lo fac 1 Δ Hi Lo fac 1 Δ add. param. ign. timing MDL eng_speed cyl_charge inlet_cam outlet_cam lambda add. param. ign. timing A) Approach based on 3-D lookup tables B) Data-based model approach 73 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes to start the calculation of the data. The new extra input is just an extra argument passed to this function. The architecture and the signal flow remains stable, because any input that can be varied space-fillingly can be introduced as an input to the model. 4.2 Variability Handling of Data-based Models Because of the reason that the software by itself becomes far simpler and modular it has been implemented a new variability concept in the logical architecture. The advantages of the modular software architecture are well known [6]. Figure 17: Software configuration and variability handling in ASC@ECU Figure 17 aims to show the software development workflow for the newly implemented variability concept in the ASC@ECU method. We have developed a stable principal software for the set-point of the basic ignition timing. Further changes expected to the tested principal software are minimal. The software has an additional part that is build time generated based on engine and project specific configurations. Before configuring these data a decision is made on which inputs must be chosen for the model, the order of the inputs and what the model size should be based on the project specific requirements. The software build process, using a build-time code generator, combines the principal software with this project specific configuration. The model is then extracted from ASCMO using the Model2AMU plugin via .dcm file. This ensures that the principal software does not have changes based on engine architecture and also every project has a model with model size that is optimized based on the accuracy requirements. This solves the problem of having multiple variants of software which have to be maintained for various engine architectures. The procedure also ensures control on what inputs are used to build the model. Any change in Project and engine specific function configuration Software deliverables .hex and .a2l Model (meta-) data extracted as .dcm Input parameters, model size, etc. Stable principal software Software build-time code generator Offline modelling in ASCMO Common SW for all projects 74 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes the inputs to the model or the size of the model has to be clarified between both the calibration engineer and the function responsible. 4.3 Runtime and Scheduling Although the improvement of accuracy is a strong argument for the introduction of data-based models on the ECU, the implementation of such models must not increase the run time in a resource stricken engine ECU. The following section aims to compare the various possibilities that exist in scheduling data-based models in the AMU and how the scheduling of the model was done for the time critical basic ignition timing scenario. The scheduling of the data-based model on AMU can be done in three different modes: 1) Synchronous Mode: In this mode of operation the main core of the controller (in this case the Bosch MDG1 controller) calls the AMU to start processing the data model with the current value for a set of inputs. After the call of the AMU, the main core of the controller waits for the AMU to return the result of its calculations which the main core then uses in further calculation chains. 2) Parallel Mode: In this mode of operation the main core of the controller calls both the AMUs in the MDG1 ECU to start processing the model with the current value for a set of inputs. The computation of the same model is split among the two AMUs. After the call of the AMUs the main controller waits for the AMUs to return the result of their calculations. 3) Asynchronous Mode: In this mode of operation the main core of the controller calls the AMU to start processing the model with the current values for a set of inputs. After the call of the AMU, the main core does not wait for the AMU to return the result, but goes on to perform other calculations. The AMU runs in parallel to the main controller and returns the value of calculation back to the main core when queried, after the model is completely evaluated. The figure below shows a comparison of the three different modes. Figure 18: Different modes of scheduling the AMU Synchronous Mode Process_Proc ( ) { AMU call CORE WAITS AMU Return } Next_Proc ( ) { } Parallel Mode Process_Proc () { AMU call CORE WAITS AMU Return } Next_Proc ( ) { } for i=1 to N/ 2 for i=N/ 2+1 to N AMU 1 AMU 2 Asynchronous Mode Process_Proc1 ( ) { AMU call } Next_Proc ( ) { } Process_Proc2 () { AMU return } 75 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes The basic ignition timing set-point is calculated in an engine synchronous process, i.e. the functions are called every time the cylinder nears top dead center. The asynchronous mode was chosen as the preferred mode of scheduling because these processes are time critical and thus an increase in runtime must be avoided. For scheduling the data-based function in a series project, the following steps have to be followed. • Determination of required number of base points: The optimum number of points that are needed to arrive at a model accuracy that is needed for the system should be determined first. For this many compressed models were built with various number of base points and their accuracies were compared. The model size between 40-50 base points was found to be required to represent the basic ignition timing set-point with sufficient accuracy, see Figure 19. Figure 19: Determination of the required number of base points • Offline emulation of the AMU: It has to be checked how much runtime a databased model with 50 basepoints and 6 inputs takes to run on the AMU. From Figure 20 a time of 16µs per model was allocated in the AMU for running the basic ignition timing set-point structure. Using the information about the number of basepoints needed for building the model of the desired accuracy, a runtime analysis of the corresponding project is done. A slot in the process order that neither supply input parameters to the model nor need the output supplied by the model for further calculations needs to be found. Figure 20: Determination of the required computation time on AMU 0 10 20 30 40 0 40 80 120 160 200 Runtime [µs] No. of base points Runtime vs. Input dim. and No. of base points 4 5 6 required runtime for a model with 6 dimensions and 50 base points 0 0,4 0,8 1,2 0 40 80 120 160 200 Model error [°CA] No. of base points Model error (Sigma) vs. No. of base points decided trade-off between accuracy and runtime 76 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes The run time of the AMU call and return is similar to the run time of the functions that existed before and hence it was ensured that there exists no increase in runtime. Figure 21 shows the scheduling resulting from the above considerations. Figure 21: Scheduling of the model calculation(s) in asynchronous mode parallel to the main core processes 5 Conclusion We believe that the application of data-based models in the engine ECU is an absolute necessity to control complex state-of-the-art engine systems that are available in the market today. These models offer higher accuracy in representing the complex system behaviour thus ensuring that the full potential of such a system can be obtained. The software and calibration methodology that Bosch developed to implement data-based models on the ECU allows a stable software architecture which results in better software quality and drastically reduces the time to market because of a shortened V cycle. We presented that this approach can be used to solve versatile modelling problems in real-time embedded system applications. It is imperative of the users to be aware of the strengths and methods to overcome the limitations of data-based model when implementing them on the ECU. Acknowledgements The current state of ASC@ECU is based on a long-standing history, starting from the development of algorithms, hardware and drivers, to the deployment of software and model-based applications. These successes have been made possible in particular by Heiner Markert, Ernst Kloppenburg and Rene Diener, whose constant efforts in research and development have been essential in enabling this technology. Furthermore, we would like to thank our colleagues from calibration department Frank Ottusch and Olaf Dünnbier and function and software development Murali Krishna Prasad, Michael Rahn and Wolfgang Seils, for their vital support in the development of the basic ignition timing functionality. Finally, we would like to thank Thomas Obertopp, Andreas Huber and Dirk Patrick Hofmann-Mees for their constant support of our activities. AMU Call processes AMU Return processes Processes independent of Ignition setpoint structure 75us needed processes before need processes after AMU calculation of 1 st model AMU calculation of 2 nd model 43 μs buffer time 16 μs 16 μs 77 2.3 Implementation of data-based models using dedicated machine learning hardware (AMU) and its impact on function development and the calibration processes Bibliography [1] R. Diener, M. Hanselmann, T. Lang, H. Markert, H. Ulmer: Data-based Models on the ECU. Design of Experiment (DoE) in Powertrain Development , Expert Verlag 2015 [2] Lichuan Liu, S. M. Kuo, M. C. Zhou: Virtual sensing techniques and their applications. 2009 International Conference on Networking, Sensing and Control [3] E. Wilson: Virtual sensor technology for process optimization. 1997 Symposium on Computers and Controls in the Metals Industry in Iron and Steel Society [4] Automotive SPICE Process Reference and Assessment Model, Release 3.0, 16 July 2015. http: / / www.automotivespice.com/ [5] B. Hardung, T. Kolzow, A. Krugen: Reuse of Software in Distributed Embedded Automotive Systems [6] P. Leteinturier, S Brewerton, K Scheibert: MultiCore Benefits & Challenges for Automotive Applications. 2008 SAE world Congress Detroit, Michigan 78 3 Design of Experiments I 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine Yooshin Cho, Donghee Han Abstract There are many sources of pressure to innovate the calibration process efficiently. Complex technologies, reinforced regulations, various products, limited resources, shorten development period and so on. The global DoE model based calibration process is the one of the efficient processes of the engine calibration. In this research, the global DoE model based calibration process for the HMC’s V6 Lambda Engine will be introduced. That is the naturally aspirated engine, 3.0liter, Gasoline Direct Injection, Dual CVVT and 3-stage Variable Intake System. The continental EMS system is adopted. 1 Introduction Upcoming regulations such as WLTP and RDE require the OEM to reduce pollutant emissions. And the governments of the major markets declared the guide line for the fuel efficiency of the vehicle. In addition, customers demand exciting driving experience. In order to fulfill the strict regulations and to meet the high requirements about the fuel consumption and the performance, increasingly efficient engine should be developed. These reasons lead the engineers to integrate advanced technologies to the gasoline engine. Hyundai Motor Company’s gasoline engines adopted Turbo Charge System, Direct Injection System, Dual Continuous Variable Valve Timing System, Continuous Variable Valve Lift System, Variable Intake System, Exhaust Gas Recirculation System. These complex technologies contribute to enhance the gasoline engine performance and fuel efficiency. However, those systems also bring the huge calibration work. Calibration engineers should model corresponding systems and have to optimize the system to get the best result. In order to calibrate the complex engines with available engine test beds and man-hour within the development deadline, the conventional calibration processes are not adequate. In this paper, the Global Model Based Engine Calibration Process will be introduced, which is efficient and covers the engine’s operating range globally and which is the result of the effort of Hyundai Motor Company to develop the efficient calibration process. The new process is based on advanced modeling method, Gaussian process model which is supported by ETAS ASCMO. And the whole tests for measuring the engine data are conducted with the newly developed automation system. The new automation system is developed with ETAS INCA-FLOW. 79 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine 2 Target of Calibration and Global Engine Model The target engine was a naturally aspirated V6 3.0L GDI engine with a three-stage intake system, dual continuous variable valve timing, and a Continental engine management system. The calibration targets are the models for the following elements: - Air Charge - Torque - Exhaust Temperature In addition, the optimizations for the following elements: - Intake and Exhaust Camshaft Timing - Injection Timing and Fuel Pressure - Ignition angle - Lambda - Status of the variable intake system The global engine model will be consists of SPEED, MAF, state of VIS, intake cam timing, exhaust cam timing, start of injection timing, fuel pressure and lambda as input parameters and TORQUE, MAP, BSFC, COV of IMEP, Exhaust Temperature and ignition timing as output parameters. The Gaussian process modeling method is used to build the empirical engine model with ETAS ASCMO. Figure 1: The Schematic diagram of Global DoE Model Based Calibration for the Base Engine Calibration 80 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine 3 Conventional Calibration Process Figure 2: The Conventional Process for the Base Engine Calibration For the base engine calibration, the conventional process does the individual test for the each fuctions and need to measure the engine parameters for whole combinations of the relevant input parameters at various speed-load operating points. In sequence, handle the measured data to extract calibration values for the specific function. In series, individual functions are fulfilled in the similar way. Then the whole test time and the test points for the conventional calibration process is enormous. Especially, variable systems which are implemented in the engine make the calibration work increased exponentially. For example, in order to perform air charge model, measurements have to be fulfilled at sixteen speed grid points, ten torque grid points, eight incam grid points, six excam grid points and three status of the intake system. With simple multiplications, we can calculate approximately 23,000 test points is needed. Assuming that each measurement takes two minutes, the air charge model takes a total of 768 hours. 4 New Calibration Process The HMC’s new calibration process includes two new methods, the one is design of experiments(DoE) and modeling method and the other is test automation of the engien on the test bed. The test is designed to make the empirical global engine model which will simulate the behavior of the engine and provide the engine data to be used to genetate calibration data. ASCMO’s Gaussian process model shows highly accurate prediction data of the engine. Also, the global engine model will be used to optimize the engine parameters for fuel efficiency and the performance. The 81 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine measurement points for the test plans can be worked through fully automatically on the test bed with the newly developed automation system with INCA-FLOW. Fig. 3 shows the total test points and test time are reduced dramatically. Figure 3: The HMC’s Global DoE Model Based Calibration Process 5 DoE Test Plan for the Global Engine Model To get the acceptable accuracy of the empirical engine model from the smallest possible number of individual measurements, the input parameter combinations are distributed in a statistically optimum manner through the space filling method. However, to accelerate the test run, SPEED-LOAD operating points are clustered. The operating points in the low RPM range are concentrated to get more precise data in the area that is the major operating area of the 3.0L engine. In this project, the 3000 DoE test points are designed for the target engine by the ASCMO and 640 points are added to the test plan for the torque modeling. 82 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine Figure 4: The DoE Test Plan with Input Parameters 6 Automated Engine Test High performance measurement automation is the key to efficient measurement of the engine on the test bed. Especially, the repeatability of the determined ignition timing is very important. In this project the new automation system was developed with INCA-FLOW which the test engineer can script test sequence on his own purpose. The ignition timing is controlled based on the information from the combustion analyzer. This new automation system can protect the engine against the knock and the excessive exhaust temperature. The Knock is prevented by adjusting the ignition timing with the close loop controller based on the knock index from the combustion analyzer. The exhaust temperature is limited by controlling the fuel mass with the temperature from dynamometer controller. Figure 5 shows the configuration and the interaction among the ECU, the INCA, the INCA-FLOW, the combustion analyzer (AVL INDICOM) and the dynamometer controller. The test sequence is also illustrated in the Figure 5. 83 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine Figure 5: Configuration for the measurement automation system the flow diagram for the sequence control Figure 6: The flow diagram for the sequence control 84 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine 7 The Empirical Engine Model (Gaussian Processes) Based on the measured engine data, Gaussian processes is used to build the mathmatical engine model which simulate the engine behavior over the entire operating area. In this project, ETAS ASCMO is used to make the empirical engine model. The quality of the global engine model can be checked with comparing predicted model data with measured data. The model error can be checked in 3 different ways. First, it can be checked with the training data which is used to make the model. Second, there is leave-one out error which means the error between the measured data and the predicted data which comes from the model without the specified measured data. The leave-one out error is good to check the overfitting. Finally, there is the model error for the test data which is not used to make the model. This type of error is effective to check the developed empirical model has the expected accuracy. The global engine model for the target engine of this project has the quite good quality. Table 1 shows the leave one out error of the empirical engine. Figure 7: The empirical engine model Table 1: Leave-One-Out Model Error of the empirical engine model 85 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine 8 Calibration based on the engine model The empirical engine model created by Gaussian processes modeling simulates the behavior of the engine with high accuracy over the entire parameter space. On the basis of the model, the optimization calibration can be done directly in the DoE tool ASCMO with the provided optimization function. In the part load area, the input parameters are optimized for fuel efficiency with the constraint of engineering criteria for the target engine like knock, exhaust temperature, emissions, combustion stability and the gradient of the parameter change. In the full load area, the optimization target will be the full load performance with the similar engineering constraint. The calibration tables of the optimal values of the input parameters can be derived directly. Then we can get the following optimal calibration tables: - Intake Cam Timing - Exhaust Cam Timing - Status of the variable intake system - Start of Injection of 1 st injection - Fuel pressure - Lambda - Ignition timing For the ECU engine modeling calibration, the simulation data of the empirical engine model can be used as the test data which is needed to fill the calibration tables of the ECU model. With the proper fitting tool, the ECU models of the following elements can be generated: - Air charge model - Torque model - Exhaust temperature model Figure 8: Calibration process based on the engine model 86 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine 9 Demonstration of the Calibration Data The calibration quality is demonstrated with the test of overall operating area logging same as the way the conventional calibration process does. The test shows the result with under the 5% error of the air charge model and under the max(5%, 5Nm) error of the torque model. The Exhaust Temperature Model has the error under 15 ℃ over the full operating points. Moreover, the optimization results for the fuel efficiency and the full load performance are satisfied with the project target. Figure 9: The difference between the calibrated ECU model and the measured data of the air charge model 10 Conclusion In the research and development center in Namyang in South Korea, Hyundai achieved a dramatic increased efficiency with the newly designed global model based calibration process. In a specific calibration project of the target engine, naturally aspirated, V6, 3.0L and gasoline direct injection engine, the measurement effort on the test bed is reduced to a quarter of what it was using the conventional process. At the same time, it achieved the project target and acceptable calibration results for the target engine. The empirical engine model generated from the test bed measurement data using the ASCMO with Gaussian processes predicts the behavior of the gasoline engine quite well. And the optimization with this empirical model shows good result. The new automation system developed for this project also enables us to measure all test data efficiently. 87 3.1 The Global DoE Model Based Calibration and the Test Automation of the Gasoline Engine Figure 10: The efficiency increase with the global DoE model based calibration process 47.200 9.400 1.361 326 0 200 400 600 800 1.000 1.200 1.400 1.600 0 5.000 10.000 15.000 20.000 25.000 30.000 35.000 40.000 45.000 50.000 Conventional Global Model Based Calibration Efficiency increase through model-based calibration Test Points (ea) Test Time (hrs) 88 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration Jan-Christoph Goos, Matthias Brumm, Sebastian Weber, Frank Kirschbaum, Yagiz Dursun, Thomas Koch Abstract An optimization procedure is presented which allows optimizing the map and curve sampling point positions in addition to their values at the same time. Past contributions in the field of model-based calibration of ECU parameters with empirical model approaches concentrate especially on new DoE methods, mathematical model approaches and approaches to optimize mapand curve sampling point values of the ECU parameters fast and accurately. With these new methods, the global powertrain system behavior can be modeled within a wide operating range with high quality. However, all optimization approaches currently only use discrete operating points on which these empirical models are evaluated during the optimization of the ECU maps. In addition, these approaches are limited to the optimization of map and curve sampling point values only. The sampling point positions are not changed during the model based optimization. In contrast, the new optimization approach, described in this paper, allows optimizing the sampling point locations in addition to the values. Therefore, the sampling point values and positions are set to get optimal empirical models of the target criteria and the map and curve roughness. The approach uses the whole information of the global target models. In contrast to a model evaluation of discrete operating points, an average model value is estimated inside the map operating range with an integral evaluation. This value is one component of the target function minimized during the optimization. Besides the typical constraints used during the optimization of ECU maps and curves with model based approaches, additional constraints for the map and curve sampling point positions can be defined. The new approach is applied to an example in the field of drivability calibration. Kurzfassung In diesem Beitrag wird eine neuartige Optimierungsmethodik beschrieben, mit der neben den Stützstellenwerten von Steuerkennfelder und -kennlinien auch gleichzeitig die Positionen der Stützstellen optimiert werden. Bisherige Arbeiten im Bereich modellbasierter Kalibrierung von Steuerparametern auf Basis von empirischen Modellansätzen beschäftigen sich insbesondere mit neu- 89 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration en DoE-Methoden, mathematischen Modellierungsansätzen sowie Ansätzen zur schnellen und exakten Optimierung der Kennfeld- und Kennlinienwerten der Steuerparametern. Diese neuen Ansätze bieten die Möglichkeit das Systemverhalten global, über einen weiten Betriebspunktbereich mit guter Qualität abzubilden. Sämtliche Ansätze bei der Optimierung der Steuerkennfelder auf Basis dieser empirischen Modelle verwenden jedoch bis dato lediglich diskrete Betriebspunkte, an denen diese Modelle während der Optimierung ausgewertet werden und beschränken sich zudem stets auf die schnelle Berechnung optimierter Stützstellenwerte der Kennfelder und Kennlinien ohne jedoch die Positionen dieser Stützstellen während der modellgestützten Optimierung zu verändern. Das in diesem Beitrag beschriebene Optimierungsverfahren ermöglicht dagegen erstmals ebenfalls die Optimierung der Stützstellenpositionen, indem die Positionierung und die Werte der Kennfelder so ermittelt werden, dass diese hinsichtlich der zugrundeliegenden empirischen Modelle der Zielgrößen sowie der Kennfeld- und Kennlinienrauheit optimiert werden. Der Ansatz hierfür ist die erstmalige Ausnutzung des gesamten Informationsgehalts von globalen Modellen der Zielgrößen, indem im untersuchten Kennfeldbereich nicht nur einzelne diskrete Betriebspunkte betrachtet werden, sondern ein durchschnittlicher Zielmodellwert mittels einer integralen Auswertung berechnet wird. Dieser wird als Komponente der Zielfunktion bei der Optimierung genutzt. Weiterhin werden bei diesem Verfahren neben den Nebenbedingungen, die bei den bisherigen Ansätzen ebenfalls Verwendung finden, zusätzliche Bedingungen bei der Optimierung berücksichtigt, die sich auf die Stützstellenpositionierung beziehen. Das beschriebene Verfahren wird an einem Beispiel aus dem Bereich der motorseitigen Fahrbarkeitskalibrierung angewendet. 1 Introduction The model-based calibration of engine or transmission control units (ECU/ TCU) is divided into several process steps (see Figure 1 and [1, 2, 3, 4, 5]). The common process includes the steps of the test planning, the measurements on the test bench, the data analysis and modeling, the optimization, the calibration map and curve calculation and the validation of the optimized calibration settings [1, 2, 3]. These steps are discussed briefly in this chapter. During the test planning, variations of calibration parameters sets are defined. This variation list is usually calculated by the use of Design of Experiment (DoE) [6, 7]. Thus, after the measurements are done on the test benches the empirical models of the system’s target criteria can be estimated with an adequate quality. Several DoE methods concentrate on generating a test plan for polynomial models, like D-Optimal DoE test designs [8, 9, 10]. According to that, polynomial models, which can be already utilized during many calibration tasks like in drivability or engine emission calibrations with local operating point models, are still one of the most common model approaches because of their simplicity [1, 4, 6, 11, 12, 13]. Those local operating point models are estimated from local DoE test plans, in which the operating point (OP) of the powertrain is set to fixed values and only the calibration parameters are varied. Thus, the often nonlinear influences of the powertrain OP variables on the 90 3.2 modele eling [6 modelbehavio Fi Neural archica 5, 6, 14 equate respect OP var culated ied in c A furthe standa sequen ori with calibrat 19]. Aft mated. based o first mo the mo Adding desired DoE te calibrat priori to informa bration models Optimization ed powertr 6, 10, 11]. based cali or. igure 1: Sta networks, al local mod 4, 15]. It ha ly with the t, global m riables and d with resp combinatio er improve rd DoE ap ntially in dif h statistical tion param ter the me Afterward on the crite odel. This del estima variations d model qu est plannin tion engine o get a we ation of the paramete s and onlin n of ECU Ma ain calibra . However bration pro ate of tech Gaussian del trees (H as been sh ese model modeling is d the calibr pect to the n with the ement of th pproaches, fferent pro l criteria. W meter varia easuring fo ds, a new c eria of the variation i ation is upd s and upda uality is re g and mod eer does n ell modeled e powertra ers in a w e DoE app ap Sampling ation target r, several a ocess and hnology for process m HiLoMoT) hown that approache s to model ration para OPs and t calibration he process the test p cess steps With online ations and or a small combinatio online DoE s automat dated with ating the m eached. In deling, less ot need to d system b in system ide OP ra proaches. Point Values t criteria ar advancem the capab r model-bas models (G are for exa the global es in a goo the system ameters [1 the calibra n paramete s are so-ca planning, m s so that th e DoE app the mode initial test on of calib E approac tically set a the additio model esti contrast t s knowled o determine behavior [5 behavior is ange by us s and Positio re not take ments were bility to mo sed calibra PM) or loc ample non system be od approxi m’s target , 17]. Henc ation param ers during t alled online measurem he test pla roaches, t eling are d plan is fin ration para h in order and measu onal inform mation are to the stan ge about t e a proper 5, 19]. Thu s available sing the co ons with Mod en into acc e develope del a more ation at the cal model n linear regr ehavior can mation [1, criteria as ce, the Do meters and the measu e DoE app ent and m n optimalit he plannin done simul nished, a f ameters an to improve ured on th mation of th e automati ndard sequ the system amount of s, valid an e in depend ombination del-Based Ca count durin ed to impro e complex e Daimler A networks li ression mo n be mode 5, 13, 16] s a function oE test plan d the OPs a ring. proaches. W modeling ar ty is define ng of new O ltaneously first model nd OPs is e the qualit he test ben his variation ically done uential pro m is require f measurem nd compre dence on t n of sophis alibration ng modove the system AG ke hierodels [1, eled ad- ]. In this n of the n is calare var- With the re done ed a pri- OP and [5, 18, is estichosen ty of the nch and n result. e until a ocess of ed. The ments a hensive the calisticated 91 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration The next process step of the model-based calibration is the optimization of the calibration parameters as seen in Figure 1. Therefore, the model predictions of several target model criteria need to be considered. Usually some target model criteria are part of the objective function and are optimized while others are defined as constraints during the optimization [6, 16, 20]. The optimal settings for the calibration parameters of the powertrain’s ECUs mostly depend on and vary with the OPs of the powertrain. Hence, target model criteria are always optimized on several OPs [6, 20]. In the case of local OP models, several different OP models are used during the optimization and the weighted sum of these OP model values is optimized in most cases [6, 16, 20]. For global target models, which are valid inside an OP range, the calibration engineer needs to define a list of OPs, on which the weighted sum has to be optimized usually with a similar approach than for local OP models. These optimized OPs are often either chosen based on a calculation of representative cluster OPs, which are important during the certification, or they are defined by the sampling point (SP) positions, on which the calibration parameter maps and curves are defined [6, 16, 21, 22]. After the optimization, the optimized calibration parameters on the target model OPs are used to calculate the actual objectives of the model based calibration process, the calibration parameter maps, curves or scalars [3, 7, 16, 20, 23]. This sequential process is often time consuming because the two steps of optimization and map calculation usually have to be conducted multiple times to get an adequate trade-off result that optimizes the target model objectives and the calibration parameter map and curve smoothness. In addition, there are usually still deviations between the optimized calibration parameters on the OPs and the interpolated OP values inside the smoothed maps and curves [12]. These deviations might cause non-optimal target objective criteria or even violations of the optimization constraints. Rather than optimizing OP calibration values and calculating the corresponding maps, curves and scalars in two separate sequential steps, the optimization can be performed directly onto the actual calibration parameter maps, curves and scalars used in the ECU functions [11, 12, 16, 22, 24, 25, 26, 27]. Therefore, the trained empirical models of the target model criteria on the defined OPs are used to get optimal calibration maps, curves and scalars so that this calibration can be directly applied inside the ECU (see Figure 1). Besides the target model criteria, the map and curve shape regarding to smoothness can also be considered during the optimization within this approach to meet the requirement to have the calibration parameter maps and curves as smooth as possible for reasons of drivability. The disadvantages of the sequential optimization and map calculation steps are reduced or even avoided with this approach (cf. Figure 1). In [12] the optimization problem of the calibration parameter maps and curves SP values is adjusted to get accurate target model objective estimations inside the maps and curves. Moreover, with this approach, the optimization degree of freedom is reduced to accelerate the optimization. In addition, an analytical estimation of the gradient and the hessian for the objective and constraint functions are used to gain optimization speed with gradient-based optimization algorithms [12]. With these advancements, it is possible to optimize the calibration parameter map and curve SP values efficiently [12, 28]. 92 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration Nevertheless, the current approaches focus only on the optimization of SP values. The SP positions of the calibration maps and curves are fixed to the initial values and, thus, are not optimized. An approach of calculating SP positions and values, which finds an optimal set of SP values and positions for a given discrete set of OP values, was already introduced [14]. Thus, this approach can be used to calculate map SP positions and values from optimized OP values within the sequential approach and the two separate steps of target criteria optimization and map calculation. However, a direct optimization of calibration maps and curves on the basis of modeled target criteria is not possible with this approach. Thus, this sequential process would again lead to the mentioned disadvantages of a slow sequential process with inaccuracies due to deviations between the optimized OP values of the calibration parameters of the first step and the interpolated values inside the maps of the second step [12]. This paper presents a novel evaluation approach for global models, which enables the optimization of SP values and positions. This approach uses the entire information of global models with a calculation of an average target model criterion value inside an OP range based on an integral model evaluation. Contrary to current optimization approaches of SP values with a target model evaluation on discrete OPs, the optimization problem gets defined continuously. With this novel approach, the optimization of calibration map SP values and positions can be done simultaneously. Inside the following chapter, the basic idea of the integral evaluation of global target models is introduced. The OP range is analyzed based on the DoE design space in chapter 3 to get a valid evaluation, before the objective function of calibration map SP optimization is described in chapter 4. In chapter 5, the new approach is applied to a calibration use case of drivability functions inside the ECU of a passenger car. The benefits and drawbacks of the new approach are discussed. In chapter 6, the results are summarized and an outlook on the further development is given. 2 Integral evaluation of global target models The integral evaluation approach is used to get an average value of the empirical target model criterion within a defined OP range of the powertrain. An empirical global model criterion depends on the values of different ECU function calibration parameters and the powertrain OP variables. In general, the number of OP dimensions depend on the calibration use case. Usually, the OP variables regarding powertrain calibration are defined by the values of speed and load. This standard case is presented in this paper. Thus, the value of a global model criterion can be calculated in equation (1) by using the two OP variables and : ( ) , ... , , ... , , , 1 q k t t w w w v u f f = (1) With this equation it is possible to get the estimated value for the target criterion with a given set of values of the calibration parameters on a fixed OP. Each calibration parameter value is defined by a curve or a map in dependence on the OP variables. Calibration parameter SP values are stored on defined SP positions. The calibration value on an arbitrary OP is usually linear interpolated inside a curve and bilinear interpolated inside a map. For a calibration map, the interpolation is computed with the four surrounding SP values , the locations of each SP and and the location variables of the OP and (see Figure 2 and equation (2)). 93 3.2 Fig ( = k u w Empiric uous a gral va be calc values penden calibrat The int the em get mo suming tions, th val of , F t j i = To get value is (4)). Conseq range d averag is secti the ove Optimization gure 2: Exa ( ) ( 1 1 , , , − ⋅ − = j j k k a x x v u ( x u a , for cal model and continu lue can be culated ins (see Figur ncies chan tion param terpolation pirical mod odel is calc g that all th he integral to a ( 1 1 u f i i j j y y x x t ∫ ∫ + + the averag s divided b f quently, th depends o e model va onally app erall surfac n of ECU Ma ample of bi ) , 1 , , + + ⋅ − j i i i j k k k z b y y ) j j k k x x = + 1 , approache uously diffe e calculated side a calib re 2). On t nge to diffe meter value function fo del equatio culated with he calibrat l over the t nd to ( , , , , 1 x v u w v u ge target m by the sur 1 , , j i t j i F S f = he average on the locat alue for m plied to an ce of the O ap Sampling ilinear inte ( ) , 1 1 , , + ⋅ − ⋅ j i i k k z a b z z j j k k k x x x u − − = + 1 es like poly erentiable d on the in bration par he edges erent SPs. is not con or each ca on (1). With h the two O ion param two OP dim can be c ) ,..., , , 1 w z y x model valu rface on w , t j i F with e target m tions and v ore than a amount of P range [2 Point Values rpolation in , 1 , , 1 , + + + ⋅ + j i j i j i k k k b z z j k and ynomial mo regarding nterval of th rameter m of the qua Thus, the ntinuous an alibration pa h this appr OP variabl eter maps mensions i calculated ( , , , , y x v u w k ue inside th which the in h ( j j i x S = + , model valu values of e a single ma f quadrant 29, 30]. s and Positio nside a cal ( ) , 1 , 1 1 1 , + + + ⋅ − j i j i k k z a z ( i i k k y y v b , , odels, GPM the input he two OP ap quadra adrant, cali e bilinear i nymore (se arameter m roach, the les as the s and curve inside a qu with equa ) ( ,..., , u w z q k he quadran ntegration )( i j y x − + + 1 ue inside t every calib ap or curve ts (see cha ons with Mod libration pa ) ( 1 1 , , + ⋅ ⋅ + = j k z b a v u w ) i k y y v − = + + 1 1 Ms or HiLo argument P variables. ant enclose bration par nterpolatio ee equation map or cur integral of variables o es have th uadrant def tion (3) [29 , , , , z y x v u q nt of integr is calculat ) i y − + 1 the define bration para e OP quad apter 3) an del-Based Ca arameter m 1 , 1 , , , + + j i k k k k z z y x i i k k y y − oMoTs are ts. Thus, t . This integ ed by the f rameter va on formula n (2) and [ rve is inser f the empir of integrat he same S efined on th 9, 30]. )) d d v u ration, the ted (see e ed powertr ameter. To drant, equa nd then div alibration map ) (2) e continhe integral can four SP alue defor the 29]). rted into rical tarion. As- SP posihe inter- (3) integral equation (4) rain OP o get an ation (3) vided by 94 3.2 In man nation (with a quadra rithm fo Kronrod With th model c trast to amoun finite s range. 3 A In orde tionally empiric Figur In the p the cal es, this can be change calibrat togethe ramete position Optimization t f ny cases, t of the bilin n output tr ant. Since t or the num d quadratu his approa criteria by o prevalen t of OPs is um of mo Analysis er to calcul y (see chap cal model i re 3: Calibr previous c ibration pa s condition e evaluated es on each tion map a er. In Figu ers, which ns covered n of ECU Ma 1 1 1 ∑ − = = n i t S f his integra near interp ransformat the implem erical eval ure formula ach, it is p calculating nt approac s used, the del evalua s of the v late a valid pter 2) and s extrapola ration exam d chapter the arameters n cannot be d, have to h location and curve re 3, the p are stored d up by bot ap Sampling 1 1 1 , ∑ − = m j t j i F al function olation for tion). Thus mentation is uation of a a, is used [ possible to g an avera ches, in w e new appr ations to g valid ope d model va d inside th ated. mple with t different sa e assumpti have iden e satisfied be analyz of a map SP locatio procedure d in calibra th maps ar Point Values with S = cannot be rmula and s, this func s done in M a double in [31, 32, 33 o use the age model which a su roach of in get the true erating ra alue, the in e design s two calibra ampling poi ion was in tical SP lo . Thus, the zed. Since or curve S ons of all c is shown ation maps re marked s and Positio ( )( 1 y x x m n − = e solved an the used e tion is num MATLAB, t ntegral, wh 3]. entire info value insid um of mod ntegration e average ange ins ntegration space of th ation param int position ntroduced t ocations. In e sections e the (bi-)l SP of each calibration by an exa s with diffe . These ve ons with Mod ) 1 y m − nalytically d empirical m merically in the standa ich is base ormation of de an OP r del evalua can be inte model va ide the d can only b he DoE tes meters stor ns that all ma n many ca , over whi inear inter h calibratio parameter ample of tw erent SP p ectors of SP del-Based Ca due to the model appr ntegrated f ard MATLA ed on a Ga f global e range [29]. ations on erpreted a alue inside design sp be evaluat st. Otherw red in maps aps and cu alibration u ch the inte rpolation e on parame rs are con wo calibrat positions. T P positions alibration (5) combiroaches for each AB algoaussianmpirical In cona finite s an inan OP pace ted secwise, the s with urves of use casegration equation eter, the nsidered tion pa- The SP s define 95 3.2 the lim and bla values The de tion. A also ap parame A varia and a s In the n tion qu combin be foun can be tive-set OP has hull de this poi tic of ex the con id edge as well OPs, w age tar age va eter ma chapte 4 Np Figure The go maps, tem. Th Optimization its of each ack edge p for each O esign space popular m pplied duri eter variati ation includ set of linea next step, uadrant are nation on a nd that sa found by t strategy s at least sign space int (cf. the xtreme val nvex hull d e OPs. Co and will n will be integ rget model lue is the b ap and cur r. Novel op point val 4: Distinct and gro oal of the curves or he present n of ECU Ma h potential points in F OP dimens e of the Do method for ng this ap ons for wh des the ca ar inequality A the OP po e inserted an edge po atisfies the solving a l [34, 35, 36 one valid e and the black edg ues inside esign spac nsequently ot be extra grated with l criterion v basis of th rve SP pos ptimizatio lues and tion betwee oup samplin model-ba scalars wh ted novel a ap Sampling section o Figure 3. O ion, equati oE test ha the desig pproach. F hich a mea alibration p y constrain [ w v u A ,.. , , 1 ⋅ osition valu into equat oint can on inequality inear prog 6]. If a sol solution fo empirical t e points of e at least on ce is also v y, the integ apolated. A h the appro value of th e optimiza sitions and on appro d position en individu ng point ax ased calibr hich are op approach o Point Values of numerica Over each ion (3) can as to be co n space e irst, one c asurement parameters nts defines w w q k ,..., ., ues and tion (6) (se nly be eva y equation gramming p ution for th or the calib target mod f Figure 3) ne map or valid inside gral model All quadran oach of equ he valid OP ation object d values, w oach for ns ual samplin xis for two ration proc ptimized re optimizes S s and Positio al integrat quadrant n be calcula onsidered t evaluation convex hul was done s and the s the bound ] b T ≤ for each ee Figure luated if an s [29]. Th problem w he problem bration pa del criteria ). The edge curve rega e the quad evaluation nts, which a uation (3) P range (s tive functio which will b ECU ma ng point ax calibration cess is to egarding th SP position ons with Mod ion and ar bounded b ated [29]. to get a va is the con l is calcula during the OP variab ds. h edge OP 3). A calib ny solution e feasibilit with the first m exists, th rameters i will not be e OPs hav arding the rant of fou n over this are surroun and summ ee equatio on for the c e presente ap and cu xis per calib n paramete find calib he target c ns of ECU del-Based Ca re shown by two uni alid model vex hull, w ated based e global D bles (globa P of every bration pa n for to ty of this p t phase of he analyze inside the e extrapol ve the char OP values ur surround s quadrant nded by fo med up to a on (5)). Th calibration ed in the fo urve sam bration par er curves bration pa criteria of t maps and alibration as grey que SP evaluawhich is d on all oE test. al DoE) (6) integrarameter can problem f the aced edge convex ated on racteriss. Thus, ding valis valid our valid an averis averparamollowing mpling rameter rameter the sysd curves 96 3.2 in addit maps a amoun optimiz Two di need to tion pa the so vector defined change bration SP loca gle SP which a and cu have d Fi Figure with se curve v map an an opti range f mized SP valu ramete initial v the SP the val Optimization tion to the and curve t of optimi zation appr fferent typ o be consid rameters i called “gro changes th d separate e in this SP map or c ations per axis use are the ad rve SP val ifferent ind igure 5: Ex cali 5 shows a eparate SP values and nd curve v imization a for map or SP values ues to be er has four values and locations id range a n of ECU Ma SP values es are add ization deg roaches (se pes of SP dered diffe nside a sin oup sampl he SP loca ly for each P vector o urve (see OP dimen case, ever dditional inp ues of eac dividual val xample of o ibration pa an example P axes. Th positions. alue which argument r curve , t s. In this ex optimized, SP values are no opt that are in are kept fix ap Sampling s [30]. Thu ded to the gree of fre ee for exam axis are c erently duri ngle ECU ing point a ation of all h calibratio nly change Figure 4). nsion is ad ry calibrati puts for th ch calibrati ues for ea optimized s arameters w e of the op he valid O It is detec h is located . If n the neighb xample, th , one is no s consider timization nside the v xed. For the Point Values us, the SP e inputs o eedom incr mple [12]). commonly ng the opt function of axis”. A ch calibration on parame es the loca For the g dded to the on parame he optimiza on parame ach SP of e sampling p with individ ptimization OP range cted by the d inside or o value is bored outsi he map of ot varied. T red during arguments valid OP ra e map of t s and Positio positions f the optim reases in c . used insid imization. ften use a ange of a n paramete eter map a ation of the group SP u e optimizat eter has its ation algor eter are inp each map a point locatio dual sampl inputs for is crucial e method d r on the ed located o ide-located the first ca The curve the optimi s. For the o ange are o the first ca ons with Mod of the calib mization a compariso de an ECU Maps and n identical position in ers. An ind nd curve, e correspo use case, tion argum s separate ithm. Furth puts for the and curve. ons and va ling point a r four calib for the op described i dge of this on the edg d value is a alibration p of the fou zation. Tw optimized S optimized. alibration p del-Based Ca bration pa algorithm a on to conve U and bot curves of SP positio n a group dividual SP so that a onding sing a single v ments. For e position v hermore, t e optimizat alues for fo axes ration para ptimized m n chapter valid OP r e of the v added to t parameter rth calibra wo remain SP positio SPs on e arameter, alibration rameter and the entional th types calibraon axis, SP axis P axis is position gle caliector of the sinvectors, he map tion and our ameters map and 3. Each range is alid OP he optihas 19 tion paon their ns, only dges of two SP 97 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration positions for the dimension and three SP positions for the dimension will be optimized. For the other map and curves, the amount and location of the optimized SP positions are different. Other SP values that are not used as optimization inputs can be fixed to the corresponding reference value or calculated in dependency of the optimized SP values and positions. A similar approach like in [12] is used to find the smoothest solution for the current optimized SP variables. The following descriptions are limited to the use case of individual SP axes for the sake of clarity. The optimization’s objective function contains two different weighted components [30]. A target value cost component and a roughness cost component . These components are already used in several other publications [12, 27, 28, 29]: ( ) , , ,..., , , ,..., , , 1 , , ,..., , , ,..., , , opt opt opt opt opt opt opt 1 opt 1 opt 1 R obj opt opt opt opt opt opt opt 1 opt 1 opt 1 Tg obj obj ⎟⎠⎞ ⎜⎝⎛ ⋅ − + ⎟⎠⎞ ⎜⎝⎛ ⋅ = q q q k k k q q q k k k z y x z y x z y x S z y x z y x z y x S S ϕ ϕ (7) This objective function is optimized with the described degree of freedom: min arg obj , , ,..., , , ,..., , , opt opt opt opt opt opt opt 1 opt 1 opt 1 S q q q k k k z y x z y x z y x (8) The first component of the objective function is based on various average empirical model target criteria values of the powertrain calculated with the integral approach (see chapter 2). E.g. such a value can represent the average fuel consumption or an average emission of the engine. Each average model target criterion value is divided by the value evaluated with the reference calibration parameter maps and curves to analyze a relative change of each average target model value. In addition, the resulting relative values of the average target criteria values can be weighted by t ϕ : , , ,..., , , ,..., , , 1 reference , opt opt opt opt opt opt opt 1 opt 1 opt 1 Tg ∑ = ⎟⎠⎞ ⎜⎝⎛ ⋅ = T t t q q q k k k t t f z y x z y x z y x f S ϕ (9) It has to be mentioned, that the integration quadrants for the approach introduced in equation (3) and (5) can consist of smaller sections than defined by each calibration parameter SP position vector (see also chapter 3). On every SP location of one of the calibration parameter maps or curves, other calibration map or curve values without a SP are interpolated with their surrounding SP locations and values [29, 37]. The roughness cost component of the objective function contains for each calibration parameter map or curve a value that describes its roughness. It is based on the values of the gradient and curvature in and -direction on each SP that are individually squared and then summed up (compare [11, 12, 27, 38, 39]). These roughness values depend on the calibration parameter SP values but also on the SP locations. In another paper, the calculation was defined with a vector of SP values and a constant matrix [11, 12, 27, 38]. With variable SP positions, this matrix also depends on the location vector [30]. Equation (10) defines the roughness value for each map and curve . ( ) ( ) , , , opt opt opt opt opt opt opt R k k k k k k Tk k k k k z z y x R z z z y x c ⋅ ⎟⎠⎞ ⎜⎝⎛ ⋅ = ⎟⎠⎞ ⎜⎝⎛ (10) 98 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration Like for t f , each calibration parameter roughness value is divided by the corresponding reference value evaluated for the reference calibration map or curve. The weighted sum of these relative roughness values of all calibration maps and curves defines the total roughness cost value . , , , , 1 reference opt reference opt reference opt reference R opt opt opt R R R ∑ = ⎟⎠⎞ ⎜⎝⎛ ⎟⎠⎞ ⎜⎝⎛ ⋅ = q k k k k k k k k k k z y x c z y x c S ϕ (11) In addition, constraints can be considered during the optimization. Other average empirical model target criteria, which are calculated with the integral approach, model target values on specific OPs or constraints for roughness values of the calibration parameter maps or curves can be defined. The optimization approach can satisfy the design space based on the global convex hull on every SP location of every map and curve inside the valid OP range. Therefore, the convex hull is evaluated on every SP of every map and curve with equation (6) (see chapter 3). These are the extreme points for at least one calibration parameter curve or map with a salient point and crucial for the convex hull evaluation. Due to the fact that every calibration parameter map and curve can have different SP locations, the number of convex hull evaluations is equal to the number of optimized SP positions of every calibration parameter plus the number of edge OPs of the valid design space (see chapter 3). In contrast to equation (6), the convex hull evaluations become nonlinear inequality constraints for the optimization because of the interpolated values that have to be calculated for the maps and curves that do not have a SP position located on the considered OP. Another new type of constraints for model-based calibration addresses the limits of the optimized SP positions [30]. The varied locations of each calibration parameter map and curve need to stay in the initial sorted order so that the locations do not pass each other and stay inside the boundaries defined by the edges of the valid OP range. For the map of the first calibration parameter in the example of Figure 5, the following constraints for the optimized SP positions are considered additionally during the optimization. A minimum distance between the optimized SP locations can be parametrized: ub opt 1 opt 1 opt 1 lb ub opt 1 opt 1 lb 3 2 1 2 1 , , , v v u u b y y y b b x x b < < < < opt 1 opt 1 opt 1 opt 1 opt 1 opt 1 3 2 2 1 2 1 y y y y x x < + < + < + ε ε ε (12) The optimization objective function and the constraints are implemented in MATLAB and different optimization algorithms can be chosen by the calibration engineer. A swarm intelligence based algorithm called particle swarm optimization (PSO) can be used, which is able to find the global minimum for calibration tasks with suitable optimization hyper parameter settings [40]. The disadvantage of such a global optimization approach is the required high amount of function calls and iterations. Thus, another optimization algorithm is fmincon from the MATLAB Optimization Toolbox, which uses gradient information to find a local optimum efficiently [41, 42, 43, 44]. This algorithm optimizes a nonlinear objective function of equation (7) and can consider the formulated nonlinear and linear inequality constraints. To lower the risk of finding an inadequate local optimum for the calibration parameter map and curve SP 99 3.2 position calibrat other a 5 C As a u car's E shown. proach and cu referred If the d load ch betwee can osc of the v induce low acc With se fect the Tip-in d sition s Such a are dis speed sponse parame Figure wit It has a can be calibrat Optimization ns and val tion param algorithms f Calibrati se case, a ECU which . This exa introduced urve SP va d to [28] fo driver requ hange occ en the eng cillate duri vehicle. In negative e celeration everal calib ese drivabi driving man step from z a maneuve abled, a ti to oscillate e of a simi eter setting e 6: Tip-in th enabled already be successfu tion approa n of ECU Ma lues, the a meter map for this cal ion test c a calibratio h influence mple has d in [12]. I alues effic or a detaile ests a pos curs on the gine and th ng the load addition, b effects on [37]. Thus bration par lity functio neuvers ar zero to a d er can be s p-in cause e with the lar tip-in m gs focused to 60% pe d and disab en shown, ully calibra aches [13] ap Sampling algorithm c ps and cur ibration op case on of the l es the resp already be ts focus w iently. The d descripti sitive torqu e powertra he driven w d change w backlashes drivability. , ECU func rameters d ns and the re usually defined end seen in Fig es the sum e natural fr maneuver w on comfo edal positio bled drivab , that the p ated on veh ]. The deve Point Values can be sta rves. Neve ptimization longitudina ponse of t een prese was on find e calibratio ion. ue with the ain. Due to wheels an with a neg s inside th . Those ba ctions are depending e resulting used to ca d value ca gure 6. If t m of the dri requency o with enable rt is also s on and the bility ECU f parameters hicle powe eloped pro s and Positio arted multip ertheless, problem in al dynamic the powert nted in [28 ing optima on exampl e accelerat o the limite d tires, the ative impa e transmis acklashes used to re on the OP impact on alibrate the auses com the corresp ven half s of the pow ed drivabil hown. load chang functions (c s of the loa ertrain test ocess is als ons with Mod ple times w it is plann n future. cs function train durin 8] with the al calibratio e is introd tor pedal d ed stiffness e drivetrain act on comf ssion and t need to be educe thes P values, it drivability. ese parame prehensive ponding dr haft torque wertrain. T ity function ge reaction cf. [28] and ad change benches w so used du del-Based Ca with variou ned to imp ns of a pas ng load ch e optimizat on paramet duced brief during coa s of the dr n and the fort and dr the final dr e traversed se effects [ t is possib . eters. A pe e oscillatio rivability fu es and the The powert ns and cal n of a pow d [13, 37, 4 e related fu with mode uring the p alibration us initial plement ssenger ange is tion apter map fly. It is asting, a rivetrain vehicle rivability rive can d with a [13, 37]. le to afedal poons [13]. unctions e engine train relibration wertrain 45]) unctions el-based present- 100 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration ed test case and the criteria are calculated from the dynamic measurements with the software tool described in [13]. Four scalar criteria characterize the quality of the powertrain response to a tip-in maneuver in the different target dimensions. The calibration engineer faces the trade-off between a comfortable load change reaction and a dynamic response behavior of the vehicle with a fast acceleration. The first criterion that describes the comfort dimension is . This criterion rates the powertrain oscillations inside the frequency range that is perceived negatively by the human passengers [13]. The second criterion, the jerk, is the maximum gradient of the vehicle acceleration during the tip-in. The two criteria that evaluate the agility dimension are the torque buildup and the critical time [13]. is the required time to reach a specific percentage of the requested final acceleration. With the torque buildup value, the deviations between the real engine torque and the engine torque which is requested by the pedal position step of the driver are calculated. The deviations are summed up over the time and scaled to the torque buildup criterion. A global space-filling test plan with 433 different tip-in maneuvers was estimated in [28]. Five calibration parameters of the corresponding functions were varied. Two of them are defined within maps. The three other calibration parameters are defined by curves. In addition, the OP variables, the engine speed at the beginning of the tip-in and the pedal position step, were varied. Like in [13], the four criteria are modeled properly in dependency of the calibration parameters and OP variables with third order polynomial models. Table 1: Optimization results in comparison to the reference calibration novel optimization approach of SP locations and values optimization approach of SP values used in [28] Optimized target criteria Torque buildup Torque buildup Torque buildup Torque buildup Calibration smoothness none medium none medium Torque buildup +11.98% +11.3 % + 5.94% + 4.39% +32.98% +32.37% +33.83% +23.32% Jerk +14.47% +19.72% + 5.63% + 3.55% + 0.00% + 0.01% + 3.64% + 5.74% The calibration parameter maps and curves are optimized regarding the torque buildup criterion. Thus, the agility of the load change reaction is the optimization goal. During the optimization, the integral evaluation approach is used (cf. chapter 2). With this approach, the SP positions and values of calibration parameter maps and curves are optimized inside the valid design space of the OP variables (cf. chapter 3 and 4). All calibration parameter maps and curves have individual SP axes to be optimized (cf. chapter 4). In addition to the convex hull design space constraints on the curve and map SPs, another nonlinear design space model is considered. The additional target criteria , jerk and are also used as constraints during the optimization. The limits are set to the corresponding values that are reached by a proper preexisting reference calibration (cf. chapter 4). The nonlinear design space model and the additional modeled target criteria constraints are based on the integral evaluation 101 3.2 approa In conc torque mizatio tings (c Figure The firs the cali tical to tion pa tive fun erage t referen conside the opt In addi optimiz of chap criterio Optimization ach of chap clusion, th buildup wi on algorithm cf. chapter 7: Optimiz st optimiza ibration pa the optimi rameter m nction target mod nce calibra erably imp timized cal ition to the zation, the pter 2 are s n was opti n of ECU Ma pter 2. Thu e calibratio ithout impa m used for 5). zation resu tion pa ation result arameter m zation resu map or curv ( 0). del criteria ation param roved by a ibration do e result rea average t shown for mized bas ap Sampling us, the corr on goal is airing the c r this calibr ults for the arameter m t is shown maps and c ults shown ve smoothn Table 1 s in compari meter map approximat oes not imp ached by t arget crite the optimi sed on the Point Values responding s to optimiz comfort cri ration test criterion a map and cu in Figure curves are n in [28]. Fo ness was i shows the ison to the ps and cu tely 12 pe pair the co the novel ria values ization res sum of rep s and Positio g average ze the agi iteria a case is the average tor urve smoo 7. In orde rescaled. or the first ncluded in optimizatio e correspon rves. The rcent. All c mfort criter approach calculated ults of [28] presentativ ons with Mod model valu lity criterio nd jerk (an e fmincon w rque buildu thness r to preser The applie optimizatio nside the o on results nding value average constraints ria. of SP pos d with the ]. In [28], t ve OPs. Th del-Based Ca ues are res on of the a nd ). T with stand up and no c rve confide ed scaling on run, no optimization regarding es reached torque bu s are satisf sitions and integral ap the torque he map an alibration stricted. average he optiard setcalibraentiality, is idencalibran objecthe avd by the ildup is fied and d values pproach buildup nd curve 102 3.2 SP loca by 5.7 approa calibrat ment o [28]. Figure A seco buildup conside change um leve calibrat percen By cho map an These can be Howev Optimization ations wer percent wi ach of [28] tion was g of the sum e 8: Optimiz leve ond optimiz p. In additio ered durin ed to get a el of smoo tion maps t in compa osing the w nd curve sm two optimi achieved er, the com n of ECU Ma re not optim ith the new . It is note iven as 7. of the torq zation resu vel of calibr zation is d on, the sm ng optimiz a solution o othness. Th and curve arison to th weighting moothness zation resu with the n mputation ap Sampling mized. The w approach ed, that th 43 percen que buildu ults for the ration para done with t moothness o zation. Th of the calib he results es, the av he referenc the ca s for the ca ults of the ew optimiz of the opt Point Values e average t h in compa e improve t in [28]. T p model o criterion a meter map the same of the calib hus, the o bration para are shown verage torq ce calibrati alibration e alibration. calibration zation app timizations s and Positio torque bui arison to th ement in co This value on the disc average tor p and curve target crite bration par objective f ameter ma n in Figure que buildu ion. All con engineer ca n test case proach of th s that were ons with Mod ldup value he results o omparison was based rete optim rque buildu e smoothn erion of th rameter ma function w aps and cu e 8. Despit up is still i nstraints ar an decide show the he SP valu e run for th del-Based Ca e can be im optimized w n to the re d on the im mized OPs up and a m ness he average aps and cu weighting urves with te those sm improved re satisfied a suitable improvem ues and po his calibra alibration mproved with the eference mproveused in medium e torque urves is is a medimoother by 11.3 d again. level of ent that ositions. tion ex- 103 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration ample lasted between 1.8 to 8 days. In contrast, the optimization approach of SP values used in [28] which is designed for a fast optimization only took about 7.5 minutes. Several advancements will be implemented in future to reduce the optimization time. These advancements include an improvement of the code efficiency, parallel computing features and pre-optimization methods to reduce the optimization effort during the main optimization approach. Nevertheless, the improvement that can be achieved with the new approach will vary according to the calibration test case. The trade-off between the additional effort and the improvement has to be analyzed for further calibration use cases. In addition, it is pointed out again that the optimization function fmincon was used. With this gradient-based algorithm, one local optimum is found. The chance of finding an adequate optimization result is higher with the choice of a suitable starting point. Nevertheless, as already suspected, the analysis of multiple optimization runs with various optimization settings confirms the multimodality of the SP values and positions optimization problem. Some optimization runs result in local optima that lead to inferior objective criteria in comparison to the results of Table 1. An optimization metaheuristic has to be used to lower the risk of finding a local optima (cf. chapter 3). The disadvantage of such an optimization algorithm will be a lower speed of convergence and a higher computation time. Thus, the usage is not sensible until the above-mentioned improvements of the optimization speed are implemented. The optimization also often stops with inferior local solutions, if nonlinear constraints on local operating or SPs are defined. Such a constraint can be for example a limit for the mentioned nonlinear design space model on the map and curve SPs (cf. chapter 4). This type of constraint can violate the requirement of continuous differentiability in case of a SP position variation for calibration maps and curves with individual SP axes (cf. chapter 4). In this case, the optimization algorithm may stop with an inferior local solution. In contrast, the suggested optimization metaheuristics are not limited to continuously differentiable constraints so that additional optimizations have to be executed in future. That may lead to solutions that also satisfy the local constraints on the SPs. 6 Conclusion and outlook An integral evaluation of empirical models has been developed to calculate average values of target criteria that are used during the optimization. Therewith, the entire available information for global models is used. The novel approach allows to optimize the value and position of sampling points for calibration parameter maps and curves simultaneously. Besides empirical models of the target criteria, the smoothness of the calibration parameter maps and curves can be considered during the optimization. The approach also allows to define constraints for other target criteria. Both types of SP axes for calibration parameter maps and curves that are used in engine and transmission ECUs of passenger cars can be optimized. It was shown by an example of drivability calibration that the calibration can be improved with the SP values and positions optimization approach in comparison to approaches that just optimize the SP values. Nevertheless, the computational complexity increases with the additional degree of freedom during optimization, the integral 104 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration evaluation and the more complex calculation of the optimization constraints. The trade-off between the improvement and the additional effort is case-dependent and the benefit has to be analyzed with more use cases and experience. Further improvements of the approach will concentrate on the enhancement in optimization speed. Parallel computing, code improvement of the objective functions and pre-optimization techniques will be implemented. In addition, more suitable optimization algorithms for this multimodal optimization problem will be tested. The current approach calculates the average target criteria with an equal weighting of the entire OP range that is considered during integration. Methods of kernel density estimation can be used to reduce a driven test cycle to a continuous and continuously differentiable function that maps the importance of the OPs inside the OP range used for integration. This function can be multiplied with the target criteria before the integral is calculated. This integral value leads to a representative test cycle prediction without the consideration of dynamic influences, which can be optimized. References [1] Berger, B.: Modeling and Optimization for Stationary Base Engine Calibration, Technical University Munich. PhD Thesis. Munich 2012 [2] Furch, D.; Link, M.: PAoptimizer, Ein neues Werkzeug zur modellgestützten Kennfeldoptimierung von Motorsteuergeräten. MTZ 63 (2002) 12, pp. 1046- 1050 [3] Mitterer, A.; Zuber-Goos, F.: Modellgestützte Kennfeldoptimierung - Ein neuer Ansatz zur Steigerung der Effizienz in der Steuergeräteapplikation. ATZ 102 (2000) 3, pp. 188-194 [4] Stuhler, H.; Kruse, T.: Parametrierung von Motorsteuerungen. In: Isermann, R. (ed.): Modellgestützte Steuerung, Regelung und Diagnose von Verbrennungsmotoren. Berlin, Heidelberg: Springer 2003, pp. 22-38 [5] Hartmann, B.: Lokale Modellnetze zur Identifikation und Versuchsplanung nichtlinearer Systeme, University Siegen. PhD Thesis. Siegen 2014 [6] Kötter, H.; Sequenz, H.: Stationäre Motorvermessung mit verschiedenen Methoden und Modellen. In: Isermann, R. (ed.): Elektronisches Management motorischer Fahrzeugantriebe. Elektronik, Modellbildung, Regelung und Diagnose für Verbrennungsmotoren, Getriebe und Elektroantriebe. ATZ/ MTZ- Fachbuch. Wiesbaden: Vieweg Teubner 2010, pp. 130-166 [7] Poland, J.: Modellgestützte und evolutionäre Optimierungsverfahren für die Motorentwicklung. Berlin: Logos-Verlag 2002 [8] Töpfer, S.: Statistische Versuchsplanung. In: Isermann, R. (ed.): Modellgestützte Steuerung, Regelung und Diagnose von Verbrennungsmotoren. Berlin, Heidelberg: Springer 2003, pp. 120-130 [9] Mitterer, A.: Optimierung vielparametriger Systeme in der Kfz- Antriebsentwicklung. Statistische Versuchsplanung und Künstliche Neuronale Netze in der Steuergeräteauslegung zur Motorabstimmung. Fortschritt- Berichte VDI. Series 12, Verkehrstechnik/ Fahrzeugtechnik, Nr. 434. Düsseldorf: VDI Verlag 2000 105 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration [10] Kuder, J.; Kruse, T.: Parameteroptimierung an Ottomotoren mit Direkteinspritzung. MTZ 61 6, pp. 378-384 [11] Linssen, R.: Modellbasierte Analyse und Optimierung von Dieselmotoren, TU Aachen. PhD Thesis. Aachen 2010 [12] Goos, J.-C.; Linssen, R.; Kirschbaum, F.; Koch, T.: Computing Optimized Calibration Maps Including Model Prediction and Map Smoothness. In: Röpke, K.; Gühmann, C. (eds.). Design of Experiments (DoE) in Powertrain Development. 2015 [13] Pillas, J.; Kirschbaum, F.; Jakobi, R.; Gebhardt, A.; Uphaus, F.: Model-Based Load Change Reaction Optimization Using Vehicle Drivetrain Test Beds. In: Bargende, M.; Reuss, H.-C.; Wiedemann, J. (eds.). 14. Stuttgarter International Symposium, Vol. 2. Wiesbaden: Springer 2014, pp. 857-868 [14] Nelles, O.; Fink, A.: Tool zur Optimierung von Rasterkennfeldern. Automatisierungstechnische Praxis (atp) 42 (2000) 5, pp. 58-66 [15] Rasmussen, C. E.; Williams, C. K. I.: Gaussian Processes for Machine Learning. Adaptive Computation and Machine Learning. Cambridge: MIT Press 2006 [16] Sequenz, H.: Emission Modelling and Model-Based Optimisation of the Engine Control, Technical University Darmstadt. PhD Thesis. Darmstadt 2013 [17] Kianifar, M. R.; Campean, L. F.; Richardson, D.: Sequential DoE Framework for Steady State Model Based Calibration. SAE International Journal of Engines 6 (2013) 2, pp. 843-855 [18] Hartmann, B.; Nelles, O.: Adaptive Test Planning for the Calibration of Combustion Engines Methodology. In: Röpke, K. (ed.). Design of Experiments (DoE) in Engine Development. Renningen: expert verlag 2013, pp. 1-16 [19] Klein, P.; Kirschbaum, F.; Hartmann, B.; Bogachik, Y.; Nelles, O.: Adaptive Test Planning for the Calibration of Combustion Engines - Application. In: Röpke, K. (ed.). Design of Experiments (DoE) in Engine Development. Renningen: expert verlag 2013, pp. 17-30 [20] Fortuna, T.; Mayer, M.; Pflügl, H.; Gschweitl, K.: Optimierung von Motorsteuerungen am Motorenprüfstand. In: Isermann, R. (ed.): Modellgestützte Steuerung, Regelung und Diagnose von Verbrennungsmotoren. Berlin, Heidelberg: Springer 2003, pp. 251-274 [21] Linssen, R.; Henrich, C.: Cluster Analysis Using an Extended k-Means Process. In: Röpke, K. (ed.). Design of Experiments (DoE) in Engine Development. Renningen: expert verlag 2011, pp. 183-204 [22] Castagné, M.; Bentolia, Y.; Hallé, A.; Nicolas, F.; Sinoquet, D.: Engine Calibration: Towards an Integrated Approach. In: Röpke, K. (ed.). Design of Experiments (DoE) in Engine Development III. Renningen: expert verlag 2007, pp. 56-90 [23] Knödler, K.: Methoden der restringierten Online-Optimierung zur Basisapplikation moderner Verbrennungsmotoren. Berlin: Logos-Verl. 2004 [24] DE102004026583 B3 106 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration [25] Nessler, A.; Haukap, C.; Roepke, K.: Global Evaluation of the Drivability of Calibrated Diesel Engine Maps. In: Proceedings of the 2006 IEEE Conference on Computer Aided Control System Design. 2006, pp. 2683-2688 [26] Schloßer, A.; Linssen, R.; Bozelie, P.: Model-Based Calibration of SI-Engines: Map Optimization. In: Röpke, K. (ed.). Design of Experiments (DoE) in Engine Development III. Renningen: expert verlag 2007, pp. 145-164 [27] Dursun, Y.: Optimization of Calibration Maps With Coupled Model Evaluation, Karlsruhe Institute of Technology. Diploma Thesis. Karlsruhe 2013 [28] Goos, J.-C.; Linssen, R.; Koch, T.; Dursun, Y.; Kirschbaum, F.; Brand, K.; Gebhardt, A.: Neue Methode zur Parameteroptimierung in der modellbasierten Fahrbarkeitskalibrierung. In: . Schwingungen in Antrieben 2015. 9. VDI- Fachtagung. VDI-Berichte, Vol. 2262. Düsseldorf: VDI-Verlag 2015 [29] Weber, S.: Kennfeldoptimierung mit eingekoppelter Auswertung globaler Modelle, University Stuttgart. Master Thesis. Stuttgart 2015 [30] Brumm, M.: Kennfeldstützstellenoptimierung auf Basis einer integralen Zielfunktion von empirischen Modellansätzen, Karlsruhe Institute of Technology. Karlsruhe 2016 [31] Shampine, L. F.: Matlab Program for Quadrature in 2D. Applied Mathematics and Computation 202 (2008) 1, pp. 266-274 [32] Shampine, L. F.: Vectorized Adaptive Quadrature in MATLAB. Journal of Computational and Applied Mathematics 211 (2008) 2, pp. 131-140 [33] MATLAB. Ver. 2013b. The MathWorks, Inc. [34] The Mathworks Inc.: Matlab 2013b Optimization Toolbox Documentation - Quadratic Programming Algorithms. 2013 [35] Gill, P. E.; Murray, W.; Wright, M. H.: Numerical Linear Algebra and Optimization. Redwood City: Addison-Wesley 1991 [36] Gill, P. E.; Murray, W.; Saunders, M. A.; Wright, M. H.: Procedures for Optimization Problems with a Mixture of Bounds and General Linear Constraints. ACM Transactions on Mathematical Software 10 (1984) 3, pp. 282-298 [37] Pillas, J.; Kirschbaum, F.: Model-based Calibration of a Load Change Reaction Control Function with Hybrid State Space Models. In: FKFS (ed.). Proceedings of the 12th Stuttgart International Symposium Automotive and Engine Technology. 2012 [38] DE102009021781 A1 [39] D’Errico, J.: gridfit. Model 2-D Surfaces from Scattered Data. Ver. 29.07.2010. 2005 [40] Sauermann, R.; Boja, D.; Kirschbaum, F.: Particle Swarm Optimization for Automotive Model-Based Calibration. In: Trächtler, A.; Abel, D. (eds.). Advances in Automotive Control. IFAC Proceedings. IFAC, Elsevier 2010, pp. 733-738 [41] The Mathworks Inc.: Matlab 2013b Optimization Toolbox Documentation fmincon. 2013 107 3.2 Optimization of ECU Map Sampling Point Values and Positions with Model-Based Calibration [42] Byrd, R. H.; Gilbert, J. C.; Nocedal, J.: A Trust Region Method Based on Interior Point Techniques for Nonlinear Programming. Mathematical Programming 89 (2000) 1, pp. 149-185 [43] Byrd, R. H.; Hribar, M. E.; Nocedal, J.: An Interior Point Algorithm for Large- Scale Nonlinear Programming. SIAM Journal on Optimization 9 (1999) 4, pp. 877-900 [44] Waltz, R. A.; Morales, J. L.; Nocedal, J.; Orban, D.: An interior algorithm for nonlinear optimization that combines line search and trust region steps. Mathematical Programming 107 (2006) 3, pp. 391-408 [45] Goos, J.-C.: Messdatengestützte simulative Untersuchung der Eignung von Prüfständen zur straßennahen Abbildung von Lastwechseln, TU Darmstadt. Master Thesis. Darmstadt 2013 108 3.3 Dynamic Route-Based Design of Experiments (R-DoE) Thomas Mayer, Mohamed Ayeb, Ludwig Brabetz Abstract As an enhancement of the classic DoE (Design of Experiments), the route-based design of experiments (R-DoE) plans the entire adjustment route for automated measurements on a test rig rather than individual measurement points. The presented R- DoE algorithm calculates a basic adjustment route which provides quasi-stationary measurement points for stationary data-driven simulation models, with the shortest possible total measurement time. In addition user requirements, e.g. point dependent prediction accuracy, are taken into account. Furthermore, a priori knowledge is included, e.g. constraints of the system to be examined. Such an adjustment route is the optimal solution of a multi-objective optimization problem and is calculated using the evolutionary algorithm (EA) implemented in the R-DoE algorithm. The quasi-stationary adjustment route, in a first instance optimized for a quasistationary measurement, is superimposed with a sinusoidal oscillation in a second step. With the method proposed, it is possible to determine the necessary coefficients for the stationary model and the time constants for the dynamic model used in a Hammerstein model, which is a stationary model with an added dynamic model, for automated measurements on a test rig. Kurzfassung Als Erweiterung der klassischen DoE (Design of Experiments) plant die routenbasierte Versuchsplanung (R-DoE) nicht einzelne Messpunkte, sondern die gesamte Einstellroute für eine automatisierte Messung am Prüfstand. Der vorgestellte R-DoE Algorithmus berechnet eine erste Einstellroute, aus deren Messwerten stationäre datengetriebene Simulationsmodelle erstellt werden können. Hierbei werden neben einer möglichst kurzen Messdauer auch zusätzliche Anforderungen des Anwenders (z.B. betriebspunktabhängige Prognosegüten) und a priori Wissen (z.B. Randbedingungen des Systems) berücksichtigt. So ist eine Einstellroute die Lösung eines Mehrgrößen-Optimierungsproblems, die mit Hilfe eines Evolutionären Algorithmus gefunden wird. Die zunächst für eine quasistationäre Messung optimierte Einstellroute wird in einem zweiten Schritt mit einer sinusoidalen Schwingung überlagert. Mit dem vorgestellten Verfahren ist es möglich, im Rahmen eines Messdurchlaufs für ein Hammerstein Modell, einem stationären Modell mit ergänzendem Dynamik-Modell, die notwendigen Zeitkonstanten des dynamischen Verhaltens und auch die Koeffizienten für das stationäre Modell zu bestimmen. 109 3.3 Dynamic Route-Based Design of Experiments (R-DoE) 1 Motivation and concept Data-driven simulation models support the process of developing technical systems. Thus, the concepts can be tested already in the early development phase by simulation. This generally requires simulation models of individual components. For characterizing a component the data needed is measured on a component test rig. Statistical design of experiments (DoE) for setting appropriate measuring points is state of the art. The procedure described in this paper plans the entire adjustment sequence for measurements on a test rig as an enhancement of the well-known DoE procedure. Since this adjustment sequence is quite similar to a route in the multi-dimensional space, this method has been called route-based design of experiments (R-DoE). The R-DoE provides a measurement plan which contains the entire adjustment route, including the operating point-dependent maximum adjustment speed necessary for a quasi-stationary system behavior. Operating limits of the system, as well as prior knowledge of system behavior and user requirements are taken into account. An R- DoE adjustment route is optimal with regard to the requirements formulated in a goal function. The route-based DoE was presented at first time in [7], the entire procedure for quasi-stationary adjustment routes and the associated algorithm are described in [8]. The modeling of a large family of nonlinear dynamic system behavior is possible by connecting a stationary simulation model and a linear dynamic (Hammerstein model) [9][11]. To evaluate the dynamic behavior, the method presented in this paper superimposes sinusoidal oscillations on to the adjustment route calculated initially for quasi-stationary measurement (see also [3] or [5]). The stationary behavior and the time constants of the dynamic component can then be calculated from the dynamic measurement profile. 2 The R-DoE algorithm 2.1 System analysis and model assumption Before starting the design of experiments, a system analysis is necessary to determine the relevant influencing variables and disturbances as well as the target variables to be modeled for the system. For the data-driven model, a polynomial approach is used here. This is very suitable for data-driven models in the field of thermal management investigated here. It allows a simple representation of the connections between the influencing variables and the target variables, as well as including previous knowledge about the system behavior. In thermal management usually a third order polynomial function is sufficient for model assumption for the system behavior. In our example here we have two influencing variables and a third order polynomial for model assumption: = + + + + + + + 110 3.3 Dynamic Route-Based Design of Experiments (R-DoE) 2.2 Operating limits and constraints For the creation of a route-based experimental design with the R-DoE algorithm, a stable adjustment process and a quasi-stationary response behavior of the target variable must be provided for the adjustment route. Therefore, the maximum possible adjustment speed for each influencing variable must be determined and taken into account for creating the route-based experimental plan. The user can specify preferred operating points to the R-DoE algorithm, which can be used optionally in the adjustment route as so-called steady points . At these steady points during the measurement, the adjustment process stops for a predefined time, for example 30 seconds, in order to observe potential dynamic system behavior at constant influencing variables despite expected quasi-stationary adjustment. The position of the optional steady operating points can be defined by the user himself or can be calculated with a suitable algorithm such as, for example, the classic DoE. A fast approximation for finding D-optimal points within the operating space is the sequential point selection [2]. A space filling design, or a random distribution of a defined number of steady points in the operating space can also be used. 2.3 Adjustment routes The adjustment routes created with the R-DoE algorithm are composed of sub-routes which each connect two steady points. The specified steady points have to be included in the system and the optional steady points may be added if so wished. Each steady point may be present in the route once and only once. One steady point can be defined by the user as the starting point of the entire route. The sub-routes are described with cubic polynomial functions in the R-DoE algorithm. Thus, the adjustment route, which consists of these sub-routes, covers the operating area very well. Lower-order polynomial functions are not so flexible. Polynomial functions of higher than third order increase the complexity of the mathematical description without higher benefit. 2.4 Goal function for optimizing the adjustment route The R-DoE algorithm optimizes the following goal function: = + + An evaluation of a created adjustment route is carried out with the help of this goal function, whereby the values (total length of the adjustment route according to the adjustment time), (evaluation of the expected prediction quality of the simulation model) and (possible violation of the limits of the operating space specifications by the adjustment route) are already included weighted in the target function. 111 A main measu areas o with t of the r examp in an a achieve in (a) a as A an fr The mu gramm from de goal fu plemen the des values then ze With th function target v by simp 3 n goal is the red values includes of the oper the expect regression le shown b area , her ed in an a a suboptim nd B, and i Figure1: (a specificat rom dark b ultiobjectiv ming). Thro esired targ unction nted for the and sired targe of ero or almo he target v n. The con values can ple calcula 3.3 Dyna e optimizin s accordin the varian rating spac ted standa and below, the re called p rea outs al solution in (b) a sol (a) A subop tions. The blue for goo ve optimiza ughout the get values = e individua dominat et values a and ost zero. N value spec ntinuity of n easily be ations from amic Route-B ng of the p g to the u nce of the ce: Var = ard deviatio the vector user expe precision ar side the ma that does ution that ptimal and prediction od predictio ation uses e optimiza are minim + l objective te the tota are not rea are almo Negative in cification, it the target determine reference Based Desig rediction q user’s requ predicted ( on of the m r with the t ects to obta rea, and a ain interes s not reach is optimal (b) An opt quality in on quality target valu ation the d mized. In t + es a l target fun ached. Thi ost comple ndividual ta t is quite e t functions ed either o values of n of Experim quality of th uirements. of the re ) measured terms of th ain at least a slightly lo sting opera h the requir with regard timal soluti the operat to red for p ues for the deviations the algorit , a targe and nction s occurs u etely fulfille arget funct easy for th does not on the basi a classica ments (R-DoE he model c So the ev egression values, he model a t a defined ower predic ating points red predict d to the us ion with reg ing space poor predic e individual of actual e hm used f et value sp such that as pena until the us ed. Their fu ion values he user to have to b s of previo l experime E) calculated w valuation f model in d the design assumption prediction ction qualit s. Figure 1 tion quality ser specific gard to the is colored ction qualit l goals (Go evaluation for optimiz pecification two target alty terms ser-desired unction val s are not p define the be respecte ous knowle ental plan. with the function different n matrix n. In the n quality ty to be 1 shows y in arecations. e ty. oal Provalues zing the n is imt scores in case d target ues are ossible. e target ed. The edge, or 112 2.5 E Evolutio rithms. from b ment ro sidered tions re viduals senting the ind crossov problem matical the imp 2.6 R Figure operati used o cision a 3 Evolution onary algo The princ iological e oute corres d to adapt elated to th s compete g a better ividuals ad ver and m m. The se lly in figure plemented Figur Restrictio 3 shows t ng space ptional ste area A to th 3.3 Dyna ary algori orithms (EA ciple of op evolution. S sponding t to their en heir goal f against ea solution to dapt thems utation and quence of e 2. A com EA for the re 2: Conce ons on ope two examp and (b) fo eady points he less im amic Route-B thmus A) are a s ptimization Single ind o a possib nvironment function va ach other. T o the optim selves to th d develop f the imple mon proce e shown Rept of the i erating sp ples of the or highly d s are mark portant are Based Desig et of iterat in an EA ividuals, e ble solution t. The indiv alues. Insid The individ mization pr he requirem towards th emented e ess of an E -DoE are d implement pace and o e use of th different m ked with w ea B is ma n of Experim tive, rando is based each repre n of the opt viduals are de their co duals with roblem. In ments of th he optimal evolutionary EA can be described i ted evolutio operating c he R-DoE aximum a hite squar rked with b ments (R-DoE om-based o on method esenting a timization e measure ommon po better fitne the cours he environm solution o y algorithm found in lit n [8]. onary algo conditions process (a djustment es and the black dash E) optimizatio ds that ar complete problem, a ed by fitnes opulation, t ess survive se of gene ment by m of the optim m is show terature, e orithm. s a) for a re speeds. T e edge of t hed lines. on algoe taken adjustare conss funche indie repreerations, means of mization n schee.g. [10], estricted The unthe pre- 113 3 E 3.1 D Modelin of such ranging comple model ments a After se stimula model sive fo been re stimula method In this nonline describ of the 3 Enhance Dynamic r ng nonline h systems g from the ex dynamic structures and static electing a ation and o parameter rm of the esearched ation signa ds of class paper, the ear static fu bing the pa inputs. In 3.3 Dyna Figure (b) D ement to route mod ear dynami . Many ap e extension c neural ne , which ar nonlinearit Figure non-linear of a suitab rs. Here, th model str and appli als such as ical experi e Hammer unction an arameters o the particu amic Route-B 3: (a) Res Different m nonline deling c systems pproaches n of conve etworks. Th re serial c ties (Figure 4: Structu dynamic m ble experim he known ructures ca ed over th s the amp mental des rstein mod nd a linear of the linea ular case o Based Desig stricted ope maximum s ar dynam s is a challe have bee entional lin he former onnections e 4) [9] [11 ure of a Ha model struc mental des methods f annot be u he last few litude-mod sign [1]. del structur dynamic t ar transmis of an elem n of Experim erating spa setting spe mic syste enge beca en suggest near dyna include e.g s of linear ]. mmerstein cture, the q sign in ord for static c used read years, wh dulated PR re, which transmissio ssion elem ment as a ments (R-DoE ace and eds. ems ause of the ted and te mic system g. Wiener a r dynamic n model question a er to reliab correlation ily. Many ich have e RB signals, is a series on elemen ment as a n linear tran E) e diverse b ested in th m descrip and Hamm transmiss arises of a bly determ due to the approache either inves , or extend s connecti nt, is expan non-linear f smission e behavior he past, tions to merstein ion elesuitable mine the e recures have stigated ded the on of a nded by function element 114 3.3 Dynamic Route-Based Design of Experiments (R-DoE) with a state-dependent (here input-dependent) time constant, this represents a sensible extension to describe a wider family of nonlinear dynamic systems from practice. Figure 5 represents such a structure with three inputs and a static quadratic nonlinearity. A PT1 element with an input-dependent time constant is provided as a dynamic transmission element. The time constant is again a quadratic function of the input signals. The signal sequence obtained in the previous section with the R-DoE algorithm is used as a stimulation signal for the static part of the system. In order to stimulate the dynamic behavior of the system, the PT1 element is stimulated by means of a superimposed sinusoidal oscillation. Figure 5: Hammerstein model with input-dependent time constant 3.2 Model with undisturbed output of the measured system The static part of the system can be expressed as follows 2 3 10 2 2 9 2 1 8 3 2 7 3 1 6 2 1 5 3 4 2 3 1 2 1 x x x x x x x x x x x x y Stat ϑ ϑ ϑ ϑ ϑ ϑ ϑ ϑ ϑ ϑ + + + + + + + + + = and the input-dependent time constant T: 2 3 10 2 2 9 2 1 8 3 2 7 3 1 6 2 1 5 3 4 2 3 1 2 1 x x x x x x x x x x x x T α α α α α α α α α α + + + + + + + + + = This results in the following description of the output of the disturbance-free model in recursive form: ( + 1) = ( ) + ( + 1) − ( ) 1 − = ( + 1) + ( ) − ( + 1) denotes the sample point at time, ∙ with = 1 the selected sample time. 115 3.3 Dynamic Route-Based Design of Experiments (R-DoE) If the sampling time is small enough compared to the time constant of the system, the following simplification can be carried out: 2 3 10 2 2 9 2 1 8 3 2 7 3 1 6 2 1 5 3 4 2 3 1 2 1 1 x x x x x x x x x x x x e T β β β β β β β β β β + + + + + + + + + ≈ − This results finally in a 4th order polynomial representation for : ( + 1) = ( ) ∙ + ( + 1) + ( + 1) ( + 1) + ⋯ + ( + 1) + + ( + 1) + ( + 1) + ( + 1) + ⋯ + ( + 1) ∙ 1 − − ( + 1) − ( + 1) − ( + 1) − ⋯ − ( + 1) Proceeding from a disturbance-free system, the parameters of the static and dynamic part can be determined using the classic Least-Square method: ( ) measured T T Y F F F 1 − = Θ with the matrix with the regression vectors and the vector with the measured output values at time . Figure 6 illustrates the example of a simulated disturbance-free system and the difference between the simulated and the estimated dynamic output. Figure 6: Comparison of real (blue) and estimated (red) output of the system 0 500 1000 1500 2000 2500 3000 3500 4000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Zeit [s] y y_dyn 116 3.3 Dynamic Route-Based Design of Experiments (R-DoE) The time constant depending on the input factors can also be determined well, as shown in Figure 7, despite the simplification introduced above. Figure 7: Comparison of real and estimated values of the input-dependent time constant 3.3 Model for disturbed output of the measured system In real world applications the system output is often disturbed leading to the following system equations description: ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) 1 1 1 1 1 1 1 ... 1 1 1 1 ... 1 1 1 1 int _ _ 1 1 1 1 int _ int _ 2 3 10 3 4 2 3 1 2 1 1 1 2 3 10 3 4 2 3 1 2 1 + + + = + ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ − + + = + + + + + + + + + + = + + + + + + + + + = + + − + − + − k k y k y e k y e k y k y k x k x k x k x e k x k x k x k x k y ern Dyn extern dyn k T Stat k T ern Dyn ern dyn k T stat ε β β β β β ϑ ϑ ϑ ϑ ϑ with ( ) the disturbances at the output of the system which are assumed to be uncorrelated with the system output and inputs. 0 500 1000 1500 2000 2500 3000 3500 4000 12 14 16 18 20 22 24 26 28 Zeit [s] T_real T_estimated 117 3.3 Dynamic Route-Based Design of Experiments (R-DoE) Since one does not have access to the internal variable , the system equations can no longer be expressed in a known undisturbed regression vector on the one hand and a measured value on the other. Here the Predictive Error Method (PEM) [6] [12] can be used to determine the parameter vector iteratively: ( ) ( ) ( ) predicted measured T T Y Y F F F i i − + Θ = + Θ − 1 ) ( 1 μ with iteration index , iteration step size μ and estimated system output calculated out of the parameter vector obtained from the previous iteration step. The Least-Square method can be used as starting value for the iterations. Figure 8 illustrates again the example of a simulated disturbed system and the comparison of the simulated and the estimated dynamic output. Figure 8: Comparison of real (blue) and estimated (red) output of the disturbed system Figure 9 shows the quality of the estimation of the input-dependent time constant using the predictive error method. 500 1000 1500 2000 2500 3000 3500 0 0.1 0.2 0.3 0.4 0.5 0.6 Zeit [s] y y_dyn 118 3.3 Dynamic Route-Based Design of Experiments (R-DoE) Figure 9: Comparison of real and estimated values of the input-dependent time constant for the disturbed system 4 Conclusion Within the scope of statistical design of Experiments (DoE) an algorithm is presented which creates a continuous adjustment sequence for the quasi-stationary measurement of a component. This adjustment sequence, called route-based experimental design (R-DoE), is an enhancement of the point-discrete classical statistical experimental design. In the R-DoE adjustment route, several, even conflicting requirements of the user can be considered due to the multi-criteria optimization. The Optimization is carried out with an evolutionary algorithm. The measured values of this adjustment route are appropriate for creating stationary data driven simulation models. If the quasi-stationary adjustment route is superimposed with a sinusoidal oscillation during the measurement, dynamic simulation models can be generated with the measured values due to the systematic dynamic excitation of the system. An example demonstrates how to identify the parameters for the static and dynamic part of a Hammerstein model. For undisturbed system outputs the model parameters are determined using the Least-Square method and for disturbed system outputs iteratively, based on the PEM method. 500 1000 1500 2000 2500 3000 3500 14 16 18 20 22 24 26 Zeit [s] T_real T_estimated 119 3.3 Dynamic Route-Based Design of Experiments (R-DoE) Literature [1] Ayeb, M.; Theuerkauf, H.: Integration der D-Optimalität in neuronale Modellierungsverfahren . In: Röpke, K. (Hrsg.): Design of Experiments (DoE) in Engine Development. Renningen : expert Verlag, 2003, S. 185-194. [2] Bandemer, H. ; Bellmann, A. ; Jung, W. ; Richter, K.: Optimale Versuchsplanung. Zürich : Harri Deutsch Verlag, 1976. [3] Baumann, W. ; Klug, K. ; Köhler, B.-U. ; Röpke, K.: Modeling of Transient Diesel Engine Emissions. In: Röpke, K. (Hrsg.): Design of Experiments (DoE) in Engine Development. Renningen : expert Verlag, 2009, S. 41-53 [4] Fahrmeir, L. ; Kneib, T. ; Lang, S.: Regression: Modelle, Methoden und Anwendungen. 2. Auflage. Berlin, Heidelberg : Springer Verlag, 2009. [5] Godward, T. ; Schilling, H. ; Schaum, S.: Use of Dynamic Testing and Modeling in the Engine Calibration Process. In: Röpke, K. (Hrsg.) : Design of Experiments (DoE) in Engine Development. Renningen : expert Verlag, 2013, S. 211-230. [6] Ljung, L.: System identification theory for the user. Upper Saddle River, New Jersey, Prentice-Hall, 1999. [7] Mayer, T. ; Waschatz, U. ; Ayeb, M. ; Brabetz, L.: Process for Energy-efficient Operation of a Refrigerant Circuit. In: Gühmann, C. (Hrsg.) ; Riese, J. (Hrsg.); Wolter, T.-M. (Hrsg.): Simulation and Testing for Automotive Electronics V. Renningen : expert Verlag, 2014, S. 403-408. [8] Mayer, T. : Optimale, routenbasierte Versuchsplanung (R-DoE) zur Charakterisierung technischer Systeme. Kassel, kassel university press GmbH, 2016 [9] Narendra, K.S. ; Gallman, P.G.: An iterative method for the identification of nonlinear systems using a Hammerstein model. IEEE Transactions on Automatic Control, 11(7): 546-550, 1966. [10] Pohlheim, H.: Evolutionäre Algorithmen: Verfahren, Operatoren und Hinweise für die Praxis. Berlin : Springer Verlag, 2000. [11] Schetzen, M.: The Volterra and Wiener Theories of Nonlinear Systems. Wiley, 1980. [12] Söderström, T. ; Stoica, P.: System Identification. Prentice-Hall International, Hemel Hempstead, UK, 1989 120 4 Poster session 4.1 System for Real-time Evaluation of Real Drive Emission (RDE) Data Rajesh Reddy, Sven Meyer Abstract With the introduction of EURO 6c in 2017, emission measurements under real drive conditions in addition to measurements in the stationary emission laboratory are necessary for homologation according to in-use compliant measurement of Light Duty Vehicles (LDV). This is based on changes in the legislation given by the European Commission “[…] in order to consider changes in the properties of vehicles and the drivability […]” and in order to assure, that for the “[…] type approval verification measured emissions will be similar to the emissions under real-world conditions […]”. Emissions measured in the World Harmonized Light Vehicles Test Cycle (WLTC) will be used as reference. The difference measured will be a multiple of reference CO 2 emissions in single, percentage-wise defined sections (Motorway, Rural, Urban cycle). For this reason, a visualization would be helpful to provide the driver with information about the vehicle parameters and the sections in order to assure the compliance to the legal requirements. The main focus of this work was optimization of RDE driving. For this purpose, an assistance system has been developed to provide the driver with information about the setup, execution and the evaluation of the RDE drive. The focus was to relieve in-vehicle measurements of the compliance to EU legal requirements. During RDE test drives, drivers will be provided with important information about the driving route, already driven sections, and the exhaust emissions as well as boundary conditions. The aim of this approach is to increase efficiency and reproducibility of RDE test drives with a PEMS system. In order to enable calibration engineers to work with their conventional calibration tool chain, the solution is fully integrated with ETAS INCA. 121 4.1 Real Drive Emission (RDE) Data 1 Introduction Starting in the early sixties, first emission legislations for passenger vehicles have been introduced. Especially Los Angeles, California, suffered from pollution and smog. The major pollution source have been the emissions emitted by combustion engines, such as carbon monoxide, nitrogen oxides, hydrocarbons, and particulates. In order to develop legislation assuring clean air, the California Air Resources Board (CARB) was founded. Besides toxic carbon monoxide, nitrogen oxides are especially harmful to humans since they are irritating the airways and can even damage them. During the last decades, emission legislation has been brought into action by the European Union, Japan, China, India, and most other countries worldwide following their examples. Compared to the beginning, current emission laws are much more stringent. Figure 1 shows the different emission regulations that have been in place since 1995. Figure 1: Worldwide emission standards for passenger cars and light duty trucks (Source: Delphi) In the past, vehicles to be certified had to pass a standardized emission test based on a driving cycle that was performed on a roller test bench with special emission measurement system. The driving cycle is country-specific, e.g., US type testing uses FTP75 (Federal Test Procedure) whereas in Europe NEDC (New European Driving Cycle) was taken as reference. The problem of this test approach has been, that the test results are not consistent to the real driving emissions. For the reason of harmonization of the driving cycles and to improve emission and consumption under real driving conditions, the European Union introduced two new procedures: • A new driving cycle named WLTP (Worldwide Harmonized Light Test Procedure) for the roller test bench. • An additional emission measurement with a specific equipment named PEMS (Portable Emission Measurement System) on the road which is compared with the test results according to WLTP. 122 4.1 Real Drive Emission (RDE) Data Real driving emission (RDE) tests should not replace laboratory tests according to current NEDC and future WLTP but will complement them. RDE shall ensure that nominal emissions comply with those emitted on the road. Marking a major leap in emission measurements, on-the-road testing will be first introduced in Europe. From September 2017, RDE legislations applies for type approvals of new and from 2018 for any vehicle. It allow for a transition phase by tightening the conformity factors that define the tolerated deviation from the target emission values. 2 Motivation Under RDE test conditions, the test car will be driven on public roads within a wide range of conditions including: • High and Low altitudes • Year-round temperatures • Additional vehicle payload • Upand down-hill driving • Urban roads (low speed) • Rural roads (medium speed) • Motorways (high speed) Figure 2 is showing these test conditions in detail. Attribute Old NEDC (until 2015) New RDE regulation EU from 2017 ff Test environment vehicle dyno veh. dyno, real road tests Altitude range of tests 0..2400m different emission factors for <700m, 700m-1300m Mean road grade 0 m/ km 0 m/ km < mean grade <1200m/ 100km Max. road gradient +-0% no limitation Vehicle weight nominal cert. weight nominal weight ~ 90%* veh+max.passengers Ambient temperatures 18..24°C different em. factors: 3°C..30°C; 0..35°C Engine start temperature 18..24°C was: > 7°C; new: coldstart included Share (of total trip distance) Fixed (results from NEDC cehicle speed trajectory) urban (v ≤ 60km/ h, avg v = 15-40km/ h): 34% (29% - 44%) rural (60km/ h < v < 90km/ h) : 33% (± 10%) motorway (v > 90km/ h): 33% (± 10%) Fuel certification fuel fuel corresponding to owners manual (e.g. >= ROZ95) Test duration 20 min 90 - 120 minutes Vehicle Speed & Acceleration Very moderate limit of propulsion power by complex formula based on vehicle speed x positive acceleration may be exceeded in 5% of the test Figure 2: Broad Range of test conditions based on RDE requirements 123 4.1 Real Drive Emission (RDE) Data By RDE legislation, ranges of test conditions are given, not specific test cases. This makes testing to be a challenge for the driver. How can he make a successful RDE drive which satisfies all criteria? Of course there exist already tools in the automotive development are which are capable of verifying the compliance of RDE drives with regard to the conditions. However, they are confirming it after the RDE driving cycle has been performed. The idea of the approach described below targets to give the calibration engineers a tool which validates conformance during an RDE drive. 3 Development goal The idea was to develop a real time evaluation of RDE data during a test drive. The main focus is to provide the driver visually with information such as: • Driving behavior • Online visualization of NO X , particle number (PN), CO, CO 2 on the road in addition to regular (WLTP) emission tests on a roller test bench • Estimation during the drive if RDE boundaries can still be fulfilled • Synchronization of PEMS and ECU data • Advises the driver for distance and time yet to be travelled for the different drive cycles (Motorway, Rural, Urban) • Displays the (cumulated) status of the MAW (Moving Average Windows) calculation • Monitoring of different sensor data: emission sensor, GPS, PEMS Status, Memory Faults • Monitoring of RDE legislative boundary conditions such as: o Time, distance, and type of drive cycle o Environmental and engine conditions during the test o Vehicle speed and acceleration • Notification of emission measurement start • Prediction about the validity of the drive • Identification of emission critical engine map ranges Additionally, it was important to provide those visualization information for the calibration engineers in a standard calibration environment such as ETAS INCA. 124 4.1 Real Drive Emission (RDE) Data 4 Realization As already mentioned, the realization based on a system model representing the standard development environment of prototype cars equipped additionally with a standard PEMS system. The software model was directly integrated on standard calibration interfaces such as CAN. Both systems will be described in the following chapter. 4.1 System model INCA is connected to the Electronic Control Unit (ECU) by an ECU interfaces. This corresponds to the traditional calibration setup. To enable INCA to acquire NO X , PN, CO, and CO 2 PEMS data, PEMS hardware is connected to INCA over CAN. The PEMS system reads defined measurement parameter via the OBD interface, e.g., engine and vehicle speed. Besides this information, additional parameters such as GPS position and emission data such as NO x , will be send via CAN to the INCA application running on the laptop. In addition, the driver of an experimental vehicle will have access to the ECU via an ETK interface or by an XCP connection. These data are also send to INCA. 125 4.1 Real Drive Emission (RDE) Data 4.2 Software model Figure 3 shows the software architecture of the RDE online evaluation solution. PEMS measurement signals on CAN are collected by a signal preparation and buffered in a dispatcher/ scheduler serving the RDE validation algorithms. Figure 3: Software architecture of the RDE online evaluation. 4.2.1 Signal preparation In order to create a uniform interface for the evaluation algorithms, this module serves as an interface. Since the development of PEMS equipment is progressing fast and data transfer via CAN will be replaced by more efficient means, the module should be as flexible as possible regarding bus connections. The PEMS equipment and the ECU are used as sources of real data and signals. Additionally a simulation module could also serve as signal source for a later or repeated evaluation of the RDE drive. Today, CAN is still the prevailing vehicle bus. For this reason, this work is limited to the CAN use case. However, connection to other vehicle networks is considered in the design of the software. 4.2.2 Dispatcher/ Scheduler The dispatcher serves as a link between the interface module and the evaluation algorithms. That is, the module retrieves the data and transfers it to the evaluation. This module assures that all PEMS related data will be collected and sent to the RDE algorithm for evaluation in time. 4.2.3 RDE Algorithms In this module of the implementation, the formulas prescribed by the legislation for the calculation of the results of RDE drives are used. The formulas and all the framework conditions are implemented based on the state of current legislation. These algorithms cover the EMROAD and CLEAR legislations. 126 4.1 Real Drive Emission (RDE) Data 4.2.4 RDE visualization in INCA Based on the algorithms, the INST.DK Instrument Development Kit for INCA has been used to create own measurement displays. The gauges have been created with the goal to provide the driver with all necessary information about the fulfilment of the RDE requirements. Figure 4 is showing some examples of these gauges. Figure 4: Different INCA instruments for online visualization of RDE measurements 127 4.1 Real Drive Emission (RDE) Data 4.2.5 Recording measurements during RDE test drives In the implementation of the soft real-time system, the storage of the RDE measurement run is done by the INCA integrated recorder function. Reliable storage of measured data is possible in the 10 ms range and thus sufficient for the RDE evaluation that is performed in a 1 s raster. 5 Summary This system provides drivers with information about execution and evaluation of RDE-test drive focusing on optimization of RDE test drives. During RDE test drives, important information about driving route, already driven sections, and exhaust emissions, as well as boundary conditions to ensure compliance with legal requirements are provided visually to the driver. This includes predictions about the validity of the route, advice to the driver about distance yet to be travelled in respective cycle (motorway, rural, or urban), display of current percentage of valid windows, monitoring of NOx sensor, GPS, PEMS status, memory faults, and battery status of PEMS. The solution is fully integrated with INCA to allow calibration engineers to work with their conventional calibration tool chain. Figure 5: Different instruments for online visualization of RDE results in the INCA experiment. Literature: [1] Real Driving Emissions in the EURO 6 regulation on emission from light passenger and commercial vehicles (RDE3) [2] Delphi Worldwide Emissions Standards for Passenger cars and light duty 128 4.2 System optimization for automated calibration of ECU functions André Sell, Frank Gutmann, Tobias Gutmann Abstract Design of Experiments (DoE) is an important basis for the modelling of complex systems in modern calibration methods with the aim of increasing efficiency. However, DoE does not provide any generally applicable answers to the question of transferring the models to real-time ECU functions. These are often implemented by a plurality of dependent calibration parameters, which must be adapted successively in an iterative process. A new approach to system optimization allows the automated calibration of ECU functions on the basis of measured or modelled data by simultaneously optimizing all calibration parameters. For this purpose, first a reference model is generated on the basis of the data. The ECU function with the parameters to be calibrated is also available in the form of a flow chart representation. An optimization then appropriately adjusts all the parameters to be calibrated in such a way that the deviation between reference and ECU system behaviour is minimized taking into account limits, smoothness and priority criteria. This innovative method has been successfully applied by multiple OEMs in cooperation with the SGE Ingenieur GmbH. Functions such as load calculation, torque model or exhaust temperature model have been optimized with the help of the SGE ModelArtist software. The necessary development resources were more than halved. At the same time, potentials for future ECU development were demonstrated and verified by the automated and optimal calibration of alternative function proposals. Kurzfassung Design of Experiments (DoE) bildet eine wichtige Grundlage für die Modellierung komplexer Systeme in der modernen Applikationsmethodik mit dem Ziel der Effizienzsteigerung. DoE bietet aber keine allgemein anwendbaren Antworten auf die Frage nach der Übertragung der Modelle in echtzeitfähige Steuergerätefunktionen. Diese werden häufig durch eine Vielzahl abhängiger Applikationsparameter implementiert, welche nacheinander in einem iterativen Prozess angepasst werden müssen. Ein neuer Ansatz zur Systemoptimierung erlaubt die automatisierte Applikation von Steuergerätefunktionen auf Basis von gemessenen oder modellierten Daten durch gleichzeitige Optimierung aller Applikationsparameter. Dazu wird zunächst anhand der Daten ein Referenzmodell gebildet. Ebenso steht die Steuergerätefunktion mit den zu applizierende Parametern in Form einer Flussdiagrammdarstellung zur Verfügung. Eine Optimierung passt anschließend alle zu applizierenden Parameter 129 4.2 System optimization for automated calibration of ECU functions gemeinsam so an, dass die Abweichung zwischen Referenz- und Steuergeräte- Systemverhalten unter Berücksichtigung von Grenzen, Glattheits- und Prioritätskriterien minimiert wird. Diese innovative Methode wurde mit Hilfe der SGE ModelArtist Software erfolgreich von verschiedenen OEMs in Zusammenarbeit mit der SGE Ingenieur GmbH angewendet, um z.B. Funktionen wie Lasterfassung, Momentenmodell oder Abgastemperaturmodell zu optimieren. Die notwendigen Entwicklungsressourcen konnten dabei mehr als halbiert werden. Gleichzeitig wurden Potentiale für eine zukünftige Steuergeräteentwicklung aufgezeigt und durch die automatisierte und optimale Applikation von Funktionsalternativen belegt. 1 Motivation Modern development methods form the basis of projects with technical and economic success. Therefore Design of Experiments (DoE) is an important tool for modelling complex systems with the aim of increasing efficiency. However, so far this method has not provided any generally applicable answers to the question of transferring the models to real-time ECU functions. These are often implemented with a variety of interdependent calibration parameters, which have to be calibrated in an iterative process. 1.1 The challenge of calibration The aim of calibration is to create a reliable working system under all ambient conditions over production tolerances and ageing behaviour. For this purpose, application-oriented, empirical methods are often used. The effort required for calibration thereby increases roughly proportionally with the number of variants and development stages as each time data is generated, evaluated and processed into calibration parameters in an identical manner. Only little synergy effects result for repeated calibrations. For this reason, specific software tools are developed on a regular basis that are tailored to the calibration of particular ECU functions. Calibration is also made more difficult because an ECU function cannot yet represent the chemical-physical correlations of an internal combustion engine in a complete and real-time manner. Thus, often only main effects are implemented so that an entirely exact setting of the calibration parameters is no longer feasible and a compromise has to be made through prioritization. Finally another challenge appears with calibration parameters whose output values cannot be determined directly or not in the entire operating range. For example, Otto engine torque models or exhaust-gas temperature models often contain parameters which apply to 100% ignition efficiency and lambda one. If these operating ranges cannot be measured for reasons of knocking limits or component protection, only an indirect determination is possible. 2 System optimization A new approach of system optimization permits the automated calibration of ECU functions by simultaneously optimizing all calibration parameters. The aim is to minimize the deviation of the ECU system behaviour from a reference behaviour. 130 4.2 System optimization for automated calibration of ECU functions The reference behaviour describes the target state that the ECU function should map. It may be provided in the form of measured or simulated data or in the form of a data-based model. The ECU function to be calibrated includes at least one, but in general, several calibration parameters, as well as one or more input and output signals in each case. These same signals also form the interfaces of the reference system. Therefore, it is only necessary to map the inputs and outputs of the function. Information about signals within the function is not necessary, consequently the above-mentioned problem of signals which cannot be directly determined is no longer relevant. To determine the deviation between the ECU function and the reference system, the input signals of the reference system are applied to the ECU function and the resulting output signals are compared to those of the reference system. In the simplest case, the absolute differences for all the output signals are computed to a mean value that is used as a scalar quality criterion, which is then minimized by the optimization. Since the calculation of the quality criterion is freely definable, the user has the possibility to bias the optimization result. For example, weights can be defined depending on the operating point, the sign or the magnitude of the deviation. It is also possible to introduce empirical criteria to evaluate the calibration when it can be calculated from the input and output signals of the function. In this way, the welldefined calibration of functions is made possible, even if they cannot accurately represent the physical relations of the underlying system and thus do not permit a trivial evaluation on the basis of the deviation. The quality criterion is minimized by means of an optimization algorithm that is simultaneously varying all the calibration parameters included in the function. If some parameters are not to be optimized, they can be excluded. An existing initial calibration can be taken into account as the starting point of the optimization, thus reducing the required time of the optimization. Limitations and smoothness requirements can be considered for the calibration parameters as well. Limits are also used to achieve a reproducible result for underdetermined functions. This need exists when calibration parameters are summed up or multiplied and an infinite number of combinations exists that provide the same result. 2.1 Implementation The described method of system optimization was implemented as an extension of the SGE ModelArtist software. This tool, which is designed for visualizing, modelling and optimizing complex systems, supports the use of various data formats as well as extensive calculated channels. Thus, it offers direct access to the measurement and simulation data usually without data preprocessing. Gaussian process models are then used as the basis for the optimization algorithms. The essential components of the system optimization are the reference system, the ECU function to be calibrated and the quality criterion. The reference system may either be represented by the loaded data or by a model. While a model is generally preferred for its positive properties regarding plausibility, the data can be accessed directly in the case of very large data sets that cannot be modelled. To avoid unwanted effects resulting from extrapolation, the model can be restricted to the 131 4.2 System optimization for automated calibration of ECU functions parameter space of the data. The ECU function to be calibrated is depicted in the form of a flow chart with its calibration parameters, which are set with initial values. Thus, the flow chart can be simply derived from the representation that is usually used for the ECU software development. It is also possible to directly integrate a Simulink system. The quality criterion is implemented in the same way and thus permits the described influence on the optimization results. Before starting the optimization, the user configures smoothness criteria and limits for the calibration parameters using fixed values or maps/ curves depending on the input signal values. A progress bar permits a quick evaluation of the current progression of the optimization. In addition, at any time during the optimization there is access to interim results of the calibration parameters to be optimized, which can be viewed and stored. 3 Application example For an ECU torque model, the method of system optimization was used to calibrate the function, which is shown in Figure 1. The torque is depicted as a function of the ignition angle. The shape and position of the optimum of the function are to be calibrated in the form of maps and curves. Solely the friction map MAP_FRICT was excluded from optimization because it was retrieved from stationary measurements. 1 torque 1 eng_load 2-D T(u) u1 u2 MAP_MIOPT 2 ign_angle 3 eng_speed 2-D T(u) u1 u2 MAP_ZWOPT 2-D T(u) u1 u2 MAP_COEF 2-D T(u) u1 u2 MAP_FRICT 1-D T(u) CURVE_SQUARE 4 lambda 100% 1-D T(u) CURVE_ZWOPTLAM 2 delta_ign_angle 3 ign_efficiency 100% eng_speed eng_load torque_opt torque_friction Figure 1: Torque model ECU function with parameters to calibrate 132 4.2 System optimization for automated calibration of ECU functions As a reference system, a Gaussian process model was trained within the input range of the data (16000 data points). Figure 2 and 3 show the models intersection and quality visualization. Because of a dynamic measurement procedure of only 0.5s measuring time per setting, the data is superimposed by some variation. This is compensated by the model and did therefore not affect the calibration optimization. Figure 3: Torque model quality visualization Figure 2: Torque model intersection visualization 133 4.2 System optimization for automated calibration of ECU functions The quality criterion was implemented by integrating the Simulink system describing the ECU function and calculating the deviation from the model reference. The resulting signal is evaluated by the optimization to a scalar quality criterion. The calibration parameters are adjusted in such a way that this criterion is minimized. For this function with 5 parameters and a total of 800 individual values to be calibrated, the optimization needed 9 hours of computing time. The result of the optimization created plausible maps, which could be transferred directly into the ECU after post-processing and extrapolation into non-measured areas. As an example figure 5 shows a comparison of the initial and optimized+postprocessed calibration of the map MAP_ZWOPT. The efficiency gain when using system optimization increases disproportionately with the number of calibration variants, since the user only needs to make adjustments to the reference system and the quality criterion as well as post-processing of the calibration parameters if needed. The optimization itself is carried out independently without the user's intervention. In the present project the necessary development resources could be more than halved to achieve the high quality requirements regarding torque model accuracy. At the same time, potentials for a future ECU development were shown and verified by evaluating alternative functions with an automated and optimized calibration. torque_delta +- ECU function implemented as integrated Simulink system Deviation torque_ecu Reference torque "torque_ref" SIMULINK simulation (ECU_Fcn_TORQUE) "lambda" "ign_angle" "eng_load" "eng_speed" DCM (MAP_ZWOPT) DCM (CURVE_ZWOPTLAM) DCM (CURVE_SQUARE) DCM (MAP_COEF) DCM (MAP_MIOPT) ECU torque Input signals Calibration parameters Figure 4: Quality criterion calculation for torque model optimization 134 4.2 System optimization for automated calibration of ECU functions 4 Summary With the presented method of system optimization, the calibration engineer and software developer is provided with a universal tool for the automated calibration of ECU functions. The integration of the system optimization into a tool chain for modelbased calibration ensures seamless processing of measured data into optimized calibration parameters. The application of this method leads, besides significantly reduced effort, to an increasing quality and reproducibility of the calibration. Typical challenges of manual calibration processes are solved and there is no longer a need to develop specific software tools for single functions to calibrate. This innovative method has been successfully applied by several OEMs in cooperation with SGE Ingenieur GmbH to optimize ECU calibration. It significantly reduced development resources and also showed the potential of further software development. Figure 5: Calibration parameter MAP_ZWOPT. Comparison of initial and optimized+postprocessed calibration. 135 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Kento Fukuhara, Daniel Rimmelspacher, Wolf Baumann, Karsten Röpke, Yutaka Murata, Yui Nishio, Masato Kikuchi, Yukihisa Yamaya Abstract According to increasing system requirements, the complexity of modern Diesel engines has been significantly increased. Model-based calibration approaches have been successfully applied in the calibration of base engine maps with the help of Design of Experiment (DoE) methods. On the other hand, the calibrations of transient functions is mostly done within the vehicle due to its system and task complexity. In this paper, a model-based transient calibration approach is introduced, which makes use of a driving cycle simulation with prediction of engine out emissions. The emission cycle simulation contains Electronic Control Unit (ECU) as well as transient Diesel engine models. In order to generate nonlinear dynamic engine models, dynamic tests are designed and data driven models are trained. Actual ECU functions are integrated in to the simulation. Three transient functions are calibrated with the proposed calibration process. The new process features a newly developed simulation-optimizer interface as well as parallel computation, in order to increase the process usability and to compensate the increased computational load. As a result of the optimization, NOx improvement has been observed without any engine performance losses. Kurzfassung Aufgrund der steigenden Systemanforderungen hat sich die Komplexität von modernen Dieselmotoren zunehmend erhöht. Modellbasierte Verfahren wurden insbesondere für die Bedatung stationärer Kennfelder mit Hilfe von Design of Experiment (DoE) erfolgreich angewendet. Die Bedatung transienter Korrekturfunktionen erfolgt mangels geeigneter Ansätze und aufgrund höherer Komplexität derzeit noch überwiegend im Fahrzeug. In diesem Beitrag wird ein modellbasierter Ansatz für die Bedatung transienter Kennfelder vorgestellt, der auf dem Einsatz von Zyklussimulation und der Vorhersage des transienten Emissionsverhaltens basiert. Die Simulationsumgebung umfasst sowohl Steuergerätefunktionen als auch das dynamische Modell eines Dieselmotors. Für die Erstellung des Motormodells wurden geeignete Versuchspläne erstellt und am Motorprüfstand vermessen. Anschließend wurden datengetrieben Modelle trainiert. Für die Emulation der Steuergerätefunktionen wurde die echte Ansteuerlogik als Simulink Modell in die Simulation integriert. 136 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Drei transiente Funktionen wurden mit dem vorgestellten Verfahren optimiert. Das neue Verfahren zeichnet sich insbesondere durch eine neu entwickelte Schnittstelle zwischen Simulation und Optimierer als auch durch Parallelberechnung aus, um die Verwendbarkeit zu verbessern und die erhöhten Rechenaufwände zu kompensieren. Mit Hilfe dieses Optimierungsansatzes wurde eine NOx-Verbesserung ohne relevante Motorleistungsverluste erzielt. 1 Introduction Engine calibration is a process to determine the optimal system control strategy with defined hardware, software, schedule and resources. It aims at achieving defined requirements such as legislative requirements (i.e. emission norms), system requirements (e.g. durability, component protection, etc.) and customer requirements (e.g. drivability, noise, robustness, etc.). Recent enhancement of emission norms notably added further complexities of the Diesel engine calibration: I. Tightened emission limit values for more clean exhaust gas emissions II. Modification of testing procedure for a better correlation to market conditions: a. Introduction of new test cycles (e.g. WLTC, RDE) b. Enlargement of test conditions (e.g. temperature, altitude) Under these circumstances, Model-Based Calibration (MBC) has been identified as a valid methodology for increasing calibration efficiency and quality in order to compensate the increasing process complexity to be handled [1], [2], [3]. Together with Figure 1: V-Model in passenger car development and applied methodologies Design (virtual) Concept Vehicle Verification (real) Combustion concept Verification Requirement definition Software construction Specification setting System design Function design Parts design Prototyping parts Control design Engine unit function System integration Calibration (optimization) MBC MBD(Design) CAD/ CAE Fuel consumption, emissions, drivability, NV, power, OBD Planning CAE Plant model Controller Combustion CAE Engine parts Software design Software test Engine unit ECU implementation System test Vehicle Engine system Engine hardware and software Engine frame CAE Design CAE Evaluation CAE Countermeasure CAE Requirement analysis for engine software Cooling, lubrication, durability, NV Constraints analysis Data mining system (database) MBD: Model based development 137 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Design of Experiments (DoE) methodologies, efficiency of the system identification process and the model quality have been significantly improved. For early detection of specification changes and prevention of process step-backs, the V-model was introduced in the passenger car development with the aim to link the design (left-side of the V) and the verification phase (right-side of the V) at each system level (figure 1). Front-loading is an effective way to re-define the resource distribution between early and late development stage. Inefficient and cost-expensive vehicle tests (system tests) are being replaced by efficient and cost-effective laboratory tests (unit tests). Despite above mentioned all efforts, there is still plenty of room to improve for upcoming challenges with respect to tightening emission norms. One key note is that a huge part of emission calibrations, including calibration of exhaust aftertreatment system behaviour, is still performed on vehicle, although base emission calibrations are conducted on engine test bench prior to vehicle tests. First reason for this situation is that there is still a gap between vehicle system tests and laboratory unit tests. Although the combustion engine unit was calibrated according to defined requirements (which are normally sub-defined from whole system requirements), adjustments are usually needed when it comes to a whole system. This is mainly caused by lacking consideration of interactions from other sub-systems such like exhaust aftertreatment systems. Second reason is the rather neglected consideration of transient system behaviour; unit tests are usually performed under steady state mode and transient behaviour of the system is mainly examined in vehicle tests. Considering the dynamic emission modes to come (e.g. WLTC, RDE), optimization of transient system behaviour is the issue of greatest concern. Early validation of cycle emission performance is therefore desired in unit tests. In order to address above mentioned problems, a new model-based calibration process is proposed. While commonly used model-based calibration process utilizes steady state combustion engine models, utilization of transient models is essential to optimize and validate the transient engine behaviour. Together with combustion engine models, actual engine control and exhaust aftertreatment system behaviours shall be considered. There are three main challenges to be addressed in this paper: 1. Generation of accurate transient combustion engine model, 2. Set-up of transient emission cycle simulation with actual ECU logics, 3. Model-based calibration with the generated simulation model. In section 2, the model-based calibration process is presented. Calibration tasks handled in this paper are explained in section 3. Section 4 presents the set-up of the emission cycle simulation and generation process of each submodel. In section 4, calibration results are presented. 2 MBC approach for transient combustion optimization 2.1 Application of transient emission cycle simulation in MBC Main focus of this paper is the emission calibration of Diesel engine units. Ultimate goal of emission calibrations is to match emission norms with an engineering margin and given hardware and software specifications by minimizing fuel consumption. 138 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Testing procedure and limit values are defined by emission norms. In the case of light weight passenger cars, emission tests take place on vehicle (on chassis dynamometer) under controlled conditions. A number of tests are conducted to fulfill all requirements by adjusting combustion engine and exhaust aftertreatment system behaviors. Prior to vehicle tests, combustion engine units are evaluated on the engine test bench. In this phase, basic operation of the combustion engine as well as ECU virtual sensors are calibrated. At the same time, emission calibrations are conducted under steady state condition. Calibration targets are normally sub-defined from emission norms, considering defined cycle modes and limit values. Model-based calibration approach is commonly used in this case (details are explained in section 2.2). For the assessment of cycle emissions, several approaches exist prior to standard chassis dynamometer or road tests (table 1). Compared with cost intensive Vehiclein-the-Loop (ViL) or Powertrain-in-the-Loop (PiL) simulation, Engine-in-the-Loop (EiL) approach, meaning the vehicle simulation with a real combustion engine, is one effective approach to simulate transient vehicle behaviors on the engine test bench. This approach is however still time-consuming and not appropriate for the emission base calibration work, which handles over 500 calibration parameters at the same time. This is the reason why fully virtual approaches, meaning the vehicle simulation with combustion engine plant model and ECU models/ software, are preferred. Target of our work is the realization of a full virtual calibration environment with respect to transient engine and actual engine control unit behaviors. 2.2 MBC setup for optimizing transient engine combustion 2.2.1 MBC setup for optimizing steady state engine behaviour Emission calibration of today’s engine control software under steady state conditions was accomplished to a great extent by applying well-established methodologies from DoE. This MBC approach enables to address complex development targets by applying nonlinear optimization technique to steady state DoE engine models (figure 2). The actual engine control function is usually not part of the optimization loop, as its Table 1: Variation of vehicle simulations Office Laboratory Engine test bench Powertrain test bench Chassis dyno Real driving on road SiL / MiL HiL EiL Pil ViL Engine Virtual Real Aftertreatment Virtual Real Transmission Virtual Real Control unit Virtual Real Vehicle Virtual Real Driver Virtual Real Environment Virtual Real Road Unit test Lab Simulation 139 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization influence can be neglected due to the steady state operation of the engine. Since the optimization is performed offline, the method mainly benefits from its very low computational requirements, allowing the optimizer trying out many more input combinations (a) Comparison of accumulated NOx emissions in hot WLTC extra high phase (b) Comparison of model prediction behaviors between steady state and dynamic engine models over hot WLTC extra high phase Figure 3: Prediction of mode cycle emission (hot WLTC extra high area) with different simulation set-ups 100.0 97.0 157.0 211.9 0 50 100 150 200 250 Measured (sec-to-sec) Dynamic DoE model (sec-to-sec) Steady state DoE model (sec-to-sec) Steady state DoE model (weighted sum) Cumulated NOx in % 1450 1500 1550 1600 1650 1700 NOx in mg/ s Time in sec Measured Dynamic DoE model Steady state DoE model Cumulated NOx Figure 2: Standard MBC setup for optimizing steady state control software functions Combustion engine model (steady state) Grid points Grid points DoE software Simple look-up tables Map Map Map Map Engine outptus Optimizer NE TQ TQ NE TQ NOx Fuel 140 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization than any real-time setup would permit. In the case of modern Diesel engines, over 10 combustion parameters are to be calibrated for optimizing emission behaviors. This results in handling of over 500 optimization parameters since the base map of each combustion parameter may contain over 100 grid points. In the steady state emission calibration, target emission cycles and limit values are considered, even with the limitation that the models are not capable to predict the transient behavior. This approximation has been successfully realized by using clustering technique and application of the weight on each selected point. This approach however has certain limitation on the prediction accuracy. Figure 3 (a) shows accumulated NOx results from different simulation set-ups (hot WLTC extra high phase, values are shown in % compared to measurement data). While the prediction with dynamic and steady state DoE model (sec-to-sec) was conducted by feeding measured model input traces, the prediction with steady state DoE model (weighted sum) was performed by applying generated weight map from the same input traces. The diagram shows that not only the reduction of time-resolved trace information, but also the elimination of transient influence has big impact on the cycle predictive accuracy. Figure 3 (b) displays the predictive capacity of dynamic DoE model under transient conditions (also hot WLTC extra high phase). While the steady state DoE model tends to predict high peak values, the dynamic DoE model is capable to predict proper transient engine behavior considering model input trajectories from previous time steps. This is the motivation of transient model application in extended MBC. 2.2.2 Extended MBC setup for optimizing transient engine behaviour For accurate prediction of mode cycle emissions, it is desirable to extend the MBC methodology and benefit from its advantages in similar fashion. This requires three fundamental modifications compared to the standard MBC setup (figure 4): I. Replacement of the steady state engine model by a suitable dynamic model II. Implementation of closed-loop transient ECU functions for the simulation of real engine behavior under transient mode. III. Optional: Implementation of other sub-systems like exhaust aftertreatment systems for searching the best possible system settings. Figure 4: Extended MBC setup for optimizing transient control software Trace inputs Emission cycle simulation Use of standard calibration tool • ETAS INCA • ETAS INCA Flow Utilization of IAV Optimization Framework (System optimization tool) Electronic control unit model Combustion engine model (dynamic) Exhaust aftertreatment model Trace data inputs (NEDC, WLTC, RDE) 141 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization The first modification can be realized by applying results from recent investigations which provide suitable methods for dynamic system identification as well as modeling in the context of respective engine applications [4], [5], [6]. The second modification, implementation of the actual software function in the optimization loop, however, can be achieved in different ways. Considering the computational load, Model-in-the- Loop (MiL) or Software-in-the-Loop (SiL) solutions are possible choices to take [7], [8], [9]. The third modification, implementation of other sub-systems like exhaust aftertreatment system, can be realized by applying latest studies of physical-based catalyst modelling [10], [11], [12]. Details of simulation set-up are discussed in section 4. Challenges of those modifications are the preparation of transient emission cycle simulation and increased simulation load. Necessary modifications in the MBC process are therefore discussed in the following section. 2.3 Extended MBC process 2.3.1 MBC process for transient calibration Despite necessary extensions of the MBC set-up, usability of a well-established MBC process should remain and operations should be performed in similar manner. However, additional tasks are inevitable. Following the required MBC process of the transient engine calibration is shown. Main differences to the standard process can be found in the model preparation and optimization steps: I. Creation and validation of sub-models a. Transient combustion engine plant models b. Actual engine control logics c. Exhaust aftertreatment models (optional) II. Integration of sub-models into a simulation model III. Import simulation model into optimization software IV. Set-up and execution of optimization While steady state DoE engine models are created and optimizations are directly performed with DoE software in the standard MBC process, new process requires additional effort for preparing a cycle simulation model. The simulation model consists of several sub-models: combustion engine plant models, actual engine control models and additional sub-component models. The optimization is performed with much powerful and flexible optimization software to handle any kind of simulation setups, since there is now no standard for the simulation model structure. 2.3.2 Optimization environment and approach Optimization with actual combustion control models brings additional challenges. The main challenge is the increased computational load. While computationally expensive simulations have been used in the hardware design or concept development, rather simple and practice-oriented solutions have been preferred in the field of engine calibration. This is because of the number of optimization parameters to be handled (already discussed in section 2.1). In terms of the optimization problem, problems handled in the engine calibration have much higher degree of complexity. In this regard, following fundamental measures are applied in the extended MBC process: 142 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization I. Approximation of the optimization problem a. Approximation of optimization map values to map surface b. Approximation of simulation model (sub-modelling) II. Acceleration of the simulation a. Application of the parallel computing technique b. Acceleration of simulation speed via appropriate system simplification III. Improvement of optimization efficiency via intelligent algorithm a. Limitation of search design space via system input constraints First measure is the approximation of the optimization problem. In order to compensate the increased computation load, number of optimization parameters should be reduced. For this purpose, parameterization of map surface is an effective approach. Further, sub-modelling of the target simulation model (so called meta-modelling) is a well-applied optimization strategy in model-based development. Second measure is the acceleration of the simulation itself. In this regard, utilization of the parallel computing technique is the most straightforward solution. Another important aspect is the simplification of the simulation model. This point is discussed in section 4.3.2. Third measure is the efficient handling of the optimization procedure. Since one function evaluation requires higher computational load, the decision, either the evaluation of given set of optimization parameters is worth doing or not (e.g. due to inacceptable calibration map gradient or smoothness), becomes fundamental. For optimization, IAV optimization framework has been used. The optimization framework was designed to handle various types of simulation models and calibration tasks. Figure 5 shows the basic idea of the optimization framework. Different from standard MBC process, the target simulation system is built in MATLAB Simulink environment. This optimization environment permits the utilization of external software like MEX-Files and DLLs. Two important requirements for the simulation model are: 1. all calibration labels (map, curve or scalar value) are available in model/ base workspace, 2. All simulation signals required for the calibration task are recorded. This allows the optimization framework for tuning parameters and evaluat- Figure 5: Idea of IAV optimization framework MATLAB SIMULINK model Workspace NOx Speed Load BaseMap CorrMap Software MATLAB & SIMULINK (Simulation and optimization environment) Optimizer x = fmincon(fun,x0,A,b,Aeq,beq,lb,ub,nonlcon) Minimization of cumulated NOx , application of smoothing constraint for BaseMap and CorrMap MEX DLL External software (MEX-Files, DLLs) 143 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization ing simulation results. After importing the simulation model, the user can select any kind of calibration labels available in the simulation and build optimization objectives from existing simulation signals. For the calibration labels, commonly used map smoothness and gradient constraints can be applied. Another important feature is the handling of fixed point systems. Regarding this issue, automatic configuration of minimal parameter step sizes is implemented. 3 Calibration of transient functions 3.1 Target Diesel engine and calibration function The new calibration procedure was applied to a four cylinder Diesel engine with production ECU. The engine equips single Variable Geometry Turbocharger (VGT), high pressure and a cooled low pressure Exhaust Gas Recirculation (EGR) circuit. For the control of the combustion and the optimization of the emission performance, following control parameters may be calibrated: I. Main injection timing II. Pilot injection timings III. Pilot injection amounts IV. Rail pressure V. Boost pressure VI. Fresh air amount VII. Low pressure EGR fraction Combustion control mainly consists of two parts. First part is so called open-loop control part. In this part, demand values of combustion parameters are stored in base calibration maps (typically over engine speed and injection quantity in Diesel engine). Second part is the closed-loop part. In this part, determined combustion demand values in the open-loop part are modified considering detected transient condition by equipped sensors or estimators. Calibration of open-loop part is usually called steady state or base calibration. The latter is called transient calibration since the control logic is rather designed to improve transient engine behavior. Although majority of emission performances is determined by the base calibration, further improvement and adjustment of emission performance is possible by the transient calibration. 3.2 Target transient calibration tasks Proposed MBC process is demonstrated with three transient calibration tasks. Three tasks are selected since those transient calibrations require representation of transient engine behavior and actual combustion control logics. In this study, WLTC was selected as target emission mode. In the following section, the three calibration tasks are explained. 144 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 3.2.1 Calibration of smoke limit function The applied ECU contains special functions for controlling the engine response during load transients, when the combustion is impacted by a distinct lambda drop, leading to high amount of unfavorable emission [13]. Control of the exhaust gas composition can be achieved either by fuel path or by air path control [14], [15]. A typical function for fuel path control is given by the smoke limiter, which cuts of the fuel amount in case of very low lambda values. This, however, also reduces the engine torque, which is usually not desired by the driver. The situation can be relaxed by using transient air path control, in order to avoid these low lambda values beforehand. For the given control software, it is respectively possible to reduce the EGR rate when lambda falls below a certain level, leading to additional amount of fresh air. The limitation of the EGR is applied according to the minimal lambda determined by the current engine operating point and estimated EGR mixture. Determined minimal lambda is then compared with actual lambda detected via equipped sensor. Demand EGR rate is then modified from its original value coming from EGR base map. The function's behavior can be adjusted by changing values in minimal lambda definition maps. Figure 6 illustrates the situations described above for a typical driving situation from the WLTC, containing several positive load steps. The diagrams show simulated WLTC with 121 different smoke limit calibrations. The line color is synchronized between diagrams and sorted by cumulated NOx mass flow in observed phase; lighter line color means higher NOx emission. While the engine speed and fuel delivery trajectories are constant (upper two diagrams), trade-off between NOx and Soot generation can be seen by observing line color index (third and fourth diagrams). This Figure 6: Simulated transient trajectories controlled by EGR limit (Hot WLTC t = 760…790 sec) Engine speed Fuel quantity NOx mass flow Soot mass flow Lambda HP-EGR valve flow LP-EGR valve flow Fresh air mass flow 145 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization trade-off was caused by different lambda limit calibrations (fifth diagram). By tuning minimal lambda, EGR and fresh air mass behaviors can be controlled (lower three diagrams). 3.2.2 Calibration of injection timing correction function Second calibration task is the calibration of injection timing correction function. Injection timing of a Diesel engine is calibrated to control the combustion centre over crank angle. An appropriate setting of combustion centre realizes a good compromise between combustion performances (e.g. combustion efficiency, NOx, Soot emission). The optimal injection timing is however influenced by in-cylinder gas state, since the application of higher EGR slows down the combustion speed, resulting in late combustion centre. Purpose of this function is the adjustment of combustion centre under transient operation. This function is essential for the transient combustion optimization, since base injection timing is defined over engine speed and load (in base map), while in-cylinder gas state can be impacted by boost pressure and EGR delay even at the same engine speed/ load operating point. Figure 7 illustrates the functionality of the injection timing correction in the same fashion as figure 6. The injection timing correction is adapted based on current EGR rate (engine feedback). The correction behavior can be controlled by two calibration maps. Figure 7: Simulated transient trajectories controlled by main timing correction (Hot WLTC t = 760…790 sec) Engine speed Fuel quantity NOx mass flow Soot mass flow Main timing HP-EGR valve flow LP-EGR valve flow Fresh air mass flow 146 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Figure 8: Behavior of LP-EGR delay compensation function Figure 9: Simulated transient trajectories controlled by LP-EGR compensation (Hot WLTC t = 760…790 sec) 0 1 2 3 4 5 6 7 8 9 10 EGR mass flow Time in sec Fuel mass flow EGR mass flow Injection amount step, engine speed constant, target LP-EGR fraction 100% Actual HP-EGR Actual LP-EGR Actual total EGR Target total EGR Engine speed Fuel quantity NOx mass flow Soot mass flow Lambda HP-EGR valve flow LP-EGR valve flow Fresh air mass flow 147 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 3.2.3 Calibration of LP-EGR delay compensation function Equipping long LP-EGR loop, delay of EGR gas delivery is inevitable under steep load increase. This is why a LP-EGR delay compensation function is introduced in the applied ECU [16]. The function applies HP-EGR to compensate the shortage of EGR gas when LP-EGR delay is detected (figure 8 and 9). The behaviour of LP-EGR compensation can be controlled by adjusting PID controller factors which determine the estimation of delayed EGR gas amount. Those factors are mapped in two calibration maps. 4 Set-up of transient emission cycle simulation 4.1 Target system overview and simulation set-up For the demonstration of transient calibrations, an emission cycle simulation was constructed in Simulink environment. Since only the longitudinal vehicle dynamic is considered, input to the system is the target vehicle speed defined by target driving cycle. The vehicle speed information will then be translated in to engine speed and acceleration pedal values via vehicle and driver systems. Giving those information, ECU torque coordinator determines target injection quantity which imaginably corresponds to engine break torque. According to the determined injection amount, base demand values of airand fuel-path parameters are determined. Those values are then modified and actuator set values are determined in order to achieve desired combustion. Actual engine input values are the results of the actuator control loop. Engine outputs are given accordingly. The engine outs are observed via internal and external sensors. Some sensor values are then fed back to ECU (closed-loop control). Exact representation of above explained system has several difficulties and some degree of simplification is therefore required. First simplification is required due to high computational effort of vehicle and powertrain simulation including transmission and driver systems. This problem can be relaxed with an assumption that the calibration takes place with fixed target emission cycle, driver and vehicle specifications. By applying this assumption, simulation input can be replaced by engine speed and engine break torque. However, the engine control unit requires ECU load parameter (= demand fuel quantity). In order to simplify the simulation, the vehicle torque coordinator was replaced by DoE engine model. Beside above simplification, real behaviour of fuel path actuator was eliminated with the assumption that demand value can be realized without any deviation. Sensor systems are modeled together with combustion engine models. In this simulation, an exhaust aftertreatment system is not implemented. 4.2 Dynamic DoE engine model 4.2.1 Model overview In order to consider the dynamic system behaviour, Diesel engine models were constructed containing two main parts (figure 10). First part is the air path actuator block 148 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization and second part is the combustion block. The actuator block simulates transient air path behaviours and predicts actual combustion set values (e.g. boost pressure, fresh air mass, LP-EGR fraction) according to actuator valve positions. As model inputs, combustion parameters like engine speed, torque, main injection timing and rail pressure are also used. Those parameters are the representative image of energy flow through the turbine which determines the air path behaviours. In the simulation, this block is controlled via ECU closed-loop controls. The combustion block simulates the combustion engine and sensor behaviours. This model includes emission models which are necessary in emission calibration tasks. 4.2.2 Test design Dynamic DoE test plans are generated for the identification of dynamic engine behaviour. Different to steady state DoE test plan, dynamic DoE test plan (excitation signals) is designed in time domain, defining demand combinations of system inputs at each time step. Not only the model input design space (i.e. parameter variation range), but also the frequency domain (i.e. parameter variation speed) of system inputs are considered at dynamic DoE test planning. While the design space defines the area in which models are capable to predict system behaviour accurately (e.g. covered engine operation area), the frequency domain defines the degree of dynamics models are capable of (e.g. less dynamic NEDC or more dynamic WLTC mode). Chirp and APRBS (Amplitude modulated Pseudo Random Binary Sequence) test design algorithms are two well applied test planning algorithms [5]. While chirp test design generates sinusoidal curves in defined frequency domain, APRBS test design creates signals consisting of a series of ramp and hold sequences. Chirp test design has the advantage of smooth signals and well defined frequency domain. APRBS test design is appropriate for the improvement of steady state predictive accuracy of identified models, since its signals contain ω = 0 Hz domain (hold phase). In order to obtain accurate models under both transient and steady state mode, chirp and APRBS test design are combined. At the test planning, the parameter design space was examined by boundary detection tests in advance. Test results were evaluated and the design space was specified by using table and convex hull based constraints (figure 11 (a)). Frequency domain of chirp signals was defined considering target emission cycles to be simulated. Figure 11 (b) shows an example of the generated excitation signals (chirp test plan). Figure 10: Specified transient engine model structure Rail pressure VGT position Main injection timing Intake flap valve position HP-EGR valve position LP-EGR valve position Boost pressure EGR rate LP-EGR fraction Engine speed Fuel flow Pressures Emissions Temperature Lambda Noise Engine brake torque Air path actuator models Combustion engine models 149 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization (a) Applied table based design space constraints (b) Generated excitation signals over time Figure 11: Example of chirp test plan Main timing Rail pressure Boost pressure EGR rate LP-EGR fraction Engine speed Time in sec Engine load Main timing Rail pressure Boost pressure EGR rate LP-EGR fraction 150 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 4.2.3 Nonlinear dynamic modelling After executing transient DoE measurements, engine models are trained on the basis of the parametric Volterra series approach (figure 12). The model structure consist of polynomial nonlinearity, Finite Impulse Response (FIR), linear combination and Infinite Impulse Response parts (IIR). Prior to model identification works, synchronization of measured emission signals was performed considering gas travel times. By feeding obtained DoE measurement data, model coefficients are estimated using least squares regression. Model structure, like polynomial order of nonlinearity and input delays of FIR, were adjusted for each model output to obtain the best compromise between fitting and validation RMSE. For the assessment of model accuracy, Root Mean Square Error (RMSE) and normalized Root Mean Square Error (nRMSE%) shown with the following equations are applied. For the normalization of the RMSE, minimal and maximal value from fitting and validation measurement data are used respectively. ∑ (1) % ∑ · 100 (2) In general, appropriate model prediction accuracy was obtained with all modelled engine outputs. Air path models are important components, since the representation of transient engine behaviour is the main focus of this study. Emission signals are another important component for the emission calibration. While rather good model accuracy was obtained from NOx and CO2 emission model, improvement of Soot emission model is still desired. Figure 13 shows some examples of predicted engine out signals compared with real measured signals. Figure 12: Schematic expression of parametric Volterra series model 2 · · 1 2 · 1 · 1 1 2 · · 1 1 z -1 z -1 z -1 + + + + + + + Polynomial function z -1 z -1 z -1 + + + + · 3 · 2 · 1 Polynomial nonlinearity Finite impulse response (FIR) Linear combination Infinite impulse response (IIR) 151 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 4.3 ECU model 4.3.1 ECU model overview Simulink ECU model was implemented in the emission cycle simulation. The used models are originally designed for software development and coding purposes. Since the models are not compiled as a software, used simulation solution can be recognized as MiL. For computational speed purpose however, control functions are compiled as S-Function. Fixed data type is handled with the given models and calibration data (e.g. map values) are stored in MATLAB workspace. Figure 13: Comparison between measured and predicted engine out signals (WLTC) Vehicle speed 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 2.27 : 1.19 Simulated Measured Cumulated sum : 30788.43 (+2.4%) : 30067.95 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Fresh air mass flow nRMSE% RMSE Statistics : 2.27 : 1.19 Simulated Measured Cumulated sum : 30788.43 (+2.4%) : 30067.95 Cumulated Fresh air mass flow Cumulated Fresh air mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 2.63 : 0.08 Simulated Measured Cumulated sum : 871.86 (+0.5%) : 867.61 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Fuel mass flow nRMSE% RMSE Statistics : 2.63 : 0.08 Simulated Measured Cumulated sum : 871.86 (+0.5%) : 867.61 Cumulated Fuel mass flow Cumulated Fuel mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 2.89 : 2.02 Simulated Measured Cumulated sum : 9585.39 (+0.8%) : 9506.36 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 NOx mass flow nRMSE% RMSE Statistics : 2.89 : 2.02 Simulated Measured Cumulated sum : 9585.39 (+0.8%) : 9506.36 Cumulated NOx mass flow Cumulated NOx mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 3.23 : 0.20 Simulated Measured Cumulated sum : 147.44 (-24.2%) : 194.44 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Soot mass flow nRMSE% RMSE Statistics : 3.23 : 0.20 Simulated Measured Cumulated sum : 147.44 (-24.2%) : 194.44 Cumulated Soot mass flow Cumulated Soot mass flow Time in sec 152 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 4.3.2 Function selection The ECU model consists of several control parts (e.g. torque coordinator, air path control, fuel path control, virtual sensor, aftertreatment control, etc.) and each control part has a number of functions. The ECU model originally links engine sensor feedbacks and actuator demand values. However, some modifications on model interface (ECU model inputs and outputs) and elimination of ECU functions were necessary due to the applied simplification discussed in section 4.1. Selection of necessary ECU functions was performed with an analytical approach. Considering the necessary ECU model outputs (= engine model inputs), the necessary ECU functions are automatically searched backward step by step. As a results of the analysis, a set of necessary functions and unknown signals (= signals which are not provided due to the elimination of ECU functions) are identified. Through this process, the number of ECU functions was reduced and the simulation duration was drastically reduced. For the unknown signals, either a constant value or a signal calculated by DoE engine model was applied. Based on the implemented ECU model, the behavior of transient functions can be now simulated. 4.3.3 Simulation results Figure 14 shows the comparison between measured and simulation signals. The simulation signals were obtained by only giving engine speed and torque traces in the simulation. Observing engine model input parameters (fig. 14 (a)), appropriate behaviour of ECU models can be observed. However, some deviations can be observed from EGR valve flows. This was caused by lacking accuracy of plant model sensor feedback. Figure 14 (b) shows the measured and simulated engine outputs. Improvement and degradation of emission prediction accuracy was caused by ECU model behaviour. 153 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization (a) Simulation results of engine model input parameters Figure 14: Comparison between measured and simulated signals (WLTC) Vehicle speed 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 4.10 : 0.64 Simulated Measured Cumulated sum : -4518.79 (-3.3%) : -4673.42 0 200 400 600 800 1000 1200 1400 1600 1800 Main injection timing nRMSE% RMSE Statistics : 4.10 : 0.64 Simulated Measured Cumulated sum : -4518.79 (-3.3%) : -4673.42 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 2.88 : 39.81 Simulated Measured Cumulated sum : 1095341.80 (+1.1%) : 1083220.50 0 200 400 600 800 1000 1200 1400 1600 1800 Rail pressure nRMSE% RMSE Statistics : 2.88 : 39.81 Simulated Measured Cumulated sum : 1095341.80 (+1.1%) : 1083220.50 Measured Simulated Time in sec 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 13.87 : 13.69 Simulated Measured Cumulated sum : 134403.85 (-6.5%) : 143709.17 0 200 400 600 800 1000 1200 1400 1600 1800 VGT position nRMSE% RMSE Statistics : 13.87 : 13.69 Simulated Measured Cumulated sum : 134403.85 (-6.5%) : 143709.17 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 15.52 : 11.00 Simulated Measured Cumulated sum : 63221.75 (+20.8%) : 52327.82 0 200 400 600 800 1000 1200 1400 1600 1800 HP-EGR valve pos. nRMSE% RMSE Statistics : 15.52 : 11.00 Simulated Measured Cumulated sum : 63221.75 (+20.8%) : 52327.82 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 10.52 : 6.08 Simulated Measured Cumulated sum : 32729.87 (+6.3%) : 30801.46 0 200 400 600 800 1000 1200 1400 1600 1800 LP-EGR valve pos. nRMSE% RMSE Statistics : 10.52 : 6.08 Simulated Measured Cumulated sum : 32729.87 (+6.3%) : 30801.46 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 16.49 : 14.57 Simulated Measured Cumulated sum : 109878.07 (-1.1%) : 111129.31 0 200 400 600 800 1000 1200 1400 1600 1800 Intake flap valve pos. nRMSE% RMSE Statistics : 16.49 : 14.57 Simulated Measured Cumulated sum : 109878.07 (-1.1%) : 111129.31 Measured Simulated 154 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization (b) Simulation results of engine model output parameters Figure 14: Comparison between measured and simulated signals (WLTC) Vehicle speed 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 4.76 : 2.48 Simulated Measured Cumulated sum : 28576.51 (-5.0%) : 30066.70 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Fresh air mass flow nRMSE% RMSE Statistics : 4.76 : 2.48 Simulated Measured Cumulated sum : 28576.51 (-5.0%) : 30066.70 Cumulated Fresh air mass flow Cumulated Fresh air mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 3.12 : 0.09 Simulated Measured Cumulated sum : 881.25 (+1.6%) : 867.62 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Fuel mass flow nRMSE% RMSE Statistics : 3.12 : 0.09 Simulated Measured Cumulated sum : 881.25 (+1.6%) : 867.62 Cumulated Fuel mass flow Cumulated Fuel mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 5.93 : 4.13 Simulated Measured Cumulated sum : 6763.42 (-28.8%) : 9505.67 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 NOx mass flow nRMSE% RMSE Statistics : 5.93 : 4.13 Simulated Measured Cumulated sum : 6763.42 (-28.8%) : 9505.67 Cumulated NOx mass flow Cumulated NOx mass flow 0 200 400 600 800 1000 1200 1400 1600 1800 nRMSE% RMSE Statistics : 2.98 : 0.18 Simulated Measured Cumulated sum : 184.45 (-5.1%) : 194.44 Measured Simulated 0 200 400 600 800 1000 1200 1400 1600 1800 Soot mass flow nRMSE% RMSE Statistics : 2.98 : 0.18 Simulated Measured Cumulated sum : 184.45 (-5.1%) : 194.44 Cumulated Soot mass flow Cumulated Soot mass flow Time in sec 155 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 5 Calibration of transient functions 5.1 Calibration approach By using the generated simulation model, presented transient calibrations were performed. At the optimization, four calibration maps and two curves from three transient calibrations were optimized simultaneously, with respect to cycle simulation results (e.g. accumulated NOx, CO2 and Soot of WLTC). The total number of optimization parameters reaches over 200. Due to high computational load of the generated simulation model (approximately 2 times faster than real time), direct optimization of all map grid points is unrealistic. Hence the sub-modelling approach was applied and an optimal calibration setting was found from generated sub-model. In the sub-modelling approach, calibration maps were approximated by giving offset values to its base map value. This reduces number of optimization parameters down to 6 (each calibration map has one offset value). Behavior of emission cycle simulation model was sub-modeled by steady state DoE technique. Figure 15 (a) shows approximated calibration map surfaces (base map values + offset). The offset values were varied considering calibration range and model validation area. Figure 15 (b) displays pairwise view of the generated steady state spacefilling DoE test plan. The test plan contains 6 input offset parameters and total 1,200 DoE test points. One advantage of meta-modeling approach is that simulations can be fully parallelized, since all test points are defined beforehand. The simulations and subsequent signal evaluations were carried out by using the optimization framework (section 2.3.2). Thanks to applied parallel computing technique (using 24 workers), 1,200 WLTC simulations were performed within 20 hours. 5.2 Results Figure 16 (a) … (c) show the result of 1,200 DoE simulations. Each figure contains a plot with simulated signals for whole WLTC (lower west), an enlarged plot (lower east) and a plot with accumulated signals (upper west). The line color is indexed by cumulate NOx and synchronized over plots. Figure 17 summarizes the simulation results with colored markers. Calibration trade-off between three parameters can be seen (NOx, Soot and fuel consumption). While the impact of calibration setting on fuel economy is rather small, clear trade-off between NOx and Soot emission can be observed. By using obtained simulation results, steady state sub-models were generated by using Gaussian process model. For the finding of optimal NOx/ Soot trade-off line, optimizations were performed to minimize NOx emission with different Soot constraints. Figure 18 and table 2 describe the optimization results showing the optimal trade-off front. By applying introduced calibration procedure, NOx emission was reduced by 2% without any engine performance losses. 156 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization (a) Calibration maps and applied offsets Figure 15: 6-dimensional steady state DoE test plan for sub-modelling (b) Pairwise view of space filling DoE test plan for map offsets Figure 15: 6-dimensional steady state DoE test plan for sub-modelling Smoke limit lambda correction factor EGR rate LP-EGR delay characteristic map 1 Main injection timing correction factor Engine speed Main injection timing correction offset Smoke limit lambda correction offset LP-EGR delay characteristic map 2 Smoke limit lambda correction factor LP-EGR delay characteristic map 1 Main injection timing correction factor Main injection timing correction offset Smoke limit lambda correction offset LP-EGR delay characteristic map 2 LP-EGR delay characteristic map 1 Main injection timing correction factor Smoke limit lambda correction offset LP-EGR delay characteristic map 2 157 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization (a) NOx mass flow (b) Soot mass flow (c) Fuel mass flow Figure 16: Results of 1,200 emission cycle simulations (DoE simulation) Time in sec NOx mass flow Cumulated NOx mass 50 sec Time in sec Soot mass flow Cumulated Soot mass 50 sec Time in sec Fuel mass flow Cumulated fuel mass 50 sec 158 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization Figure 17: Obtained optimal setting with applied MBC process Figure 18 Obtained optimal results with simulation meta-models Table 2: Summary of optimization results Cumulated NOx mass flow over WLTC cycle in % Cumulated Soot mass flow over WLTC cycle in % Cumulated fuel mass flow over WLTC cycle in % Base setting Obtained trade-off scatter range Cumulated NOx mass flow over WLTC cycle in % Cumulated Soot mass flow over WLTC cycle in % Cumulated fuel mass flow over WLTC cycle in % Base setting Base 1 2 3 4 5 Optimization result # Base 1 2 3 4 5 Model outputs Cumulated NOx mass flow [%] 100.0 98.1 94.7 92.2 90.2 90.0 Cumulated Soot mass flow [%] 100.0 100.0 105.0 110.0 115.0 120.0 Cumulated fuel mass flow [%] 100.0 100.0 100.0 100.0 100.0 100.0 Input settings (offset values) Smoke limit lambda correction factor 0 0.072 0.058 0.034 0.004 0.004 Smoke limit lambda correction offset 0 -0.999 -0.999 -0.917 -0.876 -0.911 LP-EGR delay characteristic map 1 0 -0.0299 -0.0299 0.0656 0.0894 0.0990 LP-EGR delay characteristic map 2 0 -0.0019 -0.0019 -0.0021 -0.0021 0.0029 Main injection timing correction factor 0 -0.100 -0.100 -0.100 -0.100 -0.100 Main injection timing correction offset 0 -0.524 -0.926 -1.000 -1.000 -1.000 159 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization 6 Conclusion In this paper, a new model-based calibration process was proposed and applied for three transient calibrations of a Diesel engine. The new model-based process utilizes the emission cycle simulation which contains actual ECU functions and a transient engine plant model. The optimization of calibration maps was conducted with submodelling approach. Transient engine models were generated by using dynamic DoE approach. Excitation signals were designed considering design space and frequency ranges, and they are measured on an engine test bench. Coefficients of nonlinear data-driven models were adjusted to respond to the measured training data. As a results, models with appropriate accuracy were obtained for the optimization work. Optimization was performed with sub-modelling approach. This means, the full emission cycle simulation was sub-modelled by using conventional steady state DoE approach in order to reduce simulation load. By using identified sub-models, three transient maps were calibrated. As a result of the optimization, NOx improvement was seen without any engine performance losses. Predictive accuracy of the dynamic engine model is a key for the successful calibration with the presented process. While some engine output models (e.g. NOx emission, fuel mass flow, air mass flow) are well modelled and show good correlation to measured signals, improvement is necessary for other quantities such as Soot emission. Simplification of ECU structure is another key issue when emission cycle simulation is handled in the optimization. For the real calibration work, acceleration of simulation model is fundamental since the number of calibration parameter rises up to 500. While sub-modelling approach is well-recognized optimization procedure, the current optimization approach shall be improved by direct optimization. Although the exhaust aftertreatment system was not implemented in this paper, the component must be considered when true system optimum is desired. In the future, achievement of further fuel efficiency and fulfilment of target emission limits under real driving conditions becomes essential. Optimization of transient system behaviour is the key technology. Considering this point, development of advanced calibration methodologies is fundamental for future engine calibration approaches. Literature [1] W. Baumann, T. Dreher, K. Röpke, S. Stelzer: DoE for Series Production Calibration, 7th Conference on Design of Experiments (DoE) in Engine Development, 2013 [2] C. Haukap, B. Barzantny, K. Röpke: Model-based calibration with data-driven simulation models for non-DoE experts, 6th Conference on Simulation and Testing for Automotive Electronics, 2014 160 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization [3] Y. Murata, Y. Kato, T. Kanda, M. Sato: Application of Model Based Calibration to Mass Production Diesel Engine Development for Indian Market, 8th Conference on Design of Experiments (DoE) in Engine Development, 2015 [4] W. Baumann, K. Klug, B.-U. Köhler, K. Röpke: Modeling of transient diesel engine emissions, In: DoE in Engine Development IV. expert verlag. 2009 [5] W. Baumann, S. Schaum, K. Röpke, M. Knaak: Excitation Signals for Nonlinear Dynamic Modeling of Combustion Engines, IFAC World Conference, Seoul, Korea, 2008 [6] T. Godward, H. Schilling, S. Schaum: Use of Dynamic Testing and Modeling in the Engine Calibration Process, 7th Conference Design of Experiments (DoE) in Engine Development, 2013 [7] D. Rimmelspacher, W. Baumann, P. Klein and R. Linssen: Transient Calibration Process using Chip Simulation and Dynamic Modeling, 7th Conference Design of Experiments (DoE) in Engine Development, 2013 [8] M. Simons, M. Feier and J. Mauss: Using Chip Simulation to optimize Engine Control, 7th Conference Design of Experiments (DoE) in Engine Development, 2013 [9] J. Mauss, M. Simons: Chip simulation of automotive ECUs, 9th Symposium on Automotive Powertrain Control Systems, Berlin, Germany, 2012 [10] G. Koltsakis, N. Margaritis, O. Haralampous, Z. Samaras: Development and Experimental Validation of a NOx Trap Model for Diesel Exhaust, SAE transactions: Journal of Fuels and Lubricants SAE 2006-2001-0471, 2006 [11] G. Koltsakis, N. Margaritis, O. Haralampous: NOx Trap Modeling and Experimental Validation Using Ultra-fast Analyzers, Topics in Catalysis, Vol. 42-43, Issue 1-4, 2007, Pages 65-69, 2007 [12] N. Margaritis, O. Haralampous, G. Koltsakis: Modeling of the NOx trap and experimental validation using ultra-fast NOx analyzers, Topics in Catalysis 42- 43(1): 65-69, 2007 [13] D. Alberer, L. del Re: Optimization of the transient Diesel engine operation, 9 th International Conference on Engines and Vehicles, Naples, Italy, 2009 [14] C. D. Rakopoulos, E. G. Giakoumis: Diesel Engine Transient Operation, Springer- Verlag, London, England, 2009 [15] L. Guzella, C. H. Onder: Introduction to Modeling and Control of Internal Combustion Engine Systems, Springer, Berlin, Germany, 2010 [16] Y. Nishio: Application of Model Based Control to the Mass Production Diesel Engine using Dual EGR system, HORIBA CONCEPT JAPAN 2014, Japan, 2014 Glossary MBC Model-Based Calibration MBD Model-Based Development DoE Design of Experiments 161 4.3 Dynamic MBC Methodology for Transient Engine Combustion Optimization MiL Model in the Loop (simulation) SiL Software in the Loop (simulation) HiL Hardware in the Loop (simulation) EiL Engine in the Loop (simulation) PiL Powertrain in the Loop (simulation) ViL Vehicle in the Loop (simulation) FIR Finite Impulse Response IIR Infinite Impulse Response RMSE Root mean Square Error NEDC New European Driving Cycle WLTC Worldwide harmonized Light vehicles Test Cycle RDE Real Driving Emission VGT Variable Geometry Turbocharger EGR Exhaust Gas Recirculation NOx Nitrogen oxide APRBS Amplitude modulated Pseudo Random Binary Sequence 162 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU Simon Wunderlin, Antoine Hurstel, Ernst Kloppenburg Abstract Modern Diesel engines and their control are becoming more and more complex (number of actuators and degrees of freedom). Thus, heuristic or physical modelling methods reach limits regarding accuracy and calibration effort. Data-based modelling of complex and high-dimensional dependencies with modern machine learning methods and efficient acquisition of the database using DoE methods is state-of-theart. The universally applicable method ASC@ECU allows to apply such models in electronic control units under real-time conditions using the hardware acceleration module AMU which is available on Bosch MDG1 engine ECUs. Furthermore, the method supports efficient software development and calibration by providing a seamless tool chain. This paper explains the procedure using the example of the exhaust gas temperature of a Diesel engine, it discusses the obtained results regarding complexity handling, functionality and the efficiency improvement in software development and calibration. Additionally, it shows the further potential of the method. Kurzfassung Künftige Anforderungen an die Motoren führen zu steigender Komplexität (Anzahl der Freiheitsgrade, Steller) der Systeme, mit der Folge, dass heuristische oder physikalische Modellierungsmethoden hinsichtlich Genauigkeit und Applikationsaufwand an Grenzen stoßen. Die datenbasierte Modellierung komplexer hochdimensionaler Zusammenhänge mit modernen Machine-Learning-Verfahren und effiziente Gewinnung der Datenbasis mittels DoE-Verfahren ist gängige Praxis. Die universell einsetzbare Methode ASC@ECU erlaubt mittels der auf Bosch-Motorsteuergeräten der Generation MDG1 vorhandenen Hardwarebeschleunigung AMU, solche Modelle direkt unter Echtzeitbedingungen für Steuergerätefunktionen einzusetzen, sowie eine effiziente Softwareeinbindung und Bedatung durch eine geschlossene Toolkette. Am Beispiel der Funktion zur Modellierung der Abgastemperatur eines Dieselmotors werden das Vorgehen und die erzielten Ergebnisse bezüglich Beherrschbarkeit der Komplexität, Funktionalität und dem Effizienzgewinn in Softwareentwicklung und Applikation erläutert, sowie das weitere Potenzial der Methodik dargestellt. 163 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU 1 Introduction The Bosch engine ECU 1 software calculates a model for the exhaust gas temperature at engine outlet valve. This modelled temperature signal is used as a virtual sensor for other software functions (such as EGR 2 temperature model or boost control functionality) because the temperature sensor is not able to show fast temperature changes due to its slow dynamic behavior. Additionally, the modelled value is used for plausibility check of the temperature sensor. The exhaust gas temperature at engine outlet valve is influenced by various parameters: On the one hand by air system parameters which define the physical state in the intake manifold, on the other hand by the injection system whose numerous degrees of freedom affect the exhaust gas temperature. Especially the conditions in the combustion chamber mostly cannot be measured which makes it difficult to describe them by a physical model or to calibrate such physical models. Therefore, very often empirical models are used for the exhaust gas temperature. The current Bosch ECU exhaust gas temperature model is a data-based model and consists of look-up tables for the temperature behavior under reference conditions and additional corrections for the local deviations from these reference conditions. The reference conditions are mainly defined by engine operation point (engine speed and load) and engine operation mode. Hence, many parameters which affect the exhaust gas temperature are not considered explicitly, but only implicitly depending on the reference conditions. During the calibration process for a project, the air and injection system calibration is usually optimized and changed several times. Therefore, the reference conditions for the exhaust gas temperature model change as well, which makes it necessary to recalibrate the temperature model. Additionally, modern Diesel engines and their control become more and more complex due to future legislation. New degrees of freedom will also affect the exhaust gas temperature. However, it is difficult to extend the existing model. Due to these reasons, a new exhaust gas temperature model is proposed which considers the relevant influencing parameters explicitly. The new model shall have an improved accuracy with reduced calibration effort compared to the current model. Additionally, it shall be extendible easily to fulfill future requirements. 2 ASC@ECU 3 : Technical approach For the new exhaust gas temperature model, Bosch has decided to apply the Gaussian process regression algorithm (GPR), [1]. The GPR algorithm is a nonparametric regression method for modelling non-linear dependencies with many influencing values. Contrary to parametric methods (such as map based models), nonparametric models do not require a predefined function structure. The complexity of the function results from the provided data. 1 Electronic control unit 2 Exhaust gas recirculation 3 ASC: Advanced simulation for calibration 164 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU The software tool ASCMO by ETAS GmbH uses the GPR algorithm in order to model and display multivariate interrelations. It is commonly used for the development and calibration of ECU functions and for the optimization of engine operation or components. With the current engine ECU generation MDG1, Bosch provides the possibility to export multivariate GPR models from ASCMO to the ECU and to calculate them in real time. The calculation of GPR models requires to compute a high number of exponential functions which makes it very runtime consuming. For this reason, each MDG1 4 ECU contains an additional hardware acceleration chip, the so-called AMU 5 , [2]. GPR models for complex interrelations may exceed the limits of available ECU resources (runtime, memory). For such cases, it is possible to compress the GPR model by training a substitute model with significantly fewer data points. As the resulting substitute model uses the same type of evaluation formula as the original GPR model, it can also be computed on the AMU. Figure 1: Tool chain for implementing ASC models on engine ECU 3 Approach applied to temperature model 3.1 Data base The data base for the development of the new exhaust gas temperature model is taken from DoE 6 measurements which have been executed at a 4 cylinder Diesel engine on the engine test bench. The measurements consist of 650 stationary measurement points with the following signal variations: 4 Device 3 or higher 5 Advanced modelling unit 6 Design of experiment 165 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU ● Engine speed: 750 - 3600 rpm ● Fuel injection quantity: 0 - 58 mg/ stroke ● Rail pressure: 300 - 2170 bar ● Start of energizing of the main injection: 17 °CA 7 before TDC 8 - 8 °CA after TDC ● Air flow sensor mass flow: 150 - 1000 mg/ stroke ● Intake manifold pressure: 970 - 2740 hPa ● Intake valve upstream gas temperature: 28 - 80 °C ● EGR split factor (low pressure EGR fraction of total EGR): 0 - 100 % ● Fuel quantity of the post injection: 0 - 4.5 mg/ stroke ● Start of energizing of the post injection: 26 - 69 °CA after TDC The measurement points are distributed in this 10 dimensional data space according to a space filling experimental design. The design space has been constrained to plausible operation conditions. 3.2 Selection of model input parameters In order to achieve a good model, it is essential to consider the right input signals. For determining the optimal input combination for the new model, a superset of potentially relevant input values has been defined. This superset consists of the following 13 values which have been selected based on physical consideration and system knowledge: ● Engine speed ● Fuel injection quantity ● Rail pressure ● Start of energizing of the main injection ● Air flow sensor mass flow ● Intake manifold pressure ● Intake manifold gas temperature ● Fuel quantity of the post injection ● Start of energizing of the post injection ● Total EGR ratio ● EGR split factor (low pressure EGR fraction of total EGR) ● High pressure EGR valve position ● Low pressure EGR valve position According to formula (1), 8191 different combinations exist to select an arbitrary input vector of above 13 input values. 8191 )! 13 ( ! ! 13 13 13 1 13 1 = ⎟⎟⎠ ⎞ ⎜⎜⎝ ⎛ − ⋅ = ⎟⎟⎠ ⎞ ⎜⎜⎝ ⎛ ∑ ∑ = = i i i i i (1) 7 Crank angle 8 Top dead center 166 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU For each of the 8191 different input combinations, an ASCMO model has been created for evaluating the model accuracy. The result is shown in figure 2 where the “leave-one-out 9 ” root mean square error (RMSE) of each model is plotted as function of the number of input signals. The figure contains two plots: In the foreground all models are shown, whereas the background plot only shows the models with the best input combination for each number of inputs. Figure 2: Model error as function of model input number Conclusions of these results: - Adding an input generally improves the model quality because it provides additional information about the system. - At some point, the model quality reaches a level which cannot be improved any further by adding more inputs. This point is reached with 8 inputs in the given example. With more inputs, the model error even increases slightly which indicates that some of the input parameters are not independent of each other and that the model is over-determined. - The quality of the correlation between the input values and the exhaust gas temperature is much more important for the model accuracy than the number of input values. E.g., the model error for the best combination of 4 input values is 63 % below the model error for the worst combination of 10 input values. - The best combination with N input parameters does not necessarily contain all the input values of the best combination with N-1 parameters. - As the GPR algorithm creates the model based on the data, ignoring any physical relations, the best input combination obtained for a model must always be reviewed with respect to physical plausibility. The best results obtained for given training data do not necessarily provide the best results for independent test data. 9 Test each data point against a model containing the other N-1 data points (N is the to-tal number of data points) 167 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU For the given example, the best combination with the following 8 input values has been chosen for the new exhaust gas temperature model: ● Engine speed ● Fuel injection quantity ● Rail pressure ● Start of energizing of the main injection ● Air flow sensor mass flow ● Intake manifold pressure ● Intake manifold gas temperature ● Fuel quantity of the post injection This input list seems to be plausible. The fuel injection quantity has the highest influence on the exhaust gas temperature as it contains the energy which is burned in the combustion chamber. The physical state of the gas in the intake manifold is represented by the mass flow, the pressure and the temperature. The start of injection influences the position of the combustion which affects the efficiency of the combustion and thus the temperature. The engine speed influences the reaction time in the combustion chamber and the rail pressure affects how the fuel is sprayed into the combustion chamber. 4 Results 4.1 Training data from the engine test bench The accuracy of the new model is to be compared with the accuracy of the current map based model. In order to do this, the latter was calibrated using the data from the engine test bench DoE measurements which are described in Chapter 3.1. Figure 3 shows the model error of the current model (left part) and the new ASC@ECU model. The accuracy of the new model is much improved. Both models are data based models, but the new ASC@ECU model has two crucial advantages: 1. The current model only considers 6 input values explicitly. Other influencing parameters are considered implicitly based on the reference conditions. On the other hand, the new model is taking 8 input values into account which makes it to be more robust against deviations from the reference conditions. Furthermore, it can be easily extended for the case that an additional degree of freedom is to be considered because the GPR algorithm derives the model structure directly from the data. However, for extending the current map based model, requires to find a suitable mathematical relation between the existing and the added part which is not always easy to find. 2. The ASC@ECU based model considers all interdependencies between the in-put values, the model is covering the complete multivariate data space. On the other side, the effect of the corrections in the current model (e.g. mass flow correction) are independent of the other model inputs. 168 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU Figure 3: Comparison of the accuracy of current and new model 4.2 Test data with vehicle The data from the engine test bench measurements were used at the same time as training and test data. More realistic results are achieved if different data are used for training and testing. For obtaining test data, stationary and transient measurements have been executed at a vehicle with an equal engine. 4.2.1 Stationary vehicle measurements The vehicle has been operated on a test track using a brake trailer to assure stationary conditions. Around 130 stationary operating points have been measured. Figure 4 shows the measured and modelled exhaust gas temperature for each of these operating points. The accuracy is also quite good, although lower than at the test bench engine. But this was expected since the engine test bench data have been used for training the model. At some operation points, the model still can be improved. For example at low loads, the model is underestimating the temperature. This is due to the fact that those particular operating conditions were not included in the training data. The average model error for the shown points is 19 K, whereas the RMSE is 29 K. 169 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU Figure 4: Measured vs. modelled temperature at stationary vehicle measurements 4.2.2 Transient vehicle measurements Figure 5 shows the sensor based exhaust gas temperature (dashed line) and the modelled temperature (solid line) during a road measurement with the same vehicle. The temperature sensor cannot measure fast temperature changes, it shows a relatively slow transient behavior. However, the exhaust gas temperature can change very rapidly, mostly in cases of fast changes of the injection quantity which can change from one combustion cycle to another. As the model is implemented as quasi-stationary model, it takes quick changes of the input values into account. Hence, the modelled value shows a faster transient behavior than the sensor value. 170 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU Figure 5: Measured vs. modelled temperature at transient vehicle measurements, without modelling the sensor dynamics Figure 6: Measured vs. filtered modelled temperature at transient vehicle measurements 171 4.4 Implementing a real time exhaust gas temperature model for a Diesel engine with ASC@ECU The exhaust gas temperature model can be considered as virtual sensor because the modelled value provides the fast transient behavior of the temperature which cannot be measured directly. On the other hand, modelled and measured temperatures cannot be compared due to the different transient behaviors of model and sensor. In order to provide a better comparability, the modelled value is low pass filtered with a mass flow dependent time constant. The result is shown in figure 6 where the low pass filtered value of the model output and the measured exhaust gas temperature are shown. The filtered modelled temperature has the same transient behavior as the measured temperature and the accuracy is very good. 5 Summary and outlook The current Bosch engine ECU generation MDG1 contains an additional HW acceleration unit (AMU) which allows to calculate GPR models or RBF nets in real time on the ECU. Bosch additionally provides a seamless tool chain for creating, analyzing and ex-porting such models into the ECU software. This technique has been implemented for developing a new exhaust gas temperature model. As the new model is considering more system parameters than the previous model, it works in an extended data range. In addition, the input data space is completely covered, whereas the old model only provided some base calibration maps with local corrections. These points lead to a better accuracy, especially in operating conditions which deviate from the reference conditions. Furthermore, the calibration effort can be reduced because there are no calibration recursions necessary if the reference calibration is updated. The new model structure allows to extend the model easily in case of new system parameters need to be considered. Measurements at the vehicle show that the new model is also well transferrable to an-other system. In stationary and also transient operation, the new model provides good results. The exhaust gas temperature model is just one example of applying the technical approach ASC@ECU. There are many applications possible, especially for multidimensional problems where the relations cannot be easily described with physical formulas and/ or where intermediate values cannot be measured. This applies to many processes which happen in the combustion chamber. It is favorable to use the gathered DoE measurements also for other purposes. For ex-ample, with a global DoE measurement of an engine, models for temperatures, raw emissions, set points and more can be derived. Bibliography [1] C. E. Rasmussen, C. K. I. Williams: "Gaussian Processes for Machine Learning"; The MIT Press, 2006. ISBN 0-262-18253-X. [2] K. Röpke, C. Gühmann (editors), in cooperation with 65 co-authors: "Design of Experiments (DoE) in Powertrain Development", p.227-241: R. Diener et al: "Data-based Models on the ECU"; ISBN: 978-3-8169-3316-8 172 5 Dynamic 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Taro Shishido, Jing He, Masataka Kaihatsu, Carsten Haukap, Thomas Dreher, Michael Hegmann Abstract Recently, various OEMs and suppliers focus on Model Based Calibration methodologies to cope with the increasing calibration workload due to the increasing demands of the emission legislations. This paper gives an example of dynamic modelling calibration process for a common Gasoline DI engine. Dynamic models allow the introduction of global and dynamic optimization methods for the daily calibration work, in combination with a drastically reduction of test bench time for the model building process. This paper describes the data driven, dynamic DoE model process. Further, the derived model will be used to compare two base calibrations and optimization methods: the common global steady state approach and the new global dynamic approach. For the dynamic optimization, the Optimization Framework and a common dynamic drive cycle is used. One challenging part of the dynamic model building process are the test bench measurements and its automation. In experiments, a sinusoidal chirp excitation input is used for speed, load, spark, valve timing, A/ F, fuel pressure and injection timing. With the usage of the sophisticated test bench automation, the reaction of limit violation could be executed within the frequency of system sample rate. As result, the generated simulation model can estimate dynamic engine output from input parameters with high accuracy. This allows efficient optimizations for the calibration. Still challenging are the optimization approaches and methodologies. 1 Introduction In recent years, much effort has been expended to improve the model based calibration (MBC) methodologies to cope with the increasing requirements of the emission legislation. Due to this development and in combination with the introduction of efficient calibration processes the workload of the calibration engineer was reduced. Even more important, the required test bench time was very much reduced, too. Furthermore, it became more and more important to model the transient behaviour of the engine to build up an environment to simulate the entire vehicle. Here, a simulation environment for vehicle and cycle simulations for the optimization of emissions and operation and aftertreatment strategies will be introduced. 173 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Even though the field of dynamic modelling is still a focus topic of methodology development, this paper introduces the dynamic MBC process for the daily calibration work: a process based on the measurement of time resolved data and the introduction of a sophisticated model building process for dynamic engine models as well as a process for global map optimization. The benefits of this development is versatile, a massive gain of quality in terms of optimisation results, traceability and efficiency in combination with a reduction of cost and test bench time. 2 Calibration process using dynamic DoE models The calibration process using dynamic DoE models consists of seven steps: the first step is the task definition to define the overall target of the calibration task and known limitations and conditions. A short test bench phase follows to retrieve engine constraints and boundaries. During test plan phase the entire experimental designs will be created, necessary for the given task. The actual engine identification is done in a second short test bench phase to collect the data followed by the dynamic modelling phase. The introduced transient MBC process allows a very flexible handling. Where as the global map optimization or cycle simulation is the main focus, the steady state information and models are automatically retrieved. Thus the process is flexible for both approaches, steady state and dynamic calibration work, refer to Figure 1 [1]. Figure 1: Flowchart of transient calibration process 2.1 Task definition The overall task is the base calibration of a direct injected 4 cylinder gasoline engine. Whereas the derived dynamic models can be used for the majority of base test bench calibration tasks, such as air charge, torque and temperature model 174 5.1 Dynamic Modelling for Gasoline Direct Injection Engines calibration, this paper describes the example of a camshaft optimization. The optimization task shall be a global closed loop optimization of a WLTC cycle for a given vehicle and engine combination. To meet realistic engine operation, the dynamic models are combined with a model of the ECU in a Simulink environment. The given WLTC defines the operating range of the dynamic models, too. Besides, the WLTC cycle, replicated on the engine test bench, is used to retrieve frequency and gradient information for the later test design and for the validation of the final DoE models. The control parameters of the DoE models are chosen to be engine speed and air charge, intake and exhaust valve timing, rail pressure and start of injection, lambda and finally spark timing. For this examination, the boost pressure controller identification has decided not to be an explicit task. Therefore the turbo charger wastegate and the throttle position are the result of the boost pressure controller. This constraint is important as the final models are dependant of the controller calibration. Finally, the results of the global optimization using dynamic models will be compared with the conventional steady state MBC process. This comparison is easily possible, because the steady state DoE model can be easily retrieved form the dynamic DoE models. Figure 2: Input and output parameters for engine model 2.2 Test bench set-up Figure 3 shows the engine test bench set-up. The test bench system iTest operates the engine and dynamometer, auxiliary systems as well as all measurements devices. INCA is used to operate the ECU and is hooked up to the engine via an ES910. The test bench system ORION in the main component to control the engine. ORION is connected to iTest using the ASAP3 High Speed Server and is connected to INCA using the Direct INCA Interface. This set-up allows parallel control of the dynamometer as well as the access to all engine maps within a time frame of 10Hz. Furthermore, this set-up in combination with the sampling rate is fundamental to be able to react on limit violations, especially knocking. Here, the knock detection is realized using the measurement device KIS4. 175 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 3: Engine test bench layout 2.3 Constraints and boundary detection To ensure optimal test design for dynamic DoE model building, information of the engine’s design space and frequency is mandatory. Here, to detect the steady state parameter ranges, the ORION Boundary Finder function is used to automatically screen the multidimensional operation range of the engine [2] [3]. Figure 4 shows as example the result of the engine dependencies of the engine parameter as three dimensional convex hulls. Figure 4: 3-D convex hull of the engine design space 176 5.1 Dynamic Modelling for Gasoline Direct Injection Engines The information of the engine and parameter operation range is extremely important to avoid limit violations within the test design. Using the Boundary Finder results, the danger of exceeding an engine limit while exciting the engine dynamically may not be completely avoided, because of the steady state nature of the methodology, but very much limited. For the definitions of the static and dynamic test plan boundaries, the later purpose of the calibration task is essential. For an unknown engine it may be important to define full parameter ranges for the test design to be able to cover the complete engine operation range for a base calibration or to evaluate engine potentials. If the calibration task is more specific or to improve the model quality, it may be advantageous to limit the parameter range towards a reasonable offset of the input domain. Figure 5 shows two examples of a design space for a full parameter range and a limited range using an offset for given engine maps. Figure 5: Example of input parameter ranges for the test design (Left: Full parameter range, Right: Limited range using an offset to a given map) Finally, for dynamic engine excitation the setting speed of the parameters must be defined. Here, the DMT automatically estimates the maximal gradients using the power spectral density (PSD) for each defined input parameter of a given measurement. Typically any valid engine cycle such as the WLTC or FTP has been proved to provide sufficient information of dynamic DoE models. Figure 6 plots the PSD and frequency domain of the input parameters. Figure 6: The spectral analysis of input parameters 177 5.1 Dynamic Modelling for Gasoline Direct Injection Engines 2.4 Experimental design The Dynamic Toolbox (DMT) is used for the design of the test plans. The objective of the test design is to stimulate the system engine in way to be able to identify a time dependant, empirical data driven model. For this, the test design must follow basic boundary conditions. The system must be stimulated in the entire frequency domain of interest. And the excitation must be twice as fast as the requested time resolution of the later model (Shannon theorem). To obtain the required time dependant system response of the engine, a sinusoidal pattern is used as test design. Here, the pattern is an amplitude modulated chirp signal, whereas the amplitude represents the parameter range and the chirp frequencies represent the later frequency resolution of the time dependant model (parameter gradients). Refer Figure 7 a). Figure 7: Test design types of the DMT To improve the steady state prediction capabilities of the model, it is recommended to introduce a test design with a steady state excitation, as shown in Figure 7 b), because a sinusoidal excitation does not cover the very low frequency regime for w ω 0. Finally to cope with large time scales of the system reaction, a stepwise excitation as shown in Figure 7 c) may be used as option, too. Overall temperature changes for dynamic cyclic excitation are understood as system responses with large time scales. Further typical examples are the boost pressure rise (delay) for a fast engine acceleration. The latter design types: the ramp and hold excitation as well as the step excitation, are special cases of the known APRBS signal (amplitude modulated pseudo random signal). Here, the DMT offers the advantage to adopt the test design more easily and efficient to the desired model task. The final test plans are calculated to provide a space filling or optimal design space coverage. Main influencing parameter to control the design space coverage in is the duration of the excitation sequence, especially for the sinusoidal test plans or the number of ramp and hold points. 178 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 8: Splitting of the design space into zones The introduction of individual zones for the test design bears additional advantages: Main advantages are a better design space coverage, flexibility in case of different engine control modes (varying injection patterns, boost pressure control modes, etc.). Finally, the test design should not exceed a time limit to avoid large result files and to allow emission bench purging and calibration and engine conditioning. Figure 8 shows an example the zoning introduced for this investigation. Table 1: Test designs and test plan duration planned for this investigation Model Purpose No Excitation Type Zone Duration Total Time[h] Cycle Simulation 1 Sinusoidal I 4 × 45min 3.0 II 4 × 45min 3.0 III 4 × 45min 3.0 IV 4 × 45min 3.0 V 4 × 45min 3.0 VI 4 × 45min 3.0 VII 1 × 45min 0.75 2 Ramp and Hold I 2 × 45min 1.5 II 2 × 45min 1.5 III 2 × 45min 1.5 IV 2 × 45min 1.5 V 2 × 45min 1.5 VI 2 × 45min 1.5 VII - - 3 Step Excitation I - - II 1 × 45min 0.75 III 1 × 45min 0.75 IV 1 × 45min 0.75 V 1 × 45min 0.75 VI 1 × 45min 0.75 VII - - Sum 31.5 179 5.1 Dynamic Modelling for Gasoline Direct Injection Engines 2.5 Measurements & limit violations and limit reaction for dynamic testing The highest priority while operating the engine on the test bench is to prevent the engine from damages due to restricted engine setting. Further, the later model quality strongly depends on the quality of the collected data. Therefore is mandatory to install a test bench automation which is able to operate and excite the engine dynamically and being able to watch and react on limit violations. Especially to prevent a Gasoline engine from knocking, over temperature and misfire is extremely challenging. Two methodologies for the reaction of limit violations shall be introduced shortly. 2.5.1 Safe track methodology A straight forward solution for a system reaction in the case of a limit violation is shown in Figure 9 and names ‘Safe Trace’. The save trace corresponds to a parameter combination, which is known to be safe, for example the base settings form the maps for the given speed and load combination. In case of a limit violation, the automation system steps all parameters towards the safe trace within a definable step with until the limit violation is removed. After a short stabilization time, the automation systems ramps back all parameter towards the normal demand trace. The disadvantage of this methodology is, that typically all base maps are optimized for steady state engine operation. Hence, the danger of violating a limit while moving and operating the engine on the safe track may not be excluded. Figure 9: Limit reaction using a ‘Safety Trace’ 2.5.2 Limit reaction using offsets For a gasoline engine operated at high speed and load, the ‘safe trace’ methodology turned out not to be flexible and fast enough to react on engine knocking, over temperature and misfire in parallel. The limiting factor is, that the methodology does not allow an evaluation of the root cause of the violation. The reaction always follows the same procedure. In order to cope with an individual handling of different violation cases, ORION offers up to three limit groups, which can be mapped individually to a given violation. This 180 5.1 Dynamic Modelling for Gasoline Direct Injection Engines allows a reaction to a violation without provoking another limit violation. Furthermore, the limit reaction of the systems allows a reaction within the time frame of one sample, here ~10Hz of the used ASAP3-HS interface. This turned out to be fast enough to operate the engine safely 2.6 Data post processing and model building Two model types are available in the DMT, the parametric Volterra series polynomial model and the Gaussian process models. Further, the model building process is very much simplified, the latest version even allows to find optimal model parameters settings automatically. The advanced user is, of course, still able to tune the models individually varying various settings, such as output transformation, model orders, feedback, filter types or model limits and restrictions. Indeed, the quality of the modes strongly depends on a careful data post processing. Especially the alignment of the channels is very important. For this, the DMT offers two approaches. At first, the gas travel or dead times of the measurement is determined automatically. Using this information, the models are trained. In a second step the models are optimized to handle rise times. Using this technique the later models turn out to be optimal to cope with complex system behaviour, such as thermal inertia of the engine and probes, analyser characteristic, etc. Figure 10 shows an example of a system response with a large dead and rising time Figure 10: Delay time The models are available as common Matlab m-files as well as Simulink models. Further, a static steady state model can easily be retrieved to be available in the IAV EasyDoE environment. 2.7 Model validation For the verification of the models a vehicle cycle, reproduced on the engine test bench, was used. Here, the cycle has been the same as for the determination of the parameter gradients for the test design. The advantage of a replicated vehicle cycle is that any influences of the measurement devices can be neglected. To rate the model quality, various statistical characteristics are taken into account, such as the absolute and relative RMSE (root 181 5.1 Dynamic Modelling for Gasoline Direct Injection Engines mean square error, MSE (maximal square error) R² (coefficient of determination), MAD (mean absolute deviation) and MAD (mean absolute deviation). For a visual validation, Figure 11 shows on the left hand side a time based comparison of the measurement and on the right hand side a predicted over observed plot. The plot allow a fast evaluation of the overall robustness, outliers and the limitations. In addition, but not shown in this paper, the design space of the model is checked against the parameter ranges of the cycle to avoid or determine model extrapolation. Figure 11: Validity check 2.8 Model results In the following, the model results will be briefly introduced and discussed on the base of a comparison of a cycle measurement and the cycle simulation of the models. From this, three cycles have been replicated on the test bench: An FTP, NEDC and WLTC, refer Figures 12 - 16. While evaluating the models one has to take one issue into account: The models are trained for fired engine operation, only. Fuel cut conditions are not taken into account. Therefore all results show large deviations for this engine operating mode. Therefore the shown models results are the extrapolated. All calculated errors (RMSE) include a model extrapolation. Fuel mass flow model The overall normalized RMSE of the fuel mass flow model is ~ 1.3% for the set of all three vehicle cycles, showing very small deviations against the measurement on a time resolved scale. The steady state prediction capability is very good. 182 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 12: Fuel mass flow model THC model The overall normalized RMSE of the THC model is ~ 3.5% for the set of all three vehicle cycles, showing acceptable deviations against the measurement on a time resolved scale. The steady state prediction capability is acceptable. Figure13: THC model NOx model The overall normalized RMSE of the NOx model is ~ 3.5% or 195ppm absolute for the set of all three vehicle cycles, taking the Fuel cut conditions into account, the results would me much better. The steady state prediction capability is very good. 183 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 14: NOx model Soot and PN models The overall normalized RMSE of the soot and PN models are ~ 3.5..4% for the set of all three vehicle cycles. Here, the overall errors assumes an acceptable model quality, where as an evaluation on sec-by-sec base show large deviations, especially for the acceleration phases of the vehicle. Although the trends comes out correctly, the secby-sec accuracy is not acceptable. (R² = 0.85 …088). Low reproducibility of the soot emission measurement devices may be a reason. Figure 15: Soot model 184 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 16: Particle number model 2.9 ECU map optimization and verification Two optimization approaches shall be discussed in this paper. At first a global steady state MAP optimization. These results will be compared with a global and dynamic optimization of a given CYCLE. 2.9.1 Global steady state MAP optimization The global steady state MAP optimization methods follows a two step procedure. Start point for the global optimizations are the calculated local optima for a given distribution of speed and load on the base of steady state models. In order to take smoothness of input as well as output model parameters into account, the gradient for each node will be calculated upon its surrounding and is used as constraint for the second optimization loop. Main advantage of this procedure is the process robustness and calculation speed. Here, the steady state models are retrieved from the dynamic models. Not used for this evaluation, the method would allow to define a cumulated output values as additional constraint or map dependant weights. The following Table summarizes the constraints used for the optimization. Table 2: Constraint condition of calibration Setting Condition Parameter name Value Unit Target Output Min fuel-mass g Constraint Max HULL 1 - Constant lambda 1.00 - Gradient Input dx 10 - Gradient Input dy 10 - Max COV 3.0 % Max Soot 1.0 mg/ m3 Max PN 1.00E+06 #/ cc Optimize Setting start value 3 - - Global opt √ - Verification 1: Comparison of steady state maps Figure 17 compares the original measured base maps of the engine (left) with the optimized maps that has been optimized with respect to BSFC. 185 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 17: Comparison of measured base maps (left) and optimized maps (right) In general, both calibrations, the original base maps and the optimized maps, show the same tendency of the emission and fuel consumption levels. Further, the results allow a qualitative comparison only, because the cost function for the optimization was defined as optimal in terms of BSFC, but not set up in a way to allow a quantitative comparison. 186 5.1 Dynamic Modelling for Gasoline Direct Injection Engines The main deviations are close to full load for BSFC and COV and maximal speed and load for the NOx emissions. The root cause for the described deterioration is a different calibration strategy of enrichment. Further, large deviations can be explored in the very low load area. Here, the root cause is a different calibration of the fuel pressure and start of injection, refer Figure 18. Figure 18: Fuel pressure and λ Due to the large uncertainty of the PN/ PM models, the particles will not be taken into account to evaluate the optimization procedure. But, the results show the efficiency and the transferability of the optimization process for base calibration or the engine. Especially, for the very important task of camshaft can be done very fast. Verification 2: Comparison of cumulated CYCLE results Figure 19 shows the cumulated emission and fuel consumption for the given WLTC validation cycle. For a constant fuel consumption (fuel-mass) the optimized maps show significant higher levels for the THC and NOx emissions of the optimization compared to the base maps. The soot emission are almost constant whereas the particle number is doubled. As already explained previously, the particle models are not reliable and thus they will not further be discussed. 187 5.1 Dynamic Modelling for Gasoline Direct Injection Engines Figure 19: Comparison of base calibrated and the global steady state optimization for a cumulated cycle Parts of the significantly higher cumulated THC and NOx emission can be explained due to the different operating strategies of the cycle replication and the cycle simulation. For the cycle simulation, the fuel cut has not taken into account, here, the models are always over predicting the emission. Further, the model quality for idle and low load are not satisfactory, this may explain additional deviation, too. Where as the engine operating strategy as well as sufficient models can be provided for idle etc. additional deviations are the result of dynamic engine operation. And this can not be explained by the steady state models. 2.9.2 Global dynamic CYCLE optimization In order to meet optimal results for real engine or vehicle behaviour the optimization has to be done on the base of a given cycle. Thus the target is to optimize the static engine maps of the ECU towards an optimal results for dynamic engine driving. For this, dynamic engine models are mandatory. Figure 20: Model overview 188 5.1 Dynamic Modelling for Gasoline Direct Injection Engines The Optimization Frame Work (OFW) is an optimization toolbox that allows complex dynamic, cycle based map optimization. In the example shown in this abstract, the target cycle of the WLTC shall be the base for the map optimization. In order to increase the accuracy of the results, the dynamic models of the engine are combined with a Simulink model of the ECU. Figure 20 shown the set up of the model structure to perform the global dynamic map optimization for a given vehicle cycle. To meet real engine behaviour the accuracy of the ECU model is mandatory. To evaluate the ECU, Figure 21 shows for the given WLTC the fuel pressure (Pf), camshaft positions (IVT, EVT) and the spark (IG) of a bench measurement and of the used ECU model. The overall accuracy is very good, only very deviation may be found due to not supported control modes. Figure 21: Verification of the ECU model accuracy Finally, Figure 22 shows the optimized maps for a camshaft optimization for the given WLTC. As constraints, the same conditions have been applied as stipulated in Table 2 of the steady state optimization. Figure 22 : ECU map generated by OFW 189 5.1 Dynamic Modelling for Gasoline Direct Injection Engines 3 Conclusion This paper introduced and explained the process for dynamic model building for gasoline engines. Starting from the task definitions and test design and model building, the requires for the automation system have been discussed. Two options for a limit reaction in case of a limit violation were shown. Further, the results of a global steady state map optimization and a global dynamic cycle optimization have been shown and discussed. As outlook, the dynamic models in combination with the introduced calibration processes proved to be ready for base calibration work for the daily mass production calibration work. Reference [1] Kaihatsu, M.; He, J.; Shishido, T.; Haukap, C.; Dreher, T.; Hegman, M.: Dynamic Modelling Calibration for Gasoline Direct Injection Engine, Presentation in Powertrain Calibration Conference 2016 (in Japanese), Akihabara, Tokyo [2] He, J.; Kakimoto, F.; Sato, F.; Haukap, C.; Dreher, T.: Steady State Calibration for Catalyst Heat-up Optimization on Gasoline Direct Injection Engine, DoE in Engine Development, Berlin, 2015 [3] He, J.; Kakimoto, F.; Sato, F.; Haukap, C.; Dreher, T.: Steady State Calibration for Catalyst Heat-up Optimization on Gasoline Direct Injection Engine, Keihin Technical Review, Vol. 4, pp. 17-27, 201 (http: / / www.keihin-corp.co.jp/ technology/ tec_report_201512.html) 190 5.2 Excitation Signal Design for Nonlinear Dynamic Systems Tim Oliver Heinz, Mark Schillinger, Benjamin Hartmann, Oliver Nelles Abstract In this work a new method to generate optimized signals for nonlinear dynamic system identification is proposed. Similar to popular existing signals (chirp, multisine, pseudo random sequences, etc.) an ‘OptiMized Nonlinear InPUt Signal’ (OMNIPUS) is passive, so no online adaption is performed and only the dominant time constants of the process is required for the optimization. The objective is a uniform data distribution in the input space spanned by delayed process inputs and outputs. Since the behavior of the output is unknown a priori, the dominant time constant is used to approximate the output roughly through a linear dynamic model. To overcome the challenging task of optimizing all 𝑁 samples of the excitation signal simultaneously, the optimization is done sequentially. This iterated optimization appends one sequence in each step. Typical limitations like amplitude constraints and rate constraints can be accounted for the optimization. Even an classification function which predicts the sequences to be feasible or infeasible can be considered as well. The advantages of the novel excitation signal generation are demonstrated on a real high pressure fuel supply system. Kurzfassung In dieser Arbeit wird eine neue Methodik zur Erzeugung von optimierten Signalen für die Identifikation nichtlinearer dynamischer Systeme vorgestellt. Ähnnlich zu weitverbreiteten Signalen (Chirp, Multi-Sinus, pseudo zufälligen Sequenzen, etc.) ist ein OMNI- PUS (OptiMized Nonlinear InPUt Signal) passiv, sodass keine Online-Adaption an den Prozess stattfindet. Für die Optimierung wird lediglich die dominante Zeitkonstante des Prozesses benötigt. Zielsetzung ist eine homogene Datenverteilung im Eingangsraum, welcher durch verzögerte Prozessein- und Ausgänge aufgespannt wird. Der unbekannte Ausgang wird durch ein lineares Modell auf Basis der dominanten Zeitkonstante grob approximiert. Um die anspruchsvolle Aufgabe alle 𝑁 Amplitudenwerte gleichzeitig zu optimieren zu umgehen, wird das Signal sequenziell optimiert. Diese iterative Optimierung hängt in jedem Schritt eine optimierte Sequenz an. Typische Einschränkungen wie die Beschränkung der Amplituden und Verfahrgeschwindigkeiten können ebenso berücksichtigt werden wie beliebige Klassifikationsfunktionen, welche eine Sequenz als erlaubt oder unerlaubt klassifizieren. Die Vorteile der neuen Signalgenerierung werden anhand eines Systems für die Hochdruck Kraftstoffzufuhr demonstriert. 1 Introduction The identification of nonlinear dynamic systems is of growing interest especially in the automotive sector where tightening emission regulations call for new modeling tech- 191 5.2 Excitation Signal Design for Nonlinear Dynamic Systems niques. The identification with black box models becomes more important since powerful modeling approaches like Gaussian process models (GPM) and local model networks (LMN) demonstrated impressive performances. These models are based on the measured data of the system. The gathered information contained in the datasets are decisive for the accuracy of the identified black box model. Thus, the excitation of the process plays the main role in the identification task. Usually processes which are to be modeled are subdivided in static and dynamic processes. The design of experiments (DoE) can be classified accordingly. By dealing with static processes without knowledge of the underlying process, space filling designs are commonly used. This is possible since each input can be manipulated independently [16]. The design of dynamic experiments (dynamic DoE) is a more challenging task. Besides the operating points defined by the amplitude levels of the excitation signal, the transitions in between have to be considered. Additionally the important frequency range of the underlying process should be incorporated in the excitation signal [9]. The DoE of either static or dynamic processes can be divided in model-based and model-free approaches. Model-free dynamic DoE focuses on specific excitation signals. Well established are sinusoidal-based signals like chirp and multisine [1, 13]. But also step-based signal types like Amplitude Pseudo Random Binary Sequences (APRBS) are useful for the identification of nonlinear processes [12]. Modifications of APRBS to meet input constraints of the process are discussed in [4, 7]. In [17], different signal types were combined to achieve a better identification result. Model-based dynamic DoE utilizes the knowledge of the underlying process to generate an input signal. These optimized excitation signals minimize the variance of the parameter estimation of the given model. An iterative optimization of the signal as proposed in [6] has the ability to focus on and consequently improve the worst performing model characteristics and may incorporate with any case of input and output constraints as well. There are some shortcomings: (i) A model has to be identified before which means for black-box modeling, a suitable excitation signal has to be chosen first. (ii) The approach is time-consuming due to the model update and the nonlinear optimization of the design. (iii) The obtained excitation signal is only optimal for the chosen model structure (i.e. nonlinear approximator, order of the dynamic model). So changing the model structure leads to a non-optimal design. Due to these specified shortcomings, this paper focuses on model-free approaches. This contribution is organized as follows. In Sect. 2 some well-established excitation signals are analyzed and discussed. The exposed properties motivate the generation of a new excitation signal described in Sect. 3. The superiority of the new signal is demonstrated on a real high pressure fuel supply systems (HPFS) which is described in Sect. 4. The identification procedure and the results are given in Sect. 5 and in Sect. 6. Finally, a conclusion is given in Sect. 7. 2 Popular Excitation Signals For the modeling of static processes, a space filling design is desirable, if no prior knowledge is available. Also for dynamic models this is a desirable property. By utilizing a 192 5.2 Excitation Signal Design for Nonlinear Dynamic Systems Nonlinear AutoRegressive with eXogenous input (NARX) structure, the inputs of the nonlinear static approximator are delayed inputs and outputs (see Fig. 1). u(k) y(k) ˆ y(k) q 1 q 1 q 1 q 1 q 1 q 1 u(k 1) u(k 2) u(k m) nonlinear dynamic model nonlinear static approximator f( · ) y(k 1) y(k 2) y(k m) Figure 1: External dynamics approach in NARX structure. For a model 𝑓 (⋅) with 𝑛 process inputs and a common order 𝑚 , the predicted model output is given by ̂ 𝑦(𝑘) = 𝑓 (𝑢 1 (𝑘 − 1), … , 𝑢 1 (𝑘 − 𝑚), … , 𝑢 𝑛 (𝑘 − 1), … , 𝑢 𝑛 (𝑘 − 𝑚), 𝑦(𝑘 − 1), … , 𝑦(𝑘 − 𝑚))) . (1) Here 𝑢 𝑖 (𝑘) is the 𝑖 -th input of the process and 𝑦(𝑘) is the output of the process at the discrete time step 𝑘 . Since the delayed versions of the process output act as the inputs for the nonlinear approximator, the data distribution in the input space of a NARX model depends strongly on the underlying process. Figure 2 shows the space filling properties of some popular excitation signals in the input space for a nonlinear process in Hammerstein configuration. The nonlinearity is formed by an atan -function followed by a first order dynamic system with unity gain. 𝑦(𝑘) = 0.2𝑓 (𝑢(𝑘 − 1)) + 0.8𝑦(𝑘 − 1) (2) 𝑓 (𝑢(𝑘 − 1)) = atan(8𝑢(𝑘 − 1) − 4) + atan(4) 2atan(4) (3) As input signals a ramp, a chirp, an APRBS, and a multisine signal are investigated. For a thorough discussion see [1, 12, 13, 17]. While the data distribution of a ramp design shows more density in the center of the input space, the chirp signal generates more points near the boundary. Similarly to the ramp design, the multisine generates most of the data points at the inner input space. The most homogeneous data distribution is provided by the APRBS. The input space can be classified in different areas: (i) Data points close to the characteristic curve are low frequent. (ii) High frequent excitations lead to data points distant to the characteristic curve. The APRBS combines low frequent components (piecewise constant) and high frequent components (step shaped amplitude changes). With this combination, the APRBS reaches high dynamic areas (upper left and lower right corner) of the input space as well as low frequent areas near to the equilibrium. So a step-based signal is able to fill the whole input space with data points. 193 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 0 100 200 300 0 0.5 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1 0 100 200 300 0 0.5 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1 0 100 200 300 0 0.5 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1 0 100 200 300 0 0.5 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1 k u(k) u(k 1) y(k 1) k u(k) u(k 1) y(k 1) k u(k) u(k 1) y(k 1) k u(k) u(k 1) y(k 1) a) b) c) d) Figure 2: Input signals and input spaces for an artificial saturation type process of first order in Hammerstein configuration: a) Ramp, b) chirp, c) APRBS, and d) multisine. 3 A New Excitation Signal for Nonlinear System Identification In this section, the optimization of a discrete-time excitation signal for nonlinear system identification is described. A discrete signal is a time series consisting of a sequence of amplitude values. A simultaneous optimization of all 𝑁 amplitude values is not reliably possible for large 𝑁 . Some simplifications make this computational expensive optimization problem feasible in a suboptimal manner. By implementing input constraints, the generated signals are suitable for many industrial applications. An excitation signal optimized in this manner is referred to as OptiMized Nonlinear InPUt Signal (OMNIPUS). 194 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 3.1 Pseudo Input Space The estimation of the function 𝑓 (⋅) in (1) is the main task of nonlinear system identification. The model inputs 𝑢(𝑘 − 1), … , 𝑢(𝑘 − 𝑚), 𝑦(𝑘 − 1), … , 𝑦(𝑘 − 𝑚) span the input space of the nonlinear approximator. An excitation signal for nonlinear system identification should cover this input space equally with data points, similar to a space filling design for a static problem. In a model-free dynamic DoE, the output 𝑦 is not available, thus a proxy output has to be provided. Here we use a linear dynamic system to supply a proxy output ̃ 𝑦 . The choice of the linear model is important for the optimization of the excitation signal. Thus, this proxy model should have a similar time constant as the underlying process and represents the prior knowledge necessary to construct a meaningful excitation signal. The space spanned by the delayed inputs and the delayed proxy outputs will be referred to as the pseudo input space. The use of a proxy model implies that the proposed excitation signal generation is model based. But all model-free dynamic designs rely on a rough knowledge of the process. For example the hold time of an APRBS is chosen according to the dominant time constant. And even the frequency range of a chirp and/ or multisine signal is selected corresponding to the dynamic behavior of the process. In the opinion of the authors, this rough knowledge about the process is not comparable to the information used in model-based dynamic DoEs. Thus, we consider our approach as model-free. 3.2 Optimization of Sequences The objective of the optimization is an equal data distribution in the pseudo input space. A common choice to assess the quality of the data distribution is the minimax distance [8]. This objective penalizes big holes in the data distribution. convergence append to signal sequence optimize sequence OMNIPUS yes no optimized signal initial signal Figure 3: Flowchart of the sequential optimization. The loop converges, if the user specified signal length is reached. By optimizing all amplitude levels of an excitation signal of length 𝑁 , an 𝑁 -dimensional problem arises. Due to the computational expensive minimax loss function and the number of nonlinear parameters, a simultaneous optimization is unreliable. Here we propose a sequential composition of the excitation signal (see Fig. 3). The key idea is splitting the 𝑁 -dimensional optimization into low-dimensional subproblems. This global suboptimal solution is more robust compared to a concurrent optimization of all levels of the excitation signal. The repeated optimization of single sequences 𝑢 with length 𝑙 requires a new objective function since the minimax distance is not well suited for the sequential optimization. The new optimized sequence should fill the biggest hole of the already existing cloud of 195 5.2 Excitation Signal Design for Nonlinear Dynamic Systems data points 𝑋 old in the pseudo input space. Therefore the sum of the nearest neighbor distances of the new sequence in the pseudo input space 𝑋 new (𝑢) to all existing points is maximized: arg max 𝑢, 𝑙 ( 1 𝑙 𝑙 ∑ 𝑖=1 𝑑 𝑁𝑁,𝑖 (𝑋 old , 𝑋 new (𝑢(𝑖)))) . (4) The function 𝑑 𝑁𝑁,𝑖 (𝑋 old , 𝑋 new (𝑢(𝑖))) calculates the smallest distance of the point 𝑋 new (𝑢(𝑖)) to the points of 𝑋 old . In general there is no restriction to the shape of the sequence. As mentioned in Sect. 2, a piecewise constant signal showed the ability to reach every point of the (pseudo) input space. With this observation it is reasonable to use the amplitude value and the sequence length as two optimization parameters. The amplitude range of the optimized signal is given by the user, e.g., the actuation limits of the underlying process. The sequence length is limited in order to avoid very high frequent signals ( 𝑙 → 1 ) as well as very low frequent signals ( 𝑙 → 𝑁 ): 𝑎𝑇 ≤ 𝑙 ≤ 𝑏𝑇 . (5) A reasonable choice for the limiting factors are 𝑎 = 1 and 𝑏 = 3 with 𝑇 being the approximate time constant of the process used for the optimization. k u(k) 0 0.5 1 0 0.2 0.4 0.6 0.8 1 a) b) ˜ y(k 1) u(k 1) 20 40 60 80 100 0 0.2 0.4 0.6 0.8 1 1 2 3 1 2 3 Figure 4: a) Optimized excitation signal and b) the corresponding data points in the pseudo input space. Figure 4 reveals the relation between the excitation signal and the pseudo input space. The last three sequences are highlighted ( 2 , 3 , and å ). All the sequences were optimized sequentially. Thus, each sequence itself fits in the biggest hole of the current data distribution. It is noticeable that the data distribution refines with increasing excitation signal length. This means that the space filling property is always fulfilled, only the density of the data increases with increasing signal length. This is a major advantage over common excitation signals, where the signal length has to be known a priori. Furthermore, the OMNIPUS algorithm can take into account data points from an existing signal. Existing data points can be simply considered in 𝑋 old . 196 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 3.3 Implementation of Constraints An excitation signal for industrial use should comply with the constraints imposed by the process. Common are restrictions for the amplitude level and signal gradients, respectively. But also areas of critical operation should be omitted. Figure 5 shows the progression inside the optimization (Fig. 3) of a sequence by respecting the user given constraints. The gray shaded area highlights the implementation of the three input constraints treated in this contribution. permissible optimal sequence transition speed adjust rate sequence yes no no yes convergence optimize sequence initial signal adjust amplitude length & Figure 5: Flowchart of the optimization inside the iterated OMNIPUS generation. The gray shaded area highlights the implementation of three different input constraints. Due to actuator limits, the amplitude levels of the optimized sequence should lie between minimum and maximum values defined by the actuator. When there are rate constraints given by the user, the step transition between two amplitude levels has to be adapted, for example with ramp [4] or sine transitions [7]. After adjusting the amplitude and transition speed, the permissibility is checked by a classification function. Forbidden sequences are omitted during the optimization. Each permitted sequence is evaluated by (4). After convergence of the optimization, the best sequence is appended to the already optimized part of the signal. This optimization is iterated until the desired signal length is reached. Figure 6 shows two OMNIPUS (upper) and the resulting pseudo input spaces restricted to 𝑢(𝑘) ∈ [0, 1] . The maximum rate of change is limited to a) a strong constraint |Δ𝑢| ≤ 0.02 and b) a weaker constraint |Δ𝑢| ≤ 0.2 . The transition between two amplitude levels is realized with a sinusoidal function, thus the rate of change lies between the given constraints. Increasing restrictions of the transition speed leads to lower frequent signals. Hence the feasible areas in the pseudo input space shrink towards the equilibrium. This indicates, as mentioned above, a low frequent signal. Otherwise a higher frequent signal creates more points in the outer regions. This assumption that signal a) is lower frequent compared to signal b) is verified by a power spectral density estimation (PSD) of the two signals in Fig. 6 (lower). Presented are only symmetric constraints, which allow an equal, maximal increase or decrease of the input signal in one time step. But the OMNIPUS generation can also in- 197 5.2 Excitation Signal Design for Nonlinear Dynamic Systems corporate unsymmetric constraints ( Δ𝑢 ∈ [−𝑎, 𝑏] with 𝑎 ≠ 𝑏 ). This is of major importance because many industrial applications behave differently for increasing and decreasing input values. 0 500 1000 1500 2000 2500 0 0.5 1 0 1 2 3 4 5 6 7 8 10 -3 -10 0 10 20 30 frequency [Hz] PSD [dB/ Hz] u(k) k 0 0.5 1 0 0.2 0.4 0.6 0.8 1 u(k 1) y(k 1) a) 0 0.5 1 0 0.2 0.4 0.6 0.8 1 u(k 1) y(k 1) b) Figure 6: First 2500 samples of the OMNIPUS using a sine transition to realize the rate constraints (upper) a) |Δ𝑢| ≤ 0.02 and b) |Δ𝑢| ≤ 0.2. The corresponding pseudo input space is shown in the middle. Power spectral density estimations (PSD) of the input signals highlight the dominating frequencies (lower). 198 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 4 Example: The High Pressure Fuel Supply System The main components of a HPFS system are the high pressure rail, the high pressure fuel pump, and the ECU (Engine Control Unit; compare Fig. 7). The pump is actuated by the crankshaft of the engine. A demand control valve in the pump allows to control the delivered volume per stroke. A pressure-relief valve is also included in the pump, but should never open, if possible. Hence, we want to limit the maximum pressure during the whole measurement process. The pump transports the fuel to the rail, which contains the pressure sensor. From there, it is injected into the combustion chambers. See [15] for more details. The system has three inputs and one output. The engine speed ( 𝑛𝑚𝑜𝑡 ) affects the number of strokes per minute of the pump and the engine’s fuel consumption. The fuel pump actuation (Mengensteuerventil, 𝑀𝑆𝑉 ) gives the fuel volume which is transported with every stroke of the pump. It is applied by opening and closing the demand control valve accordingly during one stroke of the pump. The injection time 𝑡 inj is a variable calculated by the ECU, which sums up the opening times of the single injectors and is, thus, related to the discharge of fuel from the rail. Injection Time Rail Pressure Engine Speed Actuation Figure 7: Main components of the HPFS system, inputs (continuous lines), and output (dashed line). The figure is taken from [18]. During the measurement procedure, the injection time is not varied manually but set by the ECU. The permissible times depend on many factors and a wrong choice could extinguish the combustion or even damage components. The engine load would have a major influence on the HPFS system via the injection time, but is omitted to prevent the necessity of a vehicle test bench. 5 Identifying the Rail Pressure System When comparing excitation signals for nonlinear system identification the question of quantifying the signal quality arises. In general, good excitation signals lead to informative data, which lead to good models. Thus, in this paper the model quality may asses the quality of the excitation signal. For this reason, the process is modeled by local model networks (LMN) and Gaussian process models (GPM) as well. Both models use the external dynamic approach in NARX structure to represent the dynamic behavior. 199 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 5.1 Local Model Networks A LMN can be divided in (i) the partitioning and (ii) the local models. In a fuzzy interpretation, the partitioning describes the rule premises (IF part) and the local models describe the rule consequents (THEN part). The input space is partitioned using validity functions Φ 𝑖 (𝑧) . The local models are described by ̂ 𝑦 𝑖 (𝑥) . The inputs 𝑥 and 𝑧 are subsets from all available process inputs. Each process input can be assigned to (i) 𝑥 , (ii) 𝑧 , (iii) 𝑥 and 𝑧 , or (iv) none. This property is beneficial especially for dynamic models where the operating point can be described by a few model inputs [2, 3]. The output of the LMN can be calculated as sum of 𝑀 local models ̂ 𝑦 𝑖 , weighted with their validity function Φ 𝑖 : ̂ 𝑦 = 𝑀 ∑ 𝑖=1 ̂ 𝑦 𝑖 (𝑥)Φ 𝑖 (𝑧) , 𝑁 ∑ 𝑖=1 Φ 𝑖 (𝑧) = 1 (6) The LMNs in this contribution are constructed using the hierarchical local model tree (HILOMOT) algorithm [10]. This incremental growing tree construction algortihm divides the input space with axes-oblique splits. All local models are chosen to be of affine type in this paper. The validity functions are generated by sigmoid splitting functions that are linked in a hierarchical, multiplicative way, see [10] for more details. 1st Iteration 2nd Iteration 3rd Iteration 4th Iteration initialization opt. opt. opt. z 1 z 2 z 1 z 2 z 1 z 2 z 1 z 2 z 1 z 2 z 1 z 2 z 1 z 2 Figure 8: The first Iterations of HILOMOT using a two-dimensional 𝑧-input space. The procedure of the HILOMOT algorithm can be explained with the help of Fig. 8. Starting with a global affine model, in each iteration an additional local affine model is generated. The local model with the worst local error measure (gray areas in Fig. 8) is split into two submodels, such that the spatial resolution is adjusted in an adaptive way. The linear parameters of the new submodels are estimated locally by a weighted least squares method. This is computationally extremely cheap and introduces a regularization effect which increases the robustness against overfitting, as stated in [11]. The axes-oblique partitioning is achieved by optimizing the current split direction and position in each iteration. Only the new split is optimized, all already existing splits are kept unchanged. The initial split direction for the optimization is either one of the orthogonal splits or the direction of the parent split (dotted lines in Fig. 8). In this contribution, the 𝑥 -input space consists of all model inputs. In this contribution, the 𝑧 -input space is reduced to one time delayed process inputs and output: 𝑧(𝑘) = [𝑢 1 (𝑘 − 1), … , 𝑢 𝑛 (𝑘 − 1), 𝑦(𝑘 − 1)] . (7) 200 5.2 Excitation Signal Design for Nonlinear Dynamic Systems This choice takes into account, that the actual operating point is defined by the level of the actual process inputs and output. The first (and higher) derivatives of the model inputs and output are assumed to be insignificant to describe the operating point. 5.2 Gaussian Process Models Gaussian Process models (GP models) are a nonparametric bayesian modeling approach. It is assumed, that the output 𝑦(𝑥) of a stationary system is generated by a latent function 𝑓 (𝑥) and additive Gaussian noise 𝜖 𝑦(𝑥) = 𝑓 (𝑥) + 𝜀, 𝜀 ∼ 𝒩 (0, 𝜎 2 n ). (8) Opposed to parametric models like LMNs, GP models represent the system’s latent function 𝑓 (𝑥) by a stochastic process. A joint zero-mean Gaussian distribution with covariance function 𝑘(𝑥 𝑖 , 𝑥 𝑗 ) is assumed for any set of function evaluations. Predicting the system output 𝑦 ∗ (𝑥 ∗ ) at the test point 𝑥 ∗ using a GP results in the predictive distribution 𝑝(𝑦 ∗ |𝑥 ∗ , 𝒟 ) = 𝒩 (𝜇 ∗ , 𝜎 ∗ ), with { 𝜇 ∗ = 𝑘 T ∗ (𝐾 + 𝜎 2 n 𝐼 ) −1 𝑦 𝜎 ∗ = 𝜎 2 n + 𝑘 ∗∗ − 𝑘 T ∗ (𝐾 + 𝜎 2 n 𝐼 ) −1 𝑘 ∗ . (9) Thereby, 𝒟 = {𝑥 𝑗 , 𝑦 𝑗 |𝑗 = 1, … , 𝑛} denotes the 𝑛 input-output pairs used for training, 𝑘 ∗ ∈ ℝ 𝑛 the vector of covariances between the latent function value 𝑓 (𝑥 ∗ ) at the test point 𝑥 ∗ and the training function values, 𝐾 ∈ ℝ 𝑛×𝑛 the covariance matrix of the training function values, 𝜎 2 n the noise variance, 𝐼 the 𝑛 × 𝑛 identity matrix, 𝑦 ∈ ℝ 𝑛 the vector of training outputs, and 𝑘 ∗∗ = 𝑘(𝑥 ∗ , 𝑥 ∗ ) the prior variance of 𝑓 (𝑥 ∗ ) . The covariance function is usually parameterized by a set of hyperparameters Θ . For example, when using a squared exponential covariance function 𝑘(𝑥 𝑖 , 𝑥 𝑗 ) = 𝜎 2 0 exp (− 1 2 (𝑥 𝑖 − 𝑥 𝑗 ) T Λ −1 (𝑥 𝑖 − 𝑥 𝑗 )) , with Λ = diag(𝑙 2 1 , … , 𝑙 2 𝐷 ), (10) the hyperparameters are the prior variance 𝜎 2 0 , the lengthscales 𝑙 1 , … , 𝑙 𝐷 , and the noise variance 𝜎 2 n . They can be learned by maximizing the log marginal likelihood log 𝑝(𝑦|Θ) = − 1 2 𝑦 T (𝐾 + 𝜎 2 n 𝐼 ) −1 𝑦 − 1 2 log|𝐾 + 𝜎 2 n 𝐼 | − 𝑛 2 log 2𝜋. (11) More information can be found in [14]. To perform the modeling, the software ETAS ASCMO was used. In order to reduce the computational complexity, it utilizes a sparse Gaussian Process implementation instead of the classic algorithm presented above. See [5] for more information regarding dynamic modeling in ASCMO. 6 Results After preprocessing the data, the results of the models are compared. Besides the quantitative analysis based on the error values on a test data set, the simulated data are visualized and compared to the measured data. 201 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 6.1 Operating Point Depending Constraints For the design of an excitation signal for an unknown system, the operating range has to be explored. In the first instance, the amplitude constraints were determined based on (previous) stationary measurements. The feasible amplitude levels are given in (12) and (14). The maximum rate of change were chosen by expert knowledge ((13) and (15)) based on the given actuator limits. 𝑛𝑚𝑜𝑡 ∈ [1200, 4000] min −1 (12) d d𝑡 𝑛𝑚𝑜𝑡 ∈ [−1000, 4000] min −1 s −1 (13) 𝑀𝑆𝑉 ∈ [0, 50] mm 3 (14) d d𝑡𝑀𝑆𝑉 ∈ [−500, 500] mm 3 s −1 (15) To validate the constraints of the process, a combined signal consisting of ramp and chirp sequences (ramp-chirp) was generated according to (12)-(15). All the operating points defined by this input signal were feasible for the process. The coverage of operating points depends strongly on the chosen excitation signal type. By changing the signal type of the process, the coverage of operating points changes as well. Consequently, more extreme rail pressure values may occur compared to the ramp-chirp excitation. These operating point dependent effects to the rail pressure call for a further tightening of the constraints. A first test with an OMNIPUS optimized according to the constraints in (12)-(15) unveils these operating dependent constraints by forcing the engine into infeasible operating conditions. By decreasing the maximum rate of the engine speed from d d𝑡 𝑛𝑚𝑜𝑡 max = 4000 min −1 s −1 to d d𝑡 𝑛𝑚𝑜𝑡 max = 1000 min −1 s −1 , the rail pressure stayed inside an uncritical pressure range during operation. The more conservative constraints for d d𝑡 𝑛𝑚𝑜𝑡 ∈ [−1000, 1000] min −1 s −1 (16) are used in all further measurements. 6.2 Data Acquisition The OMNIPUS is compared to a signal combination consisting of ramp and chirp sequences proposed in [17]. The measurement time for each signal is limited to 𝑡 max = 10 min. The sampling frequency is 𝑓 0 = 100 Hz, thus a signal length of 𝑁 = 60 000 results. The OMNIPUS is optimized over the whole signal length. For the generation of the proxy output, two first order systems are used. The poles in the 𝑧 -plane are chosen equally to 𝑝 1 = 𝑝 2 = 0.99 which correspond to a time constant of 𝑇 ≈ 1 s. Within the constraints, all generated sequences are permissible, thus no classification function is used. The transition between piecewise constant parts of the excitation signals are adjusted to the rate constraints using sine transitions. This optimized signal is compared to a ramp-chirp signal. By generating the ramp-chirp combination, the first half of the 202 5.2 Excitation Signal Design for Nonlinear Dynamic Systems signal consists of a ramp sequence, the second half consists of a chirp sequence. The high frequent chirp together with a low frequent ramp sequence seems to be a reasonable excitation of the process. All signals are generated according to the above mentioned constraints. To test the achievable model quality for both excitation signals, an independent test signal is generated. For a fair comparison, the test signal consists of a chirp-ramp signal and an OMNIPUS one half each. 6.3 Accuracy of the Simulation Results The dynamic order of the models is chosen to be three, thus the model inputs are three time delayed process inputs and outputs: ̂ 𝑦(𝑘) =𝑓 (𝑢 1 (𝑘 − 1), … , 𝑢 1 (𝑘 − 3), 𝑢 2 (𝑘 − 1), … , 𝑢 2 (𝑘 − 3), … 𝑢 3 (𝑘 − 1), … , 𝑢 3 (𝑘 − 3), 𝑦(𝑘 − 1), … , 𝑦(𝑘 − 3)) (17) 𝑢 1 = 𝑛𝑚𝑜𝑡, 𝑢 2 = 𝑀𝑆𝑉 , 𝑢 3 = 𝑡 inj , 𝑦 = 𝑝 r (18) The function 𝑓 (⋅) will be approximated using LMNs and GPMs. 6.3.1 Qualitative Analysis For the qualitative analysis, the model output of both models is compared to the measured data. Figure 9 and10 show the simulated output on the test data set, based on the models which are trained with either the ramp-chirp or the OMNIPUS data. y(k) , ˆ y(k) 0 1 2 3 4 5 6 10 4 0 5 10 15 20 25 k measured OMNIPUS ramp-chirp Figure 9: Output signal of the ramp-chirp GPM and the OMNIPUS GPM on test data. The measured output is given as reference. The GPMs do not show dramatical mismatches between process and the identified models (Fig. 9). In the last half of the signal the ramp-chirp model seems to be slightly worse because of some slightly discrepancies. But in the overall the GPMs seem to be valid in the range of operation. For the LMN, the mismatch between model and process is in most cases small (e.g. Fig. 10 a) and Fig. 10 d)). But there are also some significant mismatches for example 203 5.2 Excitation Signal Design for Nonlinear Dynamic Systems -20 -10 0 10 20 30 6 8 10 12 14 16 18 -5 0 5 10 15 5 10 15 20 Part 1 Part 2 Part 3 0 1 2 3 4 5 6 10 4 -20 -10 0 10 20 30 k a) b) c) d) OMNIPUS ramp-chirp measured measured y(k) , ˆ y(k) Figure 10: Output signal of the ramp-chirp and the OMNIPUS HILOMOT model. The measured output is given as reference. a) good model fit for both models on the ramp-chirp sequence b) plant model mismatch of the OMNIPUS model on the ramp-chirp sequence c) plant model mismatch of the ramp-chirp model on the OMNIPUS sequence d) good model fit for both models on the OMNIPUS sequence. in Fig. 10 c). The nonlinear behavior of the process is not well identified by the rampchirp LMN. This major mismatch between process and model indicates a poor modeling most likely because informative data in this area of operation are missing. Figure 10 b) shows a minor mismatch between process and the OMNIPUS LMN. 6.3.2 Quantitative Analysis For the quantitative analysis the normalized root-mean-squared errors (NRMSE), based on the simulated data, are used. For a thorough analysis, the errors were calculated for the following 6 different signals: • 𝑢 t : Whole test signal - displayed in Fig. 9 and Fig. 10 . • 𝑢 t, ramp : Ramp sequence of the test data set - Fig. 10 part 1. • 𝑢 t, chirp : Chirp sequence of the test data set - Fig. 10 part 2. • 𝑢 t, OMNIPUS : OMNIPUS sequence of the test data set - Fig. 10 part 3. 204 5.2 Excitation Signal Design for Nonlinear Dynamic Systems • 𝑢 OMNIPUS : Training data of the OMNIPUS model. • 𝑢 ramp-chirp : Training data of the ramp-chirp model. Table 1: Normalized root-mean-squared error on simulation data of the HILOMOT model. model type test data training data HILOMOT 𝑢 t 𝑢 t, ramp 𝑢 t, chirp 𝑢 t, OMNIPUS 𝑢 OMNIPUS 𝑢 ramp-chirp OMNIPUS model 0.0244 0.0282 0.0428 0.0198 0.0061 0.0396 Ramp-chirp model 0.0781 0.0294 0.0153 0.1091 0.0931 0.0130 In Tab. 1, the results of the LMN models are highlighted. Since the partitioning of a HILO- MOT model depends on the data point distribution in the 𝑧 -input space, the excitation plays an important role of the achievable model quality. By comparing the simulation error on the training data sets, the model based on the OMNIPUS unveils the benefit against the ramp-chirp model. The simulation error of the ramp-chirp model on the OM- NIPUS data ( 𝑢 OMNIPUS ) is 15 times higher compared to the OMNIPUS model. In contrast, the error of the OMNIPUS model evaluated on the ramp-chirp data set ( 𝑢 ramp-chirp ) is only 3 times higher compared to the ramp-chirp model. On the test data set, the OMNIPUS model is in most cases superior to the ramp-chirp model. Since the ramp-chirp model has a strong mismatch for the OMNIPUS test sequence (see Fig. 10), the difference between the two models is maximal on the sequence 𝑢 t, OMNIPUS . The chirp sequence 𝑢 t, chirp in contrast is better determined by the ramp-chirp model, thus the OMNIPUS model performs worse. Table 2: Normalized root-mean-squared error on simulation data of the GP model. model type test data training data GPM 𝑢 t 𝑢 t, ramp 𝑢 t, chirp 𝑢 t, OMNIPUS 𝑢 OMNIPUS 𝑢 ramp-chirp OMNIPUS model 0.0343 0.0344 0.0268 0.0427 0.0349 0.0376 Ramp-chirp model 0.0302 0.0258 0.0208 0.0389 0.0383 0.0197 The simulation errors of the GP models are shown in Tab. 2. The difference of the two GP models evaluated on the training data sets is much smaller compared with the HILO- MOT models. Even the simulation errors on the whole test data set are comparable. Actually, the ramp-chirp model performs slightly better on the test data set. Surprisingly the ramp-chirp model shows a better generalization ability on the OMNIPUS sequence 𝑢 t, OMNIPUS of the test data compared with the ramp-chirp model. In general this is an abnormal behavior which needs further investigation. By comparing the test data results of Tab. 1 and Tab. 2 it is noticeable, that the HILO- MOT model has the best quality with OMNIPUS excitation. The worst model quality is 205 5.2 Excitation Signal Design for Nonlinear Dynamic Systems also achieved by a HILOMOT but with the ramp-chirp excitation. The quality of the two GP models are in between the HILOMOT models. Obviously the generalization ability depends (i) on the data point distribution of the trainings data in the input space and (ii) on the chosen model structure. It seems that the quality of the GP model is less sensitive to the data point distribution. The exact root cause for this desirable behavior is an issue for future research. In contrast, the HILOMOT model is very sensitive to the data point distribution. An optimization of the excitation signal creates most benefit for the HILOMOT model. 6.4 Data Coverage of the Input Space The objective of the OMNIPUS Main components of the HPFS system equally distributed data in the input space. Figure 11 shows a reduced input space of the identification problem assuming a first order MISO system. The input 𝑢 3 is fixed through the engine control and therefore irrelevant for the dynamic DOE. Thus, only projections to the remaining model inputs are considered. This figure reveals the data coverage of the different excitation signals in this subspace. The ramp-chirp signal shows an equal data distribution in the 𝑢 1 - 𝑢 2 -projection with scattered data in this view (Fig. 11 middle). The OMNIPUS operates on discrete amplitude values, thus the data distribution concentrate on a few input levels. The benefit of the OMNIPUS becomes visible by taking the output 𝑦 into account. The OMNIPUS generates more points at the extreme values of the output (outer views of Fig. 11). In contrast, the ramp-chirp signal generates no points in this areas. Thus, the data set contains no information at the boundary of the input space. A model based on such a data set has an increased risk of wrong extrapolation behavior. 0 10 50 20 2000 30 3000 4000 0 1500 2000 2500 3000 3500 4000 0 10 20 30 40 50 4000 3000 2000 0 5 10 15 20 60 25 40 20 0 u 1 (k 1) u 2 (k 1) y(k 1) y(k 1) u 1 (k 1) u 2 (k 1) u 2 (k 1) u 1 (k 1) OMNIPUS ramp-chirp OMNIPUS ramp-chirp Figure 11: Reduced input space of the identification problem, with 𝑢 1 ̂ = engine speed, 𝑢 2 ̂ = fuel pump actuation and 𝑦 ̂ = being the measured rail pressure. The OMNIPUS data (blue) set shows a good coverage of the reduced input space compared to the rampchirp data (red). It is most likely, that the bad data coverage of the ramp-chirp signal in the input space is responsible for the strong plant-model mismatch in Fig. 10 c). 206 5.2 Excitation Signal Design for Nonlinear Dynamic Systems 7 Conclusion and Outlook This contribution proposes a new optimization strategy to generate excitation signals for nonlinear dynamic systems. The optimization may take typical industrial constraints into account like rate and amplitude limitations. But also inaccessible sequences can be avoided during optimization, thus a plant friendly measurement is possible. The efficiency of the newly proposed signal generator was demonstrated on a high pressure fuel supply system. First tests with statically determined rate constraints forced the engine to operate in critical operating areas caused by the high dynamic excitation OMNIPUS. The excitation signals were adapted to more conservative constraints to prevent the engine from critical operations. The OMNIPUS was compared to a combined signal consisting of ramp and chirp sequences. Based on local model networks (LMN) and Gaussian process models (GPM), the generalization abilities of the different excitation signals were investigated. The quality of a GP model was only slightly affected by the novel signal. In comparison to that, more informative data in the OMNIPUS data set lead to huge benefit of the model quality when modeling with LMNs. References [1] W Baumann et al. “Excitation signals for nonlinear dynamic modeling of combustion engines”. In: Proceedings of the 17th World Congress, The International Federation of Automatic Control, Seoul, Korea . 2008. [2] Julian Belz, Tim Oliver Heinz, and Oliver Nelles. “Automated Order Determination Strategies for Nonlinear Dynamic Models”. In: Computational Intelligence, 2015 IEEE Symposium Series on . IEEE. 2016. [3] Julian Belz and Oliver Nelles. “Order Determination and Input Selection with Local Model Networks”. In: Proceedings of the 20th IFAC World Congress . Toulouse, France, July 2017, [accepted]. [4] Michael Deflorian and Susanne Zaglauer. “Design of experiments for nonlinear dynamic system identification”. In: IFAC 18th World Congress, Milano 2011 . 2011. [5] Tobias Gutjahr et al. “New Approaches for Modeling Dynamic Engine Behavior with Gaussian Processes”. In: Design of Experiments (DoE) in Engine Development . Ed. by Karsten Röpke. Expert Verlag, 2013. [6] Christoph Hametner et al. “Optimal experiment design based on local model networks and multilayer perceptron networks”. In: Engineering Applications of Artificial Intelligence 26.1 (2013), pp. 251-261. [7] Tim Oliver Heinz and Oliver Nelles. “Vergleich von Anregungssignalen für Nichtlineare Identifikationsaufgaben”. In: Proceedings 26. Workshop Computational Intelligence . Ed. by F. Hoffman, E. Hüllermeier, and R. Mikut. KIT Scientific Publishing, Nov. 2016. [8] Mark E Johnson, Leslie M Moore, and Donald Ylvisaker. “Minimax and maximin distance designs”. In: Journal of statistical planning and inference 26.2 (1990), pp. 131-148. [9] Lennart Ljung. System identification . Springer, 1998. 207 5.2 Excitation Signal Design for Nonlinear Dynamic Systems [10] Oliver Nelles. “Axes-oblique partitioning strategies for local model networks”. In: 2006 IEEE Conference on Computer Aided Control System Design, 2006 IEEE International Conference on Control Applications, 2006 IEEE International Symposium on Intelligent Control . IEEE. 2006, pp. 2378-2383. [11] Oliver Nelles. Nonlinear system identification . Springer, 2001. [12] Oliver Nelles and Rolf Isermann. “Identification of nonlinear dynamic systems classical methods versus radial basis function networks”. In: American Control Conference, Proceedings of the 1995 . Vol. 5. IEEE. 1995, pp. 3786-3790. [13] Rik Pintelon and Johan Schoukens. System identification: a frequency domain approach . John Wiley & Sons, 2012. [14] Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian processes for machine learning . The MIT Press, 2006. ISBN: 026218253X. [15] Robert Bosch GmbH, ed. Ottomotor-Management. Systeme und Komponenten . 3rd. Friedr. Vieweg & Sohn Verlag, 2005. ISBN: 3-8348-0037-6. [16] Thomas J Santner, Brian J Williams, and William I Notz. The design and analysis of computer experiments . Springer Science & Business Media, 2013. [17] Nils Tietze. “Model-based Calibration of Engine Control Units Using Gaussian Process Regression”. PhD thesis. 2015. [18] Nils Tietze et al. “Model-based calibration of engine controller using automated transient design of experiment”. In: 14th Stuttgart International Symposium . Wiesbaden: Springer Fachmedien, 2014. DOI: 10.1007/ 978-3-658-05130-3_111 . 208 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Matthias Hekrenz, Daniel Reppel, Michael Hegmann, Jochen Broz, Christoph Brands Abstract The increasing complexity of internal combustion engines increases demands on engine systems and components. Multi-physics simulations play a decisive role in their development. Both manufacturing tolerances and variations of input data can be considered by simulation of minimum and maximum parts in order to obtain limiting values of the characteristic values of the systems of interest. However, the nonlinear system behavior leads to the problem that minimum and maximum parts do not necessarily result in the limiting values since other combinations in the tolerance bandwidth might lead to higher variations of the results. At Schaeffler, parameter combinations of engine components and systems are analyzed and optimized during system simulation by using DoE (Design of Experiments) methods. In close collaboration with IAV GmbH a process has been developed that allows the engineer to carry out a parameter screening to detect the most influencing parameters of physical models, use these in RBF (radial basis functions) -cluster models, optimize the system considering given limit values of the response results by “design centering”, and create tolerance fields. The screening is the most important part of the process and - split in several steps - allows the analyst to investigate system models with a high number of inputs. The optimization of the inputs is carried out using a multi-criteria particle swarm optimization and the output tolerance distribution is evaluated using Monte-Carlo simulations. In this presentation, the process is applied to the simulation of engine systems, the process steps are explained in detail, and it is shown how this method can be used in an early phase of development to contribute essentially to the robust design of engine components and systems. Kurzfassung Mit der zunehmenden Komplexität von Verbrennungsmotoren steigen die Anforderungen an die verbauten Komponenten und Systeme. In der Entwicklung solcher Systeme spielen daher multiphysikalische Simulationen eine entscheidende Rolle. In diesen können bereits während der Auslegungsphase sowohl fertigungsbedingte Toleranzen als auch Streuungen der Eingangsdaten durch Simulation von Minimalbzw. Maximalteilen berücksichtigt werden, um so Grenzkurven der interessierenden System-Kennwerte zu erhalten. Das Verhalten dieser Systeme ist meist nichtlinear und somit ergibt sich das Problem, dass Minimal-und Maximalteile nicht notwendigerweise die Grenzkurven darstellen, weil andere Kombinationen innerhalb der Toleranzen zu größeren Variationen führen. Daher werden bei Schaeffler in der Systemsimulation von Motorkomponenten und -systemen Vergleiche unterschiedlicher Parameterkombinationen 6 Design of Experiments II 209 6.1 Application of a DoE based robust design process chain for system simulation of engine systems modellbasiert und mit Hilfe von DoE- (Design of Experiments) Methoden vorgenommen und hiermit die Systemparameter optimiert. In Zusammenarbeit mit IAV GmbH wurde eine Prozesskette entwickelt, die das Screening einflussführender Parameterkombinationen von physikalischen Modellen, eine Modellbildung als Cluster-Modell von RBFs (radialen Basisfunktionen), eine Optimierung der Eingangsgrößen unter Berücksichtigung der Grenzvorgaben im Ausgangsraum („Design Centering“) sowie die Erstellung von Toleranzbändern erlaubt. Das Screening als wesentlicher Bestandteil der Prozesskette wird dabei in mehreren Teilschritten durchgeführt und erlaubt auch die Untersuchung von Systemmodellen mit einer hohen Anzahl an Eingangsparametern. Die Ergebnisse des Screenings werden bei der anschließenden Modellerstellung berücksichtigt. Für die Optimierung der Eingangsgrößen wird ein Partikelschwarm-Optimierungsalgorithmus verwendet, welcher eine multikriterielle Optimierung erlaubt. Die Toleranzverteilung wird mit der Monte- Carlo-Methode ermittelt. In diesem Beitrag wird die entwickelte Prozesskette auf die Systemsimulation von Motorsystemen angewendet, die einzelnen Schritte im Detail beschrieben und gezeigt, wie dadurch bereits in einer frühen Entwicklungsphase ein entscheidender Beitrag zur robusten Auslegung von Motorkomponenten und -systemen geleistet werden kann. 1 Introduction The emission limits introduced by the Euro 6c stage and the requirement that those are fulfilled in more dynamic test cycles as the WLTC (Worldwide Harmonized Light Duty Test Cycle) and on-road measurements (RDE Real Driving Emissions) attach more importance to the gas exchange of internal combustion engines. Schaeffler offers a wide range of products for the optimization of the gas exchange [1]. The product range consists amongst others of optimized standard valve train components, systems for cam profile switching and cylinder deactivation, hydraulic and electromechanical camshaft phasers [2] for optimized valve timing and the fully variable valve train system “UniAir” [3,4]. During the product development phase of such systems, physical simulation models are used already in an early stage to simulate the system behavior and quantify the influence of system and design parameters and their tolerances under different operating conditions. Furthermore, uncertainties of input data at the system interfaces can be considered. The general relation of the input and output parameters of a system simulation model is shown in Figure 1. Figure 1: Input and output of a system simulation model 210 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Depending on the complexity of the physical system, the simulation model can contain a vast number of design and system parameters. The significance of each input parameter with respect to the performance parameters can be very different or even show an influence only in combination with other input parameters. Therefore, a sensitivity analysis is necessary to assess single parameter influences as well as the interdependency of multiple input parameters on the system performance. For the systematic simulation of minimum and maximum tolerance parts the interdependency of the tolerances of different parameters needs to be considered as well. Due to the nonlinear system behavior minimum and maximum parts do not necessarily result in the limiting output values since other combinations in the tolerance bandwidth might lead to higher variations of the results. A test plan is created using DoE methods in order to take a large amount of significant parameters and the variation of their tolerances into account. From these results a data driven model is created. Due to their fast computation time these models are well suited for complex optimization tasks and allow for the creation of tolerance bands and the optimization of system and design parameters. In close collaboration between Schaeffler and IAV a process chain has been developed which is capable to capture these steps for the system simulation of engine systems. At the beginning of the process chain, a parameter screening is performed in order to detect the most influencing parameters of physical models in the design space. Based on the results of the screening, a data driven model as a cluster of RBFs is built. This model can be used for an optimization of input parameters considering given boundary values of the response results (“design centering”) using a particle swarm optimization algorithm that allows for multi-criteria optimization. Finally, the tolerance distribution of the system parameters is evaluated using this model in order to create tolerance fields. 2 Screening 2.1 Problem A commonly asked question when dealing with physical systems or models is “Which system parameters (model inputs) are real ly relevant? ” The greatest chal lenge when searching or screening these relevant system parameters of defined influences for a system with several input parameters is the large number of the resulting parameter combinations. Table 1 illustrates this challenge. Table 1: Number of included parameter combinations System parameter Number of included parameter combinations 3 7 5 31 10 1.023 25 33.554.431 50 1.126* 10 15 100 1.267* 10 30 211 6.1 Application of a DoE based robust design process chain for system simulation of engine systems In classic screening procedures all factors are varied individually between minimum and maximum values. Only factors that produce significant changes in the performance parameters (output parameters) are considered later on. Interactions (influences that only become significant when more than one factor is changed) are not considered when assessing the relevance of a parameter. This means that all interactions relating to this parameter are ignored if the effect of the parameter individually is small. 2.2 Requirements The objective of the development of a screening method to determine non-linear, impact-related parameter combinations in high-dimensional design space was to include interactions of insignificant input parameters. The most important point here is quite clearly to minimize the number of the parameter combinations to be tested, Table 1. As a further provision, it should be possible to perform partial steps of the method in parallel and possess a freely definable termination condition. Finally, no prior knowledge about the system or the underlying physics needs to be supplied during the process, the resulting model is a black box model. 2.3 Methodology The design space was normalized to the range [-1 , +1] for the screening. This normalized average need not necessarily correspond to the centroid of the design space. The investigation begins at a reference point, e.g., +1, and examines the design space in detail at different reference points. Each observation of a reference point results from the impact caused to the normalized average of the design space. Figure 2: Normalized design space with different reference points The original function is represented in terms of increasing multiplicity of the inputs: (1) The values x, y, z are exclusively the reference values. The process is divided into 5 steps: The first step is to create a "Low Dimension Library“. To this purpose a full -factorial test plan to determine the direct influence and the dual interactions is defined, multiplicity 1 and 2. This step facilitates the search for higher multiplicity terms in the later steps. To determine to which accuracy the system behavior at the reference point can already be described using the identified parameter combinations, a verification is performed with a given number of random parameter combinations. The number of individual parameters of a parameter combination to be tested corresponds to half the 212 6.1 Application of a DoE based robust design process chain for system simulation of engine systems number of system parameters of the system. The analysis of the response of the performance parameters is performed via residual consideration. Here the result is already broken down into its already known influences. If the remaining residual influence is significant, the response of the performance parameter could not be described adequately. This parameter combination is classified as a "combination with unknown influences" to be investigated in subsequent steps. A threshold value (eps) is defined when initializing the screening method to indicate the significance of an influence. The eps value can be defined as follows:  a constant, pre-defined value (from engineering knowledge)  calculated using pre-defined equation  determined from a characteristic diagram  dynamic adjustment of the value during screening using already determined influences In the second step all parameter combinations, which still contain significant influences, are saved in a separate list ( “ garbage collector ” ). This list is searched for frequently occurring patterns of parameter combinations. For the next step a test plan is created with parameter combinations that contain yet unknown influences. These test plans can be calculated separately and independently and can be summarized as follows:  Parameter patterns determined from the "garbage collector"  Parameter combinations that have been detected but cannot be fully explained at the current iteration stage  Measuring points randomly distributed in the design space with high variation of the individual parameters In the third step , a so called “parameter regression” is conducted on the newly mea sured test plan points. Each parameter combination of the test plan is considered as an individual group and reduced gradually. Here the group is initially halved and a check is run whether the influence could be isolated in one of the new groups. If not and the group contains a high number of system parameters, the group is re-mixed. If this does not lead to fewer parameters then the parameters continue to be reduced gradually. A model run is performed every time one parameter is removed to determine its influence. The following illustration clarifies the approach. 213 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Figure 3: Parameter Regression For a search parallelization, the test plan can be divided into individual test plans and these can then be processed by different processors (CPUs) using the “ parameter regression ” . Here it is best to use correlations to perform the division of the parameter combinations. The time to process the test plan depends on the number of parameter combinations to be examined and the contained interactions. It is recommended to make restrictions and stop the search of a processor after a defined number of model runs has been completed. Individual test plan parameter combinations that have not yet been considered can possibly already be described using the results of other processors otherwise they are considered in further iterations. In the fourth step the results of the individual test plans of all processors are collected. Influences still unknown are calculated as far as possible using the newly determined parameter combinations. After that the parameter combinations with an unknown, significant residual influence are saved in the "garbage collector". A check is now conducted again to determine how much design space of the reference points can already be described by using the determined parameter combinations as described in the first step. A termination condition is defined at the beginning of the screening. For example, the algorithm shall continue to search until it can describe at least 95% of the influences of the design space. For additional reliability this process can be repeated a number of times. In the fifth step all the determined parameter combinations with influences for the next reference point from Figure 2 are calculated. A check is conducted again to determine how much design space of the reference point can already be described using the determined parameter combinations, first step. If the influences can only be described insufficiently, see termination condition of fourth step, then further relevant influencing factors are searched for and the process is continued from the second step onward. The process of this method is shown schematically in Figure 4. 214 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Figure 4: Screening Method 2.4 Results of the screening method The screening method was assessed on various physical models. Figure 5 shows the identified parameter combinations with the corresponding influences of a physical heat exchanger model with 100 system parameters. Combination 14 [with the system parameters 9 23 24 26 34 35] is a recognized term of multiplicity 6 in the relative parameters. The influence shown refers to the maximum variation of the performance parameter. . Figure 5: determined system parameter combinations with relevant influences A further illustration shows the identified influences in relation to the corresponding parameters. The influences shown could be reproduced and confirmed in the physical model. It shows clearly that there are system parameters that become significant only in higher multiplicity, e.g., parameter 36 has insignificant influence up to multiplicity 3. 215 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Figure 6: Relevant influences of system parameters 3 Data driven meta modelling 3.1 Radial basis function network Since a complex physical simulation can be very costly in terms of computational time, it is advantageous to replace the original simulation by a data driven model that evaluates very quickly. In the following approach, the data driven model is realized by a radial basis function network (RBF). This is a special type of a feed forward neural network with one hidden layer and a radial basis function 𝜑 as activation function. If the input is described by a vector 𝐱 ∈ ℝ 𝑛 , the output is a scalar and is given by (2) where 𝑁 is the number of hidden neurons, 𝐜 𝐢 is the center vector of neuron 𝑖 and 𝑤 𝒊 is the weight of the 𝑖 -th neuron entering the linear output neuron. This equation can be written in matrix form (3) 3.2 Subspace decomposition Since the number of hidden neurons necessary for an adequate description of the physical system increases geometrically with the number of input variables, a naive model training may not be possible for a system with as many as 100 system parameters at reasonable computational cost. However, the numerical effort can be drastically reduced if the high dimensional input space can be dissected into a number of (at least partially) orthogonal subspaces. This means, that the original function 𝑓: ℝ 𝑛 → ℝ is replaced by a sum of single RBF nets defined on lower dimensional subspaces. For a suitable decomposition, the results of a previously conducted influence analysis are used. The subspaces are defined on the basis of groups of identified influential parameter combinations. The parameter combinations are grouped in a way that both a minimum number of single RBF nets and a minimum number of inputs for each net are achieved. 𝑓(𝒙) = ∑𝑤 𝑖 𝜑(‖𝐱 − 𝐜 𝑖 ‖), 𝑁 𝑖=1 [ 𝜑 11 𝜑 21 ⋯ 𝜑 1𝑁 𝜑 2𝑁 ⋮ ⋱ ⋮ 𝜑 𝑁1 ⋯ 𝜑 𝑁𝑁 ] [ 𝑤 1 𝑤 2 ⋮ 𝑤 𝑁 ] = [ 𝑓(𝑥 1 ) 𝑓(𝑥 2 ) ⋮ 𝑓(𝑥 𝑁 ) ]. 216 6.1 Application of a DoE based robust design process chain for system simulation of engine systems After setting up this base model structure, each single RBF net is trained whereby additional simulation runs may be necessary [6,7]. Figure 7: Train the RBF models The resulting models are then combined to an overall model. Since the subdomains may partially overlap the models cannot be simply added. So only the center vectors 𝐜 𝐢 of each sub model will be kept and the weight vectors will be recalculated. From equation 2 it becomes clear, that this can be done by a simple linear regression. Figure 8 illustrates the model merging process in the case of two RBF nets. Since the subspaces are orthogonal in this case, the resulting overall regression matrix has a block diagonal structure. Figure 8: Merging 2 RBF nets defined on orthogonal subspaces into overall model 3.3 Results This model building strategy was evaluated by using a physical nonlinear heat-transfer model with 100 system parameters (see Chapter 2). The amount of connected RBF models to one large cluster varies in accordance to the performance parameter. For this example, between 6 and 10 RBF models with up to eight inputs each got connected to one large RBF cluster. The validation based on 270.000 points shows very small validation error, Table 2. For t he performance parameter max_Temp_difference_degC‘ 98% of all measuring points possess a validation error smaller 1%. To calculate these points a standard laptop took only 30s. 217 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Table 2: Error histogram of three important performance parameter 4 Tolerance Fields The data driven simulation model based on a RBF network can be used for a huge amount of model evaluations. Since the computational effort is significantly reduced compared to physical models, the system simulation using the data driven model allows to perform parameter variations for a high number of system parameters including their tolerances. Besides the nominal system performance, which is typically represented by a graph, a tolerance field can be created for each performance parameter. This tolerance field shows the scatter band of the performance parameter, when the system parameters are varied within their minimum and maximum values. Figure 9 shows a schematic representation of the tolerance field of a performance parameter which was created by the data driven model. The black curve represents the nominal system performance. For the analysis of the tolerance field, the values assumed for the system parameters can be varied according different distributions within the tolerance range. Figure 9: Tolerance field of a performance parameter, created by RBF network model. integral_total_pressure' Max_piston_temp_degC' max_Temp_difference_degC' RMSE Error 0,39 0,76 0,33 Arithmetic mean error 0.20% 0.50% 0.22% Max error 18.61% 11.47% 4.09% Error >25% 0 0 0 25%-15% 2 0 0 15%-8% 30 27 0 8%-4% 276 535 1 4%-1% 4763 35774 5206 1%-0.4% 27858 72941 30632 0.4%> 235918 159570 233008 218 6.1 Application of a DoE based robust design process chain for system simulation of engine systems 5 Design Centering 5.1 Objective In the design centering or yield maximization process, we are looking for a nominal design point that allows for a widest possible tolerance range in the domain of input parameters without violating a given set of constraints in the co-domain. In other words, we are seeking for the hypercube with maximum volume that is centered at the design point and fits into the constrained input space. The task is illustrated for a two-dimensional input and response space in Figure 10. It should be noted that although the constrained area is rectangular in the space of performance parameters, the constrained input domain may be non-convex or even non-contiguous. 5.2 Particle swarm optimization The tolerance ranges depicted in the right hand side of Figure 10 are described by the coordinates of the center point 𝐜 = (𝑐 1 , 𝑐 2 , … , 𝑐 𝑛 ) and by the tolerance 𝐭 = (𝑡 1 , 𝑡 2 , . . , 𝑡 𝑛 ) in each input variable. Hence, an optimal design, that is the hypercube with maximum volume, can be obtained from the following optimization task 𝑉 𝑜𝑝𝑡 = arg min 𝒄,t {− ∏ 𝑡 𝑖 + 𝜑(𝒄, 𝒕) 𝑛1 }, (4) where the constraints on the performance parameters have been replaced by a penalty function 𝜑(𝒄, 𝒕) . This penalty function is evaluated by randomly placing a fixed number of points near the surface of the current hypercube and checking the resulting performance parameters for each point against the given boundaries. For the actual optimization task, a particle swarm optimization (PSO) was used that is described in great detail in [8]. This algorithm uses a population of candidate solutions called particles that move around in the search space, that is defined by 𝒄, 𝒕 ∈ ℝ 𝑛 in the current optimization task. The movement of the particles is governed by their own best known position and by the current optimum of the entire swarm. The penalty function described above is evaluated for each iteration step and each particle. To lower the computational cost, detected limit violations are stored. Figure 10: Illustration of the design centering task. 219 6.1 Application of a DoE based robust design process chain for system simulation of engine systems Figure 11 gives an example for a possible outcome of the described optimization algorithm. Here, the nominal design point for three system parameters of a valve train system has been determined. A constraint was defined by giving an upper limit on one of the performance parameters of the system. 6 Summary and Outlook System simulation using physical models plays an important role in the development process of engine components and systems. These simulation models can contain a large number of input parameters that need to be considered including their tolerances. Each parameter can have an impact on the system performance individually or in combination with other parameters. A process chain that uses DoE methods has been presented. In the first step, a parameter screening is performed to determine those parameters and parameter combinations that have a significant influence on the system performance. These parameter combinations are used in the next step of the process chain to create a data driven meta model. This model consists of a RBF network, shows very small validation errors and evaluates very quickly. Since the model can be used for a huge amount of model evaluations it is well suited for further analysis and optimization of the system. The creation of tolerance fields and visualization as a scatter band for a certain output of the model has been shown as typical application example. Furthermore, this model can be used for the "design centering". That means the nominal design point is optimized to allow for the widest possible tolerance range of the input parameters. For this task, a particle swarm optimization algorithm is used and constraints on the output parameters are taken into account. The robust design process chain for the system simulation of engine systems has been used for the analysis and optimization of several systems. The applications showed Figure 11: Result of a design centering for 3 input variables optimized and a limit on one performance parameter. Detected limit violations are marked by blue dots. 220 6.1 Application of a DoE based robust design process chain for system simulation of engine systems that the integration of the process chain in the existing simulation environment is a key factor for usability and enables the simulation engineer to make use of it in his daily work. The modular interface of the process chain allows the connection to further simulation tasks beyond system simulation, e.g. for multi body dynamic simulation (MBS). The methods for data analytics used in the process chain could also be used to analyze data from sensors in general - no matter if this data is produced during the product development, the manufacturing or during the lifecycle of a product. In connection with the Digital Platform [9] - a central data hub for big data analyses - this holistic approach enables Schaeffler to test innovations before the first prototype is even built and to permanently optimize our products throughout their lifecycle. Literature [1] Scheidt, M., Lang, M.: Pure Efficiency - Developing combustion engines from the perspective of a supplier. Solving the Powertrain Puzzle - Schaeffler Symposium, Baden-Baden, 2014 [2] Solfrank, P., Dietz, J.: Potentials of Modern Camshaft Phasing Systems. In: MTZ 77 (2016), Nr. 11 [3] Haas, M., Rauch, M.: Elektrohydraulischer Vollvariabler Ventiltrieb. In: MTZ 71 (2010), Nr. 3 [4] Haas, M., Piecyk, T.: Get Ready for the Combustion Strategies of Tomorrow. Solving the Powertrain Puzzle - Schaeffler Symposium, Baden-Baden 2014 [5] Haukap C., Hegmann M., Köhler B-U.: Strategies for improving the process of Test Design and Test Plan computation for high dimensional designs, 6th conference “Design of Experiments (DoE) in Engine Development”, 2011. [6] Baumann W., Dreher T., Röpke K., Stelzer S.: DoE for series production calibration, 7th conference “Design of Experiments (DoE) in Engine Development”, 2013. [7] Haukap C., Barzantny B., Röpke K.: Model-based calibration with data-driven simulation models for non-DoE Experts, 5th conference “Simulation and Testing for Automotive Electronics”, 2014. [8] Clerc M.: Particle Swarm Optimization, ISTE Ltd, 2006 [9] Schaeffler AG: Digitalization - Shaping the future with the “Digital Agenda”, Schaeffler “At a glance”, 2017, http: / / www.schaeffler.com/ atagla nce 221 6.2 Application of Emulator Models in Hybrid Vehicle Development Justin Seabrook, Pascal Revereault, Mike Preston, Markus Grahn, Johan Bringhed, Björn Lundberg Abstract The trend for greater use of simulation in vehicle and powertrain development has led to new applications of statistical methods. DoE has long been used to build “emulator” models of more complex simulation models, thus creating fast models that can be used in a more flexible optimisation environment. Indeed, stochastic or Gaussian process models, which have gained considerable prominence in automotive DoE, are an evolution of emulator methods from the 1990s. In this paper, some aspects of the use of emulators are explored, taking as an example the development of a hybrid vehicle. For this 48V mild hybrid passenger car, the DoE variables were a mix of hybrid strategy, aftertreatment hardware and aftertreatment control parameters. The interactions between variables, especially those associated with the hybrid supervisor, meant that experiment design constraints were required and, even with these constraints, there is the potential for discontinuous response characteristics. The responses of interest were tailpipe emissions on various RDE cycles, and the optimisation objectives were to minimise CO2 hardware cost while meeting targets for emissions over the chosen cycles. 1. Introduction With pressure on the automotive industry to meet increasingly stringent legislative tailpipe emissions and fleet average CO2 targets, the importance of thorough analysis techniques into alternative powertrains has never been higher. The new WLTC and RDE cycles present a further challenge, with the emissions control envelope being stretched significantly relative to the NEDC. Presented with these challenges, and aware of the limitations of depending solely on internal combustion engine technology advancement for reaching the legislative targets, OEMs are increasingly exploring and investing in alternative powertrains [1,2]. Although limited in operational flexibility, the conventional powertrain has retained its majority share of the passenger vehicle market due to advances in aftertreatment technology. This has brought significant improvements in de-NOx capability. The move to electrification, with its associated increases in drivetrain efficiency and waste energy recycling, has the ability to reduce both engine CO2 and NOx. 222 6.2 Application of Emulator Models in Hybrid Vehicle Development In addition, the hybrid powertrain opens up synergies with the aftertreatment system, enabling the optimisation of tail pipe emissions with more flexibility over exhaust temperature management than is possible with a conventional powertrain. The 48V hybrid architecture represents an interim step in the direction of full electrification, making available all the hybrid operating modes of a full hybrid vehicle, but limited in terms of energy storage and electric motor power. The 48V architecture adds significant operational flexibilities to powertrain and aftertreatment system strategies and operation, enabling greater opportunities for optimisation and bespoke operation under challenging and diverse real world driving conditions. 2. Emulator Models DoE was “invented” in the 1930s [3,4] for agricultural research, and was adopted by the chemical and process industry in the 1950s [5]. The automotive industry did not embrace DoE until the 1990s [6] after Taguchi [7] brought his ideas to the US and Europe. Most of these DoE applications concerned collecting real data from a test and creating a statistical model of the physical system. But, as computers came into widespread use, the idea of making a “model of a model” was born. The term “emulator” refers to the use of DoE models to emulate or replicate a computer model. The reasons for creating emulators mostly revolve around the speed of the emulator model compared to the original computer model, which is often computationally expensive. The computer model may take seconds, minutes or hours to run, whereas the emulator is likely to have an evaluation time of only a few milliseconds. This greater speed gives the engineer flexibility to explore the system more efficiently, be that through a model viewer or with an optimisation algorithm. The first emulator models were simple linear models fitted to classical 1950s designs. But for more sophisticated computer models, the limitations of polynomial emulators are self-evident and statisticians started to design experiments explicitly for computer experiments [8]. The landmark paper for non-linear emulation with a stochastic process model was Sacks et al [9], which demonstrated the power of the emulator approach for the optimisation of transistor electronics. In subsequent years, the method was adapted for use with empirical data [10,11] and, more recently, Gaussian stochastic process models have become the pre-eminent modelling method for automotive DoE. 3. Hybrid Application Two 48V parallel hybrid configurations were considered for the XC90 vehicle platform which, using the standard hybrid terminology, were “P0” and “P2” applications (Figure 1). The P0 layout had a 48V Belt Starter Generator (Motoring 12kW, Generation 15kW) located on the 2L TDI engine FEAD (front end accessory drive) with a pulley ratio of 3.6, and an 8Ah battery. The P2 layout, capable of ‘Electric Vehicle’ or EV mode, had an Integrated Starter Generator (Motoring: 15kW, Generation: 17kW) driven at a ratio of 3.6 off the vehicle-side clutch plate, and a larger 10Ah Battery. 223 6.2 Application of Emulator Models in Hybrid Vehicle Development Figure 1: Schematic of P0 and P2 Vehicle Architectures P0 and P2 shared the same initial aftertreatment system layout, as depicted in Figure 2. As discussed later, catalyst volumes were variables in the DoE and the final optimised values varied from the standard baseline system, and between P0 and P2 applications. In addition to this, an electric heater (also a DoE variable) was mounted before the inlet of the LNT, the size of which was likewise included in the DoE. Figure 2: Standard A/ T layout in all cases (volumes varied) Only P2 could operate in EV mode, but both architectures could operate in four other modes as follows: (1) Torque Assist - where the engine is assisted by the electric motor to power the vehicle (2) Regenerative Braking - the capture of braking energy and storage thereof in the battery (3) Generation - the increased loading of the engine with the electric motor run as a generator (4) Electric Exhaust Heating - use of the 48V circuit to power an electric heater located before the intake of the LNT. The modelling of both architectures was achieved using the Ricardo modelling tool V- SIM (Vehicle SIMulation). V-SIM is a forward facing Simulink model, and a key com- 224 6.2 Application of Emulator Models in Hybrid Vehicle Development ponent in the Ricardo IMBD (Integrated Model Based Development) process. V-SIM is able to model vehicle hardware, controllers and hybrid control algorithms. For the purpose of ensuring adequate model accuracy and robustness, initial model setup and correlation was done using data gathered on an existing non-hybrid XC90 vehicle. After validation of the baseline model, the hybrid systems, functionality and controls were added to emulate the P0 and P2 architectures. To support the DoE investigation and optimisation processes, the models were configured to allow easy access and setting of all model parameters of interest. 4. Design of Experiments 4.1 Design A necessary step in the emulation process is the selection of variables to include in the DoE. The computer model may have hundreds or thousands of parameters, but the vast majority of these are not of great interest to the engineer, who is usually concerned with a subset of these that can be defined or varied independently. There were are large number of variables that could potentially be included in the hybrid optimisation DoE. In such a situation, the options are either to “brainstorm” the candidate variables or to conduct a preliminary “screening” DoE. In this instance, it was expected that some creativity would be needed to reduce the number of variables to a manageable level, so a brainstorming session was held including simulation and DoE specialists. The variables cover five areas for the hybrid applications: Considering first the selection of hybrid mode, Figure 3 shows the maximum torque curve for e-motor alone, for engine alone and for the two together. For a wide range of operation, the hybrid supervisor must choose how to provide the torque requested by the driver and other vehicle systems. It takes into account the battery state of charge and the efficiency of the engine and electrical systems. Specifically, one of the key choices at any operating point is whether the e-motor provides torque (torque assist mode) to the wheels or charges the battery (generation mode). Both of these modes change the engine operating torque and therefore the efficiency of the engine. The boundary between the torque assist and generation modes is shown by the black line in Figure 3. Above this line, torque assist is active and the amount of torque assist increases with distance from the line. Below the line, generation mode is active and likewise the amount of generation increases with distance from the line. The shape and location of this line are calibration variables. For the purpose of the DoE, 225 6.2 Application of Emulator Models in Hybrid Vehicle Development the shape of the curve was maintained but it was allowed to move up and down according to a scaling variable “TA-Gen Line Scaler”. Figure 3: Engine and E-motor Torque with TA-Gen Line (illustrative) Figure 4: Active Range of EV, TA, Gen and LNT Heat Modes versus SOC 226 6.2 Application of Emulator Models in Hybrid Vehicle Development The hybrid supervisor strategy mode entry and exit conditions also take into account battery State of Charge (SOC), as illustrated in Figure 4. Thresholds for enabling: • Electric Vehicle Mode • Torque Assist Mode • Generation Mode • LNT Heater Mode are calibrateable parameters and were included in the DoE. It was necessary to handle several tables and curves within the control strategy in a similar manner to “TA-Gen Line Scaler”. For example, rather than define a curve with 6 calibrateable elements as 6 DoE variables, the characteristic shape of the curve was retained and modified by scaling factors and/ or offsets. Eight of these scalers/ offsets were included as DoE variables: Parameter Min Max TA-Gen Line Scaler 0 1.5 EV Active Scaler 0.5 1 EV SOC Offset 0.02 0.3 TA Scaler 0 3 TA SOC Offset 0.02 0.3 Gen Scaler 0 3 Gen LNT SOC Offset 0 0.45 Gen Max SOC 0.3 0.55 The units of these variables are dimensionless and ranges are strategy-specific but included in the table for completeness. The initial battery state of charge was included so that the vehicle could be optimised for a range of battery states to ensure that vehicle performance was robust to initial SOC: Parameter Min Max Initial Battery SOC (%) 30 70 Five parameters relating to the LNT, two hardware and three control parameters, were also included as DoE variables: Parameter Min Max LNT Size (L) 0.2 2 LNT Heater Power (W) 0 3000 LNT Heater Off Temp (°C) 180 250 LNT Regeneration Threshold (%) 75 95 LNT Regeneration Window (°C) 50 200 The regeneration window refers to the width of the window, not the absolute temperature range which is always centred on 300°C. For example, a window of 100°C means that regeneration is enabled between 250 and 350°C. 227 6.2 Application of Emulator Models in Hybrid Vehicle Development Finally, two hardware parameters relating to the SCR system completed the list of DoE variables: Parameter Min Max SCR Size (L) 1.1 3 SCRF Size (L) 3 5 In total, 16 variables were defined for the DoE. Some responses were expected to be highly non-linear with respect to some variables so the DoE was “over-sized” with 400 cases. Figure 5 shows the design. It can be seen that the shape of the design space is a pure hypercube. No design constraints were required. Figure 5: Design 4.2 Simulation Figure 6: RC130 RDE Cycle Each of the hybrid vehicle configurations represented by the 400 DoE cases were run in VSIM over two test cycles - WLTC and RDE. The RDE cycle (Figure 6) was the 228 6.2 Application of Emulator Models in Hybrid Vehicle Development Ricardo City 130 (RC130) which represents a congested city in the urban phase with low average speed, followed by moderately aggressive rural and motorway phases with maximum speed of 130km/ h. Each WLTC run was approximately 2 minutes and the total run time was 13 hours for the 400 cases. The RDE cycle is longer and the run time per case was 5 minutes, or 33 hours for all cases. 4.3 Modelling For each of P0 and P2, SPMs were fitted to the data for the following responses, all in mg/ km unless stated otherwise: • WLTC Engine Out NOx • WLTC Tailpipe NOx • WLTC CO2 [g/ km] • WLTC SOC Delta [-] • RC130 EO NOx • RC130 TP NOx • RC130 Urban EO NOx • RC130 Urban TP NOx • RC130 SOC Delta [-] • RC130 CO2 [g/ km] The models for P2 are given in Figures 7 and 8. Both the RDE and WLTP cycles represent close-to-real world driving, and the responses vary similarly with respect to the DoE input parameters. The greatest sensitivity seen, and represented in both cycles, is to the Generation Max SOC - which shows a significant NOx and CO2 penalty as this parameter increases. Figure 4 shows the reason for this trend (depicted in green): as Gen Max SOC is increased, the battery SOC window for all other hybrid operational modes is restricted. A variable that shows significant emissions control benefit is the LNT Heater Power, which can be seen to reduce tail-pipe NOx significantly through rapid heating of the LNT and down-stream SCR. But, also apparent is an associated CO2 cost with running a higher power LNT heater. The higher battery drain for LNT heating leaves less opportunity for EV mode, and greater need for battery SOC maintenance (generation). Another input with significant impact is the EV SOC Offset parameter. As implemented in the strategy, the greater the offset, the less EV mode is used in the cycle. Conversely, greater use of EV in the cycle has more benefit in terms of both engine-out NOx and CO2. The final input variable with significant influence on tail-pipe NOx, is the LNT Size. Larger LNT monolith masses require greater energy to heat, so there is a CO2 and engine out NOx penalty, but a significant benefit in tail-pipe NOx. This is due to the associated greater NOx storage capacity and greater conversion efficiency at higher exhaust gas space velocities. 229 6.2 Application of Emulator Models in Hybrid Vehicle Development Figure 7: Model Viewer with P2 RDE Cycle Responses Figure 8: Model Viewer with P2 WLTC Responses 4.4 Optimisation & Validation A key advantage of the emulator approach is the speed of optimisation. Optimisation with the original VSIM models would entail a lengthy iterative process of simulation > evaluation > adjustment and back to simulation. With the emulator model, the optimisation can be set up in a few minutes and runs in a few minutes. Separate optimisation processes for WLTC and RDE produced very different optimum configurations. So it was decided to generate a pareto curve for WLTC CO2 and RDE CO2 with constraints on NOx and change in battery state of charge on both cycles. Change in SOC was constrained to be net positive. Figure 9 shows the opti- 230 6.2 Application of Emulator Models in Hybrid Vehicle Development miser setup and Figure 10 shows the output from this process. Additional pareto curves were also generated with constraints applied to catalyst volumes to understand the cost versus CO2 characteristic in more detail. Once the optimum configuration had been selected, a V-SIM run was undertaken to confirm the emulator prediction. Figure 9: Optimiser Configuration P2 WLTP CO2 [g/ km] Figure 10: Pareto Curve for WLTC and RDE CO2 231 6.2 Application of Emulator Models in Hybrid Vehicle Development 5. Conclusions Hybrid vehicles continue to grow in popularity and market share, and 48V systems are a cost effective interim technology. Selection of hybrid hardware and architecture is complicated by the interactions with supervisor controller calibration and the aftertreatment system. Using DoE to emulate more computationally-intensive physical models enables the engineer to incorporate a large number of variables in the optimisation process. This gives flexibility at the optimisation stage to evaluate more hardware options on an “optimised supervisor controller” basis. In the emulator application presented, 16 input variables were handled in a 400 case DoE. The non-linear modelling capability of the stochastic process model was maintained with a reasonable number of test runs, even with the relatively high dimensionality involved. 6. References [1] Peter Mertens & Mats Andersson (2016), “Volvo Cars Strategy for Electrification and Positioning for the Future”, CTI Symposium, Berlin [2] Nikolce Murgovski, Markus Grahn, Lars Mårdh Johannesson & Tomas McKelvey (2015), “Automated Engine Calibration of Hybrid Electric Vehicles” IEEE Transactions on Control Systems Technology (Volume: 23, Issue: 3) [3] Fisher, R A (1935), The Design of Experiments, Oliver and Boyd, Edinburgh [4] Yates, F (1935), "Complex experiments", Supplement to Journal Royal Statistical Society, Vol. 2, pp 181-247 [5] Box, G E P & K B Wilson (1951), "On the experimental attainment of optimal conditions", Journal of the Royal Statistical Society, Series B, Volume 13-1 [6] Grove, D M & T P Davis (1992), Engineering Quality and Experiment Design, Longman Scientific & Technical, UK [7] Taguchi, G (1987), System of Experimental Design, Volume 1, UNIPUB/ Kraus International Publications, New York, USA [8] M. D. McKay, R. J. Beckman and W. J. Conover (1979), “A Comparison of Three Methods for Selecting Values of Input Variables in the Analysis of Output from a Computer Code”, Technometrics Vol. 21, No. 2 [9] Sacks J, Welch W J, Mitchell T J & Wynn H P (1989), “Design and Analysis of Computer Experiments”, Statistical Science [10] Justin Seabrook, Simon Edwards, Tomasz Salamon & Ian Noell (2003), “Comparison of Neural Networks, Stochastic Process Methods and Radial Basis Functions for the Optimisation of Engine Control Parameters”, DoE in Engine Development, Berlin [11] Thomas Kruse, Holger Ulmer & Ulrich Schulmeister (2007), “Use of Advanced Modelling and Optimization for Diesel and Gasoline Engine Calibration”, 4th conference on DoE in Engine Development, Berlin 232 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Owen Parry, Julian Dizy, Vivian Page, Amit Bhave, David Ooi Abstract Virtual engineering that combines physico-chemical models with advanced statistical techniques offers a robust and cost-effective methodology for model-based IC engine calibration and development. The probability density function (PDF)-based Stochastic Reactor Model (SRM) Engine Suite is applied to simulate a modern diesel fuelled compression ignition engine. The SRM Engine Suite is coupled with the Model Development Suite (MoDS) to perform parameter estimation based on the engine measurements data at representative load-speed operating points. The fidelity of the SRM Engine Suite is further tested by carrying out blind tests against measurements for combustion characteristics (heat release rate, in-cylinder pressure profile, etc.). The comparison between the calculated (software evaluated) and the measured combustion characteristics shows good agreement. Furthermore, the model evaluations for engine-out soot and NO x emissions agree relatively well with those measured over the entire engine load-speed operating window. Fast response computational surrogates are generated using High Dimensional Model Representation (HDMR), the performance of which are then compared with the validated SRM Engine Suite. The associated global sensitivities are also evaluated to understand the influence of the engine operating conditions and the model parameters on combustion metrics such as ignition delay, maximum heat release rate, maximum pressure rise rate and peak in-cylinder pressure. The benefits of the proposed methodology, comprising formulation of physico-chemical model-based surrogates and its application to software-in-loop (S-i-L) and control in compression ignition engines, are also discussed. 1 Introduction Demonstrating continuous reduction of vehicular/ machine and powertrain development costs and time through the use of model-based engineering analyses is key to the wider adoption of innovative virtual or digital engineering workflows that augment experimental analyses. The variability and complexities of the modern powertrain control calibration with the increasing degrees of freedom offer significant opportunities for model-based engineering to yield cost-reduction benefits, while developing vehicles and non-road machines that comply with stringent CO 2 and other gas phase as well as particulate phase emissions regulations. Traditionally, engine calibration has relied on engine dynamometers and vehicle testing. The Engine Control Unit (ECU) development has conventionally been dependent on look-up table based approaches, or on cost-intensive closed-loop control strategies that rely on several production-level sensors [1]. Measurement-driven engine calibration methodology involves the generation of engine data, generally dictated by a design of experiments (DoE) strategy. The data points are then used to build statis- 233 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines tical meta-models, or response surfaces, of various combustion characteristics; for instance, engine performance and emissions as a function of engine load and speed. Optimisation techniques are then applied to these fitted meta-models to identify optimal actuator settings at individual load-speed points, followed by interpolation in order to generate smooth actuator maps. The powertrain system complexity and the necessary design iterations within the measurement-driven calibration provide further impetus to augment the calibration with model-driven methods. Model-based methodology with calibrated zero dimensional (0D) models that incur low computational expense have been widely applied to control applications. For example, recently a 0D model was applied towards prediction and optimisation of combustion and engine performance parameters such as the angle corresponding to 50% of fuel mass fraction burnt (MFB50), the maximum incylinder pressure and indicated mean effective pressure (IMEP) for model-based combustion control [2]. The same combustion and performance characteristics were also estimated elsewhere for feed-forward control within the ECU [3,4,5]. In another study, based on test measurements from a 4-cylinder compression ignition engine, mean value models for fuel consumption and emissions were used for model-based optimal calibration [6]. However, adequate model-accuracy is crucial to the design of an optimal, fast-response and robust control system [7]. Using physics-based models is hence necessary to ensure the accuracy of the results and importantly the applicability of the model to a wide range of driving conditions, and even other engines [8]. To exploit the predictive capability of combustion characteristics offered by 3D computational fluid dynamics (CFD) models, a two-step methodology can be adopted. First, a detailed CFD simulation is run in order to populate the training datasets, followed by regression analysis to construct the response surfaces. The potential for various regression techniques, such as K-nearest neighbours, Kriging, Neural Networks and Radial Basis Functions to replace or partially substitute CFD evaluations has been studied by [9]. The efficiency of the methodology still largely depends upon the number of CPU-intensive CFD evaluations, which is dictated by the Design of Experiment (DoE) method adopted and the number of design parameters. Simulating an individual load-speed operating point using 3D CFD with detailed chemical kinetics can take up to a day or two, making it impractical to cover the whole design space. To realise efficient model-based engineering, it is vital to combine the predictive capability of simulation with practical computational overhead. Furthermore, to shorten powertrain development cycles, it is important to equip simulation engineers with systematic methodologies that make predictive simulation-based workflows more robust and more accurate. Recently, Extremum Seeking (a gradient-based optimisation approach) was used to calibrate the heat transfer coefficient in a semi-empirical, simplified engine model in order to minimise the error between the measured and modelled in-cylinder pressure [7]. Initial model parameter values were selected based on experience and on studies in the literature in order to improve the likelihood of finding a global, rather than local, minimum in the cost function. A hybrid calibration method based on a combination of physics-based models and optimisation methods has been proposed [10]. The entire IC engine simulation model was divided into sub-systems and the parameters calibrated using measurement data on the sub-system level. This step was then followed by performing calibration with a reduced number of the most dominating parameters in several loops, while using the entire IC engine simulation together with optimisation methods. In another 234 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines study, [11] has also demonstrated automated calibration of combustion and heat transfer parameters, such as the radiation coefficient, the combustion terminal angle, combustion speed, etc., using a combination of ant colony and genetic algorithms. The aim of this paper is to present a methodology for developing IC engine models with high predictive capability and low computational expense, making them ideal candidates for model-based engineering workflows such as model-in-loop or software-in-loop. To achieve this objective, the probability density function (PDF)-based SRM Engine Suite was chosen for simulating fuels, combustion and emissions in modern IC engines [12-15]. The physico-chemical simulator was then coupled with an advanced statistical toolkit, Model Development Suite (MoDS) [16-18], to calibrate the model “automatically” based on the measurements data. This represents a significant improvement over previous attempts involving manual calibration [19]. MoDS was then used to generate fast-response surrogates and perform sensitivity analysis of key combustion characteristics and emissions as a function of engine operation variables. The paper is organised as follows: First, the Tier 4 capable IC engine geometry and operating conditions are presented. Then, the two software toolkits that were used, i.e. the SRM Engine Suite and MoDS, are introduced. Both have been applied in numerous studies elsewhere, hence, to avoid repetition, only the features unique to the present paper are explained. The automated base model calibration in terms of the in-cylinder pressure profile, soot and NO x emissions performed over the representative engine load-speed operating points is presented, followed by the results of the blind tests carried out on additional engine operating points. This is followed by the discussion on the quality of the computational surrogates generated and the sensitivity of the combustion and emissions characteristics of interest on engine operating variables. 1.1 Engine The data used for model validation has been obtained from a Cat® C4.4 ACERT turbocharged Diesel-fuelled Compression Ignition (CI) engine. In total, 146 steady state operating points with single and double injection strategies were obtained during the present study. Table 1 provides the basic engine geometry data, whereas the steady state load-speed operating points are displayed in Figure 2. Please note that the load-speed points used in this study are not representative of the engine’s final calibration. Table 1: Engine geometry for the Cat® C4.4 ACERT single-turbocharged Dieselfuelled Compression Ignition (CI) engine. 235 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 1: The Cat® C4.4 ACERT single-turbocharged Diesel-fuelled Compression Ignition (CI) engine. 1.2 SRM Engine Suite The SRM Engine Suite is an advanced toolkit to simulate fuels, combustion and emissions in IC engines. Figure 2 presents a snapshot of the software graphical user interface (GUI). The GUI offers the users an intuitive workflow for creating a new engine model based upon engine templates and a fuels library. The main inputs concern basic engine geometry, operating conditions, and fuel characteristics, with advanced features to account for characteristic k-ε turbulent mixing time profiles, injection mass rate profiles, etc. Figure 2: A screenshot showing the SRM Engine Suite graphical user interface. The software has been successfully applied to simulate unsteady spark ignition (SI) combustion [20-22], conventional compression ignition [23] including dual-fuel (natural gas with Diesel pilot), partially premixed charge compression ignition (PPCI) as well as other low temperature combustion modes of engine operation [24-27]. In addition, engine breathing (intake and exhaust) and EGR (including the species concentrations) are also accounted for by the software. However, other powertrain components such as heat exchangers, turbocharger, etc. are not modelled within the SRM Engine Suite. In such instances, the SRM Engine Suite offers a coupling for cosimulation with 1D and 3D toolkits [28,29]. The SRM (Stochastic Reactor Model) Engine Suite uses a Probability Density Function (PDF) transport equation [30,31] based approach and accounts for the detailed 236 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines chemistry of fuel oxidation and the emissions formation pathways. The model also accounts for the inhomogeneities in the -T space (chemical equivalence ratio; temperature) by mimicking the sub-processes of turbulent mixing, wall heat transfer, multiple direct injections, etc. For variable density flows, an MDF (mass density function) rather than a PDF is used to describe the SRM. This MDF is then solved using a Monte Carlo particle method with a second-order operator splitting algorithm [31]. The SRM Engine Suite provides four model parameters which have been calibrated to represent the experimental heat release rate (HRR) and the in-cylinder pressure: • The liquid fuel evaporation coefficient which influences the atomisation of the injected spray into the turbulent flow (λ evap ). • The heat transfer coefficients which impact the heat transfer between the flow within the cylinder and the combustion chamber surface (C 1 , HRC 2 ) • The characteristic turbulent mixing time (k/ ε) parameter used to control frequency of mixing events (C ). • The injector spray distribution term, directly related to the injected spray angle within the cylinder (α inj ). Additional calibrated parameters include: • The turbulent mixing time during injection (C inj ). • The stochastic heat transfer constant (C HT ). • The Sauter Mean Diameter constant (SMD A ). The empirical soot sub-model has four input parameters. The rates of soot formation and oxidation are controlled by pre-exponential multipliers (C sfpe and C sope respectively) and exponential multipliers (C sfe and C soe respectively). Finally, the NO x sub-model has a single scaling coefficient (C NOx ) which was applied globally (for all operating points studied). 1.3 MoDS MoDS is a highly flexible software package designed to simplify model development using an advanced suite of numerical and statistical tools. Models can take a variety of different forms, including virtually any executable that can be run from the command line and reads (writes) its inputs (outputs) outputs from (to) file. The key features of MoDS include: • Automatic parameter calibration - Calibrate a model against an existing data set. • Uncertainty analysis - Find the parameters with the largest uncertainties and/ or propagate those uncertainties through a model. • Fast response/ surrogate models - Generate surrogates to approximate results from a detailed model, but which can be evaluated in a tiny fraction of the time. The user can choose between polynomial, HDMR and Kriging. • Sensitivity analysis - Quantify how much of the variance in the model outputs is due to each model input. • Multi-processor support - Parallelise workflows via MPI. • An intuitive GUI - Supports the design process and guides the user through the work-flow (see Figure 2). 237 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines The features of automatic calibration, surrogate construction and sensitivity analysis form much of the basis for this work and are described in more detail in the following section. Figure 3: An example screenshot of the MoDS graphical user interface. 1.4 SRM-MoDS Coupling 1.4.1 Base Model Calibration The twelve model parameters described in the previous section were calibrated automatically through MoDS using data from experiments performed at 30 operating points. The operating points were defined via 10/ 13 (single/ double) injection process conditions, . 300 points on the pressure profile, together with NO x and soot emissions data, , were used to compare the model and the experiments. The goodness of the model parameters, , was calculated using a least-squares objective function, ∑ , (1) To initialise the calibration, a high-dimensional uniform sampling method, based on Sobol sequences [32], was used to select suitable starting values for the model parameters being calibrated [33]. The local optimisation was then carried out using Hooke and Jeeves’ algorithm [34]. This algorithm was chosen in preference to gradient-based methods which were previously found to perform poorly on similar model systems. At each iteration of the optimisation process, MoDS provides inputs to the SRM Engine Suite, runs it, and evaluates Equation (1) for each output. MoDS then modifies the model inputs according to Hooke and Jeeves’ algorithm, with the aim of minimising Φ( θ ), and begins the next iteration. 1.4.2 Blind Tests To test the predictive capability of the calibrated model, its output at a further 30 operating points was compared to corresponding experimental data. 1.4.3 Surrogate Construction and Sensitivity Analysis MoDS was used to fit surrogate models to the SRM Engine Suite data in order to enable rapid evaluation at new operating points. The high dimensional model represen- 238 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines tation (HDMR) method was used to fit the surrogates, as it allows for the automatic selection of the polynomial order of the surrogate. Another benefit of the HDMR method is that the coefficients can be used to calculate the global sensitivity of each output variable to each input variable. The foundation of the HDMR method is the series decomposition of each model output (response) using orthonormal basis functions. 2 Results and Discussion This section examines results from the three main phases of the study: the calibration of the SRM Engine Suite to experimental data, testing of the model against data not included in the calibration (blind tests) and the construction of surrogates to approximate the detailed model results. In each case, simple metrics are used to compare the in-cylinder pressure profiles, raw concentrations of NO x and soot emissions. 2.1 Base Model Calibration The base model calibration yields a global set of model parameters applicable over the entire load-speed operating window. The results in this section assess the quality of the calibration by directly comparing with experimental data for in-cylinder pressure, NO x and soot emissions. This validation is a necessary step before testing the model under other process conditions. 2.2 In-cylinder pressure profile In order to quantify how well the calibrated SRM Engine Suite reproduces the experimentally-measured in-cylinder pressure profiles (as a function of crank angle), a positive fractional difference was computed for each profile. That is, the average error at an operating point with process conditions x was defined as ∑ abs , , , (2) where N 30 is the number of operating points. The torque-speed map of P err presented in Figure 4 shows that the SRM Engine Suite is able to match the experimental pressure data well - differences are less than five per cent at all operating points and typically less than three per cent in the upper halves of the speed and torque ranges. 239 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 4: A torque-speed map of the average fractional difference in In-cylinder pressure between the SRM Engine Suite and experimental data. The average is computed over 300 points in the pressure profile. The grey labelled circles show the 30 operating points used in the model calibration. 2.3 NO x emissions In this section, the raw concentrations of engine-out NO x (i.e. NO + NO 2 ) will be presented as a comparison between the experimental data and the SRM Engine Suite predictions. Figure 5 shows the experimentally-measured raw NO x output concentration from the C4.4 engine, as a function of torque and speed. There is a clear trend for increased NO x emission at low speeds. Figure 5: A torque-speed map of the experimentally-measured raw NOx emissions. The grey labelled circles show the 30 operating points used in the model calibration. Redder colours correspond to higher NOx emissions. Figure 6 depicts the raw NO x output concentration predicted by the SRM Engine Suite as a function of torque and speed. The calibrated model is able to capture the trend in the experimental data for higher NO x emissions at lower speeds and torques. 240 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 6: A torque-speed map of the NOx emissions predicted by the SRM Engine Suite. The grey labelled circles show the 30 operating points used in the model calibration. The scale used to map NOx values to colours is the same as in Figure 5. Figure 7 illustrates the difference in the raw NO x output concentration predicted by the SRM Engine Suite and that measured experimentally for the C4.4 engine. The model matches the experimental NO x data to within 40 per cent at all operating points, and to better than 20 per cent at most of the steady state load-speed operating points. Almost invariably, the discrepancies between the model and the experiments are due to an underprediction of the raw NO x by the model compared to the experimental data. Figure 7: A torque-speed map of the fractional difference in NOx emissions between the SRM Engine Suite and experimental data. The grey labelled circles show the 30 operating points used in the model calibration. 2.4 Soot emissions Following the same format as the previous section, the engine-out soot concentration will be analysed as a function of load and speed, including assessing the accuracy of the SRM Engine Suite predictions relative to the experimental measurements. Figure 8 plots the experimentally-measured soot output from the C4.4 engine as a function of torque and speed. Soot emissions are relatively constant across all oper- 241 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines ating points, increasing only slightly with engine speed. We note that the elevated emissions observed at point 27 could be explained by the fact that the engine is not being operated optimally; the operating points shown here do not exactly match those used in the final engine calibration. Figure 8: A torque-speed map of the experimentally-measured soot emissions. The grey labelled circles show the 30 operating points used in the model calibration. Note that no colour scale is included here in order to protect data confidentiality. Redder colours correspond to higher soot emissions. Figure 9 shows the soot output predicted by the SRM Engine Suite as a function of torque and speed. Similar to the experimental data, the model predictions are relatively flat across the speed-torque plane, but the calibration is able, to an extent, to capture the increased emission at operating point 27. Figure 9: A torque-speed map of the soot emissions predicted by the SRM Engine Suite. The grey labelled circles show the 30 operating points used in the model calibration. The scale used to map soot emission values to colours is the same as in Figure 7. The difference in the soot output predicted by the SRM Engine Suite and that measured experimentally for the C4.4 engine is shown in Figure 10. The model typically underestimates the measured soot output by 10-50 per cent at low speed and high torque, and by a factor of two at low torque. 242 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 10: A torque-speed map of the fractional difference in soot emissions between the SRM Engine Suite and experimental data. The grey labelled circles show the 30 operating points used in the model calibration. 2.5 Blind Tests While the model is relatively successful at matching the calibration data, it is of course important to validate its performance with data not included in the calibration. For these blind tests, 30 more operating points were selected and the model reevaluated under the new conditions. In the following section, the results of these tests are presented in the form of model-experiment difference maps for in-cylinder pressure, NO x and soot emissions. 2.5.1 In-cylinder pressure (blind tests) A torque-speed map quantifying the average model-to-experiment discrepancy for incylinder pressure, P err , (See Equation 2 for calculation) at the 30 blind test operating points can be seen in Figure 11. Comparing with Figure 4, it is clear that the model performs as well at the blind test points as it does at the calibration points. There is a similar trend for larger discrepancies at low engine speeds, but even in those cases the model remains accurate to within four or five percent. 243 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 11: A torque-speed map of the average fractional difference in in-cylinder pressure between the SRM Engine Suite and experimental data for 30 operating points not included in the model calibration (labelled grey circles). 2.5.2 NO x emissions (blind tests) Figure 12 depicts the difference in the NO x output predicted by the SRM Engine Suite and the experimental data for the 30 blind test operating points. The performance of the model at the test points is similar to the calibration points (Figure 7) except for very low torques where the model-experiment discrepancy is between 50 and 90 per cent. Figure 12: A torque-speed map of the fractional difference in NO x emissions between the SRM Engine Suite and experimental data for 30 operating points not included in the model calibration (labelled grey circles). 2.5.3 Soot emissions (blind tests) The difference in the soot output predicted by the SRM Engine Suite and the experimental data for the 30 blind test operating points is plotted in Figure 13. Both the magnitude of the difference between experiment and model and the trend with speed and torque is very similar to the calibration points (Figure 10). 244 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 13: A torque-speed map of the fractional difference in soot emissions between the SRM Engine Suite and experimental data for 30 operating points not included in the model calibration (labelled grey circles). 2.6 Surrogate and Sensitivities The HDMR surrogate was constructed using 1000 Sobol points sampled from within the process-condition space containing all 60 operating points (those used in the initial calibration, and those used to test the calibrated model). Separate surrogates were generated for the single and double injection cases. The motivation for treating the two separately was that three extra process conditions are required to describe the double injection cases. While the model could have been made to mimic single injection cases by (for example) setting one injection mass to zero, this effectively makes all single injection operating points edge cases of the surrogate model, which is not desirable from a numerical point of view. To assess the quality of the surrogates, they were evaluated at the 30 engine loadspeed points used in the original MoDS calibration. The soot and NO x results from the surrogates were then compared with the results obtained from the SRM Engine Suite. The performance of each surrogates relative to the SRM Engine Suite is shown in Figure 14 and Figure 15. The surrogates capture NO x trends relatively well across most of the 30 operating points. However, at the low speed, single injection points (13 to 15), the surrogate fails to reproduce the trend in NO x emission predicted by the SRM Engine Suite. Both surrogates provide a reasonable match to the soot emissions predicted by the detailed model across the majority of the operating points. Once again, the largest discrepancies are seen at the mid to low speed operating points (9 to 15). 245 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 14: A comparison of NOx results between the SRM Engine Suite and surrogates for both single and double injection operating points. Note that no scale is included here to protect data confidentiality. Figure 15: A comparison of soot results between the SRM Engine Suite and surrogates for both single and double injection. Note that no scale is included here to protect data confidentiality. A useful characteristic of HDMR surrogates is that the coefficients of each term in the decomposition can be used to compute the global sensitivity of each output to each input. In Figure 16 the sensitivity of the in-cylinder pressure to different process conditions is shown. Note that, since the pressure profiles comprise 300 data points, each of which requires its own surrogate model, the sensitivity values quoted here are averages across the whole profile. Figure 16 reveals that the predicted pressure profile is most affected by the initial pressure and the injection parameters, which together explain more than 90 per cent of the total variance. 246 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 16: A pie chart illustrating the fraction of the total variance in in-cylinder pressure (averaged over the entire profile) accounted for by different process conditions. Terms contributing less than five per cent of the total are grouped together in the "Other” category. Sensitivities are also shown here for NO x (Figure 17) and soot (Figure 18). Where there are two parameters shown, this represents the sensitivity of the model output to varying both of the parameters simultaneously. In both cases, emissions are shown to be most sensitive to the injected mass of fuel. In addition, NO x emissions are sensitive to the EGR fraction and to the injection parameters, whilst soot is affected more by the initial pressure. These factors account for around 70 and 80 per cent of the total variance in NO x and soot emissions respectively. Figure 17: A pie chart illustrating the fraction of the total variance in NOx emissions accounted for by different process conditions. Terms contributing less than five per cent of the total are grouped together in the "Other" category. Figure 18: A pie chart illustrating the fraction of the total variance in soot emissions accounted for by different process conditions. Terms contributing less than five per cent of the total are grouped together in the "Other" category. 247 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines 2.7 Parameter sweeps In order to rigorously assert the robustness of the calibration and the performance of the model, a set of parameter sweeps on the full load rated speed operating point has been performed. The parameters being swept include EGR fraction, injection timings and boost pressure. Due to confidentiality reasons, the ranges used in these swings and the specific configurations of each case cannot be disclosed. The output measures that will be considered in this study are, as in the previous sections: incylinder pressure, NO x , and soot. The automated calibration procedure employed for this study is described in Section 2.1; it uses 18 calibration points and 42 blind test points. The concatenated pressure profiles, together with a zoomed-in profile for a single case, are plotted in Figure 19. These are composed of all of the individual control points that are accounted for in the objective function during the calibration. From the individual profile, it is clear that the SRM model is able to capture not only peak cylinder pressure, but also the start of combustion for both the pilot and the main, the total duration and the correct amount of heat transfer throughout the cycle, in particular in the expansion stroke. To this end, a few pressure control points have been added to the objective function just before EVO in order to ensure that the total heat transferred to the walls is accurate. Figure 19: Upper panel : The in-cylinder pressure profiles for the calibration and blind test points, concatenated into a single series. The SRM Engine Suite profiles are plotted as red lines and the experimental data as blue lines and points. The vertical dashed line separates the calibration cases (left) from the blind tests (right). Lower panel : The pressure profile for case 14 as a function of crank angle. The colours and line styles are the same as those used in the upper panel. Looking at the NO x emissions, shown in Figure 20 and covering a wide range of values (the highest value shown is 1250ppm larger than the smallest one), one can observe that both the trends and the absolute values are reflected in the SRM predictions, not only for the calibrated cases, but also in the blind tests. 248 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 20: The NO x emission measured experimentally (blue lines with points) and predicted by the SRM Engine Suite (red line) for each of the 60 combinations of the sweep parameters. The vertical dashed line separates the calibration cases (left) from the blind test cases (right). Finally, considering soot emissions, it is worth noting that this heavy-duty engine utilises a highly-optimised combustion system that produces very little soot at full load, making it remarkably difficult to simulate and predict trends and absolute differences from point to point. Nonetheless, the calibrated Hiroyasu model included in the SRM Engine Suite, using acetylene (C 2 H 2 ) as a precursor in each stochastic parcel, generates very satisfactory values which contain the majority of the trends and can be scaled successfully to account for the magnitude of these differences too, as evidenced in Figure 21. Figure 21: The soot emission measured experimentally (blue lines with points) and predicted by the SRM Engine Suite (red line) for each of the 60 combinations of the sweep parameters. The vertical dashed line separates the calibration cases (left) from the blind test cases (right). 249 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines 3 Conclusions The SRM Engine Suite was used to simulate the C4.4 diesel-fuelled compression ignition engine. It was calibrated automatically via MoDS software using experimental data taken at thirty representative engine load-speed operating points. The calibrated model matches in-cylinder pressure profiles and NO x emissions well, soot emissions to a lesser extent (factor of two). Blind testing of the SRM Engine Suite produces similar results, broadly validating its use at the full load-speed operating window. One exception is NO x emissions at low torque where the errors are significantly larger than for the calibrated points. In a parameter sweep test, where boost pressure, EGR percentage and injection timings are largely perturbed, the physics included in the SRM Engine Suite account well for the underlying phenomena and predicts the aforementioned measures (in-cylinder pressure, NO x and soot) to a high degree of accuracy. Surrogates provide a reasonable approximation to the full SRM Engine Suite results, particularly for the double injection operating points. The worst surrogate performance for NO x and soot is for single injection at mid-to-low speed, where differences are approximately a factor of two. The coupling of SRM Engine Suite and MoDS to perform automatic calibration of a detailed engine model to experimental data, followed by the generation of fastresponse surrogates and associated sensitivities, constitutes a powerful tool for automotive engineers. 4 Acknowledgments This work was supported by the Advanced Propulsion Centre UK in the framework of the APC3 Project 113059 -ASCENT (Advanced Systems for Carbon Emission reduction through New Technology). 5 Definitions/ Abbreviations EGR Exhaust Gas Recirculation HDMR High Dimensional Model Representation IMEP Indicated Mean Effective Pressure MoDS Model Development Suite MPI Message Passing Interface PDF Probability Density Function SRM Stochastic Reactor Model 250 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines References [1] Rask, E. and Sellanu, M., “Simulation-based engine calibration: Tools, techniques and applications”, SAE Technical Paper 2004-01-1264, 2004. [2] Benedett, M. F., Berrone, S. and Scialo, S., “Efficient combustion parameter prediction and performance optimization for a diesel engine with a low throughput combustion model, Energy Conversion and Management , 96, 105-114, 2015. [3] Finesso, R, Spessa, E. and Yang, Y., “Fast estimation oc combustion metrics in DI diesel engines for control-oriented applications”, Energy Conversion and Management , 112, 254-273, 2016 [4] Real-time combustion parameter estimation algorithm for light-duty diesel engines using in-cylinder pressure measurement, Applied Thermal Engineering , 33-43, 2013. [5] Catania, A. E., Finesso, R. and Spessa, E., Predictive zero-dimensional combustion model for DI diesel engine feed-forward control”, Energy Conversion and Management , 52, 3159-3175, 2011. [6] Isermann, R. and Sequenz, H., “Model-based development of combustion engine control and optimal calibration for driving cycles: general procedure and application”, IFAC-PapersOnLine , 49-11, 633-640, 2016. [7] Tan, Q., Divekar, P., Tan, Y., Chen, X., and Zheng, M., “Engine model calibration using extremum seeking”, IFAC-PapersOnLine , 49-11, 730-735, 2016 [8] Sahara, H., Kakuda, Y., Kim, S., Shimo, D., Maruyana, K., Moringa, S., Hashimoto, T., Hayashibara, H., and Tetsuno, M., “Development of model-based combustion control to ensure robustness of emissions, fuel consumption and noise performances in new generation diesel engine, IFAC , 2013. [9] Shi, Y., Ge, H-W., Reitz, R. D., “Computational optimisation of internal combustion engines”, Springer , 2011. [10] Prah, I., Trenc, F. and Katrasnik, T., “Innovative calibration method for system level simulation models of internal combustion engines”, Energies 9 , 708, 2016. [11] Niu, Q. Y., Fan, C., Wang, C., Zhao, Y. W. and Dong, Y. C., “Research on the parameter calibration of the internal combustion engine work process simulation model”, Advanced Material Research 308-310 , 953-961, 2011. [12] Wang, B., Mosbach, S., Schmutzhard, S., Shuai, S., Huang, Y., and Kraft, M., “Modelling soot formation from wall films in a gasoline direct injection engine using a detailed population balance model”, Applied Energy 163 , 154-166, (2016) [13] Etheridge, J. E., Mosbach, S., Kraft, M., Wu, H., and Collings, N., “Modelling soot formation in a DISI engine”, Proceedings of the Combustion Institute 33 , 3159-3167, (2011) [14] Kittelson, D., and Kraft, M., “Particle Formation and Models”, Encyclopedia of Automotive Engineering , 1-23, (2015) [15] Bhave, A., Kraft, M., Montorsi, L., and Mauss, F., “Sources of CO emissions in an HCCI engine: A numerical analysis”, Combustion and Flame 144 , 634-637, (2006) 251 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines [16] Mosbach, S., Braumann, A., Man, P. L.W., Kastner, C. A., Brownbridge, G.P.E., Kraft, M., “Iterative improvement of Bayesian parameter estimates for an engine model by means of experimental design”, Combustion and Flame 159 , 1303- 1313, (2012) [17] Brownbridge, G.P.E., Smallbone, A. J., Phadungsukanan, W., Kraft, M., Johansson, B., “Automated IC engine model development with uncertainty propagation”, SAE Paper , SAE 2011-01-0237 [18] Brownbridge, G.P.E., Mosbach, S., Kraft, M., Smallbone, A. J., Coble, A., Bhave, A., “Automated parameterization of conventional 1D cycle simulation tools with advanced parametric and experimental uncertainty propagation”, IAV DoE Conference , Berlin, Germany, 2011. [19] Smallbone A., Bhave A., Hillman M., Saville A., and McDavid R., “Virtual performance and emissions mapping for diesel engine design optimization”, SAE 2013-01-0308 , 2013. [20] Etheridge, J. E., Mosbach, S., Kraft, M., Wu, H., and Collings, N., Modelling cycle to cycle variations in an SI engine with detailed chemical kinetics, Combustion and Flame 158 , 179-188, (2011) [21] Lauer, T., Heiss, M., Bobicic, N., Holly, W. et al., "A Comprehensive Simulation Approach to Irregular Combustion," SAE Technical Paper 2014-01-1214, 2014, doi: 10.4271/ 2014-01-1214. [22] Morgan, N. M., Smallbone, A. J., Bhave, A., Kraft, M., Cracknell, R., and Kalghatgi, G, “Mapping surrogate gasoline compositions into RON/ MON space”, , Combustion and Flame 157, 1122-1131, (2010). [23] Smallbone, A. J., Bhave A., Coble, A. R., Mosbach, S., Kraft, M., and McDavid, R. M., “Identifying optimal operating points in terms of engineering constraints and regulated emissions in modern diesel engines”, SAE Paper , SAE 2011-01- 1388 [24] Wu, H., Collings, N., Etheridge, J. E., Mosbach, S., and Kraft, M., “Spark ignition to homogeneous charge compression ignition mode transition study: a new modelling approach”, The International Journal of Engine Research 13 , 1-25, (2012) [25] Etheridge, J. E., Mosbach, S., Kraft, M., Wu, H., and Collings, N., “A Detailed Chemistry Simulation of the SI-HCCI Transition”, SAE International Journal of Fuels and Lubricants 3 , (1), 230-240, (2010) [26] Etheridge, J. E., Mosbach, S., Kraft, M., Wu, H., and Collings, N., “A Detailed Chemistry Multi-cycle Simulation of a Gasoline Fueled HCCI Engine Operated with NVO”, SAE International Journal of Fuels and Lubricants 2 , (1), 13-27, (2009) [27] Mosbach, S., Aldawood, A. M., and Kraft, M., “Real-Time Evaluation of a Detailed Chemistry HCCI Engine Model Using a Tabulation Technique”, Combustion Science and Technology 180, (7), 1263-1277, (2008) [28] Cao, L., Bhave, A., Su, H., Mosbach, S. et al., "Influence of Injection Timing and Piston Bowl Geometry on PCCI Combustion and Emissions," SAE Int. J. Engines 2(1): 1019-1033, 2009, doi: 10.4271/ 2009-01-1102. 252 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines [29] Coble, A., Smallbone, A., Bhave, A., Mosbach, S. et al., "Implementing Detailed Chemistry and In-Cylinder Stratification into 0/ 1-D IC Engine Cycle Simulation Tools," SAE Technical Paper 2011-01-0849, 2011, doi: 10.4271/ 2011-01-0849. [30] Bhave, A., Balthasar, M., Kraft, M., and Mauss, Fabian., “Analysis of a natural gas fuelled homogeneous charge compression ignition engine with exhaust gas recirculation using a stochastic reactor model,” The International Journal of Engine Research 5 (1): 93-104, 2004. [31] Bhave, A. and Kraft, M., "Partially stirred reactor model: analytical solutions and numerical convergence study of a PDF/ Monte Carlo Method", SIAM J. Sci. Comput . 25 (1), 1798-1823, 2004. [32] Sobol, I. M., “On the distribution of points in a cube and the approximate evaluation of integrals,” USSR Comp. Math. Math. Phys . 7(4): 86-112, 1967, doi: 10.1016/ 0041-5553(67)90144-9. [33] Sobol, I. M., "On the Systematic Search in a Hypercube," SIAM J. Numer. Anal. 16(5): 790-93, 1967, url: http: / / www.jstor.org/ stable/ 2156633. [34] Hooke, R. and Jeeves, T. A., ““Direct Search” Solution of Numerical and Statistical Problems,” J. ACM , 8(2): 212-229, 1961, doi: 10.1145/ 321062.321069 253 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines 6 Appendix Figures 22 and Figure 23 show in-cylinder pressure profiles for the two sets of 30 operating points used to calibrate and test the model. Figure 22: Pressure profiles for each of the 30 operating points used to calibrate the model. The blue dashed and red solid lines correspond to the experimental data and SRM Engine Suite predictions respectively. 254 6.3 Fast response surrogates and sensitivity analysis based on physico-chemical engine simulation applied to modern compression ignition engines Figure 23: Pressure profiles for the 30 operating points used to blind-test the model. The blue dashed and red solid lines correspond to the experimental data and SRM Engine Suite predictions respectively. 255 7 Big Data II 7.1 The Connected Car and its new possibilities in ECU calibration Lars Hagen, Andreas Walter, Michael Bargende Abstract Today the calibration of Engine Control Units (further called ECU) is often done based on a limited data base, which is the result of few test vehicles and a small amount of measurement data. This leads to a tailored calibration for one/ less vehicles, which cannot stand for the complete entity of vehicles caused by manufacturing tolerances, usage or lifetime. Herein, the Connected Car has its advantages since it is able to transmit vehicle data at any time in any place (assuming mobile net is available). Therein the amount of vehicles is a less limiting factor. However, there are limits for the mobile net in respect to the data transfer rate. Due to that a suitable selection of the measurement data has to be done to achieve an either vehicleindividual or vehicle-fleet optimal parameterization. This paper presents an algorithm that collects data from several test vehicles, filters it based on its information content for a special calibration task and transmits the collected data afterwards to a server/ cloud. Here the collected data of all vehicles is used together with a suitable optimizer to create an improved dataset. Right afterwards this dataset can be reloaded to the vehicle fleet. So, it is possible to create a master-dataset, a car-individual calibration or even special vehicle clusters in dependency to the use case. Exemplarily this algorithm is demonstrated offline for a physical model, which calculates the pressure in the exhaust manifold. Kurzfassung Die Kalibrierung von Motorsteuergeräten geschieht heutzutage oftmals anhand einer sehr kleinen Datenbasis, die sich durch wenige Versuchsträger und eine begrenzte Anzahl an Messdaten ergeben. Hierdurch wird die Applikation auf einen/ wenige Versuchsträger zugeschnitten, was jedoch aufgrund von Fertigungstoleranzen, Einsatzgebiet und Lebensdauer nicht der Gesamtheit aller Fahrzeuge optimal entsprechen kann. Das vernetzte Fahrzeug (auch Connected Car genannt) spielt hier seine Vorteile aus, da zu jeder Zeit an jedem Ort Fahrzeugdaten übertragen werden können (sofern die Mobilfunkkonnektivität gegeben ist). Und dies über eine große Fahrzeuganzahl hinweg. Dennoch sind bzgl. Datentransferrate Limitierungen für den Mobilfunk gesetzt, wodurch eine geeignete Messdatenauswahl am Fahrzeug erfolgen muss, um eine optimale Parametrierung entweder fahrzeugindividuell, oder für die gesamte Fahrzeugflotte, zu ermöglichen. 256 7.1 The Connected Car and its new possibilities in ECU calibration Der in diesem Paper vorgestellte Algorithmus sieht vor, von verschiedenen Versuchsträgern Daten zu sammeln, nach deren Informationsgehalt für eine spezielle Applikationsaufgabe zu filtern und anschließend auf einen Server zu übertragen. Hier werden die gesammelten Daten zur Applikation mit einem geeigneten Optimierer eingesetzt, um einen verbesserten Datenstand zu erzeugen. Dieser kann nun auf die Fahrzeugflotte zurück übertragen werden. Ob ein Masterdatenstand erzeugt wird, eine fahrzeugindividuelle Applikation oder möglicherweise spezielle Fahrzeugcluster gebildet werden, hängt von dem einzelnen Anwendungsfall ab. Exemplarisch wurde der Algorithmus offline an einem Modell für den Druck im Auslasskrümmer getestet. 1 Introduction Currently the automotive industry has to cope with many changes and challenges respectively. It can be observed that the product lifecycle is getting shorter and shorter. Moreover the market demands more and more model variants and derivatives. When analyzing the car manufacturer’s product portfolio over the last decades it is obvious that there was a continuous growth. At the same time the customer’s requirements increase in regard to fuel consumption, performance and comfort. In parallel the desire for new vehicle powertrains and concepts has increased strongly so that the pressure on the conventional combustion engine with respect to fuel consumption is rising. According to [2] the introduction of the electronics has made vehicles to the most complex consumer good. This can be observed in the multiple times increase of engine control functions in the last years. Beside the mentioned requirements above it is mandatory to keep an eye on the quality and the costs to persist on the highly competitive automotive market. The introduction of new development processes and methods seem to be unavoidable and can bring crucial market benefits. Especially the keyword “Internet of Things” (means the connection of daily things with the internet) is yielding lots of new possibilities during the product creation process of the vehicle. It is important that these possibilities are identified and exploited beneficially. Especially the calibration of ECU show a high optimization potential. Caused by the before mentioned boundary conditions, the complexity and variance of ECU- Software has increased that drastical that the manual offline calibration process cannot be done efficiently anymore. By the use of the Connected Car it is possible to automate this development step. Moreover, this technology has the potential to tailor the calibration of ECU functionalities for single vehicles or vehicle groups (e.g. influenced by driving behavior or vehicle location). That helps to improve the calibration quality and can further support to reach the car manufacturer’s goals. 2 Modelling in engine control units The combustion engine is a mechatronic system, since it is controlled by an ECU. The used electronics and software enable the possibility to make the engine more efficient and more powerful continuously. Nevertheless, at the same time the amount and complexity of the control unit’s functions has strongly increased. From 1995 to 2005 the amount of calibration parameters has been raised four times from 1800 to 7200 [5]. This is driven by the necessity of applying the Software to specific engine projects and to fulfill the requirements. According to [4] nowadays the calibration of 257 7.1 The Connected Car and its new possibilities in ECU calibration ECU functions is strongly connected to the usage of highly developed calibration methods so that the effort for the parameterization of engine control unit software can be handled in spite of the complexity of its functionalities. 2.1 The Diesel Engine and its airsystem The Diesel engine was invented at the end of the 19 th century by Rudolf Diesel and is characterized in the classical meaning by the compression of air, fuel injection into the cylinder (internal mixture formation), self-ignition and quality control [6]. The compression ratio ε for diesel engines lies in the range of 16-24. This is higher than in gasoline engines and leads in principal to a better thermal efficiency [1]. During the 125 years of its application the Diesel engine was continuously developed further. Important milestones have been the introduction of the exhaust gas recirculation (EGR) technology for reduction of the nitrogen oxides and the (exhaust turbo-) charger technology to increase the engine power. In fig.1 an actual Diesel airand exhaust system with one-stage turbocharger, highand low pressure exhaust gas recirculation and Diesel particulate filter is shown. Figure 1: The airand exhaust system of an actual passenger car 2.1.1 The exhaust back pressure model The pressure upstream the turbine and in the exhaust manifold respectively is needed for many engine functions in the model-based control. Examples are the boost-, the camshaft or the exhaust gas recirculation-control (EGR-control) that require the thermodynamical state in the exhaust manifold to have the actual and/ or the desired value calculated. A mathematical model can substitute a pressure sensor and therefore contribute to save cost for components. In the control unit-Software-development mathematical models can be distinguished in two main groups: dataand modelbased models. In case of modelbased development for the airsystem often well-known equations from the area of thermoor fluid dynamics are suitable. For a pressure calculation in the exhaust manifold, it has been pointed out that the thermodynamic state change of the gas mixture over the turbine can be used well for sys- 258 7.1 The Connected Car and its new possibilities in ECU calibration tems with exhaustgas-turbocharger with variable turbine geometry (VTG-charger). Based on the Bernoulli equation in energy form with the extension of density and temperature change for an ideal gas, the following expression can be found: . 2 2 const T c H g w p v = + ⋅ + + ρ (1) Furthermore, by applying the continuity and isentropic equation, the throttle equation can be derived. ⎟⎟⎠ ⎞ ⎜⎜⎝ ⎛ ⋅ ⋅ ⋅ = in out in eff p p RT p A m ψ 2 & (2) The equation describes the relation between the massflow over the turbine and the VTG-Position, which directly has an influence to the effective area A eff , the temperature in the exhaust manifold T3 and the pressures upstream and downstream the turbine (p3 and p4). The ψ function is called flow function. It has been found out that an isentropic exponent κ of infinity, leads to good results regarding the state change of gas over the turbine. This value is unphysical, but it simplifies the solving of the equation on the ECU. 2 2 2 3 2 4 4 3 eff A T R m p p p ⋅ ⋅ ⋅ + + = & (3) If the input values pressure downstream turbine, massflow over turbine, temperature in the exhaust manifold are taken as granted, then only dependency of the model quality is the calibration of the effective turbine area. This effective area is mainly dependent on the VTG-position and the turbine massflow, because those have a direct influence on the flow formation and flow cross section. In the software this can be realized via a 16x16 map with the input values VTG-position and turbine massflow. For calibration of this area map, measurements are often made at the so called ”golden engine” on the test bench in stationary state. Here, variations of the load over the complete operating range of the engine are done. With the help of a suitable optimizer this data can now be used to generate the area map right afterwards. Since the throttle equation does not consider any inertiaor dynamic effects, the dynamic behavior can be changed via a lowpass filter. This measurement can be done also at the test bench or in real driving mode, following strongly the “Try and Error” principle. In fig.2 an area map for the exhaust backpressure model of a 1.67 liter 4 cylinder Diesel engine is shown which has been created at the test bench. 259 7.1 The Connected Car and its new possibilities in ECU calibration Figure 2: Base area map of the exhaust backpressure model But especially in this example the problem of nowadays used calibration processes can be observed. Is the calibration at the “golden engine” really valid for the complete engine generation which might be sold a million of times? The calibration results are transferred to other modeltypes and are maybe slightly changed in the variant application. But it is obvious that the “golden engine” cannot be identical to each manufactured engine of its family and amongst others part tolerances, vehicle to vehicle spreadings, sensor tolerances are further not taken into account either. 3 The Connected Car and its new possibilities for the ECU The Connected Car describes the connection of the vehicle with the mobile net. The usage of the data channel to external sources is already heavily in use for the infotainment domain. For instance the navigation system informs about traffic jam warnings, the emergency assistant calls in case of an accident the ambulance or in upper class vehicles it is possible to stream multimedia contents like movies or music onto the vehicle display. In the domain of the engine control unit development and in series vehicles as well, new unused application possibilities are offered, see fig.3. Figure 3: Application possibilities between the Connected Car and the control unit 260 7.1 The Connected Car and its new possibilities in ECU calibration A very high potential has the field data acquisition, since the Connected Car is permanently able to measure vehicle data. This is interesting for many development areas. For the topic diagnosis it could be possible now to do the workstation diagnosis via remote access, and give a direct feedback to the vehicle driver. In times of less test vehicles it could be beneficial to test and validate functions remotely, e.g. from the desk of the engineer. Moreover it is possible to do wireless software updates which offers advantages in the development as well as in series. This paper works out a method for partial automated calibration with the Connected Car as use-case. Based upon that the development efficiency and the calibration quality can be improved. 4 Partial automated calibration with the Connected Car The below given presented algorithm for partial automated calibration can be split into two main parts: Collection of measured data from the vehicle and the optimization of the model parameters on the server side, see fig.4. Generally three stages can be distinguished here: test-bench, vehicle and server/ cloud. Nowadays the testbench stage represents the most used or rather standard application process in these days for a high amount of engine control unit functions. Based on a grid measurement wherein in equidistant intervals the engine turning speed and load is varied it is possible to generate a base calibration with a suitable optimizer. For the described example of the exhaust backpressure model, the base area map can be derived (cf. fig. 2). Figure 4: Concept for partial automated calibration at the Connected Car 261 7.1 The Connected Car and its new possibilities in ECU calibration This calibration now can be transferred to the connected vehicle fleet (now stage vehicle), the so-called flashing of the control unit. By comparison of the modeled and reference value (here the sensor value) a first assessment of the calibration quality can be done. If the calibration does not meet the requirements or in case of having some improvement potential left, new measurement data is necessary to be collected. Having these data, the existing calibration of the vehicle can be adapted specifically. For a relief of the cost intensive and further limited mobile net data channel the so-called measurement assessment function based on a sensitivity matrix (further called SeM) is used. This evaluates the excellence of a measured section based upon the parameter sensitivities, so that it can be assessed for generation of a new calibration dataset. That kind of processing ensures that only the relevant or useful measurement data of the vehicle(s) is collected, which is beneficial for the calibration task. This data will be send to the server afterwards. Now, on the server stage an optimizer builds a new dataset which fits to the new available measurement data. The optimized dataset and actual SeM are transmitted back to the vehicle(s). Because of the back transferred SeM the algorithm gets a ”mind” so that every participated vehicle in the calibration process knows, which measurement data is already available on the server stage. Thereby the transmitted measurement data gets reduced continuously. At the same time the calibration is improved until an abortion criteria is reached and the found calibration dataset can be viewed as satisfactorily. 4.1 Acquisition of measurement data The overall goal of measurement data collection is to gather the highest possible information content while having as less data as needed transmitted at the same time. Or in other words, for calibration improvement purposes only the needed data shall be collected. It has to be pointed out that the measurement effort is not decreased in that case. But the amount of data transferred through the mobile net channel can be significantly reduced. This pre-filtering of the measurement data is done by the sensitivity matrix (SeM) which requires the parameter sensitivities ∂y ∂p i . In this term y is the model output value and p i are the model parameters. According to this, the sensitivities can be derived by differentiation of the model equations with the wanted parameters. This can be done numerically or algebraically. It should be noted that a SeM can be built additively so that measurement windows with low information content can be deleted. Based upon that the ”collected knowledge” in form of the SeM can be transmitted to the vehicles without transmitting the complete measurement data. In the literature approaches can be found, pointing out relations between a sensitivity matrix of a model and its parameter estimation error. This will not be presented at this point. For an assessment of a measurement during the drive it is necessary to define measurement windows of a certain time, e.g. 50 seconds. For every measurement window the reduction of the parameter estimation error is calculated. If the parameter estimation error falls for a certain percentage, then the measurement window can be classified as relevant. In this case the measurement windows will be transmitted via the mobile net to the server. Otherwise it will be deleted on the vehicle and not transmitted. 262 7.1 The Connected Car and its new possibilities in ECU calibration 4.2 Optimization of model parameters The optimization of the model parameters is done on the server stage. As a base provided from one or more vehicles, the collected and as relevant classified data can be used. Henceforth these can be optimized together or if needed separately and vehicle individual. The used optimizer depends on the calibration task and has several impacts. Mainly the choice is influenced by the number of optimization parameters, the amount of measurement data and the boundary conditions of the optimization problem (e.g. are there many local minima ? , how much time is available? ). 4.3 Return of the SeM and model parameters to the vehicle fleet The base-SeM and optimized model parameters are returned to the vehicle fleet, after having the parameters optimized. Hereby the quality of the model calibration is improved continuously. Furthermore, the vehicles are informed about the available measurement data by the SeM. The big advantage of that proceeding is that the notation of the available measurement data can be made by one single matrix. The algorithm runs until a defined abortion criteria is met. 5 Results of the offline prototype for the calibration of the backpressure model In the target environment the ECU is connected to the so-called Connectivity Control Unit (CCU) via CAN-Bus. On this mobile-capable and in comparison to the ECU powerful gateway, additional calculations can be done. Moreover it can transmit or receive data from the cloud/ the server, see fig.5. The proof of functional capability for partial automated calibration algorithm of the exhaust backpressure model is done offline via rapid prototyping. Therefore the CCU is substituted by a rapid prototyping module from the company ETAS. On the one hand that enables to do additional calculations, like solving the measurement assessment function, but on the other hand it is not mobile capable. The cloud and accordingly the server is replaced by a standard PC with sufficient computing power, which is connected via Ethernet with the RPmodule. This handles the parameter optimizations and provides the saved SeM to the vehicle. 263 7.1 The Connected Car and its new possibilities in ECU calibration Figure 5: Networking of the Connected Car online Figure 6: Networking of the Connected Car as offline prototype For a validation of the described algorithm, a 30km round track starting at the Bosch plant in Schwieberdingen is defined (see fig.7). The test track is driven three times so that the back coupling of the optimization result and the saved SeM from the PC can be tested. 264 7.1 The Connected Car and its new possibilities in ECU calibration Figure 7: Test track round about Stuttgart (Source: Google Maps) Since a 16x16-map like the effective area map of the exhaust backpressure model has in sum 256 parameters and accordingly optimization parameters the sensitivity matrix has a dimension of 256x256. The computation of this is very inefficient, because the measurement assessment function has to run as fast as the exhaust backpressure model on the ECU in a 20ms time raster. To avoid this, the effective area is approximated by a polynomial of third order with 9 parameters (see eq. 4). The input values turbine massflow and VTG-position are same as in the area implementation. After optimization the entries of the 16x16-map can be determined by insertion of the supporting points into the polynomial equation via back calculation. 3 9 2 8 2 7 2 6 5 2 4 3 2 1 Trb Trb VTG Trb VTG Trb Trb VTG VTG Trb VTG eff m m r m r m m r r m r A & & & & & & ⋅ + ⋅ ⋅ + ⋅ ⋅ + ⋅ + ⋅ ⋅ + ⋅ + ⋅ + ⋅ + = θ θ θ θ θ θ θ θ θ (4) Fig. 8 shows the plot of the exhaust backpressure of the reference value (sensor) and the corresponding model with the initial dataset. The white areas represent the measurement windows, which have been sorted out by the measurement assessment function, because of its information content. To classify a measurement window as relevant, its parameter estimation error has to sink for at least 5 percent. The length of a measurement window is here defined with 50 seconds each. With this approach the measurement data has been reduced strongly and this leads on the target hardware to a reduction of the mobile net load. The calculated SeM as well as the relevant measurement segments are stored on the PC after the first test run. 265 7.1 The Connected Car and its new possibilities in ECU calibration Figure 8: Relevant measurement segments of test run 1 Using the reduced measurement data an optimizer on the PC can be operated. In this example the realization of the optimizer is done by the Curve-Fitting-Toolbox of the Mathworks Matlab. This fits the polynomial function in the best way to the data points found so that the mean value and the standard deviation of the model error is reduced (see Table 1). Table 1: Model error test run 1 Mean value of the model error [hPa] Standard deviation of the model error [hPa] Initial dataset -32,91 93,01 Optimized dataset -11,43 78,964 At the beginning of the second test run the SeM is transferred back from the PC to the RP-module. In dependency to the calibration strategy it is possible to transmit the SeM also to identically constructed vehicles to generate a master dataset for multiple vehicles. Hence it can be avoided that several vehicles collect similar informative data. In fig. 9 the measurement plot of the exhaust backpressure for the second test run is demonstrated. It is visible that the data amount can be reduced vastly. Figure 9: Relevant measurement segments test run 2 266 7.1 The Connected Car and its new possibilities in ECU calibration In Table 2 the comparison of the model error between the initial and optimized dataset for the first two test runs is shown. Again much data has been eliminated, but the mean value (about 28hPa) as well as the standard deviation (round 13hPa) of the model error have been improved anyway. Table 2: Model error test run 1+2 Mean value of the model error [hPa] Standard deviation of the model error [hPa] Initial dataset -41,95 99,78 Optimized dataset -14,08 86,59 For the third test run the measurement plot of the relevant measurement segments is not shown anymore, since the measurement assessment function has not found any informative measurement windows anymore. Table 3: Model error test run 1+2+3 Mean value of the model error [hPa] Standard deviation of the model error [hPa] Initial dataset -40,6 99,95 Optimized dataset (red. measurement data) -13,78 88,4 Optimized dataset (complete measurement data) -11,55 88,3 In Table 3 the result of all three test runs is represented. The data amount can be reduced strongly by the measurement assessment function and the model error has been reduced in comparison to the initial dataset significantly. Interesting is that in case all measurements of the test runs 1 - 3 are used for optimization, the benefit of the model error is neglectable in comparison to the reduced data. The standard deviation is reduced for only 0.1 hPa and the mean value for about 2.2hPa. Exemplarily in fig. 10 a short excerpt of the first test run for the exhaust back pressure of the sensor, the initial dataset as well as the optimized dataset (based on reduced measurement data) is shown. It is obvious that in nearly all areas the deviation between modeled and reference value has been reduced because of the new parameterization. 267 7.1 The Connected Car and its new possibilities in ECU calibration Figure 10: Excerpt pressure plot test run 1 6 Conclusion The Connected Car offers many possibilities, either in the development phase as well as in series applications. It can help to make further improvements of the conventional combustion engine with respect to fuel consumption, emissions and performance. But the development process for future powertrain concepts or in other vehicle domains can be done more efficiently, too. Thus the costs and the duration of the development process can be reduced. This publication has the goal to point out these possibilities shortly and to demonstrate it as an example application. A very high potential for optimization can be found in the calibration process. This results due to the constant increase of vehicle models and derivatives as well as the more and more complex control unit software and variance. On the one hand the procedure for preparation of a calibration can be simplified essentially by partial automation, but on the other hand the possibility to make a calibration tailored to single vehicles or vehicle groups is given. Hereby the mentioned advantages of the Connected Car can be reached. The example for calibration of an exhaust backpressure model can confirm this. First an algorithm based on the sensitivity matrix for reduction of measurement data is presented. That solution reduces the load of the limited data channel of the mobile net significantly. It can be shown that by usage of parts of the measurement data comparable calibration results can be reached than by usage of the complete data. Further it is possible to use the algorithm for generation of a complete new calibration in the development process, an adaption of an initial dataset to a vehicle derivative or for adaptation of the calibration in series applications to single vehicles. 268 7.1 The Connected Car and its new possibilities in ECU calibration Acknowledgement I’m very thankful for the cooperation with Dr. Christian Potthast, Dr. Fabian Jarmolowitz, Dr. Kosmas Petridis, Dr. Matthias Bitzer and Dr. Thomas Bleile for joint development of the presented algorithm. Literatur [1] Braess, H.-H. and Seifert, U., Vieweg Handbuch Kraftfahrzeugtechnik, 7 th edition, Springer Vieweg Verlag, 2013 [2] Schäuffele, J. and Zurawka, T., Automotive Software Engineering, 5 th edition, Springer Vieweg Verlag, 2013 [3] Schmidt, C. and Isermann, R., Dieselmotorenregelung: Vorhaben Nr. 407, Modellbildung, Simulation und Identifikation des dynamischen Verhaltens von Dieselmotoren ; final report, FVV Verlag, 1992 [4] Isermann, R., Elektronisches Management motorischer Fahrzeugantriebe, Vieweg Teubner Verlag, 2010 [5] Berns, S., Fehrmann, Sassen, K., Beitrag der Diesel-Elektronik zur Leistungssteigerung und Emissionsreduzierung. VDI/ VDE-Fachtagung AUTOREG, 2004, Wiesloch 2./ 3.3.04, VDI-Berichte 1828, VDI Verlag, 2004 [6] Pischinger, S., Verbrennungskraftmaschinen II, Lecture notes, 28th edition, Selbstverlag, 2011 269 7.2 Processing vehicle-related measurement data Leon Evgeni Kusnezow, Ortwin Escher Abstract Handling vehicle-based data is seen as a huge opportunity with millions of devices roaming the streets. With our longstanding expertise in the field of testing vehicles, we will share our unique perspective on big vehicle data. The testing of vehicles requires the gathering and processing of multiple gigabytes per vehicle and hour in order to detect and analyze problems. Creating a balance between the sheer size of raw data with thousands of vehicles and the cost efficient use of hardware for data analytics has resulted in a specialized data gathering, storing, processing, distributing and connecting chain. After sketching the here and now of data analytics in the vehicle test process, we will create an outlook on where we can go from here and what has to be done to reach that goal. Kurzfassung Fahrzeuge werden zunehmend zu wichtigen Datenquellen für mobile Dienstleistungen. Mit unserer langjährigen Erfahrung im Bereich von Fahrzeug- und Steuergerätetests bieten wir eine einmalige Perspektive auf Big Data im Fahrzeugkontext. Im Rahmen von Fahrzeug- und Steuergerätetests fallen viele Gigabyte an Daten pro Fahrzeug und Stunde an. Anhand dieser Daten können Probleme frühzeitig erkannt und behoben werden. Die Rohdatenmenge von tausenden von Fahrzeugen stellt mit ihren Eigenschaften eine Herausforderung für etablierte Big Data Tools dar. Daher haben wir Spezialtools für die Verarbeitungskette von Sammlung, Speicherung, Verarbeitung, Verteilung, Verknüpfung und Visualisierung geschaffen, die eine kosteneffiziente Hardwarenutzung und vielfältige Einsatzmöglichkeiten erlaubt. Neben dem aktuellen Stand der fahrzeugbasierten Datenanalyse schaffen wir einen Ausblick auf die nächsten Schritte und Ideen, wie diese erreicht werden können. 270 7.2 Processing vehicle-related measurement data 1 Use cases and data sources The collection of measurement data from vehicles has a wide range of use cases within each lifecycle stage of a car. These span from the development of single control units or component groups over field application engineering to fully-fledged vehicle fleet validation tests and even end customer vehicles. The most prominent usages are the detection of errors, improvement of components and clearance tests. The collected data is also useable for secondary projects that can lead to the development of new customer facing products. It is a compelling testbed for emerging data driven projects. These can be prototypes for features like usage statistics for insurances, wear down prognosis with pre-emptive maintenance or using the vehicles as moving distributed sensors for cloud based services beyond pure car-to-car communication. Examples for such services are traffic analysis, weather prognosis, park space search, road quality checks and many more. A new generation of small OBD units like the PACE logger empower users with a pure IT background and little vehicle knowledge to implement innovative usages of automotive data. Loggers connected directly to the different bus systems provide more opportunities for innovative features and result in scientific research using machine learning and pattern recognition. 2 Data collection Most data loggers connect to control units or bus systems like CAN, FlexRay, Ethernet/ XETK, ETK, LIN or K-Line. They collect either signal or message based data. Synchronizing audio and video recordings with the normal data collection is gaining popularity. This is especially true for the areas of assisted and automated driving tests. We support over 30 different logger types of several vendors in various projects. Many of these use custom data formats and there is no desire on the OEM side to migrate to newer models with standardized formats. Many of these formats will start to disappear and will be replaced by formats like ASAM MDF 4.1 which is gaining a lot of traction in its market share. Usual sampling rates for signal based data range from 1 Hz up to 400 kHz for timing critical information. An alternative means to collect data is a segment based approach with variable sampling rates in relation to specific values, e.g. rpm. 271 7.2 Processing vehicle-related measurement data 3 Transmission The use case defines the size of the data. This can limit the ways to transmit it to a central system. Long-term trials can create data volumes of over one hundred gigabytes per day and vehicle. A data transmission using cellular technology is currently not an option when handling these kinds of volume. Other use cases like the monitoring of VIP vehicles require cellular transmissions. A complete collection of vehicular data is not required in these cases. Figure 1: Incoming data of a single OEM in GB (compressed) Control unit development can profit from the low bandwidth required for its tests. Mobile networks permit remote interaction with the device and can therefore create new and cost efficient test setups. A research and development employee can keep in contact with multiple drivers in different climate zones and collect feedback after changing unit parameters without leaving his office. Figure 2: Live prototype monitoring 4000 5000 6000 7000 8000 9000 10000 272 7.2 Processing vehicle-related measurement data Figure 3: Direct control unit access from the engineers’ office The limiting factors for the examples shown in figure 2 and 3 is the available bandwidth and network stability. Defining strict road safety guidelines for modifications of active vehicles is a necessity in these scenarios. 4 Management and storage Small test setups handle them self without the need for large management toolkits. Organizing and scheduling large-scale operations with thousands of vehicles and loggers, custom configurations with scheduled configuration rollouts and automatic data analysis is the other side of the complexity scale. Figure 4: Managing automated large-scale fleet validation with IAV MeDaPF 273 7.2 Processing vehicle-related measurement data Measurement files are the product of a vehicle or test bench paired with a data logger. They require a configuration defining what and how to record in order to match the test case requirements. The results are useable for in depth analysis with interactive data drill downs or automated checking for known issues, patterns or events. This requires an analysis client with a matching analysis configuration. This is just a small example of the moving parts within a measurement data management system. Orchestrating distributed processes with fine-grained user access control down to a single measurement or evaluation increases the required logistics even more. The growing number of loggers and archival times of up to ten years for certain OEMs increases the storage demands. NAS or SAN based storages were not able to handle the load in terms of size and throughput with petabytes of data. While US customers readily rely on available cloud services, European and Asian OEMs distrust these due to secrecy concerns. We therefore had to create a custom private hosting for company clouds with an object storage system to match the demand of these OEMs. 5 Analysis The growing number of electronic devices within a car increase the complexity of test setups. Automated tests with an automated analysis of the resulting data is necessary for these systems. With a new device generation or new governmental guidelines, a re-validation of all existing functionality and behavior is required. Some governmental agencies require such re-validation with each new generation of firmware. Advanced data analysis is therefore inevitable and will reduce the cost to do so. 5.1 Generic vs Specific Engineers require assistance of software developers and trained data analysts during the development of control units and the vehicle application phase to create customized algorithms to assist the development and acceptance testing. The software developer or analyst has to gain a comprehensive understanding of the behavior of the device in order to assist the engineer in creating a suitable testing environment for the specific device. Creating a customizable and generic piece of analytics software that will handle most usages is a daunting task and increases the amount of knowledge required to operate it efficiently. The complexities involved in the development of generic solutions creates an environment where various parts of a company create specific solutions that differ in minor ways from each other. This has resulted in test vehicles of OEMs with multiple data loggers collecting the same information for separate departments. Generic toolkits help to prevent these situations by providing a common ground to work on. They offer a wide range of combinable analytical methods and interfaces for new tools to solve the task. 274 7.2 Processing vehicle-related measurement data Whether to use such generic tool or a specialized solution requires an elevated viewpoint on existing technologies. Identifying the right tool for the right job is a challenge in itself. Modern organizations have trained data analysts who support the engineer in achieving his tasks. They know the possible types of analytic software and help to set up generic systems within the engineers’ workplace. Figure 5 highlights a generic tool for device engineers that is configurable without programming knowledge. The user customizes the conditions and methods for his task and generates a result report for each measurement. Figure 5: Analysis process with IAV Mara 5.2 Manual vs Automated Users desire to automatize reoccurring workflows. Reaching a fully automated workflow requires several steps in advance. Not everything is easily automatable without implementing a custom piece of logic within a pre-existing software that contains the expertise of the user and can decide how to proceed or what an important bit of information is. Therefore, a manual intervention within specific workflow steps may be required. The provided workflow has to be split into parts within an analysis software until the manual steps can be automatized as well. Figure 6 contains an example of such a workflow. 275 7.2 Processing vehicle-related measurement data Figure 6: Concept of the software IncaFlow Fully automated processes do not require any human interactions to run. Transferring parts of the knowledge and expertise of the engineer into a piece of software will increase his productivity and reduces the chance of mistakes. It is an investment with a long-term profit for common tasks. Later stages of the vehicle development process can utilize the automatic validation. The checks can keep on running for loggers that return data after the control unit development was completed. The combination of a matching configuration with other mandatory information like project, model or environmental parameters can trigger the automated workflow. This can alert about undesired states of the control unit and save costs when investigating failures or uncommon behaviours. 5.3 Laptop vs Server A laptop or tablet within a prototype is one of the most important tools of an engineer in the development phase. They are required for interactions with on board devices when remote interfaces are not available. On-site hardware allows direct access to measurements directly after a test run. Server or cloud hardware performs better but requires a fast transmission of the data to the server. This is not possible in every testing environment. We employ several different approaches in order to minimize the time an engineer has to wait on his results and to provide the right tool for the right job. An analysis client on a laptop is the obvious solution. To allow other engineers to gather and analyse data and reduce the desire to delete files due to space concerns we offer small distributed data centres. They get the data through Wi-Fi, SD cards or SSD disks and store it on a local NAS. The metadata of the measurement is transferred to the cloud-based application that determines which kind of automated analytics have to be run. The travel case sized appliance does the number crunching and the engineer can retrieve his results locally while it transfers the rest of the results into the cloud for different engineers to continue their work on. Another approach is a local filtering of data. The cloud receives a minimal set of data required for the immediate test case and the complete files later when a sufficient 276 7.2 Processing vehicle-related measurement data transfer bandwidth is available. This allows real time analytics and in depth analytics later on. It is used to detect test protocol deviations that allow an early test restart in order to save valuable time. These tools provide a live view on vehicles in the field. We use these technologies for VIP vehicle monitoring with automated safety diagnostics and maintenance checks as well. Figure 7: Example for Live Data Visualization with IAV MeDaPf 5.4 Live vs Real Time There are two separate definitions of real time analytics. One is the analytical processing of data during its creation and the other is an interactive drill down of collected data. We will focus on the former one. The basis of real time analytics is the processing of gapless data with minimal delays. The processing location of real time analytics differs depending on the use case. Requiring a large set of source data implies the calculations to run in the vehicle. A large output moves the visualisation of the output into the vehicle as well. A small data source enables remote processing. Table 1: Combinations of local and remote analytics and visualisation. Case Logger location Analytics location Visualisation location 1 Moving vehicle Vehicle Vehicle 2 Moving vehicle Vehicle Cloud 3 Moving vehicle Cloud Cloud 277 7.2 Processing vehicle-related measurement data The user has to be within the vehicle in the first case. Low latencies in this kind of setup is beneficial for control unit development. This setup does not scale well with a growing number of vehicles due to the required work force. It requires a complex static setup. Moving it from one vehicle to another is time consuming as well. The second case moves the visualisation of the results into the cloud. The data is reduced to the necessary bits within the vehicle. Depending on the used loggers, additional hardware investments are required to handle this kind of setup. The third case is the most desired setup by most OEMs. The limiting factor is the bandwidth between the logger and the cloud. The analysis will be less and less realtime as soon as the generated data gets larger than the bandwidth. A data minimization by reducing the sampling rate or reducing the number of transferred signals in addition to compression is the key to success. Live data on the other hand is a reduced set of values that can contain gaps due to cellular dead spots. Live data is a monitoring tool for vehicles with a low latency of less than two seconds. The logger sends its data directly to the cloud and any consumer device with a browser can retrieve in-depth information about all active vehicles. This data is useable for some small-scale analytics as well, e.g. test protocol matching. However, the results do not represent a final state. 5.5 Predictive Analytics Predictive analytics is the evaluation of existing data. The analysis uses found patterns to predict the probability of future events based on the current state. This is measured in percent like other statistical data, e.g. there is a 90% chance that the acceptable temperature range will be exceeded when the air intake flap is 70% closed. Figure 8: Predictive analytics workflow 278 7.2 Processing vehicle-related measurement data Figure 8 highlights an overview of the workflow process. The first step is the selection of the type of prediction. The data extraction filters the required data based on the prediction type. A too wide data set causes a prediction overfitting. An overfitted prediction results in prediction errors due to the consideration of irrelevant data. An algorithm processes the extracted data. The algorithm was trained by past data sets. Some algorithm types are: • Minimal accuracy: The prediction can remain vague as long as a general tendency exists. The performance increases with a lower accuracy. • Data set driven: A low count of parameters requires simple algorithms. An increasing parameter count allows algorithms that are more complex. Complex algorithms usually increase the prediction precision. • Descriptive algorithms: The desired prediction requires a manual assumption about the root cause. Few algorithms can handle this kind of prediction. • Custom algorithms: Implementing custom algorithms is a complicated task. Existing libraries can ease the development. The quality of an algorithm is determined by existing data sets where the outcome is known. This sample size is called training data. Two kinds of training data exists: One to train the algorithm and the other to evaluate the effectiveness of the algorithm. Predictions are one kind of result that can help an engineer. The other result is the detection of unapparent dependencies between parameters and weighting parameters based on a seen outcome. Examples: - Calculating information about the driving style and environment based on RDE measurements. - Causal research on exceeding thresholds. - Trendspotting with miniscule long-term changes. - Outlier analysis on measured spikes. - Predictive maintenance. 279 7.2 Processing vehicle-related measurement data 6 Visualization and Reporting 6.1 Live and Real Time Visualization Browsers are the target environment of any modern interactive visualisation. They offer a platform independent access to data and functionality. Live data and real-time analytic views offer a glimpse into the future of browser based vehicle analytics. Figure 7 shows a live view with a backchannel into the logger. Settings like the sampling rate of the data are modifiable through the page. The shown data has a latency of less than two seconds. Figure 9: Real-time reporting visualization with IAV MeDaPf Figure 9 shows a more advanced real-time analysis and reporting prototype based on live data. A time slider allows access to past states. A playback feature visualizes existing measurements in full resolution without missing packets due to dead spots. 6.2 Interactive Visualization Another kind of reports is the interactive drill-down. It is a combination of reporting with real time analytics. We created showcases using several different big data toolkits. The most promising was Vertica in combination with Tableu in terms of performance and ease of use. Licensing costs are a showstopper for most OEMs and free tools like Kibana or Zeppelin have scaling issues with the unique complexities of a complete set of vehicular data. An in-depth look at the findings would be beyond the scope of this paper. 280 7.2 Processing vehicle-related measurement data 6.3 Report generation The report generation process differs depending on the used analysis tool and report type. Analysis tool can create basic reports. Complex reports spanning multiple measurements, a combination of data sources or a collection of vehicles require a different kind of tooling. A configurable analysis output pipeline can export the analysis results into various SQL, document based/ NoSQL or big data toolkits. From this combined storage of a selection of different analysis and inputs, either interactive reports or custom static reports can be generated. Figure 10 illustrates this architecture. Figure 10: Multi-tier architecture for report generation The result file type depends on the report implementation, e.g. PDF, PPTX, XLSX, HTML, CSV, JSON or XML. Additional customizations can include company logos and colours for corporate identity purposes. 7 Conclusion We have highlighted several ways to handle vehicle-related measurement data. The processing steps are the same in each use case but the scope and scaling is not. The use case defines what is recorded, where it is analysed, when it has to be transferred to other systems and how and where a result is shown. A data analyst helps with the comprehension of these scales and support engineers with methodology advice and a workplace, cloud and mobile setup that matches the requirements of the use case. New features and sensors increase the required collection of data. The cloud is the ideal target environment for the processing of large data volume. The main challenges for cloud usage are secrecy concerns, total cost of hardand software and 281 7.2 Processing vehicle-related measurement data mobile networks with their bandwidth limitation, blind spots and volume pricing. Each can be remedied by solutions that result in other trade-offs. Processing vehicle data for the control unit development has its focus on dynamic workflows in the early stages and fully automated workflows in the later stages of the development process. Engineers need tools that empower them without having to learn a programming language. They have to analyse specific problems and gain knowledge through predictions about the parametrization of the control unit. Fleet validation focusses on creating a large amount of mileage on real roads and test tracks. Protocols define how the driver has to behave. Analytics help with an early error detection, fault analysis and the increasing pool of knowledge can be applied to future tests. Real-time analytics help in maintaining test protocols and live views provide an immediate fleet overview. The next generation of analysis and reporting with an interactive drill-down into the data pool is near production ready. The required investment in hardand software discourages most OEMs. Trust in commercial cloud solutions will reduce the involved investment. 282 8 Data Analytics 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system Matthias Kahl, Andreas Kroll, Robert Kästner, Manfried Sofsky Abstract This contribution deals with the identification of dynamical models of diesel engine processes from data obtained during urban and interstate traffic. This means that the engine is not specifically excited for the purpose of system identification and therefore data sequences with low excitation and data distributions which are not actively generated result. Both situations are examined in this contribution. For this purpose two different data selection approaches tackling the specific problems are proposed and tested on real engine data. The first approach targets the selection of “sufficiently” excited data sequences out of long data records, the second approach the homogenization of the data distribution. Kurzfassung Dieser Beitrag behandelt die Identifikation dynamischer Modelle von Dieselmotorprozessen mit Daten, die während dem Normalbetrieb aufgezeichnet wurden. In diesem Fall erfolgt keine gezielte Anregung des Motors bezüglich der Systemidentifikation und die aufgezeichneten Daten können teilweise Sequenzen beinhalten, in denen kaum Anregung stattfindet, und die Datenverteilung wird nicht aktiv herbeigeführt. Beide Fälle werden in diesem Beitrag untersucht. Dazu werden zwei verschiedene Ansätze vorgestellt, um das jeweilige Problem anzugehen, und an realen Motordaten getestet. Der erste Ansatz zielt auf die Auswahl „ausreichend“ angeregter Datensequenzen aus einer langen Datenaufzeichnung ab, der zweite Ansatz auf die Homogenisierung der Datenverteilung. 1 Introduction Due to the strong increase of electronic control and diagnostic functions of modern internal combustion engines the design of the engine management system is associated with a high test and calibration effort. In case that a model of the considered process is available, which satisfies the high demands on approximation accuracy and computational complexity, efficient model-based methods such as hardware in the loop simulation can be used for this purpose. Describing the dynamic input-output behavior of an engine based on physical laws requires a high degree of applicationspecific knowledge. Alternatively data-driven modeling techniques can be used, as they can provide models with high approximation accuracy directly from input-output data of the considered process. 283 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system The accuracy of an identified model mainly depends on the information content of the input and output data of the considered process. Commonly, general purpose test signals, in particular broadband test signals or, in case of known model structure, Fisher information based optimally designed test signals are used to excite the process in an appropriate way. The resulting time series of the input signals are executed on engine test benches with corresponding costs. In this contribution inexpensive data which were acquired from a test vehicle during urban and interstate traffic are used, as they are commonly stored during the ECU development process and are therefore abundantly available. These data can partly have poor information content and be distributed corresponding to the specific driving situations. This means that the system is not excited specifically in the desired way and no persistent excitation or optimality of the gathered data can be guaranteed. The problem of data selection from large historical databases has been investigated in the process industries. However, these data have different characteristics compared to engine process data. Well continually operated process plants have approximately constant process variables that vary in the seldom case of reference or load changes or larger sudden disturbances, such that the data are often considered as little informative for system identification (see, e.g. [1] or [2]). In contrast, for urban and interstate traffic engine operation data process variables show most widely a permanent variation. Hence, in this contribution two simple heuristic approaches are examined. The first approach focuses on situations where the user has to choose persistently excited sequences out of a long data record in order to reduce the amount of data and simultaneously reduce the computational complexity for the identification of dynamic models. A further approach focuses on the homogenization of the data distribution to avoid overor under-representation of different dynamics or operating regimes in the model. For both approaches it is assumed that model structure of the system under consideration is unknown. Hence, the approaches should work in a model-free setting. As modern diesel engines have multiple influencing variables the approaches should be applicable for multi-input systems. For the considered case study, i.e. the identification of the pressure boost of a modern diesel engine, it is shown that by means of the proposed approaches a decimation of data without a larg reduction of model accuracy is possible, which can be desirable when dealing with large data sets or an undesired data distribution. This contribution is organized as follows. In the next section some basic concepts regarding the data requirements for system identification are briefly revisited. In section 3 data selection approaches proposed for the process industries are shortly described. Additionally, two concepts which address the specific characteristics of data originating from urban and interstate traffic are proposed. Section 4 deals with the application of these concepts in a case study before concluding in section 5. 2 Data requirements for system identification The aim of system identification consists in finding the unknown functional relationship between the explanatory variables and the output , 1, … , , (1) 284 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system based on a data set with elements, containing measurements of explanatory variables and the output of the process under consideration. The term is assumed to be an independent and identically distributed random noise sequence following a Gaussian distribution with zero mean and variance , and the subscript represents the discrete time index. This contribution focuses on parametric model structures of dynamic multiple-input single-output (MISO) nonlinear autoregressive with exogenous inputs (NARX) models, where the functional relationship (1) is described by the difference equation , , 1, … , , (2) with consisting of lagged values of the system inputs , … , and the system output 1, , … , , , , … , , , … , , , … , , T (3) and the corresponding parameter vector , , … , , , , … , , , … , , , … , , T . (4) Assuming the nonlinear function : and the maximum lags , … , and to be known (meaning the model structure is completely known), we would like to find an estimate of the parameters , which can be done by minimizing the difference between the measured output and the predicted output , , leading to the well-known prediction error minimization (PEM) framework. Throughout this contribution the common mean least squares identification criterion is considered: argmin 1 , . (5) For models, which are linear in the parameters (LIP), the solution is given by (see, e.g. [3]) , (6) with the information matrix . (7) Within the PEM framework for linear time invariant (LTI) systems it is possible to guarantee consistency of the parameter estimates, given the data set is informative 285 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system with respect to a certain model structure and that the model structure itself is identifiable (see, e.g. [3]). The former property is important to enable the identification criterion to distinguish between different models and the latter property to ensure that the minimizer of the identification criterion yields a unique solution. To ensure that the data set is informative enough w.r.t. to a set of candidate models the closely related concept of persistence of excitation was introduced, which states the necessary conditions regarding the input signals of the considered process. A formal definition for linear systems is given by: Definition 1 (persistence of excitation) [3]: Let be a quasi-stationary signal, and let the matrix be defined by 0 1 1 1 0 2 1 2 0 (8) with lim ∑ , then the sequence is persistently exciting of order if and only if is nonsingular, which is a property concerning the variation that is present in the signal . Following definition 13.1 in [3] a further interpretation of this concept in the frequency domain is, that a quasi-stationary signal is persistently exciting of order , if its spectrum Φ is different from zero on at least points in the interval . This condition is used for the design of appropriate input signals to ensure that is invertible. For a given data set one can directly check the rank of or its condition number, as numerical methods are often used to determine the parameter estimates instead of simple least square regression (see, e.g. [2] or [4]). When dealing with optimal experimental design one directly tries to maximize a scalar criterion in order to improve the condition of . Common criteria are the determinant of (Dcriterion) or the maximum of its minimum eigenvalue (E-criterion) (see, e.g. [5] or [6]). Note that in the linear regression setting corresponds to the Fisher Information Matrix (FIM). For a given data record the stated concepts regarding the conditioning of can be used in an inverse manner to check if the data are informative w.r.t. to a given model as is done in the reported data selection approaches described in section 3.1. In a model-free setting these concepts cannot be used but one can check the variation of the considered signals which is related to the concept of persistence of excitation and done in our proposed approach (see section 3.2). 3 Data selection for system identification 3.1 Reported data selection approaches Three methods which are supposed to enable the selection of informative data segments from historical data have been reported in [4], [6], and [7]. The methods are used in the process industry field, where the process is operated at a specific set- 286 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system point most of the time resulting in long periods of “flat” data which are consequently not persistently exciting and therefore not suitable for the identification of the dynamic behavior of the process. In principle the reported methods consist of two major parts, the segmentation of the data and the assessment of the data quality. For data segmentation the following three strategies are commonly used: i) sliding window, where data samples are added to a given segment until some error bound is exceeded; ii) top-down, where the time series is partitioned until some stopping criterion is reached; and iii) bottom-up, where, starting from the smallest number of segments, the segments are merged until some stopping criteria is reached ([6], [8]). For the second step (data quality assessment) criteria based on the concepts mentioned in the previous section are used. The approach reported in [4] assumes single-input and single-output (SISO) processes and that the process can be described by a linear model. The data set is traversed recursively by applying a sliding window approach and the quality of the data is assessed with different criteria for a given operating point. Starting from simple heuristics, which check if there is an activity in the input and output signals by monitoring the signals’ variability, they check if the parameters of a presumed model can be estimated for the given data sequence by checking the numerical conditioning of the corresponding information matrix. In a further step it is checked additionally if there is a significant correlation between input and output by means of a Granger causality test. In each step the calculated quantities are tested against corresponding thresholds, which have to be set by the user and which depend on the considered application. The method reported in [6] uses a bottom-up segmentation strategy, where adjacent data segments with similar information content are merged. The information content of each segment is measured by means of its FIM. The comparison is then done by a similarity measure between the eigenvectors of the FIMs obtained from each segment and segments with similar information content are merged. Finally, the resulting segments are assessed by the use of criteria known from design of experiments (DOE), like Dor Ecriteria. The design parameters, which have to be set, are the initial and desired number of data segments. A strongly related method to the ones reported in [4] and [6] can be found in [7] and [9] where also the SISO case is assumed. The quality assessment is based on the condition number of the information matrix. Furthermore, in [9] a bottom-up segmentation based on the difference in the entropy of the input and the output signal is used in order to partition the data into regions which are assumed to stem from the same linear system behavior and thus can be modelled with the same linear model. The three approaches given in the literature are not directly applicable in the case that the model structure is unknown, as the data quality assessment is based on the information matrix for an a priori given model. In contrast, we would like to do our examinations in a model-free setting. The heuristics stated in [4] can be used solely for our purpose but instead the numerical gradient of the considered signals are used alternatively, which coincides with the numerical difference quotient for time-series, in order to check the signals´ variability as stated in the next subsection. This is done in 287 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system order to deal with the instationarity of the considered signals as with the numerical gradient a mean free consideration is obtained. Furthermore, the data distribution, which has a significant influence on the identification of nonlinear models, is not directly taken into account in the reported approaches. As we would like to examine the influence of the non-uniform data distribution which is not actively generated, a distance based data thinning as discussed in subsection 3.2.2 is proposed. 3.2 Proposed data selection approaches 3.2.1 Approach I: Discarding unexcited data segments The first approach considers the problem of insufficiently excited data segments occurring in long data records acquired during urban and interstate traffic, meaning no changes occurred for a long time. These segments have only low frequency content and thus are expected to have low information content regarding the dynamics of the system. One has to be aware that these regions still contain information about the system gain which is not considered in our approach. Let us consider the engine speed measured during interstate and urban traffic depicted in Figure 1 as an 100 200 300 400 500 600 700 800 900 1000 2000 3000 4000 time in s n eng in rpm Figure 1: Time series of the engine speed measured during normal vehicle operation 100 200 300 400 500 600 700 800 900 0 5 10 time in s frequency in Hz PSD in dB/ Hz -50 0 50 100 200 300 400 500 600 700 800 900 -5000 0 5000 time in s δ Ts n eng in rpm/ s Figure 2: a) Frequency content of over time. b) Numerical gradient of a) b) 288 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system example. As can be seen from the time series there are partially stationary segments, meaning the excitation is not persistent. This becomes evident in Figure 2 a), where the frequency content of segments of length 10 seconds with an overlap of 8 seconds is calculated using a short-time FFT. Clearly, the frequency content over time corresponds to the variation of the time signal, which can be approximately characterized by using the numerical gradient depicted in Figure 2 b). Hence, the numerical gradient of the considered time signals is used to assess their variation, with the aim to discard data segments which are “insufficiently” excited. The information of the operating point is not taken into account. The numerical gradient in each point is calculated using the central difference of interior points 2 . (9) The time series of the quantities are segmented using a moving window approach. Within the windows the absolute values of the numerical gradients are used in order to calculate a scalar criterion which is compared to a threshold . (10) In this contribution the third quartile of the gradients within the windows is considered exemplarily as criterion, i.e. if there are 75% samples with a large enough gradient within the windows, these segments are rated as informative, as it reveals information about the system dynamics. Alternatively, other statistics can be used depending on the distribution of the gradients. As for diesel engine processes a lot of influencing variables exist, the engine speed and the injection fuel quantity are used as reference signals to assess the “global” excitation of the engine. This is done for simplification and is clearly not valid for all engine sub processes. Additionally, the considered system output is examined as potential reference signal w.r.t the aforementioned criterion. 3.2.2 Approach II: Distance based data thinning The second approach considers the distribution of the data. According to [10], the data should be equally distributed over the operation regime to avoid the overor under-representation of different dynamics in the model, which is relevant for nonlinear systems. On the other hand, if it is known a priori that the system behaves in a more complex way in specific operating regimes, one should gather more information, i.e. collect more data in these regions. As it is assumed, that no a priori information about the systems behavior exists, it is strived for an equal distribution of the data over the operational regime. Let us now consider the distribution of the data over the operational regime represented by and from the aforementioned measurements. As can be seen in Figure 3, the engine operates most of the time in the low ranges of and . One has to be aware that the engine was not actively excited to obtain this data distribution; rather the distribution is a consequence of the specific environmental conditions during data acquisition and thus the data may be not representative for other opera- 289 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system tional conditions, as the resulting model would focus on undesired regions with high data density. In order to get more equally distributed data, following [11], a distance-based thinning approach is used which compares the distance between two samples with a minimum allowed distance resulting from a predefined equidistant data distribution. The concept can be summarized as follows: 1. Starting from a minimum allowed distance 1⁄ between two samples defined by a necessary number of samples per dimension of a considered space defined by the quantities of interest 2. Compute the distances of the samples of the given data set by means of the Euclidean distance in this space 3. For each sample count the number of distance violations and delete the sample point with the most distance violations 4. Repeat steps 2 and 3 until no further distance violations exist Clearly, this simple algorithm leads to a thinning of data regions with high density. Again, and are used as reference signals to evaluate the distribution of the engine´s operation regime. As this is a pure static consideration the numerical gradients of the signals are also taken into account. By doing this, the considered space includes additional information about the variation of the signals. More precisely, data points corresponding to time instances with low variation defined by its numerical gradient should be closer in this space and thus be discarded. Furthermore, as in approach I, the system output is also considered as a potential decision variable. Additionally, the initial data set is partitioned using an equidistant grid over the corre- 1000 1500 2000 2500 3000 3500 4000 4500 0 10 20 30 40 50 60 q inj in mg/ stroke 0 5000 10000 n eng in rpm frequency 0 5000 10000 15000 frequency Figure 3: Coverage of operational regime 290 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system sponding reference signal in order to split the calculation of the distances in smaller parts to ensure its calculability. Thus, defined in step 1 becomes the necessary number of samples per dimension within one of the partitions called . 4 Case study: pressure boost of diesel engine 4.1 Considered model type and model structure selection Throughout this case study the cross-bilinear parametric Volterra-series NARX model NARX | 1 · · · · (11) is considered with the additional affine term . This model type is linear in the parameters, thus the simple least-squares estimator can be used for the estimation of its parameters. Furthermore, this model type is easy to handle and shows good approximation capabilities for the considered process. As mentioned before, it is assumed that the time dependencies of the considered process are not known in advance. Hence, in this contribution an established model selection approach known from linear regression is applied, namely the stepwise regression (SWR) with partial F-Test ([12]), which was already applied on this data in [13]. The SWR is a wrapper approach where model terms are successively added or removed based on their statistical significance in a regression problem (see, [12]). The significance is evaluated by a F-Test by sequentially comparing the performance of two models built of different subsets of regressors. In each step the hypotheses : 0 and : 0 (12) are tested against each other for a certain confidence level α. Rejecting the null hypotheses for a model term currently not included in the model means that the term will be added, as its parameter has a value unequal to zero. Contrarily, if there is insufficient evidence to reject the null hypothesis, for a term currently included in the model, the term is removed. The models will be used for simulation which means that the output is only based on current inputs and the previous predictions of the output. In this case is | 1 (see, e.g. [10]). To assess the generated models, the Normalized Mean Squared Error (NMSE) is used: NMSE ∑ ∑ . (13) 291 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system 4.2 System description and process data The described identification approaches were applied to the gas system of a modern supercharged diesel engine of a passenger car. The engine is equipped with a variable geometry turbine (VGT), charge air intercooler, highand low-pressure exhaust gas recirculation (EGR) and variable valve timing (VVT). An entire scheme of the considered engine configuration is depicted in Figure 4. For modeling the pressure boost , the VGT-command , the engine speed , the injected fuel quantity as well as the HPand LP-EGR commands ,HP and ,LP were used as inputs. Data of the engine were used which were collected with a test vehicle during urban and interstate traffic with a sample rate of 10 ms. The data record covers a wide operating range of the engine and consists of 92180 samples which were split into one part for identification (first 50 %) and one for validation of the obtained models (last 50 %). During the data acquisition the driver aimed to excite the engine in a dynamic fashion, i.e. a high gradient of and was targeted, which is not necessarily representative for data that is stored throughout the ECU development process. These data can more often contain stationary periods. Furthermore, the used data record is relatively short compared to data obtained during the ECU calibration process. Additionally, it is remarked that due to the experimental restrictions the quality of the used data set suffered from oversampling which may hamper the model identification and that the obtained models are restricted to the given ECU parametrization and the environmental constraints during data acquisition. 4.3 Data selection Both data selection approaches are tested on the given data set. The selection algorithms are applied to the identification data set and the resulting models are tested on the complete training data set and the validation data set. For model selection the maximum lags of the output and the inputs were chosen as 5, resulting in a candi- , , intercooler EGR-cooler exhaust aftertreatment ,LP ,HP Figure 4: Considered engine configuration 292 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system date set of 355 potential regressors. In order to obtain a sparse parametric Volterraseries model the SWR was restricted to a maximum of 50 iterations resulting in a model with a maximum of 50 regressors. 4.3.1 Approach I (Discarding unexcited data segments) The proposed data selection algorithm was implemented in Matlab and used to discard data segments with low excitation. There are two design parameters, the window length and thresholds, to choose for the considered constellation. The choice is quite intuitive. Different values of window size and thresholds were tested in order to evaluate the influence of discarding data segments with low frequency content for the given data set. The numerical gradients were standardized ⁄ with mean and standard deviation for the complete training data set. After standardization they lie in a comparable range so that identical thresholds were chosen. Furthermore, the first approach was examined when the system output is used as reference for data selection, meaning only the variation of the output is considered. Considered reference NMSE on training data NMSE on validation data dim( ) - - 0.036 0.085 50 46097 and 0.15 0.036 0.082 50 39563 0.25 0.034 0.089 50 28600 0.50 0.048 0.105 50 18729 1.00 0.110 0.079 50 11557 1.25 0.139 0.161 50 8062 0.10 0.038 0.101 50 30168 0.15 0.040 0.116 50 26404 0.25 0.041 0.104 50 19290 0.50 0.125 0.057 36 10758 0.75 0.230 0.095 50 7812 Table 1: Results for approach I 1 2 3 4 x 10 4 0 0.05 0.1 0.15 0.2 N NMSE 1 2 3 4 x 10 4 0 0.05 0.1 0.15 0.2 N Figure 5: NMSE on training data set (blue) and validation data set (red) using and as reference signal (a) and using as reference signal (b) a) b) 293 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system The thresholds depend on the chosen window length , as the values of the quartiles flatten with increasing window size. Additionally, with increasing window length it is expected that more data in a lower range of the pressure boost are selected for the given data, which strongly affects the results, as the pressure boost operates in this range most of the time. For 1000 samples good results were obtained which are given in Table 1 and Figure 5 for different values of . It can be seen that by discarding segments with low frequency content from the identification data set the resulting models are less able to describe the system behavior on the complete training data set and the validation data set with regard to the validation criterion. However, the sample size can be reduced to one half of the initial sample size without significant loss of accuracy of the resulting models for both data sets. With further decrease of the sample size the results are affected more by the specific distribution occurring on the respective data set, i.e. for the validation data set the engine operates more often in the lower pressure boost range. Comparing the results for the different reference signals it becomes evident that comparable results up to a certain remaining data size were gotten. With further increase of the thresholds different local effects occur which causes an indifferent increase of the NMSE on the respective data set. Figure 6 shows exemplarily the simulation results on the validation data set for the following two models: which is identified using the complete training data set consisting of 46097 samples and which is identified with a training data set consisting of 18729 samples corresponding to 0.5 and 1000 considering and as reference signals. 500 550 600 650 700 750 800 850 900 1000 1500 2000 p b in hPa 500 550 600 650 700 750 800 850 900 -400 -200 0 200 400 time in s residual in hPa Figure 6: Comparison of the measured (black) and simulated outputs on validation data of (blue) and (red) 294 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system As can be seen the loss of approximation accuracy for is moderate although the model was trained on a data set consisting only of less than one half of the initial sample size. For the chosen thresholds and window size a significant overemphasis of specific areas cannot be observed. However, this changes if the window length is decreased. As mentioned before in this case the resulting models focus more on dynamic areas with little improvement (e.g. the peak values of the pressure boost match more often to the measurement) whereas the stationary areas are approximated poorly. 4.3.2 Approach II (Distance based data thinning) To further examine the influence of the undesired data distribution occurring during the data acquisition with a test vehicle approach II was applied to the given data. The aim is to obtain a model which does not focus on specific operational conditions which is assumed to be better suited for general purpose applications. Approach II performs a sample wise data assessment which is a static consideration of the signals. As in this contribution dynamic models are considered, the results of the selection approaches are used for indexing the corresponding rows in the regression matrix which is used for model selection. This matrix contains the information about the dynamics up to the maximum considered lags for each quantity within its rows. Note, that this consideration can only be applied if the considered model type is LIP as there is no need to evaluate the model in a simulation. The necessary number of samples per dimension within one of the partitions has to be chosen solely for the distance based data thinning approach. As mentioned in 1000 1500 2000 2500 3000 3500 4000 4500 0 10 20 30 40 50 60 q inj in mg/ stroke 0 5000 n eng in rpm frequency 0 5000 10000 frequency Figure 7: Sample distribution of and for the initial training data set (blue) and the thinned data set (red) 295 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system section 3.2 the initial data set is partitioned using an equidistant grid over and as well as corresponding to the considered reference signals, which can be arbitrarily chosen, as it only affects the computational complexity. Here 15 intervals in the -dimension and 10 in the -dimension were used. For 10 intervals were used. The approach was also implemented in Matlab and tested for different values of . When using solely as decision variable, has to lie in a different range to get a comparable amount of remaining data. Figure 7 shows the initial and remaining data points (for 10 and the quantities describing the operation point of the engine) in the - -plane as well as the histograms of both data sets. As can be seen, the data distribution becomes more uniform and as a consequence rarely occurring operating conditions should be weighted more during the identification process. However, operating regimes with low data density of course remain sparse. The simulation results for different values of are given in Table 2 and Figure 8. Similar to the results for approach I no improvement results with a decreasing amount of data. Especially, the thinning causes that mainly data points corresponding to stationary segments are discarded, for which reason these segments are not Considered reference NMSE on training data NMSE on validation data dim( ) - - 0.036 0.085 50 46097 , , and 500 0.036 0.079 50 42863 100 0.038 0.111 50 28217 50 0.053 0.174 50 19664 10 0.064 0.204 50 7552 20000 0.049 0.165 50 24715 10000 0.059 0.235 50 20091 5000 0.069 0.273 50 15202 1000 0.061 0.168 33 6589 Table 2: Results for approach II 1 2 3 4 x 10 4 0 0.1 0.2 0.3 N NMSE 1 2 3 4 x 10 4 0 0.1 0.2 0.3 N Figure 8: NMSE on training data set (blue) and validation data set (red) using and as reference signal (a) and using as reference signal (b) a) b) 296 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system described well by the resulting models. With an increasing amount of data a higher offset error occurs for stationary segments on both data sets. This becomes more significant for the validation data set resulting in a higher value of the NMSE. Finally, Figure 9 shows the simulation results on the validation data for the following two models: which is identified using the complete training data set and which is identified with a training data set consisting of 19664 samples corresponding to 50 considering and as reference signals. As the numerical gradients of the signals are also considered for the thinning approach, especially data points corresponding to stationary segments are mainly discarded, for which reason these segments are not described well by model . Other regions are only slightly influenced by the data thinning. 4.3.3 Discussion For approach I it becomes evident that decreasing the amount of data is possible without a significant loss of accuracy provided that the window size and threshold has to be chosen in an appropriate way depending on the considered problem. For the given data set the results show a significant dependency on the distribution of the considered data so that the window length had to be chosen in a way that enough segments in a lower pressure boost range are selected. For a given window length the thresholds can be chosen by visual inspection of the chosen criterion on the gradient so that a desired value of the validation criterion or a desired amount of data 500 550 600 650 700 750 800 850 900 1000 1200 1400 1600 1800 2000 2200 p b in hPa 450 500 550 600 650 700 750 800 850 900 950 -200 0 200 residual in hPa time in s Figure 9: Comparison of the measured (black) and simulated outputs on validation data of (blue) and (red) 297 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system results. Of course other criteria can be used to evaluate the segments depending on the later use of the model. In our case no specific use was addressed. Regarding the used reference signals further examinations are needed in order to evaluate if the use of further signals improves the results or a combination of the used reference signals can reduce the occurrence of local errors. For approach II we are still confronted with the fact that the conditions on the initial training data set match more to the conditions of the validation data set than that of a thinned data set. Therefore, the results are not surprising and the question remains open, if the generalization capability of the models identified on a thinned data set is better. To ensure the obtained results for the considered application, further results are required according to the generalization capability of the models under different operating conditions. Furthermore, the fact that the proposed data thinning approach performs a sample wise and therefore static consideration of the reference signals may not match to the dynamic case. A possible improvement could be obtained if instead of a sample wise consideration the similarity of rows of the regression matrix is evaluated, so that similar segments of the given trajectory are discarded. A direct comparison of both approaches is not possible as they address different problems occurring with data from routine vehicle operation. For instance, when dealing with a huge amount of data obtained from a long measurement which cannot be handled in the identification process due to e.g. main storage restrictions one can use approach I to find excited data sequences. As mentioned before, this can be the case if one wants to use data which are stored throughout the ECU development process where gigabytes of data are acquired. This data especially can obtain longer periods of data where low excitation occurs (e.g. during interstate highway driving). For this type of data the benefit of the proposed selection method should become evident. Additionally, one can use approach II in order to obtain more uniformly distributed data if it is desirable for the later use of the model. 5 Conclusions In this contribution two situations occurring with data from routine vehicle operation were examined. The first deals with partly non excited data sequences, the second with an undesired data distribution due to the environmental conditions during data acquisition. Two simple heuristic data selection approaches applicable in a model free setting are proposed to tackle the specific situations. The first is based on a moving window segmentation combined with the numerical gradient of reference signals to select data segments which are “sufficiently” excited. The second approach should tackle the problem of an undesired data distribution resulting from normal vehicle operation. Therefore a distance-based data thinning approach was used to obtain a more uniform data distribution over the operation regime of an engine. Both approaches were tested on a data record acquired with a test vehicle during urban and interstate traffic. The results for the considered data show that there is no need in discarding data when dealing with the aforementioned problems as the results strongly depend on the specific operational conditions for the given data set. Especially, for approach II further examinations regarding the generalization capability under different operating conditions are therefore needed to make general statements. However, approach II can be used if a more uniform data distribution is desirable for 298 8.1 On the selection of appropriate data from routine vehicle operation for system identification of a diesel engine gas system the later application of the model. With the aim of the reduction of the amount of data without a significant loss of accuracy of the resulting model approach I can be used, as it shows good results for the considered data with a suitable choice of the design parameters. The benefit of this approach should become more evident if data records stored throughout the ECU development process will be considered. References [1] A. J. Isaksson, “Some aspects of industrial system identification,” IFAC Proceedings Volumes , 46(32), pp. 153-159, 2013. [2] D. Peretzki, A. J. Isaksson, A. C. Bittencourt, K. Forsmann, “Data mining of historic data for process identification,” In Proceedings of the 2011 AIChE Annual Meeting , pp. 16-21, 2011. [3] L. Ljung, System identification - Theory for the user , 2nd ed. New Jersey: Prentice Hall PTR, 1999. [4] A. C. Bittencourt, A. J. Isaksson, D. Peretzki, K. Forman, “An algorithm for finding process identification intervals from normal operating data,“ Processes , 3(2), pp. 357-383, 2015. [5] C. Hametner, M. Stadlbauer, M. Deregnaucourt, M. Jakubek, T. Winsel, “Optimal experiment design based on local model networks and multilayer perseptron networks,” Engineering Applications of Artificial Intelligence , 26(1), pp. 251-261, 2013. [6] L. Dobos, J. Abonyi, “Fisher information based time-series segmentaion of process data,“ Chemical Engineering Science , 101, pp. 99-108, 2013. [7] Y.A. Shardt, B. Huang, “Data quality assessment of routine operating data for process identification,” Computer & Chemical Engineering , 55, pp. 19-27, 2013. [8] E. Keogh, S. Chu, D. Hart, H. Pazzani, “An online algorithm for segmenting time series,” In Proceedings IEEE International Conference on Data Mining , pp. 289- 296, 2001. [9] Y.A. Shardt, S.L. Shah, “Segmentation methods for model identification from historical process data,” IFAC Proceedings Volumes , 47(3), pp. 2836-2841, 2014. [10] O. Nelles, Nonlinear System Identification , Springer, 2001. [11] N. Tietze, Model-based calibration of engine control units using Gaussian process regression , Diss., TU Darmstadt, Germany, 2015. [12] N.-R. Draper, H. Smith, Applied regression analysis , 3rd ed. New York: Wiley, 1998. [13] M. Kahl, A. Kroll, R. Kästner, M. Sofksy, “Application of model selection methods for the identification of a dynamic boost pressure model,“ IFAC- PapersOnLine , 48(28), pp. 829-834, 2015. 299 8.2 Data Plausibility at the Engine Test Bench - How important is the Human Factor in the Process? Hanno Ihme-Schramm, Alexandra Schramm Abstract This paper shows which technical domains need to be taken into account when integrating data plausibility at the engine test bench. Furthermore, the persons at the various points along the process chain who are affected by the integration can be given optimum preparation and support in terms of business psychology. This also includes actively shaping the framework conditions in the interests of smooth, sustainable introduction of data plausibility. Kurzfassung In diesem Beitrag wird aufgezeigt, welche fachlichen Themenbereiche bei der Integration einer Datenplausibilität am Motorprüfstand beachtet werden sollten. Aber auch die von der Integration betroffenen Personen können hierauf aus wirtschaftspsychologischer Sicht an den verschiedenen Stellen der Prozesskette optimal vorbereitet und unterstützt werden. Hierzu gehört auch die aktive Gestaltung von Rahmenbedingungen, die eine reibungslose und nachhaltige Einführung der Datenplausibilität ermöglichen. 1 Introduction Huge engineering efforts are currently in progress regarding all aspects of engine development in order to meet the high demands made of today's combustion engines. This includes constant testing and optimization of combustion engines at the engine test bench, particularly when it comes to combustion process development, engine testing and calibration. The result is a large number of measurement and computation variables, which test bench staff (test engineer or test bench operator) cannot check quickly and accurately for plausibility without methodological support on account of the high development pressure, huge data volume and complex engine correlations (Fig. 1). 300 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? Fig. 1: External factors impacting on test bench staff The situation of test bench staff is further aggravated by the high test bench costs which become noticeable particularly when measurements are incorrect. Incorrect measurements can be caused, for example, by the failure of sensors and instruments or creeping component damage. Such problems are often not detected immediately, resulting in further unnecessary measurements and costs that could be prevented. This frequently has a negative effect on test bench staff motivation. Fig. 2 states other reasons in favor of introducing measurement plausibility in the test bench environment. A first step here must involve the integration of important modules and methods to support test bench staff in future tasks in order to make the workload easier. Fig. 2: Reasons for introducing measurement plausibility 301 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? But for data plausibility to be introduced successfully, a second step also has to consider the people involved at various points along the process chain, together with the factors impacting on them in each case. In the end, they play a crucial role in the successful introduction and sustainable implementation of data plausibility. This therefore is the background against which this article examines for the first time the human factor in terms of business psychology. 2 Measurement plausibility at the engine test bench Meaningful integration of measurement plausibility at the engine test bench should take account of the following workflow [1] (Fig. 3). In each case, the individual modules control the measurement and computation variables of the engine in chronological order. Fig. 3: General modules of the measurement plausibility workflow The standstill test checks the current ambient state of different variables (e.g. pressure, temperature) before the engine starts. In addition, certain other variables must show the value "0" when at a standstill (e.g. engine speed, torque and fuel consumption). The subsequent reproducibility points are measured with the warmed-up and running engine, and the measurement and computation variables are compared with the previous days. It is thus possible to detect measurement errors and drift phenomena in engine operation. Only then do the actual measurements start. Subsequent offline plausibility makes it possible to check the generated set of measurement data with a suitable, user-friendly software program on the PC. This can take place directly after the measurements or many days later. The operating points with the many measurement data are imported, occurring errors can be visualized or sought and visualized using standard patterns (Fig. 4). In doing so, the test engineer needs experience and practice in using the program, due to the complex correlations involved. 302 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? Fig. 4: Offline plausibility workflow The optimum solution would be an online diagnostic tool (Fig. 5) integrated on the engine test bench for online control of engine, measurements instruments as well as measurement and computation variables. If a measurement error occurs while the engine is running, an error signal appears and warns the test bench staff. However, it is very sophisticated and elaborate to produce and implement an online diagnostic tool in view of the complex correlations. Not all measurement and computation variables can be controlled. A first step should therefore just monitor the most important variables. If an error occurs, the test engineer should know exactly how to proceed. Frequently, the cause of the error will not be immediately apparent as only the effects of the error will be visible. Besides training for the test bench staff, error management should also be integrated with recourse to measurement plausibility experts at any time so that these can offer support and advice when the need arises. Fig. 5: Implementation and difficulties of online measurement plausibility 303 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? 3 Plausibility methods The methods used for error detection can be based among others on experience with engines, physical knowledge, balance equations, comparisons, limit values, engine models or visual checks. Two concrete examples illustrate the different methods. It is very easy to check indicated data. The frictional mean pressure is derived from the computed indicated mean pressure and the effective mean pressure. According to engine physics, the frictional mean pressure is always greater than "0" and has to increase as a function of engine speed. This is revealed in both cases in the point cloud of red measurement data in Fig. 6. The blue dots (outliers) deviate clearly from the trend. Either an error occurred during indication measurement, or the effective mean pressure obtained from the torque measurement is incorrect. As described in the previous section, the cause of the error now has to be found. Fig. 6: Using frictional mean pressure to check the indicated data for a gasoline engine The second example looks at the CO 2 variable of the exhaust measurement system. The relationship between CO 2 and the air ratio (for gasoline and diesel engines) has been taken from engine physics and is shown in Fig. 7. The measured CO 2 values must be within the tolerance limit lines. Any points exceeding the limits draw attention to a possible measurement error, which must be checked and monitored by the trained test bench staff. 304 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? Fig. 7: Checking the measured CO 2 values The two methods described here can be used offline and also online. Both examples clearly show that visualization permits swift control of the measured operating points. To make it easier for the test bench staff to use the described methods and the measurement plausibility program, the workflow should be standardized and structured in the individual sub-steps. Diagrams and evaluations should always have the same appearance to facilitate familiarization and to generate a high recognition value. The test engineer should practice constantly with current measured data to continuously improve and extend his or her knowledge and skills in terms of measurement plausibility. Both standardized and as far as possible individual training measures should be used for the test bench staff. Psychological aptitude tests can be deployed to reveal the exact need for training and further development, thus warranting optimum support and mentoring for the test bench staff on the basis of the strengths shown by each employee in order to safeguard a high level of data quality. Furthermore, permanent professional support in the process should be safeguarded by internal appointments of technical experts. 4 The human factor Introducing and firmly establishing data plausibility at the engine test bench constitutes a change process than can meet with challenges and rejection. Basically, every permanent change forces us to leave our familiar comfort zone. The comfort zone is where we feel at ease, where we know precisely what to do and don't expect any surprises. If a change is now "threatened", initially this means inconvenience and uncertainty. The most frequent reaction is to reject the change and to return to the familiar work process. The test bench staff will be responsible for implementing the 305 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? changes and must therefore be taken into consideration and integrated early on in the process. Right at the start of introducing data plausibility, the test bench staff and their line managers must be made aware of the current situation. Test bench staff often receive no feedback (Fig. 8) from the development engineer about what happens with the produced measurement results or which costs were generated by incorrect measurements. It is therefore important to set up a feedback loop. The test bench staff must see how important and meaningful their work is for the whole process. The test bench staff need to know about the whole process chain involved in engine development, together with regular feedback about the costs incurred. Only then can they develop a feeling for their own responsibility in the development process and understand the impact of incorrect decisions (their own and downstream development engineers) resulting from incorrect measurement results. Fig. 8: Feedback in the data plausibility process Based on the awareness of the highly responsible role played by the test bench staff in the development process, it should become their own strong interest to ensure that the measurement process is free of errors to the greatest possible extent. As well as accepting the usefulness of supporting methods, this also includes having the courage to act on their own initiative and to take swift decisions, even with the risk of making incorrect decisions or bringing the engine test bench to a brief standstill. But before this can happen, other prerequisites have to be fulfilled first. On the one hand, the corporate culture (Fig. 9) has to be altered to encourage and allow each individual to think and act on their own initiative. This needs an error culture where mistakes are not punished but seen as an opportunity to learn and improve, sharing the lessons learned with all others to ensure future improvements. A culture of trust must be created between test bench staff, development engineer and line managers to permit open, constructive feedback between all stakeholders. Among others, this also includes measures to present achievements at short notice, 306 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? as well as showing test bench staff appreciation for the decisions they have taken. One solution here could be to introduce reward systems and incentive measures, e.g. for measurement results on a constantly high level of data quality. Fig. 9: Interaction between the various stakeholders involved in the measurement plausibility change process Both the company management and the specific line managers have an important role to play in change processes impacting on the corporate culture. Every cultural change is a long-lasting process which can only be successful if implemented consistently and sustainably. For the management, this means communicating a clear vision of what the culture should look like in future, consistently sending the right signals to the company at large. Line managers on every level are then required to implement this vision, living it as role models, expecting the most from their staff and giving them every support in the daily work process. Change processes can be prepared and monitored by external coaches to ensure they run as smoothly as possible. The coaches offer close support for the process in the initial stages, with subsequent sporadic action by being available as an interlocutor as and when the need arises. 5 Summary The issue of measurement plausibility at the engine test bench is an extensive change process that harbors challenges, takes time and causes rollout costs. The result achieved in the end consists in a sustainable improvement in measurement quality and a reduction in the costs caused by incorrect measurements. 307 8.2 Data Plausibility at the Engine Test Bench How important is the Human Factor in the Process? Going over and beyond the actual process workflow, the individual stakeholders involved in the process also play a crucial role in sustainably implementing measurement plausibility and in making a success of the whole process. The change process can only succeed if all stakeholders recognize the need for measurement plausibility with an open-minded approach to the changes involved and are willing to support and carry the process. In this context they can be provided with optimum support, preparation and training, integrating them as important players and essential mainstays. References [1] Dr. Hanno Ihme-Schramm, Andreas Flohr, „Verfahren und Vorrichtung zur Bestimmung der Messgenauigkeit einer Messeinrichtung“, Patent DE102006048730A1, Volkswagen AG 308 9 Design of Experiments III 9.1 Non-Convex Hulls for Engineering Applications Nico Didcock, Nikolaus Keuth, Andreas Rainer Abstract Every data based model suffers drawbacks linked to extrapolation. Although a rigorous definition of extrapolation does not exist, it is generally understood that data based models may perform poorly in areas that don't contain measurements. A system should only be operated in new areas while being supervised on the test bed, where one can observe failing system components. Measurement data hulls are then commonly used to re-construct the feasible region of the system or define data based constraints in the following optimisation process. If measurements are be used to train black or grey-box-models then these models may again predict poorly in extrapolating areas. The most commonly used data hull is the convex hull. For input dimensions higher than ten, the convex hull is practically computationally intractable as computational complexity grows exponentially. Another drawback is that convex shapes are often not representing the true shape of the data. This paper discusses non-convex hull concepts for engineering applications. Some methods model the boundary directly using triangulations. Others train a value function in order to introduce a notion of how far a given test point lies inside or outside with respect to the location of the measured data points. Sometimes the assumption of convexity is relaxed to absorbing data, but other relaxations exist also. Emphasis of the discussion is a setting with small to medium number of input variations, as well as a medium to high number of samples. Next to a visual comparison, the hull concepts are compared with respect to hull building and hull evaluation CPU complexity. Kurzfassung Bei Verbrennungsmotoren treten außerhalb der systemeigenen Betriebsgrenzen Phänomene wie etwa Motorklopfen oder Fehlzündungen auf, welche während des Prüfstandsbetriebs zu Not-Stopps führen. Dies hat mitunter zur Folge, dass für die nicht-fahrbaren Bereiche keine Messungen vorliegen, und der fahrbare Bereich muss daher aus einer Bereichsschätzung der fahrbaren, stabilen Messungen rekonstruiert werden. Der hochdimensionalen Bereichsschätzung kommt eine weitere Bedeutung hinzu, wenn für Kalibrationsaufgaben datenbasierte Regressionsmethoden in Betracht gezogen werden. Diese weisen aufgrund ihrer Generizität einerseits ein breites Anwendungspektrum auf, andererseits sind sie mit dem Problem der Extrapolation konfrontiert. Darunter versteht man die Verwendung eines Modells in Bereichen außerhalb der Messdaten. Werden Surrogate in der Extrapolation verwendet treten mitunter hohe Schätzfehler auf, wodurch etwaige Optimierungsresultate bedeutungslos werden. 309 9.1 Non-Convex Hulls for Engineering Applications Die in der Praxis am meisten verwendete Methode für die Modellierung des fahrbaren Bereichs ist die konvexe Hülle, diese ist aber nicht für hoch-dimensionale Kontrollsysteme geeignet, da der Rechenaufwand teilweise unermesslich steigt. Andere Methoden, etwa Distanz-basierte Methoden, können zwar schnell ausgewertet werden, reproduzieren die Daten allerdings naturgemäß schlecht, wenn Daten am Rande des fahrbaren Bereichs gemessen werden. In dieser Arbeit werden Verallgemeinerungen der gängigen Methoden der Bereichsschätzung vorgestellt, welche sich ohne Probleme auf hohe Dimensionen und mitunter große Datenmengen verallgemeinern lassen. Durch die Berücksichtigung der Datenstruktur, welche sich aus der Versuchsplanung ergibt, ist es möglich, die nicht-fahrbaren Bereiche systematisch aus den vorhandenen Daten zu rekonstruieren. 1 Introduction Infeasible regions are given by those sub-spaces of the controller space where the test-bed environment reports some kind of limit reaction. For example, this may happen due to the violation of some boundary on the temperature, or phenomena such as engine knocking or misfiring. For calibration, it is essential to model the feasible region, for example, as a constraint during an optimisation program. Additionally, many data based calibration programs use surrogate models using statistical regression methods. If an emission model is trained on a set of variables then, naturally, predictive performance of the model will be poorer the further one attempts to extrapolate from the familiar region. During the optimization process, the calibration engineer therefore needs to make a judgment of how the optimisation results extrapolate from the region contained by the data, and the engineer will include extrapolation constraints in the optimisation program in order to reflect conservativeness with respect to stable engine operation and ensure plausibility of the predictive performance of the statistical models. A generic approach to model feasible regions and extrapolation is to use data hulls of feasible measurements. In contrast with classification methods, which use measurements from both, feasible and infeasible regions, hulls use data from one single class only and differentiate between the inner versus the outer space defined by the data. Possibly the most widely used data hull is the convex hull. It is defined as the smallest convex space that contains the data. These methods are problematic in dimensions above ten, since the computation is numerically intractable. Moreover, the convex hull naturally performs poorly as a feasibility classifier if the feasible region is not convex. This paper investigates methods to compute non-convex data hulls which can be computed for large input dimensions and non-convex data shapes. Specific attention has been drawn to use methods that are able to re construct the shape of data which is intrinsic to so-called screening measurement procedures. 310 1.1 S Figure proach tween c violatio ing. In on a clo A sequ the latt fixed c screen ing poin Figur This wo a multi include method numbe Section obtaine 2 N From a inside t cation goal is measu stable r Some hull ( values, Screening 1 illustrate ed from th combinatio ons, or syst case of a ose, but st uential scre ter measur entre poin ing proced nts, and th re1: Screen ork discus -variate fe e the conve ds are num r of non-co n 3 summa ed for a sim Non-Con an abstract the shape methods, w to separa rements o region and hull metho ( ) is given , i.e. 9.1 No g Measure es a scree he centre p ons of inpu tem instab limit violat table variat eening stra rement is s nt. A globa dures perfo he screenin ning meas screen ses robust easible con ex hull, pro merically in onvex hull arises CPU mulated en nvex Dat t point of v which is i where dat ate the reg obtained fro d therefore ods use so n by those on-Convex H ement Proc ening strat point. One ut variables bilities may tion, the s tion point. ategy, whe shown in t al screenin ormed inde ng centre d urement pr ning (middl t and effici ntrol space ovide reliab tractable fo approache U complex gine mode a Hulls view, the h mplied by a of sever gions corr om contro data is on o-called va inputs w ulls for Engin cedures tegy, whe can think s following y or may n ystem may re each va the middle ng procedu ependently depends on rocedures le), global ient estima e. While tra ble models for dimensi es which h xity of the el. hull of a da the data. ral classes responding l systems nly availabl alue functio where a gi neering Appl reby each of the con dynamic r ot occur a y step bac ariation is a plot, and ure is show y from eac n the opera . Centred s screening ation of the aditional tr s for small t ions above have been methods ata set con The hull c s are availa g to differe can often e from one ons to mod ven value lications measurem ntrol variab ramps from t some po ck and con approache in this cas w on the ch other fo ating point screening (right). e support o riangulation to medium e ten. Sect introduced and predic sists of tho oncept is r able to the ent classes only be o e single cla del the fea function ment poin bles switch m the centr oint during ntinue to m ed on a ram se there ex right wher or different t. (left), sequ of feasible n methods m input size tion 2 disc d in the lit ctive perfo ose points related to e analyst, s. In engin obtained fr ass. asible regio (⋅) attains t is aphing bere. Limit screenmeasure mp from xists no re three operatuential data in s, which es these usses a erature. ormance s that lie classifiand the neering, rom the on. The s certain 311 For a g In the f Delaun tively. 2.1 D The mo three-d convex used, e of atom such th hulls ha in [8,9] each da A so-ca cations infeasib These expens inappro 2.2 B An abs been in model point. given accep following, a nay triangu Delaunay ost widely dimensiona x hull, α-sh e.g. to des ms, and the hat no poly ave been u , which are ata point. alled polar s. Here, the ble data po hull metho sive if use opriate for Bates’ Fea sorbing, no ntroduced the distanc 9.1 No ptance reg a number o ulation -bas Triangula used non al data obt hapes, as scribe the s ey are give ytope has used for th e compute r Delaunay e hull cons oints. Fig ods all req d for high dimension asibility M on-convex in [1,2] for ce of the h on-Convex H ∀ ∈ gion . of hull met ed method ations -convex hu tained via well as lo structure o en by thos an edge la he estimatio ed as the u y triangulat sists of tho gure 2: Con quire triang dimensio ns above te Model hull conc the model hull as a m ulls for Engin ( ): ( thods will b ds, spatial ull method 3D scans ocal convex of molecule se sub-com arger than on of two-d union of co tion has be ose polytop nvex hull & gulations o nal data. en. cept for hig ling of feas mapping fro neering Appl ) ∈ be discuss l , as well a ds were int s or GPS x hulls. In es from da mplexes of α. Concep dimensiona onvex hulls een used i pes that ne & alpha-sha of the data Practically gh dimens sible contro om the sph lications sed. These as absorb troduced to signals. E [6,7], α-sh ata represe the Delau ptually rela al non-con s using nea n [10] for e either conta ape a, which is , triangula sional engi ol subspac here surro e methods bing hulls , o visualize Examples hapes hav enting the unay triang ated, local nvex home arest neigh engineerin ain, nor lie s computa ation metho ineering da ces. The id unding the include respece two or are the ve been location gulation, convex ranges hbors of g applie behind ationally ods are ata has dea is to e centre 312 Let centre of the b This ca by nonand a s outside similar , whic a hull o Crucial tively. informa ing me uremen ly restr ular for Figure 2.3 S Somew models a test p easy to tions ce denote th of the data boundary d an be achi -linear reg suitable ac e the hull, i spherical ch is desira on the left. ly, the dat The choic ation or so easuremen nts neighbo ricting since r high dime e 3: Illustra Spatial Mo what more s include a point and t o construct entered at 9.1 No he bounda a. The idea data to the ∀ eved e.g. ression mo cceptance f its distan directions. able for scr ta needs to ce of these ome compu ts on a gr or to at lea e hull quer ensions gri ations of Ba s odels intuitive all models w the data. D t suitable v the data. A on-Convex H ry points o a is then fi unit spher ∈ : using radi odels. Sec ( ) = region can ce to the c By constr reening pro o be sepa e boundar utation. In rid, bounda ast one infe ries should ds are extr ates’ feasib urface inte models ar where the Density es value func A spatial h ulls for Engin of the data rst to find re surround − ‖ − ‖ ial basis fu cond, the va − ‖ − ‖ − n be chose centre exce ruction, thi ocedures. rated into ry points i [1,2] this p ary points easible me d be perfor remely tim bility mode erpolation m re given b shape is stimates ar ctions using ull is illustr neering Appl a , and a function ding to the ‖ = ‖ − unction int alue functi − ‖ − ‖, en as = eeds that o s hull is ab Figure 3 il interior an s not trivi problem ha are define easuremen rmable for e-consum el (left), a s model (righ by so-calle determined re example g a superp rated in Fig lications let deno that ma hull expan ‖ erpolation, on is given 0, ∞). A te of the data bsorbing to lustrates t nd bounda al and eit as been so ed as thos nt. This def scattered ing design patial hull ht) ed spatial d by the d es for this position of gure 3 in th ote the sc ps the proj nsions, i.e , or approx n by est point a correspon owards the the border ry points, ther requir olved by p se feasible finition is o data, as in s. (middle), a methods. distances b type but it radial bas he middle. creening jections . ximated is then nding to e centre of such respecres preperforme measobviousn particand the Spatial between t is also sis func- 313 9.1 Non-Convex Hulls for Engineering Applications Distance based models can be evaluated comparably fast using range-searching. Range searches require pre-processing the data in binary tree structures, and hull evaluation queries can use orthogonal range searches can be performed with computational complexity of order (log ) for fixed input dimension, see e.g. [4]. Although spatial hulls will typically be evaluated comparably fast, these methods will naturally perform poorly if the measurement data is close to the infeasible region. This is particularly the case for measurement procedures where limit violations will cause re-approaching towards the boundary of the feasible region. 2.4 The Surface Interpolation Model A commonly used hull is the surface interpolation model introduced for objects reconstructed from 3D scanner data, see e.g. [5]. The idea of the surface interpolation model is to model a function that rapidly declines at the border points of the feasible region. It is commonly used to model objects from 3D scanner data. Assume that each data point ∈ can be classified according to the criteria • ∈ is inside the feasible region • ∈ is on the border of the feasible region • ∈ is outside the feasible region In order to model a suitable value function, additional data is then generated on the border of the hull. The value function then interpolates (or approximates) the data given by (ℎ) = +1 ℎ = ∈ ℎ = − ℎ ∈ −1 ℎ = ∈ ℎ = + ℎ ∈ Figure 3 (right) illustrates such a value function. The corresponding hull may correspond to disconnected, non-absorbing regions. The surface interpolation model again suffers from the fact that the data needs to be separated into boundaryand interior points. 2.5 The Conic Hull In the following, the conic hull, which is a non-convex hull which can be computed fast, will be discussed. A test point lies outside the convex hull, if it can be separated from the data by a hyperplane. The test point is outside the conic hull of the data, if there is a locally separating, outward pointing hyper-cone, with separation angle at least . The conic value function is given by ( ) = max − ‖ − ‖ − cos( ) 314 The ou by , Figure Compu parame put dim An abs by then th ing con (middle respon In align hull can then ag connec picted i 2.5.1 C Compu evaluat optimis Conic h tward norm respective re 4: Illustra utation of t eters and i mension. sorbing con e conic hu nic hulls fo e), respect ding to n with seq n be mode gain, this h cts all mea in Figure 4 CPU Comp utation of t tion order sation prog hulls only r 9.1 No mal for the ely. The acc ation of ab and a s the conic v it can be e nic hull can ull is absor or each co ively. Note alway uential me eled as follo ∀ hull model surement 4 on the rig plexity of C the conic is desirabl gram, and t require sub on-Convex H n-th data ceptance r sorbing co sequential value funct evaluated u n be mode ∀ ≤ rbing towar onstant op e that the c ys contain t easuremen ows. Let th ∀2 ≤ ≤ s a single, points. A c ght. Conic Hulls value func e, for exam the solver b-linear CP ulls for Engin point is giv region is g onic hulls w conic hull tion in equ using ( ) eled as foll : = ‖ rds the cen ening ang conic hull the convex nt procedu he outward : = ‖ , connecte contour plo s ction is of mple, if the needs to e PU comple neering Appl ven by , iven by = with = with = uation requ ) floating p ows. Let th − − ‖, ntre . Figu les given is connect x hull. res from F d normal no − ‖ − ed region, ot of the re linear ord e hull is use evaluate th exity if the lications , and the n = 0,1 . (left) and (right) uires no p point opera he outward ure 4 illust by = ted and tha Figure 1, a ormals be g ‖, since the esulting val der, but so ed as a co he function amount of n-th openin = (m re-comput ations for f d normal b trates two (left) and at conic hu a sequentia given by hull subse lue functio ometimes a onstraint du n sufficientl f data is re ng angle iddle), tation of fixed inbe given absorb- = ulls coral conic equently on is dea faster uring an ly often. estricted 315 9.1 Non-Convex Hulls for Engineering Applications to those boundary points , which contribute most to the shape of the hull. Boundary points are those points that are not contained in the hull of the other data points, i.e. ∀ ∈ : ∉ ( \ ) Let ≤ denote the number of boundary points in the set . It is then possible to determine using ( ⋅ ) floating point operations, and a hull evaluation is answered using ( ) floating point operations for fixed input dimension. In the worst case, each data point will lie on the boundary, but for smaller input dimensions the number of boundary points can be significantly smaller. 2.5.2 Adaptive Opening Angles using Infeasible Data The opening angle determines the size of the hull, and therefore the degree of extrapolation. On the one hand, a comparably large hull is desirable, since the hull constraint is less restricting for the optimisation program. On the other hand, infeasible regions may be included if the opening angle is too large. Infeasible measurements can be used to determine the optimal opening angle for each data point. The idea here is to choose the largest conic hull that does not contain any infeasible data, , say. The n-th opening angle that maximises the hull then satisfies the equation ∀n ≤ N: cos( ) = max ∈ − ‖ − ‖. 3 Results This section discusses some results comparing the presented hull methods for engineering applications. First, CPU complexity is compared for growing input dimension. Second, the methods are compared with respect to predictive performance as feasibility classifiers for an engine simulation model. 3.1 CPU Comparison Computation times for the presented methods are shown in Figure 5. The top plots show CPU time for model building, the bottom three show model evaluation computation time, respectively, for the convex hull, the conic hull, as well as a surface interpolation model. Computational complexity of Bates’ model heavily depends on the function used for regression, but the surface interpolation model can be seen as a comparable benchmark. 316 For d = which smaller lation ti hull cal = 16. Similar faster f sion at in the d Figure Figure 3.2 C The aim using f hull que data se summa = 16, the s correspon r dimensio ime can be lculation is r results ho for smaller least subdata dimen e 5a: Hull B 5b: Hull E Classifica m of a feas feasible da eries for in et contains arised in Ta 9.1 No lope of the ds to a q ns, due to e seen to b s faster for old for eva r dimensio -linear. The nsion, but b Build CPU valuation C ation for an sibility mod ata only. F nfeasible m s measure able 1. on-Convex H e conic bui uadratic C the limitat be of sub-q smaller di aluation co ns, but co e surface i become nu complexity interpol CPU comp face interp n Engine S del using a For this rea measureme ements fro ulls for Engin ild CPU is CPU comp tion of non quadratic d imensions, omplexity, w nic hulls c interpolatio umerically c ty: convex lation mode F plexity: con polation mo Simulatio data hull i ason a rea ents for eac m a conve neering Appl close to tw plexity with -boundary dependenc , d ≤ 4, bu where con can be eva on model p complex fo hull (left), c el (right) nvex hull (le odel (right) n Model I is to exclud al data set ch of the d entional no lications wo on the h growing points, the cy on the d t practicall nvex hulls aluated for performs n or large da conic hull ( eft), conic ) de infeasib t has been discussed h on-road Di logarithmi data poin e conic hu data size. ly intractab can be ev any input icely indep ata sizes. (middle), s hull (middl ble measur n used to hull metho iesel. Res c scale, nts. For ll calcu- Convex ble for d valuated dimenpendent surface le), surrements perform ods. The ults are 317 9.1 Non-Convex Hulls for Engineering Applications Table 1: Summary of classification hit-rates for investigated hull methods Convex Hull Conic Hull = Conic Hull = Spatial Model Surface Interpolation Bates Model Hit-Rate 98% 93% 16% 4% 96% 96% As a benchmark, the convex hull was used since this method is the most commonly used standard and is still numerically tractable for eight input parameters. The convex hull performs well, reaching a 98% hit-rate of excluded data. For the conic hull a constant opening angle corresponding to = yields a hit-rate of 93%, and the hull corresponding to = yields a hit-rate of only 16%,. The spatial value function yields a hit-rate of only 4%. The surface interpolation model, as well as Bates’ feasibility model each result in a hit-rate of 96%. These results indicate that the spatial model suffers from the fact that the infeasible measurements are located close to the feasible data. All other proposed methods however take screening centres into account, which results in the effective exclusion of infeasible data. 3.3 Classification for an Engine Simulation Model II Predictive performance can be substantially improved if both, feasible measurements, as well as infeasible measurements are used for hull calculation. To see this, the conic hull corresponding to = , which performed poorly in the previous example, and the convex hull, which performed well, are again compared in a different setting. For the same engine simulation model, a number of measurements are iteratively performed, and at each iteration step yields either a feasible or an infeasible measurement for hull calculation. Three hulls were computed at each step, namely, a convex hull, a conic hull using feasible measurements only, and a conic hull using all data. Predictive performance for both, feasible, as well as infeasible data was computed using a reference data set of 2000 points. Results are shown in Figure 6. 318 9.1 Non-Convex Hulls for Engineering Applications Figure 6: Predictive hull performance using feasible & infeasible data. Hit-rates for feasible reference data (left) and in-feasible data (right). As a benchmark, the convex hull grows steadily with growing iterations, infeasible measurements do not influence hull calculation. As a result, hit-rates for feasible reference data is steadily growing, and hit-rates for infeasible reference data is steadily declining. The conic hull using feasible measurements is always larger than the convex hull, and the entire reference set is covered fairly fast. The conic hull using infeasible measurements does not grow as rapidly. The hit-rate for feasible reference data is always larger than that achieved by the convex hull, which indicates that the conic hull is still larger than the convex hull. For larger iterations, it can be seen that the conic hull achieves a better hit-rate than the convex hull, both, for feasible, as well as infeasible reference data. 4 Summary & Conclusion Non-convex hull methods have been investigated in order to model the feasible region of internal combustion engines. Opposed to classifiers, hull methods are in general trained using data from one single class only and are therefore often chosen as suitable feasibility models when in-feasible measurements are not available. In this work, several hull approaches have been discussed in the context of classifying highdimensional feasible regions in engineering. Many approaches that have been introduced in the literature, or are used in practice, involve convex hull algorithms or triangulations, and are numerically intractable for dimensions above ten. The investigated models include spatial hulls, Bates’ feasibility model, and the surface interpolation model. Additionally, the conic hull was discussed. This model is suitable to re-construct some characteristics of standard screening processes, and computation is performed fast. 319 9.1 Non-Convex Hulls for Engineering Applications Conic hulls, the surface interpolation model, as well as Bates’ model performed well on real data, spatial hulls are not recommended if the infeasible regions are close to feasible measurements. Literature [1] Bates, R. A., Wynn, H. P., & Fraga, E. S. (2007). Feasible region approximation: a comparison of search cone and convex hull methods. Engineering Optimization , 39 (5), 513-527 [2] Bates, R. A., & Wynn, H. P. (2004). Modelling Feasible Design Regions Using Lattice‐based Kernel Methods. Quality and Reliability Engineering International , 20 (2), 135-142. [3] Chazelle, B. (1993). An optimal convex hull algorithm in any fixed dimension. Discrete & Computational Geometry , 10 (4), 377-409. [4] Chazelle, B. (1990). Lower bounds for orthogonal range searching: I. The reporting case. Journal of the ACM (JACM) , 37 (2), 200-212. [5] Carr, J. C., Beatson, R. K., Cherrie, J. B., Mitchell, T. J., Fright, W. R., McCallum, B. C., & Evans, T. R. (2001, August). Reconstruction and representation of 3D objects with radial basis functions. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques (pp. 67-76). ACM. [6] Edelsbrunner, H., Kirkpatrick, D., & Seidel, R. (1983). On the shape of a set of points in the plane. IEEE Transactions on information theory , 29 (4), 551-559. [7] Edelsbrunner, H., & Mücke, E. P. (1994). Three-dimensional alpha shapes. ACM Transactions on Graphics (TOG) , 13 (1), 43-72. [8] Getz, W. M., & Wilmers, C. C. (2004). A local nearest‐neighbor convex‐hull construction of home ranges and utilization distributions. Ecography , 27 (4), 489- 505. [9] Getz, W. M., Fortmann-Roe, S., Cross, P. C., Lyons, A. J., Ryan, S. J., & Wilmers, C. C. (2007). LoCoH: nonparameteric kernel methods for constructing home ranges and utilization distributions. PloS one , 2 (2), e207. [10] Kowalczyk, M.: Optimierte schnelle Motorvermessung mit Online-Methoden zur Bestimmung von statischen und dynamischen Modellen. Abschlussbericht, FVV- Heft 987, 2013 320 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning Mark Schillinger, Kadir Mourat, Benjamin Hartmann, Carola Eckstein, Martin Jacob, Ernst Kloppenburg, Oliver Nelles Abstract The usage of data-based models is state of the art in modern engine calibration processes. On the one hand, calibration parameters can be optimized using engine models, resulting in improved solutions. On the other hand, automated optimization using data-based models can decrease manual tuning effort and thereby raise the efficiency of calibration processes. Yet, acquiring the necessary data is not trivial: Particularly for special operating modes like rich diesel combustion, the standard design of experiments (DoE) methods comprise a high risk of critical limit violations. The step towards global models, i.e., a variation of engine speed (nmot) and load, makes the situation even more severe. To meet these challenges, Online-DoE with Constraint Modeling (ODCM) has been developed and evaluated at Robert Bosch GmbH and Bosch Engineering GmbH. In this contribution, the method is further developed. It is applied for global measurements on rich diesel combustion. An optimization of calibration parameters is pursued using the resulting data-based models. Furthermore, two future enhancements of ODCM are presented. The first approach employs a discriminative model instead of the binary classifier used in ODCM. This further reduces the number of limit violations during the measurement process. The second approach is an active learning method based on Gaussian Process (GP) regression models. It optimizes the placement of the measurement points online during the test run, which yields an increased model accuracy. To get a better insight, the methods are evaluated on a high pressure fuel supply system of a gasoline engine in a test vehicle. Kurzfassung Die Verwendung datenbasierter Modelle ist Stand der Technik in der Motorapplikation. Die modellbasierte Optimierung von Applikationsparametern führt einerseits zu besseren Ergebnissen, andererseits reduziert sie den Arbeitsaufwand des Applikateurs. Die notwendige Vermessung des zu modellierenden Systems ist jedoch nicht trivial. Insbesondere bei der Vermessung von Sonderbetriebsarten, wie dem für den Dieselmotor untypischen Fettbetrieb, resultiert aus der klassischen Versuchsplanung ein hohes Risiko kritischer Grenzverletzungen. Mit dem Schritt zu globalen Modellen, d.h. der Variation von Motordrehzahl und Last, wird die Situation noch kritischer. 321 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning Um diesen Herausforderungen gerecht zu werden, wurde bei der Robert Bosch GmbH und Bosch Engineering GmbH das ODCM-Verfahren (Online-DoE with Constraint Modeling) entwickelt und evaluiert. In diesem Beitrag wird die Methode weiterentwickelt und auf den unterstöchiometrischen Betrieb am Dieselmotor angewandt. Im Anschluss wird, anhand der resultieren datenbasierten Modelle, eine Optimierung der relevanten Applikationsparameter vorgenommen. Weiterhin werden zwei Erweiterungen des ODCM-Verfahrens vorgestellt. Der erste Ansatz verwendet ein diskriminatives Modell statt des binären Klassifikators. Dadurch wird die Anzahl der Grenzwertverletzungen bei der Vermessung weiter reduziert. Der zweite Ansatz ist ein aktives Lernverfahren unter Verwendung von Gaußprozess-Regressionsmodellen. Dieses optimiert die Platzierung der Messpunkte im Versuchsraum online und erhöht damit die Modellqualität. Beide Methoden werden am Hochdruckkraftstoffversorgungssystem eines Ottomotors in einem Testfahrzeug evaluiert. 1 Introduction Online DoE with Constraint Modeling (ODCM) allows collecting data for global models without any prior knowledge of the system’s limits in the input space. Defining the supervised signals and their maximum and minimum values is sufficient. At the same time system limits are modeled without any supplementary measurements. The number of limit violations is substantially reduced; critical limit violations can be avoided completely. This paper further develops the ODCM algorithm first presented in [1]. In the first part ODCM for the rich diesel combustion, required for nitrogen oxide storage catalyst (NSC) regeneration, is examined. Global DoE is enabled even in rich mode. Up to now, it was very demanding to apply DoE methods to rich diesel combustion due to system limits that strongly depend on the operating point. The manual determination of operatingpoint-dependent limits involves unreasonable effort. On top of that, critical limit violations can occur with little or no prior notice and cause severe damage to the test engine and the measurement system. Thus, methods which help to avoid critical limit violations are of major importance. The method is evaluated on an engine test bed. Eleven input parameters besides engine speed and load are varied. This corresponds to a state-ofthe-art calibration project. The results are discussed and used for the optimization of calibration parameters. In the second part, two future enhancements of ODCM are presented. The first approach, a so called discriminative model, further reduces the number of infeasible measurement points. An improved model for the system boundaries is used. The second step, Active Learning, provides an online creation of the measurement samples in the input space, aiming at an optimized distribution and thus further improved model quality. The approaches are evaluated at the high pressure fuel supply system of a test vehicle with gasoline engine and compared to the original ODCM method. This contribution is the result of a cooperation between Robert Bosch GmbH and Bosch Engineering GmbH. It is organized as follows: First, the fundamentals of ODCM and the experimental setups are presented. In Section 3, a global DoE method for rich diesel combustion is presented and evaluated. Furthermore, an optimization of calibra- 322 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning tion parameters is conducted. This section is based on a project at the business unit Diesel Systems, supported by the Corporate Sector Research and Advance Engineering (CR). Section 4 introduces future developments of the ODCM method. These are benchmarked against the original algorithm. The theoretical research on this topic was pursued by CR, see [3], before it was advanced and applied to world test vehicle at Bosch Engineering. The contribution closes with conclusion and outlook in Section 5. 2 Fundamentals In this section, the fundamentals of ODCM and the two systems considered in this paper are presented. 2.1 Online Design of Experiment with Constraint Modeling Online Design of Experiment with Constraint Modeling aims at reducing the number of infeasible points during a DoE measurement. Additionally, it reduces the effort for the design of experiment. The algorithm was first presented in [1]. During measurement a classification model is permanently trained and updated The input space is automatically explored with the Online-DoE strategy Measurement points versus real limits (previously unknown) x 1 x 2 Space-filling DoE (e.g. Sobol distribution) non-drivable points are skipped the whole input space is covered x 1 x 2 x x x x x x x xx x x xx xx x x x x x x x x x x x x xx x x x x x x x x x x x x x x xx x x x x x x x x x x x x 1 x 2 x x x x x x x x x x x x x x x x x x x x x xx x x x x x x x x x x x x x x x x xx x x x x x x x x x x x x x x x x 1 x 2 x x x x x x x x x xx x x x x x x Figure 1: ODCM procedure. Planned DoE points are denoted by dots, feasible measured points by crosses and infeasible points by diamonds. The figure is taken from [1]. The algorithm works as follows (compare Figure 1 for a 2D example): In the first step, a space filling DoE is generated by the calibration engineer. For that purpose several options like sobol sequences or latin hypercube plans exist. Contrary to the classical approach, it is not necessary to manually exclude infeasible points from the plan beforehand. Once the measurement is started, ODCM successively provides the test bench automation with points to be measured. ODCM starts at a safe point and chooses the next points with increasing distance to that safe point. For each point, the test bench automation determines and returns the point’s feasibility. A point is feasible, if no limits of the system under test are exceeded during the measurement. For example, when taking measurements at a combustion engine, certain temperatures or pressures must not exceed predefined limits. The information whether a point was feasible or not is used as training data for a binary Gaussian Process classifier. The classifier uses this information to predict the feasibility of future planned measurement points. If the predicted feasibility is below a given threshold, the point is skipped and the algorithm continues with the next point. By applying ODCM, the number of infeasible points attempted to measure can be reduced significantly. Thereby, the stress on the system under test and the risk of dam- 323 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning ages or emergency shutdowns during automated measurement runs can be substantially reduced. Furthermore, the number of points falsely excluded by the calibration engineer is minimized. 2.2 Experimental Setup of the Test Engine An engine test bed is used for the experiments. The engine’s electronic control unit (ECU) settings are set via ETAS INCA. ODCM is executed in the background. The connection between ODCM and INCA is realized with a tool called ODCM2INCA, which is implemented within INCA as INCA-Instrument. The test engine is a modified diesel engine originally certified to reach Euro V emission levels. It is a four-cylinder engine with less than 1.8 l displacement and around 100 kW rated power. This engine is used for experimental purposes at Robert Bosch GmbH. The turbocharged engine is equipped with a cooled low-pressure and a high-pressure loop exhaust gas recirculation (EGR) system and a Bosch CRS2-22 common rail fuel injection system. The turbocharger has a variable geometry. The low-pressure loop EGR system was inactive during the experiments. The fuel injection system provided two pilots, a main and a post injection for the investigated engine operation points. The engine is also equipped with a throttle valve just downstream of the intercooler. The engine has a close-coupled diesel oxidation catalyst (DOC) and diesel particulate filter (DPF) downstream of the turbine. By default the engine possesses a NSC. To prevent damage to the catalyst a DOC is mounted instead. The experiments are not affected by this setup, since for the investigation only the raw exhaust gas composition is of relevance. The exhaust gas was sampled downstream of the DOC. For determining the filter smoke number (FSN) an AVL Smoke Meter was used. Gaseous components were measured with a Horiba Mexa7100D exhaust gas analyzer. An overview of the analyzers is given in Table 1. Table 1: Overview of exhaust gas analyzers. exhaust gas component measurement principle unburned hydrocarbons (UHC) flame ionization detector NO X chemiluminescence carbon monoxide (CO) nondispersive infrared analyzer oxygen (O 2 ) magneto-pneumatic analyzer Exhaust gas temperature was measured with a type K thermocouple. A Kistler transducer was mounted in a glow-plug adapter in order to measure cylinder pressure. All measured data, except FSN, is recorded time-discretely. The generated result is a continous signal. For most of the data the mean value is calculated. Exceptions are temperatures and O 2 concentration. The temperatures tend to rise during rich diesel combustion. Here the maximum value is chosen. The O 2 concentration drops continously until it reaches a minimum. Here the minimum value is chosen. 324 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning 2.3 The High Pressure Fuel Supply System Injection Time Rail Pressure Engine Speed Actuation Figure 2: Sketch of the HPFS system’s main components, inputs (continuous lines), and output (dashed line). The figure is taken from [5]. The approaches for future developments of the ODCM algorithm, presented in Section 4, are evaluated at the high pressure fuel supply (HPFS) system of a 1.4 L fourcylinder gasoline engine in a test vehicle. The HPFS system features three input and one output signal, as shown in Figure 2. For this evaluation, the injection time was set by the ECU, in order to prevent extinguishing the combustion or damaging components. Furthermore, all measurements are executed with no load, therefore no vehicle test bench is necessary. With these simplifications, two inputs remain, the engine speed and the fuel pump actuation (MSV). The rail pressure is the output to be modeled and the supervised variable. It must not exceed a given limit. 3 Global DoE via ODCM for Rich Diesel Combustion In this section a further development of the ODCM method is presented. Global engine models, using this method, are generated and evaluated for relevant outputs. An optimization regarding the output parameters is performed and evaluated. 3.1 Methodology Understoichiometric combustion in a diesel engine is a difficult operating condition: High combustion temperature, cylinder pressure or emissions can damage or even destroy the test engine and test bed equipment. As a result, in order to prevent critical limit violations, the input space in which measurements can be taken is extremely limited. In the current calibration process, the input space is carefully scanned by calibration engineers, which is time consuming and often only covers a small part of the feasible space. Figure 3 illustrates the complexity of these boundaries. On the left the feasible input space for 1136 mbar boost pressure is shown. And on the right the feasible input space for 1095 mbar boost pressure is shown. The change of boundaries is explicit. In order to reduce engineering costs and test bed time an ODCM-based methodology was developed. ODCM utilizes data-based classification in order to create an online model of the complex system boundaries. The method enables safe measurement of space-filling global DoE plans, i.e. engine speed and load can be varied throughout the 325 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning Figure 3: Feasibility for rich diesel combustion at 1000 rpm and 4 bar over air mass and EGR rate. On the left: 1136 mbar boost pressure; On the right: 1095 mbar boost pressure. Red: feasible area; Blue: infeasable area. DoE. Changes in ODCM were made regarding the training of model parameters. Those changes account for the different impact of input parameters on feasibility. To render measurements efficiently, the measurement points are arranged in a block design. For rich diesel combustion the measurement points are grouped by engine speed and load. Since changing these parameters results in a significant change of exhaust gas temperature, stabilization time between two measurements and thus measurement time rises. With this arrangement measurement of a single point takes approximately 90 to 120 s. In contrast to common DoE procedures, where the input space is restricted by the range of each input parameter, with ODCM the input space is rather formed by setting constraints on outputs. Fundamental to the procedure is the right choice of these constraints. By choosing the constraints carefully critical limit violations can be avoided. Instead of restricting system limits only, constraints can also be used to exclude certain areas of the input space which are not relevant for the investigation: e.g. O 2 concentration > 1.2 %. This can avoid unnecessary measurements. Some examples for different constraints are given in Table 2. Table 2: Examples for different constraints. criterion type value temperature before turbine (T3) upper limit 800 ∘ C maximum cylinder pressure upper limit 170 bar air-fuel equivalence ratio upper limit 1 O 2 concentration upper limit 1.2 % quantity of post injection (PoI) lower limit 2 mg/ stroke 326 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning Since generating the DoE plan is done without any prior knowledge, a very large number of unfavourable parameter combinations can arise. In the following, a particularily interesting case is described. Measurements of parameter combinations where a very low air mass was chosen have shown a tendency to worsen model quality. A closer look has revealed that the quantity of the post injection is reduced by the lambda controller to regulate the set air-fuel equivalence ratio value. The result is a rich mode without a post injection. Emissions tend to change over the course of the measurement resulting in inaccurate estimations. FSN tends to rise. Since this kind of operating mode is no favorable calibration, the classifier should mark these measurements as infeasible. The problem is solved by setting a constraint for the PoI. A lower limit for the PoI quantity is set. When the quantity drops below this threshold a limit violation occurs and the measurement is marked as infeasible. Model quality is improved significantly without any further effort. 3.2 Modeling For demonstration purposes the procedure was tested on the aforementioned test engine. Besides engine speed and brake mean effective pressure (BMEP), eleven input parameters were varied analogous to a state of the art development project, see Table 3. A space-filling distribution in combination with ODCM was used in order to generate the measurement points. Table 3: Variation range of input parameters. ECU parameter variation range min. max. engine speed 2000 rpm 2800 rpm brake mean effective pressure 9 bar 14.5 bar air mass 450 mg/ stroke 720 mg/ stroke boost pressure 1300 mbar 2400 mbar rail pressure 1200 bar 1900 bar exhaust gas recirculation (EGR) rate 0 % 16 % air-fuel equivalence ratio 0.92 0.95 start of energizing (SoE) main injection (MI) 20 °CA BTDC 6 °CA BTDC start of energizing post injection −54 °CA BTDC −74 °CA BTDC quantity pilot injection 1 (PI1) 0.8 mg/ stroke 3.5 mg/ stroke duration between SoE PI1 and SoE MI 500 μs 2000 μs quantity pilot injection 2 (PI2) 0.8 mg/ stroke 3.5 mg/ stroke duration between SoE PI2 and SoE PI1 500 μs 2000 μs 327 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning The modeling was performed using ETAS ASCMO. With this software, Gaussian Process (GP) models were generated for seven relevant outputs - UHC, CO, O 2 , FSN, NO X , T3 and noise. For each model about 2100 measurement points were used. To assess the quality of the GP regression models, the root mean square error (RMSE) 1 and the normalized RMSE (NRMSE) 2 via leave-one-out cross-validation (LOOCV) on 𝑛 training data samples 𝑦 𝑖 ∈ {1, ..., 𝑛} are utilized. For further verification of the model prediction the RMSE of a large set of test data is used, which is not contained in the training data set. The results are shown in Table 4. Table 4: Resulting NRMSE and RMSE of the training data and RMSE of the test data. output NRMSE training data RMSE training data RMSE test data UHC 7.86 % 769 ppm 775 ppm CO 4.37 % 933 ppm 952 ppm O 2 5.75 % 0.066 % 0.069 % FSN 14.42 % 0.418 0.420 NO X 5.70 % 45.8 ppm 45.8 ppm T3 1.14 % 8.55 ∘ C 8.9 ∘ C noise 1.07 % 0.99 dB(A) 1.03 dB(A) The results show that the LOOCV RMSE and NRMSE of the training data are in a very satisfying range. The NRMSE of the FSN is slightly higher, but still reasonable, with 14.42 %, which is due to the higher uncertainty of the measurement method. Furthermore the LOOCV RMSE of the training data and the RMSE of the test data correspond very well: Using ODCM it was possible to generate high quality rich mode models without any prior knowledge of the system and its limits. A further reduction of measurement points down to 1500 points does not lead to a significant decrease of model quality. 3.3 Optimization This part of the paper discusses a global optimization that has been performed using ETAS ASCMO on the aforementioned models. Aim of the optimization is to generate a feasible rich mode calibration for the investigated engine operation range. The optimization criteria are listed in Table 5. To reduce computation time only CO was maximized. For the practical applicability of the optimization results, beside the feasibility criteria, a smooth mapping is of importance. In order to realize this, smoothness criteria are used. These determine the maximal slope of an input parameter over engine speed and load. Usually rich mode calibration targets dictate CO concentration of 2 to 3%, HC concentration of less than 0.8% and O 2 concentration of less than 1%. NO X emissions usually are of less importance. Targets regarding T3, noise and soot are to be met. The feasibility model generated by ODCM was used as constraint to the optimization. Thanks to 1 RMSE = √ 1𝑛 ∑ 𝑛𝑖=1 ( ̂ 𝑦 𝑖 − 𝑦 𝑖 ) 2 2 NRMSE = RMSE 𝑦 328 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning the ODCM model all optimization results could be repeated and confirmed with the real engine. Table 5: Optimization criteria. criterion type value CO maximize - FSN hard upper bound map (2.2 to 3.4) T3 hard upper bound 760 ∘ C noise hard upper bound 90.5 dB(A) feasibility hard lower bound 0.8 The optimization results were applied on the test engine and measurements of the relevant outputs were taken. The results for CO concentration and O 2 concentration are illustrated in Figure 4 and Figure 5. Figure 4: Map of CO concentration over engine speed and load. As the figure shows, emission targets with respect to CO concentration are met in all operating points. The O 2 concentration is below 1 % for most operating points with a maximum value of 1 % at 2550 rpm and 13.5 bar. Whereas the maximal HC concentration amounts to 8496 ppm at 2050 rpm and 14.5 bar. Targets regarding FSN, T3 and noise are met for most operating points. FSN rises up to 4.2 for high engine speed and load, but is still acceptable. The optimization shows very good results. On the one hand, emission targets are met sufficiently. And on the other hand a feasible rich mode calibration is possible. 329 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning Figure 5: Map of O 2 concentration over engine speed and load. 4 Future Development of ODCM In this section, two enhancements of the ODCM algorithm are presented: replacing the binary classifier by a discriminative model (Section 4.1) and using Active Learning for an optimized design of experiment (Section 4.2). Both methods are combined to a safe Active Learning approach and evaluated at the HPFS system in a test vehicle (Section 4.3). 4.1 Discriminative Model As described earlier in Section 2.1, the ODCM algorithm features a binary Gaussian Process classifier. The classifier is used to predict whether a future measurement point is feasible or infeasible. Only points with a sufficient probability of feasibility are handed over to the test bench automation system for measurement. The classifier is trained using the binary information if a measured point was feasible or not. There is nothing in between, thus a very critical point hardly feasible and an absolutely safe point give the same information to the classifier. This also means that the classifier needs multiple infeasible training points, before it will predict any point as infeasible. The discriminative model makes use of additional training information. It is again a Gaussian Process model, but capable of using either both continuous and binary training data, or, alternatively, continuous training data only. The data used for training is a calculated inverse risk or health level of the system. This inverse risk incorporates 330 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning all signals relevant for the monitoring of the system, e.g. pressures or temperatures. It is assumed that these signals provide not only a binary information about a measurement point’s feasibility, but also tell the distance of a point to the boundary. For example, a supervised temperature with an upper limit gives such additional information. If the temperature is low, the corresponding inverse risk is high. If the temperature is almost at its limit, the inverse risk is very low. And if the temperature exceeds its limit, the inverse risk is below a critical threshold, marking this measurement point as infeasible. If multiple signals have to be monitored, their output is used to calculate a common scalar inverse risk. Using a hybrid classifier, it is even possible to incorporate supervised signals which only provide binary output together with continuous ones. This can be achieved by using a special non-gaussian model likelihood, as presented in [3]. The main benefit of using a continuous or hybrid discriminative model is that the boundary can be estimated before any infeasible points have been measured. This results in a further decreased number of infeasible points measured by the algorithm, as will be shown in Section 4.3. Even boundary exploration without any limit violations is possible. In case of the HPFS system, the rail pressure is the supervised variable. The higher the pressure at an input point, the lower is the inverse risk used as training data for the discriminative model. As the pressure can be measured in the whole operating range, a solely continuous discriminative model is used. 4.2 Active Learning Active Learning is a discipline of machine learning. This technique addresses problems where the acquisition of labeled training data is very expensive. Its basic idea is that a learning algorithm selects unlabeled data from a pool and requests an “oracle” to label this data. See [4] for an overview of different Active Learning approaches. For example, in case of stationary measurements at an engine test bench, the unlabeled data are input points for the engine. They can be generated at nearly no cost. Labeling this data means taking measurements at the test bench, which is comparatively expensive and time consuming. The goal of letting the learning algorithm choose the data to be labeled, i.e. iteratively optimizing the position of the input points, is to improve the quality of the resulting model. Alternatively, the number of measurement points necessary to obtain a required model quality could be reduced. Furthermore, it is not necessary to design a measurement plan in advance. In this contribution, an Active Learning approach based on Gaussian Process models and a differential entropy criterion is used. Thereby, a GP regression model is not only trained after the complete measurement process is done, but after each single measurement point. The next sample is chosen such that the entropy of the model increases as much as possible. It can be shown that this is equivalent to choosing the next point where the standard deviation of the model has its maximum [3]. Optimizing the measurement points online, opposed to planning them in advance, and utilizing the learned model properties are the major differences between Active Learning and classical DoE methods. 331 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning The approaches for a discriminative model and Active Learning can be combined to a safe Active Learning (SAL) algorithm [2, 3]. Thereby, two models are trained in parallel during the measurement. Model 𝑓 represents the system’s output. Model 𝑔 is the discriminative model used to separate the feasible from the infeasible area of the input space 𝕏 . The next measurement point 𝑥 𝑖+1 is chosen sequentially by optimizing 𝑥 𝑖+1 = argmax 𝑥 ∗ ∈𝕏 𝜎 𝑓 ∗ (𝑥 ∗ ) (1) s.t.: 𝜇 𝑔∗ (𝑥 ∗ ) − 𝜈𝜎 𝑔∗ (𝑥 ∗ ) ≥ 0, (2) where 𝜎 𝑓 ∗ is the predicted standard deviation of model 𝑓 , 𝜇 𝑔∗ and 𝜎 𝑔∗ are the predicted mean and standard deviation of model 𝑔 and 𝜈 is a confidence parameter. This parameter is used to tune how careful the exploration should be. The objective (1) is to find the point with maximum output model variance, constrained by the minimum probability of feasibility (2) based on the discriminative model. 4.3 Evaluation This evaluation focuses on the comparison of ODCM and SAL. Both methods were applied to the HPFS system in a test vehicle. This system is less complex compared to the rich diesel combustion and thus more suitable for a first evaluation of the SAL algorithm. In future work, more challenging applications could be tackled. 0 10 20 0 50 100 number of samples sensitivity (%) 0 10 20 0 50 100 number of samples specificity (%) 0 10 20 0 50 100 number of samples NRMSE (%) Figure 6: Sensitivity, specificity, and NRMSE for SAL (continuous lines) and ODCM (dashed lines). Sensitivity and specificity were calculated using a set of 105 space filing test points. The NRMSE was calculated using the 68 feasible test points only. After 25 measured samples, the SAL approach obtains a sensitivity of 95.6 %, a specificity of 100 % and an NRMSE of 5.9 %. ODCM results in a higher sensitivity of 98.5 %, a lower specificity of 86.5 % and a higher NRMSE of 6.4 %. Good measures to compare the classifiers are sensitivity and specificity. 3 They describe the fraction of correctly classified feasible or infeasible test points, respectively. In the left and middle plot in Figure 6, sensitivity and specificity are plotted for ODCM and SAL. The different strategies of the two classifiers are clearly visible. The binary classifier of ODCM is tuned in order not to exclude any points before it got sufficient infeasible training data. Therefore, its specificity starts at zero and raises not before the 13 th sample, which is the first infeasible one. On the other hand, its sensitivity is always close to 100 %, which means that nearly no feasible test points are excluded wrongly. SAL’s 3 sensitivity = number of points correctly classified as feasible number of all feasible test points , specificity = number of points correctly classified as infeasible number of all infeasible test points 332 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning 0 20 40 60 80 100 MSV (mm 3 ) 1,000 2,000 3,000 0 20 40 60 80 100 nmot (rpm) MSV (mm 3 ) 1,000 2,000 3,000 nmot (rpm) 1,000 2,000 3,000 nmot (rpm) Figure 7: Measured feasible (large crosses) and infeasible (large diamonds) points in case of SAL (upper plots) and ODCM (lower plots). For ODCM, the planned points are indicated by plus signs, the skipped points by squares. The algorithmic steps are shown for 5, 12 and 25 points, respectively. The current estimated boundary is plotted as red line. The small crosses and diamonds are feasible and infeasible test points. discriminative model pursues another strategy: At the beginning, only a small region of the input space is considered feasible, resulting in a very low sensitivity. Afterwards, the input space is explored gently, which results in an almost continuously increasing sensitivity. As the algorithm is tuned to be rather cautious, the specificity constantly remains at 100 %, i.e. no false positives occur. These different strategies can also be seen in the evolution of the border in Figure 7. In average over multiple test runs, SAL queried 0.5 infeasible samples in 25 planned points, while ODCM required 3 infeasible points. To compare the quality of the resulting models, the normalized root mean square error (NRMSE) is examined. The right plot in Figure 6 shows the NRMSE of models trained using different amounts of training points from SAL and ODCM. As the figure shows, the model quality evolves comparatively using both sampling methods. Finally, the NRMSE resulting from SAL is slightly better. Summing up these results, two findings can be concluded: While ODCM is already able to massively decrease the number of limit violations during a measurement compared to classical DoE methods, the discriminative model can even outperform ODCM. Almost no infeasible points are sampled using SAL. On the other hand, the Active Learning algorithm was not able to increase the model quality significantly. There is the advantage, that no initial DoE is required, but, given that its implementation is more complex and offers less degrees of freedom compared to an offline DoE, the presented AL algorithm 333 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning did not meet the expectations. This is probably due to the limitations evoked by the discriminative model. More details regarding the used SAL algorithm can be found in [2]. For the ODCM method, it is referred to [1]. 5 Conclusion and Outlook In this paper, first a further development of the ODCM algorithm and its application to the rich diesel combustion was presented. Global DoE models were successfully trained for relevant outputs. A global optimization was performed and evaluated. Emission targets were met sufficiently. A feasible rich mode calibration was provided without using any prior knowledge of the system and its limits. In the second part, two approaches for future enhancements of ODCM were described and evaluated at a test vehicle. Especially the discriminative model proved its capability to further reduce the number of limit violations during the measurement. Active Learning resulted only in a slight improvement of model quality, compared to a space filling DoE. In future research, both approaches could be further improved. An implementation and evaluation at engine test benches, e.g. for the rich diesel combustion, and more complex in-vehicle systems would make their advantages available to the calibration engineers. As for the application of ODCM for rich diesel combustion, in the next step, measurements in the complete operating range could be pursued. Possible challenges are e.g. misfires at low loads and the modeling of the system in the complete operating range utilizing standard GP models. Acknowledgments The authors would like to thank everyone who supported this research. In particular Patrick Schelkle regarding test bench automation, Jörg Pichler for the ODCM2INCA implementation, as well as Jens Schreiter, Mona Meister, Duy Nguyen-Tuong, and Benedikt Ortelt for their SAL support. References [1] Benjamin Hartmann et al. “Online-methods for engine test bed measurements considering engine limits”. In: 16th Stuttgart International Symposium . Wiesbaden: Springer Fachmedien, 2016. DOI: 10.1007/ 978-3-658-13255-2_92. [2] Mark Schillinger et al. “Safe Active Learning of a High Pressure Fuel Supply System”. In: 9th EUROSIM Congress on Modelling and Simulation . Oulu, Finland, 2016. DOI: 10.1109/ EUROSIM.2016.137. [3] Jens Schreiter et al. “Safe exploration for active learning with Gaussian processes”. In: Machine Learning and Knowledge Discovery in Databases . Springer, 2015, pp. 133-149. DOI: 10.1007/ 978-3-319-23461-8_9. 334 9.2 Modern Online DoE Methods for Calibration - Constraint Modeling, Continuous Boundary Estimation, and Active Learning [4] Burr Settles. Active Learning Literature Survey . Computer Sciences Technical Report 1648. University of Wisconsin-Madison, 2009. [5] Nils Tietze et al. “Model-based calibration of engine controller using automated transient design of experiment”. In: 14th Stuttgart International Symposium . Wiesbaden: Springer Fachmedien, 2014. DOI: 10.1007/ 978-3-658-05130-3_111. 335 9.3 Model-based iterative DoE in highly constrained spaces Stefan Scheidel, Marie-Sophie Vogels Abstract The modern powertrain calibration process is characterized by shorter development cycles and an increased number of vehicle variants. Additionally, cost reduction and resource efficiency are two major goals for the management. The base calibration as starting point of the whole process is very costly due to the high amount of testing time. Therefore, it is essential to reduce the time on the test facility while maintaining or even improving the quality of the calibration. DoE testing and model-based calibration is a proven method to deal with the high dimensional optimization problems that occur during the calibration of a modern engine. State-of-the-art DoE methods generate a test plan based on an optimality criterion (space-filling, D-optimal, etc.). When measuring that test plan on the test bed, some variation points may not be reached due to limit violations (various pressures, temperatures, turbine speed, etc.). These limit reactions lead to two unwanted effects: • The design gets distorted and the initially planned optimality is no longer given. • The resulting design space gets highly reduced, especially if a classical DoEscreening (i.e. screening of all variation parameters at once) is used. These drawbacks lead to a reduced model quality and - even worse - to the fact, that possibly the optimum is not part of the design space - and therefore may not be found during the model based optimization process. This paper presents a model-based iterative DoE method that • can be used without any previous knowledge of the system • rather dethan increases the testing time • delivers an optimal design with maximized design space in any highly constrained space • is not based on any hull calculation and therefore independent of the dimensionality of the optimization task. Kurzfassung Der Kalibrierprozess eines modernen Antriebstrangs wird heutzutage von verkürzten Modellzyklen und einer hohen Variantenvielfalt geprägt. Zusätzlich legt das Management erhöhten Wert auf Kostensenken und effizienten Umgang mit Ressourcen. Vor allem die Basiskalibrierung als Ausgangspunkt des Entwicklungsprozesses ist wegen des hohen Zeitbedarfs sehr kostenintensiv. Deshalb ist es hier besonders 336 9.3 Model-based iterative DoE in highly constrained spaces wichtig, die Prüfstandszeit zu verringern und dabei die Qualität der Kalibrierung nicht zu beeinträchtigen oder sogar zu verbessern. DoE und modellbasierte Kalibrierung sind etablierte Methoden, um den hochdimensionalen Optimierungsaufgaben Herr zu werden. Heutzutage basiert der DoE-Ablauf auf einem vor dem Test generierten „optimalen“ Versuchsplan (space-filling, D-optimal, etc). Wenn dieser Plan abgefahren wird, können einige Versuchspunkte aufgrund von Limits (Drücke, Temperaturen, Laderdrehzahl, etc.) nicht erreicht werden. Diese Limit-Reaktionen führen zu zwei ungewünschten Effekten: • Das Design wird verzerrt und die ursprünglich geplante Optimalität ist nicht mehr gegeben. • Der Versuchsraum wird stark eingeschränkt, besonders wenn ein klassisches DoE-Screening verwendet wird, bei dem alle Parameter gleichzeigt verstellt werden. Diese Nachteile führen zu einer reduzierten Modellqualität und dazu, dass ein fahrbares Optimum eventuell nicht im abgedeckten Versuchsraum liegt und damit im Zuge einer modellbasierten Optimierung nur schwer gefunden werden kann. In diesem Paper wird eine modellbasierte, iterative DoE-Methode vorgestellt, die • ohne Vorwissen über das Systemverhalten angewendet werden kann • die Testlaufzeit tendenziell eher verkürzt als verlängert • optimalen Designs und maximiertem Versuchsraum in stark Limit begrenzten Räumen erzeugt • ohne die Berechnung von Hüllen auskommt und damit auch in hohen Dimensionen angewendet werden kann. 1 Introduction Since the introduction of electronic engine controls in the late 1980s, the degrees of freedom in engine calibration have continuously grown. In the 1990s, most of the optimization tasks were still two-dimensional optimization problems and could be solved with “full-factorial testing” and “best point selection”. At the latest with the comprehensive introduction of common rail injection systems for diesel engines in the 2000s, multi-dimensional optimization problems became the daily task of diesel calibration engineers. The introduction of fully variable valve trains, direct injection and turbo-charging for gasoline engines carried a similar complexity into the every day’s work of gasoline calibration engineers as well. The state of the art method to tackle these challenges is DoE testing and model based optimization. The common workflow consists of the subsequent steps of test plan generation, test plan execution, model training and model based optimization. The test plan is generated based on an optimality criterion (space-filling, D-optimal, V-optimal, etc.) within pre-defined input boundaries. These input boundaries are either defined based on experience or as the result of a pre-investigation of limit boundaries. Several strategies for the investigation of limit boundaries have been developed, reaching from simple one factor at a time variations to highly complex screening and hull calculation [1]. 337 9.3 Model-based iterative DoE in highly constrained spaces As even an extensive boundary investigation cannot guarantee the feasibility of all generated measurement points of the DoE plan, the test plan is typically executed while monitoring limit-critical channels. In response to a limit violation, the relevant measurement point is either skipped (non-screening) or re-located to feasible settings near the limit boundary (screening procedure). Without detailed pre-knowledge of the system behavior, the only reasonable screening method is to move all parameters at once from the starting point towards the target point as illustrated in Figure 1. Figure 1: Online screening method starting in a safe center point In high dimensional spaces, pre-exploration of limit boundaries is becoming an increasingly time consuming task, as the number of edges of the variation space - i.e. the extreme combinations of the input variables - is growing exponentially with 2 dim . However, for many use cases, the maximization of the design space to the limit boundaries is crucial for the success of the model based calibration, as the optimum is likely to be found right at the limit boundaries. Practical examples are diesel engine full load calibration (optimal calibration close to limitations of exhaust temperature, turbine speed, peak firing pressure and/ or compressor outlet temperature) or gasoline engine air path optimization (best fuel consumption usually close to combustion stability boundary). Therefore, in order to maximize the design space, in a classical DoE workflow, the input boundaries have to be selected rather wide and hence the number of limit violations during the test grows. A high number of limit-violations during the test run will lead to several negative effects: • The test runtime will be increased • A shutdown of the test bed becomes more likely • The initially planned optimality of the design is distorted • The resulting design space is highly reduced The reduction of the design space is caused by the simultaneous variation of all parameters. If only one of the parameters leads to a limit violation, also the variation of all other parameters is stopped. To improve the design space in these situation, various intelligent limit reaction strategies have been suggested [2]. These strategies can improve the volume of the design space but they cannot fully restore the optimal point distribution (as non-limit 338 9.3 Model-based iterative DoE in highly constrained spaces violating points stay unchanged) and the implementation requires substantial preknowledge of the system behavior. Besides the approach to improve the DoE design based on extensive preinvestigation of limit boundaries and intelligent limit reactions, several online DoE methods have been discussed in recent years [3, 4, 5, 6, 7]. The approach of hull-calculation and test-adaption described in [1] can also be performed online [6], with the same limitations in high dimensions: Given an 11 dimensional hypercube, there are 2048 corners of the space. In those high dimensional spaces, any “hull model” based on some hundreds of measurements will not be able to describe the boundaries sufficiently. Another approach of online DoE methods sets the main focus on the improvement of the model quality of one explicit output modelled with a local model tree, whereby the model quality is determined based on statistical criteria (model error of the local models) [3, 4]. The same approach of improving the model quality of one single output model based on online criteria exists for Gaussian process models [5]. The same criterion of improving the quality of the worst model based on the GPM variance exist in combination with a space-filling criterion, choosing points out of a fixed candidate set [7]. These approaches do not take into account that even the “best” model is of little use for the calibration process, if it does not include the optimal setting for the calibration in its measured input range. Furthermore, methods where the point section is based on a fixed candidate set will reduces the likeliness to find candidates right on the boundaries of an arbitrary constrained variation space even more. 2 Concept of model-based iterative DoE To overcome the before mentioned drawbacks of a conventional DoE in highly constrained spaces, an iterative DoE strategy is suggested. The development of the method is based on the following requirements: • A pre-investigation of the limit boundaries is not required • A very wide selection of the input boundaries is possible without reduction of the quality of the resulting design • The setup and parameterization does neither require expert knowledge about DoE, modelling and optimization, nor pre-knowledge about the system behavior • The strategy can deal with a high number of limit-critical channels and a high number of input dimensions • No assumption about the limit behavior (concave, convex, polynomial, etc.) has to be made • The selection of the iterative points is not restricted to a limited number of candidates and is able to find candidates on the boundaries of the drivable variation space. Existing methods mentioned in chapter 1 involving a “hull model” can be classified as “input driven”, as they try to describe the drivable area based on the input settings of the measurements. The method described in this paper is “output driven”, i.e. based on online modelling of all limit critical responses. The iterative measurement point 339 9.3 Model-based iterative DoE in highly constrained spaces selection is based on a free (i.e. not gridor candidate-based) optimization that ensures a space-filling design within the limit boundaries given by the online models with a maximized design space. In Figure 2, an overview of the workflow is given. Figure 2: Workflow of model-based iterative DoE The start design can be a one factor at a time measurement in all dimensions or any other simple DoE design. That way, the initial search range is determined and a first simple model of the system behavior can be estimated. With this information, the iterative point distribution is started. The point search is a nonlinear optimization task, maximizing the nearest neighbor distance between the new candidate and all already measured points, while staying within the limit boundaries based on the model predictions. This optimization task is performed with a nonlinear solver, which does not depend on any candidate set. After each calculation of a new point, the relevant point is measured, the models retrained and the next point calculated. A stop criterion can be based on model quality and/ or coverage of the design space. The model type used is a neuronal network based on linear base functions with an implementation especially optimized for robust prediction, capability to deal with nonlinear system behavior and fast model-training. 340 9.3 Model-based iterative DoE in highly constrained spaces Figure 3: Comparison of classical DoE-design (left) and model-based iterative DoE design (right) Comparing the resulting designs for an exemplary two dimensional problem with only one limit critical channel given in Figure 3, the advantage of the model-based iterative DoE design becomes obvious: For the classical DoE design, a possible optimum inside the limit boundaries (blue star) is in the extrapolation area of the model, as the design space is highly narrowed due to limit reactions. Thus, a model based optimum search will hardly find that optimum if being limited to the design space. In addition, the information content of the measured points is highly non-uniform culminating in points without any additional information, as due to the limit, points placed in different locations in the initial design actually end up in the very same place. With the iterative model-based DoE method, the design space is maximized and the optimum lies within the interpolation area and can therefore be found in a modelbased optimization process. Furthermore, the model quality is improved since the point distribution is more uniform and therefore efficient, as each measurement point delivers the same amount to the model. 3 Results and Comparison The model-based iterative DoE approach was already applied in different use cases, for local as well as global models and for both gasoline and diesel engines. Based on several plots, calculations and criteria the results will be compared to the outcome of a classical DoE. 341 9.3 Model-based iterative DoE in highly constrained spaces The first application task was the optimization of a full load operating point of a heavy-duty diesel engine. 4 ECU-parameters were varied in rather wide variation ranges. Due to the wide ranges and because the starting point for the DoE variation was anyway already close to several limits, in total 5 channels were used for limitmonitoring (during the classical DoE) and as model based constraint (during the iterative DoE). For the classical DoE, 50 variation points were pre-defined by a spacefilling design algorithm, whereas for the iterative approach, the same number of measurements was used, but split in 18 initial design points and 32 iteratively calculated points. As a first result, Table 1 shows a comparison of the total test runtime, the number of limit violations and the average limit distance. Table 1: Comparison of the test execution in a local full load point Classical DoE Model-based iterative DoE Test duration 3h 51min 4h 3min Limit violations absolute 38 of 50 initial phase: 10 of 18 iterative phase: 14 of 32 Limit violations relative 76% 48% Average limit distance 61% initial phase: 67% iterative phase: 93% The average limit distance is calculated in the normalized space as the ratio of the distance between the center point to the actual measurement point to the distance between the center point and the target point as shown in Figure 4. Figure 4: Explanation of limit distance The additional time needed for the recurrent calculation of the models and the iterative selection of the next point is overcompensated by the time saved by the lower number of limit violating points. In addition, the average limit distance is significantly higher for the iterative approach. This means, even if limits were violated, the distortion of the design is much smaller. This effect is graphically demonstrated in Figure 5, showing the input-vs-input point distribution. 342 9.3 Model-based iterative DoE in highly constrained spaces Figure 5: Measured nput vs. input distribution for both methods Two main advantages of the iterative approach can directly be observed: • The covered area is in almost all directions significantly increased. • The point-distribution is more uniform, high concentrations of measurements in one spot do not occur. This observation is backed by the comparison of the histogram of nearest neighbor distances and the volume of the convex hull displayed in Figure 6. For this comparison, the distance in the normalized space of each point to its nearest neighbor is calculated. An ideal space-filling design would lead to a rectangular distribution. The resulting distribution of the iterative design gets clearly closer to that optimum. Only the points of the start-design show a significant distortion. Besides the resulting point distributions of the two measured designs, also the values of the initially planned conventional design are given. The values of the convex hull volume support the impressions given by the distance distribution: The space in the full load operating point is highly constrained by limits - more than 80% of the initially planned hypercube are not reachable due to limits. The volume of the hull of the initial design plan is therefore several factors higher than the volume of the resulting designs. But still the volume of the resulting iterative design is by 34% higher compared to the conventional approach. 343 9.3 Model-based iterative DoE in highly constrained spaces Figure 6: Nearest neighbor distance distribution and convex hull volume for a local DoE The benefit of these design characteristics for the actual engine calibration is demonstrated by using the models generated with both designs to perform the same modelbased optimization task of reducing the fuel consumption while maintaining a certain level of NOx emissions and not exceeding any component limitations (e.g. maximum cylinder pressure and exhaust temperature). Figure 7 and 8 illustrate that, based on the iterative design, a verified optimum with ~0.4 g/ kWh less fuel consumption could be found. The settings of this optimal solution are outside the design space of the conventional DoE design and therefore can hardly be found. Figure 7: Comparison of optimized BSFC values [g/ kWh] 344 9.3 Model-based iterative DoE in highly constrained spaces Figure 8: Comparison of intersection behavior including design space information Having proven its ability in a rather basic 4-dimensional optimization of a local operating point, the method was applied for the generation of a global model with 9 ECU variation parameters, so in total 11 dimensions. The intended usage of the model was the prediction of real driving emissions in an engine and vehicle simulation environment. For that purpose, the operating range was chosen as wide as possible, reaching for low idle to rated power. Also the variation ranges for the ECU parameters were set up in a way to cover all possible engine operating conditions, including altitude simulation, transient conditions and failure modes (e.g. EGR shut-off, stuck VTG and therefore low boost pressure, high-pressure pump failure causing low rail pressure, etc.) Therefore, the number of limit critical channels used for the model based iterative point distribution was raised to 7: • Intake manifold temperature • Smoke • Exhaust manifold temperature • Peak firing pressure • Turbine speed • Compressor outlet temperature • Mass fraction burned 50% For the given use case of the generation of a universal emission prediction model, the maximization of the design space and an equal point distribution to ensure sufficient model quality in all operating conditions was especially important. The outcome of the model-based iterative DoE was compared to a conventional global space-filling design set up with the very same input boundaries and executed using the same screening procedure. 345 9.3 Model-based iterative DoE in highly constrained spaces As this specific use case does not imply a direct model-based optimization task, the focus of the comparison is on the design space coverage and point distribution. First of all, Table 2 shows a summary of the test execution, giving again a comparison of the total test runtime, the number of limit violations and the average limit distance. Table 2: Comparison of the test execution for a global design Classical DoE Model-based iterative DoE Test duration 91h 43min 93h 7min Limit violations absolute 510 of 1250 initial phase: 48 of 95 iterative phase: 323 of 1155 Limit violations relative 41% 30% Average limit distance 81% initial phase: 78% iterative phase: 94% The basic trends are similar to the previously stated observations in the 4dimensional example: Lower number of limit violations, higher average limit distance and a runtime that is almost unaffected by the additional model creation and iterative point calculation. Figure 9 shows the distribution of the nearest neighbor distances. In areas where the points of a conventional DoE are highly distorted by limit reaction, the benefit of the iterative design is the most obvious. In areas with fewer constraints, the distances are very similar to the values of the conventional design. This effect can be seen as a proof of the efficient implementation of the optimization algorithm used for the maximization of the nearest neighbor distance of each new measurement point. Figure 9: Nearest neighbor distance distribution and convex hull volume for a global DoE 346 9.3 Model-based iterative DoE in highly constrained spaces 4 Conclusion The main drawbacks of the application of conventional DoE designs in highly constrained spaces are the reduction of the design space and the distortion of the design. Existing countermeasures either require a deep pre-knowledge of the system, extensive pre-testing, are candidateor hull-based and therefore limited to low dimensions or only focus on the improvement of the model quality, not on the maximization of the design space. The described iterative model-based approach has proven its ability to generate optimal designs in highly constrained spaces in the same or less time. In two relevant use cases it could be demonstrated that optimizations based on the models generated with this approach lead to better results - results which can hardly be found with models based on conventional DoE data - and in general covers a wider area with the same number of measurement points. Furthermore, the application of the method neither requires deep mathematical knowledge nor pre-testing or pre-knowledge of the system under test. For the future, the prototype implementation of the code will be transferred in the commercial software AVL CAMEO TM , extending the existing “Active DoE” strategy [8, 9] by model based limits. The current “Active DoE” strategy already uses online models for the following purposes: • Iterative point selection based on a combined input-output space-filling criterion in order to improve the quality of all nonlinear output models. • Direct measurement points in the calibration relevant output range, improving the model quality especially where it is needed. • Automatic adaption of the input-range. • Enable online optimization. The integration of model based limits will add the possibility to build more models online, as they are not taken into account for the time-consuming input-output distance calculation, but used to improve the design space coverage in highly constrained spaces and run a more robust test with less limit violations. References [1] Y. Murata, Y. Kato, T. Kanda, M. Sato: Application of Model Based Calibration to Mass Production Diesel Engine Development for Indian Market, Design of Experiments (DoE) in Powertrain Development, Berlin 2015 [2] S. Fritz, H. Hötzendorfer, M. Koller: Design of experiments in large diesel engine optimization, MTZ industrial 04 2014 [3] B. Hartmann, O. Nelles: Adaptive Test Planning for the Calibration of Combustion Engines - Methodology, Design of Experiments (DoE) in Powertrain Development, Berlin 2013 [4] P. Klein, F. Kirschbaum, B. Hartmann, Y. Bogachik, O. Nelles: Adaptive Test Planning for the Calibration of Combustion Engines - Application, Design of Experiments (DoE) in Powertrain Development, Berlin 2013 347 9.3 Model-based iterative DoE in highly constrained spaces [5] B. Raid: A strategy to employ criteria for online location selection when using Gaussian Processes, Design of Experiments (DoE) in Powertrain Development, Berlin 2015 [6] N. Sandmeier, K. Röpke: Efficient online test plan calculation with integrated boundary detection for modern model based engine calibration, Design of Experiments (DoE) in Powertrain Development, Berlin 2015 [7] N. Sandmeier, K. Röpke: Improving the usability and time effort of modern engine calibration tasks by means of an online, model-based approach, 6 th International symposium on Development Methodology, Wiesbaden 2015 [8] Rainer, A.; Koegeler, H.M.: Iterative DoE - improved emission models and better optimisation results within a shortened measurement time. In: International Journal of Powertrains, 2017 [9] Varsha, A.; Rainer, A.; Santiago, P.; Umale, R.: Global COR iDOE Methodology: An Efficient Way to Calibrate Medium & Heavy Commercial Vehicle Engine Emission and Fuel Consumption Calibration, SAE Technical Paper 348 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds Yagiz Dursun, Sebastian Weber, Richard Jakobi, Frank Kirschbaum, Stephan Rinderknecht Abstract This paper presents an approach for an automated adjusting of the road load and tire simulation on powertrain test beds. The application is shown by a practical example. The approach is based on adaptive test design and modeling. To achieve this specific characteristics of the Hilomot modeling approach are used. The objective evaluation of driving maneuvers based on the calculation of criteria from transient signal sequences which are relevant for the longitudinal dynamics behavior opens up the possibility of analyzing and calibrating engine and gearbox functions on (vehicle) drive train test beds. In order to accurately suit the driving behavior of the vehicle on the road to the test bed, a parameterization of the vehicle models in the test bed control system is required. The force between the tire and the road is implemented using the mathematical tire model according to Pacejka. So far, the scalar values of the tire parameters have been iteratively varied during certain driving maneuvers until the deviation to the road measurement is sufficiently low. This approach is difficult because of the interactions between the parameters and therefore very time-consuming. It is shown in the literature that model-based approaches reduce the duration and effort for setting up the test bench with an even better quality of the tire model parameterization. This paper goes one step further: The test planning, modeling and optimization are carried out adaptively during the entire test run. First, a very small initial test plan is applied. The optimal position of the next measurement is computed after each single measurement. On the one hand, this avoids unnecessarily long measuring times and on the other hand, a satisfactory model quality can be ensured. Kurzfassung Vorgestellt wird eine Methodik zur automatisierten Durchführung des Straßenabgleichprozesses auf (Fahrzeug-) Antriebsstrang-Prüfständen und dessen Anwendung anhand eines praktischen Beispiels. Hierbei wird die automatisierte Durchführung des Straßenabgleichprozesses durch modellbasierte Ansätze mit adaptiver Versuchsplanung und Modellbildung realisiert, wozu die spezifischen Eigenschaften der sogenannten Hilomot-Modelle ausgenutzt werden. Die objektive Bewertung von Fahrmanövern durch die Berechnung von Kriterien aus für das Längsdynamikverhalten relevanten transienten Signalverläufen eröffnet die Möglichkeit der Analyse und Applikation von Motor- und Getriebefunktionen auf (Fahrzeug-) Antriebsstrang-Prüfständen. Um das Fahrverhalten des Fahrzeugs auf der Straße realitätsgetreu auf dem Prüfstand abzubilden, ist eine Parametrierung der Fahrzeugmodelle in der Prüfstandsregelung erforderlich. Die Beschreibung der Kraft zwischen Reifen und Fahrbahn wird 349 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds an Antriebsstrang-Prüfständen mit Hilfe des mathematischen Reifenmodells nach Pacejka umgesetzt. Hierbei werden bisher die skalaren Werte der Reifenparameter bei bestimmten Fahrmanövern solange iterativ variiert, bis die Abweichung zur Straßenmessung ausreichend gering ist. Dieses Vorgehen gestaltet sich aufgrund der Wechselwirkungen zwischen den Parametern als schwierig und ist daher mit großem Zeitaufwand verbunden. In der Literatur wird gezeigt, dass modellbasierte Ansätze die Dauer sowie den Aufwand der Prüfstandsinbetriebnahme reduzieren und gleichzeitig die Qualität der Parametrierung des Reifenmodells steigern. Als Weiterentwicklung wird bei der in vorliegendem Beitrag vorgestellten Methode die Versuchsplanung, Modellbildung und Optimierung adaptiv über den gesamten Versuchslauf durchgeführt. Ausgehend von einem vergleichsweise kleinen initialen Versuchsplan erfolgt nach jeder Messung eine neue Planung für die nächste Messung. Dadurch werden zum einen zusätzlich unnötig lange Messzeiten verhindert, zum anderen kann eine ausreichende Modellgüte sichergestellt werden. 1 Overview and Introduction The standard workflow of model-based calibration can be roughly divided into the following steps: test planning measurements modeling optimization calibration map calculation During each of these steps, several decisions have to be made. Depending on the expected process complexity, different model approaches can be considered in a way, that the achievement of adequate model accuracy is ensured and the subsequent optimization can serve the purpose of the calibration of control unit functions. The major challenge in the model-based calibration is to ensure adequate model accuracy. Often the relationship between the modeled objectives and the calibration parameters is not known, because of insufficient experience. If the necessary number and the efficient placement of the measuring points are difficult to estimate an automatic algorithm for test planning and modelling like the HilomotDoE [1] approach provides benefits. A new, suitable application field is the calibration of the longitudinal dynamics of vehicles. The publications [4, 5, 6] show model-based approaches to the calibration of load-change damping function, anti-jerk control and shift operations of automatic transmissions (converters and DCT 1 ). As shown in various publications [7, 8, 9, 10] adaptive test planning approaches can reduce long measuring programs significantly during the calibration of combustion engines. All previously known approaches have in common that, starting from a relatively small initial test plan, after each measurement a new plan for the next measurement takes place. The experimental design is constructed to be adaptive over the entire test run. Even during the calibration of the longitudinal dynamics of vehicles this adaptive test planning methods can reduce the measurement effort significantly. This paper describes a method that exploits special features of Hilomot models. The prerequisite for carrying out driveability calibrations on test beds is the objectification of the subjective perception of drivability. On the basis of signal sequences of 1 DCT: Dual Clutch Transmission 350 9.4 A a trans teristic automa The de ters, wh as targ models timal se mizatio The pre determ tion adj newly d bed to 2 T Hilomo local m a proce linear, The ou membe where the num Approach for sient driving driving be atically. esign of ex hich are to get variable s, such as ettings of t on method. econdition mined on th justment b developed describe th The Hilo otDoE is an model tree) ess is divid local mode Figure 2.1 utput is giv ership func (⋅) is the mber of loc Automated A g maneuve ehavior. On xperiments o be measu es and the polynomia the applica . to achiev e test bed before the method fo he force be motDoE n approach models [1 ded into se eling appro 1: Illustratio ven by su ctions, e -th local cal models Adjusting of th er, criteria n a test be s determine ured. The c e calibratio als, Gaussi ation varia ve the sam and on th beginning or the para etween tire method h for adapt ]. Figure 2 everal sub oach descr on of a loca ummation = model and . he Road Load are calcu ed, maneuv es the com calculated on parame ian proces ables can t me desired e road is a g of the me ameterizati e and road d tive test pla 2.1 shows t bspaces in ribes the p al model n of the loc d (⋅) the d and Tire Sim lated for th vers can b mbinations criteria fro ters as inp ss models o then be fo results fo a successf easuring p on of the m dependen anning bas the basic p that a rela rocess app network tra cal model respective mulation on P he evaluati e carried o of the ap om the mea puts for em or neural n und using or the optim ul road loa hase. This model dep nt on slip a sed on Hilo principle: T atively sim proximately ined with H outputs, w e members Powertrain Te ion of the out and ev pplication p asured dat mpirical bla networks. T a numeric mized para ad and tire s work pro posited on and wheel l omot (Hier The input s mple, usual y. Hilomot [7] weighted w ship functio est Beds characvaluated parameta serve ack box The opcal optiameters simulaovides a the test oad. archical space of lly even ] with the (2.1) on. is 351 9.4 A The inn Figure from an The loc an opti addition divided global t The ba plannin motDoE gray ar This is corresp The me develop are eva point is out acc that the The pro Approach for novative c 2.2 illustr n initial mo Fig cation and imization p nal measu d again in termination asic princip ng approac E run. The reas, and transmitte ponding me easuremen ped. The m aluated. In s added. T cording to e minimum ocess is fo Automated A core of the rates the f odeling, a f gure 2.2: Ill orientatio problem. I urement po the subse n criterion ple of the ch. Figure e technica an automa ed to the a easuremen nt results a model is th n the desig The placem the S-opt m distance ollowed by Adjusting of th e method functional irst partitio lustration o on of the p n the next oint is add equent ste is reached Hilomot a 2.3 shows al impleme ated calibr automated nts. are then tr hen divide gn space ment of the timality pri of the me y re-model he Road Load is the algo principle o on of the in of the parti partition pla t step, the ded to the ep. This p d. approach c s the proce entation is ration syste calibration ransferred ed as befo of the wo e measure nciple. Th easuring po ling, partit d and Tire Sim orithm tha of the par put space tioning with ane are de e local sub sub mode rocedure i can be ex edure that done with em. The fi n system t back to M re and the orst local m ment poin e measure oints to ea ioning and mulation on P t determin rtitioning a is made h Hilomot [ etermined b-models a el with the is repeate xtended to is perform h MATLAB rst step is that autom MATLAB an e resulting model a n t in the su ement poin ach other is d measurin Powertrain Te nes the pa algorithm. e. [11] by the sol are evalua e worst rat ed until a s an adapt med during B, marked s an initial matically le nd a first m local subew measu ubspace is nt is place s maximiz ng until a est Beds artitions. Starting ution of ated. An ing and suitable tive test g a Hiloby the design. ads the model is -models urement carried ed such ed [12]. suitable 352 9.4 A termina adaptiv 3 T In [7] th tion of al mod exist in objectiv are me driveab curves acterist A suita shown units ( contact The au sets ca which n quence scalar e speeds stored ations purpos measu The po nected in Figu Approach for ation criter ve test plan Test bed he use of combustio eling. The n the devic ve, which h easured va bility calibr by using a tic scalar q ble test be in Figure 3 ECUs). El t. An efficie utomated c alibration p needs to b e in the au evaluation s. Then the together w is now pe e, the mea rement po owertrain is to the driv re 3.2. Automated A rion is reac nning algor Figure 2 d configu the Hilomo on engines basic prin ce under te has to be alues such ration, suc appropriate quantities. ed environm 3.1. The de lectrical dy ent and fas calibration parameters be optimize utomation criteria ar e criteria a with the ca erformed a asurement int is comp s installed ve axles. T Adjusting of th ched. In th rithm, deno 2.3: Operat uration a otDoE app is present nciple durin est, the te modeled. as specifi ch scalar v e algorithm ment for m evice unde ynamomet st control c system (A s to the de ed, is start system of re calculate are transfe libration va daptive du ts are tran puted. on the tes The forces he Road Load his way the oted as Hil tion mode nd contr proach for ted. Hilomo ng drivabil est environ In combus c fuel con values are ms. Therefo model-base er test is th ters are s concept is ACS) contro esired valu ted by the f the test ed from the erred as "m alues. The uring the te nsferred to st bed and occurring d and Tire Sim e Hilomot lomotDoE. of Hilomot rol conce adaptive t otDoE can lity calibrat nment and stion engin sumption a e generally ore these a d calibratio he powertra simulating presuppos ols the tes ues. After ACS, imp bed. After e collected measurem specificat est period o the Hilom the electri in real driv mulation on P algorithm i tDoE [11] ept est plannin n be used f tion is iden in the det ne calibratio and emiss y calculate are mappin on of drive ain and its the vehic sed, see [1 st procedur that, the d plemented r performin d signals, e ents" back ion of calib at the tes motDoE alg cal dynam ving operat Powertrain Te is extende ng for the for local an ntical. Diffe termination on, the ob sions. How ed from tr ng curves eability func electronic cle and tir 3]. re. First, th driving ma as a prog ng the ma e.g. mome k to the AC bration val st bench. gorithm. T mometers a tion are de est Beds ed to an calibrand globerences n of the bjectives wever, in ransient to charctions is c control e roadhe ACS aneuver, ram seaneuver, ents and CS and ue vari- For this he next are conescribed 353 9.4 A F Approach for Figure 3.1: Automated A : System c Figure 3.2 Adjusting of th configuratio 2: Forces o he Road Load on on a tes on a vehicl d and Tire Sim st bed for a le during re mulation on P adaptive te eal driving Powertrain Te est planning est Beds g 354 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds Since some forces on the test bed are not available, they must be simulated by the dynamometers. The resistances of wheel , air , ascent and acceleration counteract to the longitudinal force [14]. = + + + (3.1) The measurements required for this work are carried out without a pitch angle so that the resistance of ascent can be omitted. The wheel resistance is reduced to a rolling resistance when the vehicle is driving straight on a dry road and is defined proportional to the wheel load as follows [14]. = ∙ (3.2) The rolling resistance coefficient can be approximated as a function of the vehicle speed by means of a 4th-order polynomial [15]. = + ∙ + ∙ (3.3) The coefficient of friction parameters , , and are determined from the speed traces of roll-out maneuvers performed on a test track [14]. The air resistance can be calculated as a function of the front surface , the air resistance coefficient , the air density and the vehicle speed [14]: = 1 2 ∙ ∙ ∙ ∙ (3.4) Since there is neither a rolling resistance nor an air resistance on the test bed, these are taken into account in the simulation. In order to accelerate a vehicle, it is necessary to overcome the acceleration resistance. The acceleration resistance is the result of the translational acceleration of the mass and of the rotational acceleration of all rotating parts. = + = ∙ + ∙ (3.5) The translational component of the acceleration resistance, which is not present on the test bed, is simulated by the dynamometers. For a detailed derivation of the inertia red reduced to the wheel speeds, see [15]. The rotational inertia of the powertrain components exist on the test bed. The moment of inertia of the wheel is represented by the rotor of the dynamometer. Depending on the control concept, inertia compensation is implemented, which is a compensation of the difference between the moment of inertia of the dynamometer and that of the wheel, which is to be simulated [16]. If inertia compensation is not integrated into the control, the inertia of the dynamometers must correspond to the inertia of the wheels [17]. Figure 3.3 shows schematically the control concept of the test bed, which is available for this paper. 355 9.4 A The tor contain mentum with the the whe in long connec put var the ass measu therefo requisit iments The for The se tire to r , with th value The tire tests [1 shown traction Approach for rque of eac ning the fo m e moment eel, the tor itudinal dir cted to the riable of ea sociated d red directl ore the pow te for prop carried ou rce of frict emi-empiric road [2]. In = μ ∙ e longitud , the curv e paramete 18]. The c in figure 3 n decrease Automated A Figure 3 ch side sha rce transfe , ∙ t of inertia rque , o rection and chassis m ach wheel dyno. As th y, it is gua wer train o per function ut, a contro tion is calc cal Magic F longitudin , ∙ sin inal force ature facto ers have n coefficient 3.4. The as es down to Adjusting of th 3.3: contro aft is meas erred from , = , , , the a of the side d the force model (e.g. model an he actual aranteed t on the test ning is a su ol frequenc culated in t Formula of nal direction ∙ arctan , = , ma , , the s or and th no physical D determi ssociated s when s he Road Load ol concept o sured and tire to roa − , ∙ angular sp shaft, the e , repre . [13]). Fina nd used as angular sp that refere bed is str ufficiently f cy of 10 kH the simula f Pacejka i n and with ∙ , − , ∙ , − ax( , stiffness fa he longitud l significan ines the m slip is ofte slip increas d and Tire Sim of the test used as in ad [2] and , + , eed ω , a force , t esenting fr ally, the an s reference peed of th nce and a ressed as fast contro z is recom ation using is used for pure long ∙ , − ) actor , th inal slip rat nce, but are maximum c en referred ses. mulation on P bed [13] put value f the princip and dynam transferred iction. Eac ngular spe e value for he side sh actual value intended. A ol loop. Bas mmended [1 the equat r the forces itudinal slip arctan ∙ e shape fa tio , . e determin coefficient to critical Powertrain Te for a whee ple of angu mic radius d from tyre ch wheel m eed ω , is r speed co haft can ea e match w An importa sed on the 13]. tions 3.2 a s transferr p it looks li ∙ , actor , th ned from e of friction slip , si est Beds el model ular mo- (3.6) , of to road model is an outontrol of asily be well and ant pree experand 3.3. ed from ke (3.7) (3.8) he peak mpirical n , as nce the 356 9.4 A In the e The co surface the follo 4 R Before a corre by the t real roa the roll vehicle The pro change replicat is avail tion (4. neuver , ∙ m min − After th ually an the loa Approach for Figu equation o oefficient B es can be owing. Thi Road loa starting ca elation betw test bed sh ad measu ing resista e on the tes ocess of m e and roll-o ted on the able that c 1) using th rs of the tes m = − 1 2 − 1 , 1 , ⋮ ⋮ 1 , he rolling r nd iterative ad changes Automated A ure 3.4: Co of Pacejka describes realized by s results in ad and ti alibration o ween test hall have th rements. T ance coeffi st bed. matching b out maneu test bench calculates he Least S st track. ∙ ∙ ∙ , , ⋮ , ∙ resistance ely adjuste s of the te Adjusting of th oefficient of , the curve s the stiffne y the scalin n a simplific re simul of ECUs or bed and ro he most id Therefore, icients is a etween the uvers are p h. For dete the roll res Square me ∙ , − − , , ⋮ , coefficient ed until the st rig corre he Road Load f friction ve e shape is ess of the ng factor μ cation of th ation adj r TCUs on oad. The r entical beh an optima a prerequis e test bed performed ermining th sistance co thod from − ∙ ∙ + ∙ 2 ∙ ts are dete e time sequ espond to d and Tire Sim ersus whee s defined b tire. The r μ . The p he equatio justmen a test bed results of th havior of th al set of th site for a r and the ro on the tes he rolling re oefficients the measu ( + ∙ , , ⋮ , ermined, th uences of t those of t mulation on P el slip ratio by the para reproductio parameter E n (3.7). t on test , it is nece he measur he test veh he Pacejka realistic be oad is as f st track an esistance c based on urements o ∙ , + and con he tire para the measu the test tra Powertrain Te o [19] ameters C on of vario E is set to t beds essary to e rements de hicle regard a paramete ehavior of follows: Fir nd then the coefficients the motio of the roll- ∙ , ) nstr. : 0 ameters ar urement sig ack. This i est Beds and E. us road zero in stablish elivered ding the ers and the test rst, load e test is s, a tool n equaout ma- ) (4.1) 0 (4.2) re mangnals of iterative 357 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds approach requires high amounts of time and resources because of the interactions between the parameters. This is strongly dependent on the experience of the test engineer. In the following, a new method for determining the tire parameters B, C and D of the Pacejka model is presented. The process for the alignment between road and test bed is as follows. The limits of the Pacejka parameters are determined based on measurements made on the test track from full-load Drive-aways with slipping wheels. These restrict the design space for the test planning. A driving maneuver is performed for the parameter combinations determined in the test planning and a criterion is computed which represents the difference between the test track and the test bed. The criterion is used as a target variable and is modeled on the basis of the measured data. The parameter combinations are inputs of the model. The optimal Pacejka parameters are found by numerical optimization. The slip occurring on the wheels during the test run on the test track is calculated using the Equation (3.8) and the measured wheel speeds of the driven and nondriven axles. The driving traction coefficient occurring to the respective slip value is obtained by dividing the circumferential force by the wheel load. From the longitudinal acceleration and the total vehicle mass, the calculation of the circumferential force of the wheels of the driven axle is possible. = veh ∙ (4.3) In this case, it is assumed that the circumferential force and the axle load spread are equally on the two wheels of the driven axle. The wheel load of the driven axle is composed of a static component and a dynamic component of the reaction torque of the rotationally accelerated masses. Only the four wheels have a significant rotatory mass inertia around the vehicle's transverse axis [14]. As shown in equation (4.4) for the front axle, the wheel load is dependent on wheelbase , position , height ℎ of the center of gravity and mass inertia of wheels / . For a detailed derivation of the equation, see [14]. Some effects occurring in reality are not taken into account in equation (4.4). Thus, it is neglected that the position of the center of gravity relative to the wheel contact point changes for example during acceleration. , = ∙ ∙ , − ( ∙ ∙ ℎ + ∙ + ∙ ) , (4.4) Furthermore, the influence of aerodynamic buoyancy and wheel resistance is not taken into account. The evaluation of the Drive-aways is restricted to the first gear and thus to a range with low speeds, in which the buoyancy is negligible. The influence of the wheel resistance on the dynamic axle load is very low on a fixed track and is therefore not taken into account [14]. With equations (3.8), (4.3) and (4.4), the calculation of the slip, the wheel load and the circumferential force of the wheel at each sampling point is possible from the signal paths of the Drive-away maneuver. The determined traction coefficients μ can be mapped via the slip . In Figure 4.1, this is illustrated by the black points for some Drive-aways with different pedal values. 358 9.4 A Figure The ch the tra optima ing the m The “ls applies shows inequa and . This in which t the Pac ables is not red A drivin traction in) can this ma the Pac counte Tip-in i can occ terizatio As a cr the refe error (M Approach for e 4.1: Dete aracteristic ction-slip p lly adapts following o min‖ ( , sqnonlin” f s the “trust the curve lities is de . equality sy the measu cejka para s focused. uced with ng maneuv n-slip curve n be consid aneuver, th cejka para ract the vi s as suita cur in the on. riterion for erence and MSE) of the Automated A ermination c form of t points. It i the Pacejk optimizatio ) − μ ‖ = unction, w t-region-re resulting efined on t ystem limi urements ta ameters. In However, increasing ver is perfo e and is ev dered as a he powertr ameters ca brations in ble as a D lower gea r the qualit d the test b e signal co Adjusting of th of the para the traction s possible ka curve to on problem = min ( which is co eflective” a from the t he basis o c( ts the per ake place, n addition, it must be g restriction ormed on valuated u a maneuve rain is indu an be dete n the powe Drive-away rs, so this ty of adjus bed measu ourses of th he Road Load ameter res [3] n-slip curve e to determ o the data m with a no ( ∙ sin ontained in algorithm to traction-sli of the trac (x) 0 missible d and thus the releva e ensured t n of the pa the test be using a crit er, which i uced to vib ected. The ertrain, ha y maneuve range is t stment, the urement is he engine d and Tire Sim strictions fr e can be s mine a pa points. Th n-linear ta ∙ arctan n the Optim o solve th p coefficie tion and s esign spa excludes nt value ra that the inf rameters. ed for each terion. The is to be ev brations, in comfort fu ve to be d r for the a aken into a e calculatio suitable. T speeds is mulation on P rom Drive-a seen from rameter co his can be rget functio ∙ , ) − mization T is problem ents. A sys slip data of ce, the pa implausible ange of the fluence of h paramet e positive l valuated o n which a unctions in deactivated djustment, account du on of the d Therefore, used. Powertrain Te away mane the scatte ombination achieved b on: − μ ) Toolbox in m [19]. Fig stem of no f the Drive arameter s e combina e adjustme the param ter combin load chang objectively. great influ n the ECU d. In addit , since hig uring the p deviation b the mean est Beds euvers r plot of n which by solv- (4.5) Matlab, gure 4.1 onlinear e-aways (4.6) space in ations of ent varimeters is ation or ge (Tip- During ence of U, which ion, the gher slip paramebetween square 359 9.4 A Figure every model t In the o mum o numeri mizes t track a Figure The co of the t so that 5 A The va powert a 9-spe there a Pacejka road re In [3] is based for perf At the bed. Pe Approach for 4.2 shows -th sampli that repres optimizatio or maximum cal optimiz the differe nd the mea 4.2: Evalu mbination test bed co the determ Applicat alidation of rain is fron eed dual c are no me a, model m eference. s shown th approach forming me beginning erforming Automated A s a gray a ng point. T sents the in on step, th m of the ta zation is to nce of the asurement uation criter of Pacejka ontrol. This mined para tion and f the prese nt-transvers clutch tran easuremen manually d at the calib on a powe easuremen , Drive-aw these man Adjusting of th , = 1 area in wh The scalar nfluence of he optimal arget variab o determine signal traj t on the tes rion for the a paramete s method ta ameters re results ented meth se mounte nsmission. nts availab determined bration of t ertrain test nts on the t ways with f neuvers the he Road Load (y Prst, − hich region criterion s f the manip setting of ble is sear e a Pacejk ajectories b st bed. e comparis bed [3] ers found akes into a present a hod is perf ed and con Due to th ble from th d by the te the tire mo t bed. The test bench full-load w e Pacejka d and Tire Sim − y Str, ) n an evalu serves as o pulated va the input rched. In th ka parame between th son betwee is used to account Tip big range o formed on sists of a 4 he prototyp he test tra est enginee odel is done ese sophis h are used were perfor paramete mulation on P ation of output valu riables. variables his case, th ter combin he measure en the test parameter p-ins with h of the Pace a powertr 4-cylinder d pe status o ack. The p er, is there e successf sticated au here. rmed at th rs are dete Powertrain Te , oc ue of an e for a glob he purpos nation whic ement on track and rize the tire higher slip ejka mode rain test be diesel eng of the pow parameters efore chose fully with a tomated m he powertr ermined, w est Beds (4.7) ccurs at mpirical bal minie of the ch minithe test the test e model values, el. ed. The gine and wertrain, s of the en as a a modelmethods rain test which fit 360 9.4 A the non tion-slip and es consist as refe measu cause o On the curve is ramete These C and test pla Approach for n-linear tra p values a timated Pa tency of th rence, wer rements fr of higher d Table re e basis of t s estimate ers B , C an μ es constraints D. Figure anning with Automated A action-slipnd the curv acejka par e values r re carried o rom Drivedispersion o e 5.1: Estim eference estimated Figur he traction ed. Then co d D , as fol st = ( , , s are appl 5.2 shows h HilomotD Adjusting of th curve to th ve with the ameters of results from out at the t -aways on of the tract mated valu 12.00 11.75 re 5.1: trac n-slip value onstraints lowing: , , ) = μ max = 1 μ min = 0. μ min ied to a fu s the deter DoE, and th he Road Load he data po e estimated f the tire m m the fact test bed ra n the test t tion-slip va ues of tire m 1.60 1.50 ction-slip va es of a Dri μ max and μ est sin( est .2 ∙ μ est + 0 .8 ∙ μ est − 0 μ μ max ull grid of p rmined can he correspo d and Tire Sim oints optim d paramete model are l that the D ather than track woul alues. model para 0 0 alues and ve-away w μ min are de ∙ arctan( e 0.2 0.2 possible pa ndidate po onding trac mulation on P mally. In Fig ers is show isted in Ta rive-aways on the test ld lead to ameters B, 1.10 1.23 curve with full-loa efined, whic est ∙ )) arameter c ints, which ction-slip c Powertrain Te gure 5.1 th wn. The re able 5.1. T s, which a t track. Re poor resu , C, D ad the tract ch restrict combination h are used curves. est Beds he traceference The high re used eference ults, betion-slip the pa- (5.1) (5.2) (5.3) (5.4) ns of B , d for the 361 9.4 A Figure In this stant o and en positive evaluat modele With H reache availab tion ba for the The "le correla used fo ted as f Where sidering Hence the Hilo where y Althoug value a predict rion for Figure and Approach for e 5.2: cand application perating p ngine spee e load cha ted with th ed by Hilom ilomotDoE d a certai ble in the u sed on thi terminatio eave-one-o tion to the or models w follows [7] ( ) is the g the respe the omotDoE, y is the me gh equatio also provid ion accura r sufficient 5.2 shows d du Automated A didate point n example oint. The o d of 2000 ange (Tiphe criterion mot. E it is poss n level of usual “ope s informat n of the ex out" crosse coefficien which can : LOOCV e -th meas ective mea ( P redicte can be cal ean value o n (5.6) is v des useful acy [8]. In [ model acc s the result uring the p Adjusting of th ts used for c the three operating p rpm. For in) is perfo n , sible to sto a model rating mod tion cannot xperiment i -validation nt of deter be derived V( ) = 1 surement a asurement ed RE sidua lculated as = 1 − of measure valid only f informatio [20] is show curacy. of the Hilo rogress of he Road Load r test plann curves (righ paramete point is def each varia formed on . During r op the exp quality crit de” of Hilo t be calcu is used to (LOOCV) rmination b d from line ( ) − and ( ) ( for estima al S um of S s follows, LOOCV 1 ∑ ( ( ements. for models on for loca wn that omotDoE p f HilomotDo d and Tire Sim ning (left), c ht) ers , an fined by pe ation of , the test b run time o eriment, o terion. Bec motDoE, t lated. Con identify ad can be ap based on ear parame ( ) ( ) ) is the -t ation of the S quares) t V( ) ) − y) s that are l al model n can b procedure. oE depend mulation on P correspond d are va edal positio and th bed. Then of one vari once the H cause no he coeffici nsequently, equate mo pplied. It p validation eter estima h model ou model. that serves inear in the etworks w be used as The conti ding of the Powertrain Te ding tractio aried on o on of 100 he maneuv the mane iation Hilomot mo validation ient of det , another c odel accura provides a data and ation. It is utput witho s as a crite e paramet with respec s terminatio nuous incr number o est Beds on-slipne conpercent ver of a euver is , is odel has data is erminacriterion acy. certain can be calcula- (5.5) out conerion for (5.6) ters, the ct to the on criterease of of meas- 362 9.4 A uremen tion po The nu is calcu Where order. The am As crite of 50 is input sp Figur In this After re resultin is achie optimiz re o E Approach for nt points ca ints, chose umber of in ulated as fo (here: 3 mount of in erion for te s set. Sinc pace, re 5.2: R 2PR variation p applicatio eaching th ng model is eved by u zation are t Table eference ptimized rror in % Automated A an be see en by the H nitial points ollows acc 3) is the di nitial points erminating e the trans is filtere RESS and R 2F points use n example e maximu s now used sing a num the same a e 5.2: Optim 12 12 0, Adjusting of th n on the le HilomotDoE s is 8. The ording to [2 = mension o s is the co the test a sient behav ed. Fit of the cr d by Hilom e, d m number d to find op merical glo as those us mized valu 2.00 2.01 083 he Road Load eft side of E algorithm e required 21]: ( + ) ! ! ⋅ ! of the prob ndition for of vior of riterion motDoE out does not re r of points ptimal valu obal optim sed in the ues of tire m d and Tire Sim the figure. m, can be s minimum n blem and a first par 0.9 and a can va N,Eng mod t of the can each the p , the test es of the p mization alg test planni model para 1.60 1.58 1,25 mulation on P On the rig seen. number (here: 1) rtitioning of maximum ary due to deled by Hi ndidate po predetermi planning is parameters gorithm. Th ng (see eq ameters B, Powertrain Te ght side th of measur is the poly of the input number o partitionin ilomotDoE oints (right) ned value s terminate s , and he constra q. 5.1 - 5.4 , C, D 1.10 1.13 2,73 est Beds e variarements (5.7) ynomial t space. of points g of the (left), of 0.9. ed. The . This aints for 4). 363 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds The result in Table 5.2 shows that the method presented here is suitable for the matching between the reference and the test bench. The maximum percentage error between optimized and reference values is less than 3 %. 6 Summary and outlook The method presented within this paper describes an approach to determine the parameters , and of the Pacejka model. In contrast to the previously used procedure, the parameters are not set by iterative matching. In addition, the test engineer does not need to know the influence of the parameters of the tire model. The duration of the adjusting operation is not dependent on experience and can thus be reduced. Furthermore, in a manual and iterative procedure, the optimal parameter combination is not surely found. The presented method by numerical optimization of a target function, which maps the difference of the engine speed signals between road and the test bed measurements as a function of the Pacejka parameters, allows this. For implementation on a powertrain test bed an interface for specifying new measuring points has to be present. In addition, an online calculation of the criterion, which specifies the quality of the adjustment, is necessary. The criterion can be compared to a sensor, which detects the target values. Furthermore, no statistical knowledge is required to use the process needed for the introduced adaptive test planning. In this presented application example for HilomotDoE adaptive test planning can be combined to the model-based calibration workflow to achieve simultaneous optimization, with the precondition of an already known optimization goal. Several strategies are possible for this purpose. On the one hand, a compromise between model quality and optimization goal in test planning could lead to a faster completion of the test. On the other hand, after achieving a sufficient model quality, the optimization and subsequent validation could iteratively lead to a better model and a more precise optimization. This paper presents the application of HilomotDoE at one single operating point, which implies a local modeling of the process. This single operating point has to be selected specifically so that the slipping curve can be induced in a wide range. The defining parameters of the operating point pedal position and engine speed can be treated as variation parameters as well. This leads to a global modeling of the process and possibly to a more adequate adjusting of the tire model for the test bed. The additional use of information brought along by repetition points can provide a possibility to evaluate model quality, which allows an iterative adaptation of the termination criterion. The model quality can also be improved by online outlier detection. For the future the result of using HilomotDoE for adjusting the tire model with real test track measurements and different, maybe more robust criteria, will be investigated. References [1] Hartmann, B.: Lokale Modellnetze zur Identifikation und Versuchsplanung nichtlinearer Systeme, Universität Siegen. Siegen 2013 [2] Pacejka, H. B.: Tire and Vehicle Dynamics. Elsevier 2012 364 9.4 Approach for Automated Adjusting of the Road Load and Tire Simulation on Powertrain Test Beds [3] Weber, S.; Dursun, Y.; Bäker, B.; Jakobi, R.; Kirschbaum, F.; Körner, M.: Entwicklung einer Methodik zur Durchführung des Straßenabgleichprozesses. SIMVEC (2016) [4] Pillas, J.; Kirschbaum, F.: Model-based load change reaction optimization using vehicle drivetrain test beds. 14. Internationales Stuttgarter Symposium (2014) [5] Pillas, J.; Kirschbaum, F.: Model-based calibration of a load change reaction control function with hybrid state space models. Proceedings of the 12th Stuttgart International Symposium Automotive and Engine Technology (2012) [6] Uphaus, F.: From Road to … where? Methods and Tools in Calibrating Drivability. 10th Symposium on Automotive Powertrain Control Systems (2014) [7] Klein, P.; Kirschbaum, F.; Hartmann, B.; Nelles, O.: Adaptive Test Planning for the Calibration of Combustion Engines - Application. DoE Tagung Berlin (2013) [8] Hartmann, B.; Ebert Tobias; Kampmann, G.; Nelles, O.: LMNtool - Toolbox zum automatischen Trainieren lokaler Modellnetze. Workshop Computational Intelligence Dortmund (2012) [9] Poland, J.: Modellgestützte und evolutionäre Optimierungsverfahren für die Motorentwicklung, Universität Tübingen. Tübingen 2002 [10] Knödler, K.: Methoden der restringierten Online-Optimierung zur Basisapplikation moderner Verbrennungsmotoren, Universität Tübingen. Tübingen 2004 [11] Hartmann, B.; Nelles, O.: Adaptive Test Planning for the Calibration of Combustion Engines - Methodology. DoE Tagung Berlin (2013) [12] Montgomery, D. C.: Design and analysis of experiments. 8. edition. Hoboken, NJ: John Wiley & Sons, Inc. 2013 [13] Bauer, R.: New Methodology for Dynamic Drive Train Testing. Proceedings: Symposium on International Automotive Technology (2011) [14] Haken, K.-L.: Grundlagen der Kraftfahrzeugtechnik. Fahrzeugtechnik. 4. edition. München: Hanser, Carl 2015 [15] Mitschke, M.; Wallentowitz, H.: Dynamik der Kraftfahrzeuge, 5., überarbeitete und ergänzte Auflage. 2014 [16] Bauer, R.: Neues Regelungskonzept für die dynamsiche Antriebsstrangprüfung. 17. Steirisches Seminar über Regelungstechnik und Prozessautomatisierung (2011) [17] Dohmen, H.-P.; Pfeiffer, K.; Schyr, C.: Antriebsstrangprüftechnik. Vom stationären Komponententest zum fahrmanöverbasierten Testen. Die Bibliothek der Technik, Bd. 317. [München]: Verl. Moderne Industrie 2009 [18] Leister, G.: Fahrzeugreifen und Fahrwerkentwicklung. Strategie, Methoden, Tools. ATZ-MTZ Fachbuch. 1. edition. Wiesbaden: Vieweg + Teubner 2009 [19] MathWorks: User’s Guide Matlab. Dokumentation Least-Squares (Model Fitting) Algorithms, 2016. http: / / de.mathworks.com/ help/ optim/ ug/ least-squares-modelfitting-algorithms.html [20] Dursun, Y.; Kirschbaum, F.; Gebhardt, A.; Goos, J.-C.; Rinderknecht, S.: Approach to adaptive test planning for calibration of longitudinal dynamics of vehicles on test beds. 6th International Symposium on Development Methodology (2015) [21] Linssen, R.: Modellbasierte Analyse und Optimierung von Dieselmotoren, RWTH Aachen. Aachen 2010 365 The Authors Dr.-Ing. Karsten Röpke (Ed.) IAV GmbH Berlin Prof. Dr.-Ing. Clemens Gühmann (Ed.) Fachgebiet Elektronische Mess- und Diagnosetechnik TU Berlin Dipl.-Ing. Stefan Angermaier Robert Bosch GmbH Renningen Dr.-Ing. Mohamed Ayeb University of Kassel Kassel Prof. Dr.-Ing. Michael Bargende Institut für Verbrennungsmotoren und Kraftfahrwesen Stuttgart Dr. Wolf Baumann IAV GmbH Berlin Dr.-Ing. Jörg Beilharz IAV GmbH Berlin Tom Berghmans Toyota Motor Europe NV/ SA Zaventem, Belgium Prof. Dr.-Ing. Dieter Bestle Engineering Mechanics and Vehicle Dynamics Brandenburg University of Technology Cottbus Dr.-Ing. Amit Bhave CMCL Innovations Cambridge, United Kingdom Prof. Dr.rer.nat. Ludwig Brabetz University of Kassel Kassel Prof. Dr.-Ing. Christoph Brands Schaeffler Technologies AG & Co. KG Herzogen-aurach Johan Bringhed Volvo Cars Gothenburg, Sweden Dr. Jochen Broz Schaeffler Technologies AG & Co. KG Herzogenaurach Matthias Brumm, M.Sc. Daimler AG Stuttgart Yooshin Cho Hyundai Motor Company Namyang R&D Center Hwaseong-si, Gyeonggi-do, South Korea Seth DeLand The MathWorks GmbH Ismaning Dipl.-Ing. Nico Didcock AVL List GmbH Graz Dr.-Ing. Julian Dizy CMCL Innovations Cambridge, United Kingdom Dipl.-Ing. Thomas Dreher IAV GmbH Berlin 366 The Authors Dipl.-Ing. Yagiz Dursun Daimler AG Stuttgart Dr. Carola Eckstein Robert Bosch GmbH Stuttgart Dipl.-Inf. Ortwin Escher IAV GmbH Gifhorn Kento Fukuhara IAV GmbH Berlin Jan-Christoph Goos, M.Sc. Daimler AG Stuttgart Mukunda Gopalakrishnan Robert Bosch GmbH Schwieberdingen Dr. Markus Grahn Volvo Cars Gothenburg, Sweden Frank Gutmann, M.Eng. SGE Ingenieur GmbH Gräfelfing Dipl.-Ing. (FH) Tobias Gutmann SGE Ingenieur GmbH Gräfelfing Dipl.-Ing. Lars Hagen Robert Bosch GmbH Schwieberdingen Dr. Donghee Han Hyundai Motor Company Namyang R&D Center Hwaseong-si, Gyeonggi-do, South Korea Dr.-Ing. Benjamin Hartmann Bosch Engineering GmbH Abstatt Dr.-Ing. Carsten Haukap IAV GmbH Berlin Jing He Keihin Corporation Tochigi, Japan Dr. rer.nat.habil. Michael Hegmann IAV GmbH Berlin Tim Oliver Heinz, M.Sc. Department of Mechanical Engineering University of Siegen Dipl.-Technomath. Matthias Hekrenz Schaeffler Technologies AG & Co. KG Herzogenaurach Arvind Hosagrahara The MathWorks GmbH Ismaning Dipl.-Ing. Antoine Hurstel Robert Bosch GmbH Stuttgart Prof. Dr. Hanno Ihme-Schramm HAW Hamburg Hamburg Dipl.-Ing. (FH) Martin Jacob Bosch Engineering GmbH Abstatt Dr. Richard Jakobi Daimler AG Stuttgart Matthias Kahl, M.Sc. FG Mess- und Regelungstechnik Universität Kassel Masataka Kaihatsu Keihin Corporation Tochigi, Japan Robert Kästner, M.Sc. IAV GmbH Gifhorn Dr. Nikolaus Keuth AVL List GmbH Graz, Austria 367 The Authors Masato Kikuchi Honda R&D Co. Ltd. Japan Dr.-Ing. Frank Kirschbaum Daimler AG Stuttgart Dr.-Ing. Ernst Kloppenburg Robert Bosch GmbH Stuttgart Dr.-Ing. Mirko Knaak IAV GmbH Berlin Prof. Dr.sc.techn. Thomas Koch Institut für Kolbenmaschinen (IFKM) Karlsruher Institut für Technologie (KIT) Karlsruhe Univ.-Prof. Dr.-Ing. Andreas Kroll FG Mess- und Regelungstechnik Universität Kassel Dipl.-Ing. (FH) Leon Evgeni Kusnezow IAV GmbH Braunschweig Dr. Björn Lundberg Volvo Cars Gothenburg, Sweden Kotaro Maeda Toyota Motor Europe NV/ SA Zaventem, Belgium Dmytro Martynenko The MathWorks GmbH Ismaning Dr.-Ing. Thomas Mayer AUDI AG Ingolstadt Dipl.-Ing. Sven Meyer IAV GmbH Berlin Kadir Mourat, M.Sc. Robert Bosch GmbH Stuttgart Dr. Yutaka Murata Honda R&D Co. Ltd. Japan Prof. Dr.-Ing. Oliver Nelles Department Maschinenbau - Universität Siegen Siegen Yui Nishio Honda R&D Co. Ltd. Japan Dipl.-Ing. David Ooi CMCL Innovations Cambridge, United Kingdom Dr.-Ing. Vivian Page Perkins Engines Co. Ltd. Peterborough, United Kingdom Thiebault Paquet Toyota Motor Europe NV/ SA Zaventem, Belgium Dr. sc.nat. Owen Parry CMCL Innovations Cambridge, United Kingdom Mike Preston Ricardo UK Ltd Shoreham-by-Sea, England Dipl.-Ing. (FH) Andreas Rainer AVL List GmbH Graz, Austria Aymeric Rateau Toyota Motor Europe NV/ SA Zaventem, Belgium Rajesh Reddy, M.Sc. ETAS GmbH Stuttgart Dipl.-Ing. Daniel Reppel IAV GmbH Berlin Pascal Revereault Ricardo UK Ltd Shoreham-by-Sea, England 368 The Authors Daniel Rimmelspacher IAV GmbH Berlin Prof. Dr. Stephan Rinderknecht TU Darmstadt Darmstadt Stefan Scheidel, M.Sc. AVL List GmbH Graz, Austria Mark Schillinger, M.Sc. Bosch Engineering GmbH Abstatt Alexandra Schramm, B.Sc. Wirtschaftspsychologin Grosshansdorf Dr. Justin Seabrook Ricardo UK Ltd Shoreham-by-Sea, England Dipl.-Ing. André Sell SGE Ingenieur GmbH Gräfelfing Taro Shishido Keihin Corporation Tochigi, Japan Dr.-Ing. Manfried Sofsky IAV GmbH Gifhorn Felix Springer IAV GmbH Berlin Dipl.-Ing. Marie-Sophie Vogels AVL List GmbH Graz, Austria Dr.-Ing. Andreas Walter Robert Bosch GmbH Schwieberdingen Sebastian Weber, M.Sc. Mercedes-AMG GmbH Affalterbach Dipl.-Ing. Kassem Wehbi IAV GmbH Berlin Dipl.-Ing. ETH Simon Wunderlin Robert Bosch GmbH Stuttgart Dr.-Ing. Andreas Wurm IAV GmbH Berlin Dr. Yukihisa Yamaya Honda R&D Co. Ltd. Japan 369 Dr.-Ing. Karsten Röpke, Prof. Dr.-Ing. Clemens Gühmann (Ed.) and 65 co-authors Design of Experiments (DoE) in Powertrain Development 2015, 265 S., 191 Fig. and 25 Tab., 62,00 €, 81,00 CHF (Reihe Technik) ISBN 978-3-8169-3316-8 Zum Buch: Developing complex powertrains without model-based processes is inconceivable today. But which methods can be used to address future challenges in powertrain development such as real driving emissions (RDE) and the diversity of derivatives? Contents : A strategy to employ criteria for online location selection when using Gaussian Processes - Efficient online test plan calculation with integrated boundary detection for modern model based engine calibration - Benchmark Problem for Near Boundary Operation Control for Automotive Engine - Lookup Table Optimisation for Engineering Applications - A Dynamometer Dynamic Calibration Method for the Diesel Air-Path - Application of DoE Method on Synthetic Gas Bench for SCR studies in Early Stages of the Model-Based System Development Process - Application of Model Based Calibration to Mass Production Diesel Engine Development for Indian Market - Application of global model based calibration methodology to optimize a 2.3 litre diesel engine with SCR on WLTC cycle - Advanced Gaussian Process Modeling Techniques - Fast Engine Modelling Using On-Line Calibration Data - Steady State Calibration for Catalyst Heat-up Optimization on Gasoline Direct Injection Engine - Efficient variant calibration by automation using rules and dependencies - Efficient Calibration Process for Series Programme with Multiple Engine, Vehicle and Market Variants - Computing Optimized Calibration Maps including Model Prediction and Map Smoothness - DoE and beyond: the evolution of the model based development approach how legal trends are changing methodology - Vehicle and OEM generic HIL / SIL Model in the transmission development - Data-based Models on the ECU - Pattern recognition for classifying degradation states of lithium ion batteries - Simulation-Error Based Identification of Dynamic Calibration Models Target group: Development and application engineers from automotive manufacturers and supply industry as well as universities The Editors: Dr.-Ing. Karsten Röpke is Head of Department Development Methods at IAV. Prof. Dr.-Ing. Clemens Gühmann is Head of the Chair of Electronic Measurement and Diagnostic Technology at TU Berlin Blätterbare Leseprobe und einfache Bestellung unter: www.expertverlag.de/ 3316 Bestellhotline: Tel: 07159 / 92 65-0 • Fax: -20 E-Mail: expert@expertverlag.de