1 Introduction and motivation Wind energy is the most important renewable power source in Germany with a share of around 20 % of the overall power production [1]. In recent years the increase of rated power output per WT and a reduced levelized cost of energy are subject to research efforts. The usage of planetary journal bearings in wind turbine gearboxes Science and Research 14 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 Condition monitoring for planetary journal bearings in wind turbine gearboxes by means of acoustic measurements and machine learning Thomas Decker, Georg Jacobs, Christoph Paridon, Julian Röder* submitted: 18.09.2023 accepted: 05.07.2024 (peer-review) Presented at the GfT Conference 2023 * Thomas Decker, M. Sc. (corresponding author) Orcid-ID: https: / / orcid.org/ 0000-0002-3296-7166 Prof. Dr.-Ing. Georg Jacobs Christoph Paridon, B. Sc. Julian Röder, M. Sc. Center for Wind Power Drives Campus Boulevard 61 52074 Aachen - Germany The use of journal bearings instead of rolling bearings as planetary bearings in wind turbine (WT) gearboxes is advantageous for the power density of the drive train. In addition, they can increase the reliability of the gearboxes as they can be operated wear-free in hydrodynamic operation, i.e. with fluid friction. The dynamic loading conditions in wind turbines, as well as special conditions such as insufficient lubrication and particle contamination, can lead to mixed friction operation and consequently to wear in the journal bearings. If mixed friction is not detected, damage and failure of the journal bearing and consequently of the WT gearbox may occur. Therefore, it is required to develop a Condition Monitoring System (CMS) to detect mixed friction in the journal bearings during operation of the WT. Preliminary investigations have shown that various friction conditions can be detected by acoustic measurements in combination with machine learning classifiers. Current investigations on CMS methods for journal bearings using acoustic measurement methods are limited to component level applications. A condition monitoring methodology for wind turbine gearbox journal bearings does not currently exist. A major challenge for CMS development for journal bearings in WT gearboxes is the transfer of methods already proven at component level to gearbox applications. In preparation for CMS application at the gearbox level, this paper presents an approach for monitoring different Abstract friction conditions of journal bearings based on acoustic measurements at a component test rig. For the classification of the friction state, different machine learning (ML)-based approaches trained on the acquired acoustic measurement data are compared with respect to the achieved classification accuracy. Knowledge of the robustness of the classification method, e.g. with respect to the distance of the sensor to the bearing, provides the necessary basis for the use of the CMS at the gearbox level. The investigations are carried out under operating conditions typical for planetary bearings in wind turbines. Classification performance is evaluated using a validated elasto-hydrodynamic simulation model. The aim of the work is to develop a method that detects friction classes in the journal bearing based on structure-borne sound measurements. Here, simulation results are used to train the algorithms. Finally, the demonstrated method will be successfully applied to a test rig for wind turbine gearbox journal bearings. Based on the results, an ML approach will be selected for application in gearboxes. Keywords wind energy, journal bearings, condition monitoring, machine learning is advantageous compared to roller bearings in terms of torque density of the planetary stage and the reliability of the drivetrain and can therefore contribute to the aforementioned goals [2]. When designed and operated correctly (i.e. in hydrodynamic conditions), journal bearings have an unlimited operating lifetime. To ensure a wear safe operation of journal bearings, mixed friction states that can lead to wear of the bearing should be avoided. Mixed friction states can be detected using condition monitoring systems (CMS). CMS typically consist of a specifically designed sensor setup measuring the operating condition of the bearing and a computing unit that automatically conducts the autonomous evaluation and interpretation of the data. CMS techniques are already established for the use in gearboxes over a broad range of different industries. However, these methods are limited to the monitoring of roller bearings and gears. A reliable CMS for planetary journal bearings in WT gearboxes does not currently exist. This paper presents an approach towards the monitoring of planetary journal bearings using acoustic measurements to detect mixed friction. The experimental investigations are carried out on a component test rig for radial journal bearings. The acoustic measurements are used to train different supervised machine learning algorithms for the classification of friction states. Previous works on tribometers and small journal bearing test rigs have shown that classification algorithms can differentiate between friction states based on Acoustic Emission (AE) measurements [3, 4] and even detect wear progression [5, 6]. Different classification algorithms perform differently on the same data sets. Therefore, this paper evaluates the classification accuracy of different algorithms. Furthermore, the transferability of the method to planetary journal bearings is demonstrated on a planetary gear stage test rig with a journal bearing. Thus, the first step towards transferring the method to a gearbox in the field is shown. 2 Methods and test infrastructure The methodological approach of this work is shown in Figure 1. Acoustic measurement data is gathered from a journal bearing on the test rig under various operating conditions and merged into one data set. This data set is divided into training and validation data. The same operating conditions from the tests are also used for elastohydrodynamic (EHD) simulations to characterize the occurring friction regime. The simulation results are summarized into four different friction states that must be detected by the CMS: “fluid friction”, “mild mixed friction”, “mixed friction” and “severe mixed friction”. The simulation results are used as labels for the training data. The labels are fed into a supervised machine learning algorithm, which is trained to differentiate between the defined friction states. In a final step the algorithm uses the test data to produce a prediction based on the learned behavior. The prediction is validated using the simulation results. a. Experimental test setup The used bearing material and lubricant are typical for gearbox applications in WTs. The bearings are loaded using a hydraulic actuator with a maximum force of 216 kN, which corresponds to a specific pressure of 60 MPa in the test bearing. For this work several test bearings are used. Their specifications are given in Table 1. Science and Research 15 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 Figure 1: Workflow of the condition monitoring method used in this work Bearing width = 30 Bearing diameter = 120 Bearing surface roughness = 0.4 Bearing material 12 2 , = 100 GPa Shaft material 42 4, = 210 GPa Radial bearing clearance = 80 Lubricant 320 Table 1: Specifications of the test setup 12 2 , = 100 GPa 42 4, = 210 GPa Figure 4 shows the AE measurement results during a test procedure with a constant sliding speed v and swelling load between 15 and 60 MPa of specific pressure p̅ and the measured friction moment F R . The bottom plot shows the so-called root mean square (RMS) U RMS and the kurtosis κ Equation 1 Equation 2 with x̅ being the moving average of the raw signal and σ being the standard deviation. U RMS and κ are characteristic signal features that in previous works have proven to be effective for the differentiation between friction states [4, 9]. Both characteristics are calculated in the time domain using a moving window with a length of N = 10 6 datapoints, which at a sampling frequency of 3 MHz corresponds to a window duration of 0.33 s. Before the extraction of the signal features the raw signal is filtered using a steep bandpass filter between 0.4 MHz and 1 MHz. This frequency range was identified as suitable for the detection of abrasive wear in journal bearings [3]. The AE signals clearly correspond to the load imposed on the bearing. The extracted U RMS (t) and κ (t) give an indication for different friction states in the journal bearing. No increase in signal amplitude occurs at the load level with the lowest pressure (p̅ =15 MPa). The load interval with the highest specific pressure (p̅ = 60 MPa) results in the highest amplitude in the AE features. Strong mixed friction is assumed for this operating point. This is examined in more detail below using simulations. c. Elasto-hydrodynamic simulation of the test bearing The used classification approaches in this work all belong in the category of supervised machine learning. This means that the classification algorithm is trained to detect the different states according to labels defined by the user. The training data consists of sensor signals recorded at different friction states. For the training process a valid information (label) about the corresponding friction state (mixed or fluid) for each section of the training data set is required. These labels are generated using an EHD simulation of the test bearing in its test rig environment. The EHD simulation consists of the flexible bodies bearing and shaft. These are created by means of the finite element method and are shown in Figure 5. Previous works have shown that EHD models can predict the friction state (mixed or fluid friction) in a journal bearing if they are parametrized correctly [10]. Mixed friction in a hydrodynamic journal bearing is characteriz- = 1 = ( ) ( 1) Science and Research 16 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 The test rig is shown in Figure 2. Radial force and speed are applied by a hydraulic actuator and an electric drive. Other than for a planetary journal bearing in a wind turbine gearbox, no tilting moment acts on the journal bearing on the component test rig. Gear influences are also neglected. The specimen is mounted inside a rotatable enclose that is connected to a force transducer. This allows for the measurement of friction moment during operation, which causes the enclosure to tilt around the rotational axis. The test rig can be heated artificially to typical wind turbine gearbox temperatures (20 °C to 100 °C). b. Acoustic Emission AE has already proven to be a suitable technology to detect different friction states and wear progression in tribological systems such as pin-on-disk tribometers [3, 4] and small journal bearings [7, 8]. In this work the AE sensor (piezoelectric transducer) is bolted to the bearing cover as close to the bearing as possible, approximately 80 mm away from the load zone of the bearing (see also Figure 3). Figure 2: Experimental setup on the journal bearing component test rig Figure 3: AE sensor mounted to the bearing cover on the component test rig ed by the occurrence of asperity contact pressure p a . The asperity contact pressure is calculated using the stochastic contact model according to Greenwood and Tripp [11]. It describes the asperity contact pressure as a function of the elasticity factor K, the average elastic modulus E of the bearing and shaft and the probability distribution for the occurring contact F(H S ) as a function of the nominal gap height H S . Equation 3 With H S being the ratio between the absolute oil film height h and the root mean square summit height σ S of the surface roughness, asperity contact occurs only when the oil film height H S falls below a minimum [11, 12]. Equation 4 The asperity contact pressure p a according to Equation 3 and Equation 4 is shown in Figure 6 with different values for the elasticity factor K chosen within the recommended range [12]. According to the given contact model the simulated asperity contact pressure p a is strongly influenced by the elastic factor K, surface roughness values and the oil film height h. The latter results from the solution of the Reynold’s equation and is strongly influenced by the operating conditions (load p̅ , speed v and temperature T). Model input parameters such as the nominal bearing clearance and the roughness values of the bearing and the shaft result from surface measurements prior to the testing. The temperature of the lubricant T is measured on the test rig as well. The elastic factor K is fitted such that the resulting friction moment in the simulation corresponds to the measurement for one selected operating point (p̅ = 45 MPa,v = 0.4 m/ s). For this work the elastic factor is chosen to be K = 0.0004. Afterwards the validation of the model is carried out by a comparison of the measured and simulated friction moment for the tested procedures as shown in Figure 7. For this purpose, the force transducer of the test rig (see also Figure 2) is used to measure = ( ) ( ) = 4.4086 10 (4 ) . , < 4 0 , 4 Science and Research 17 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 Figure 5: EHD model of the component test rig Figure 4: AE measurement under dynamic load intervals and a constant sliding speed on the component test rig Figure 6: Calculated asperity contact pressure p a at different oil film heights h The occurring maximum asperity contact pressure p a,max during the swelling load procedures is given in Figure 8. Apparently, for both procedures at the first load level, no asperity contact pressure is generated. This corresponds to the signals of the AE features (see Figure 7), where no change of the AE signal amplitude can be detected in the first load level either. From that point on the level of maximum asperity contact pressure rises with each load interval. It is assumed that using the presented EHDmodel the characterization of the operating points can be done with a sufficient accuracy. Science and Research 18 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 the friction moment. Figure 7 shows two of the test procedures used in this work at 0.4 m/ s and 0.6 m/ s sliding speed. The applied load over time is equal to the procedure shown in Figure 4. The tested procedures are simulated with the EHD model and the time series of the friction moment of measurement and simulation are compared. The validation shows a qualitative agreement between simulation and measurement. For this work it is assumed that the presented parameterization and validation of the model is sufficiently accurate to demonstrate the feasibility of the CMS method shown in Figure 1. Figure 7: Comparison of the simulated and the measured friction moment Figure 8: Simulation result for the maximum asperity contact for different sliding speeds Figure 9: (a) Map of the simulated asperity contact pressure for the tested operating conditions of the bearing (Simulated bearing temperature: 55 °C) (b) Distribution of four friction states over the tested operating conditions based on the simulation results and the definition in Equation 5 (b) (a) Using the aforementioned EHD model, simulations are performed for different operating points and the resulting maximum asperity contact pressure p a,max is mapped over the specific pressure p̅ and the sliding speed v (see Figure 9 (a)). A high asperity contact pressure p a represents wear critical mixed friction, while low asperity contact pressure represent a low risk of wear. Without solid contact, pure fluid friction occurs. Using the simulation results presented above labelling of the datasets used for training and validation of the machine learning classification algorithms will be done. Therefore, the results are divided into four different classes according to the amount of asperity contact pressure p a that occurs (Figure 9 (b)). The given simulations are carried out for three different lubricant temperature levels (40 °C, 55 °C and 70 °C) to account for the temperature influence on the friction state in a typical gearbox temperature range. The simulated distribution of the friction states over the operating conditions are stored and used by the classification algorithms for the generation of labels for the training data. The training algorithm can hereby assign one of the four friction states as a training label to any operating point of the journal bearing. d. Machine learning classifiers In this work, three classification algorithms are evaluated with regard to their accuracy in detecting friction states: Support Vector Machine (SVM), Gaussian Process (GP) and Multi-Layer Perceptron (MLP) neural network. All three approaches process the same input data to ensure comparability. The input data consists of timebased signals: the extracted AE characteristics, the bearing temperature and the sliding velocity. The strengths and shortcomings of the three algorithms are discussed in the next paragraphs. With the SVM classification is achieved by defining a hyperplane that separates datapoints belonging to different classes. The hyperplane is defined such that the margin between the classes is maximized. The simple approach of the SVM is advantageous as the model’s classification decisions are comprehensible; a disadvantage resulting from the simplicity is the SVM’s limited number of adaptable parameters. Of all considered models, the SVM is the fastest to train and to run productively after training. It has already been shown that the SVM is a suitable algorithm for the detection of friction states in tribological systems [4, 9]. A GP is a probabilistic classification model. Its definition is considerably more sophisticated than for the SVM and can be found e.g. in [13]. From the user perspective, the main difference is the choice of the kernel function. The GP is comparable to the SVM in terms of time effort needed for training and execution. The MLP is a simple type of neural network. Unlike the algorithms presented before, the approach behind neural networks is not based on regression. Instead, they consist of “layers” containing multiple “neurons”. In the simple model presented here, each neuron receives input values from all neurons in the layer before, calculates the sum of all inputs plus a constant set in model training and then outputs either zero or the calculated sum, subject to whether the sum exceeds a threshold. In the final layer, the number of neurons corresponds to the number of classes to predict, and the neuron yielding the highest output value represents the predicted class. The high number of parameters configurable in training makes the MLP very adaptable to complicated relationships in the training dataset, but it also makes the model computationally expensive. All three models are evaluated with a variety of configuration parameters. Among these are multiple kernel functions for SVM and GP as well as the number of neurons and the number of layers for MLP. The performance of the best configuration of each classification algorithm is presented below. The input features are normalized to values between -1 and 1 to weight all features equally. Afterwards the samples in the dataset are randomized and split into 80 % training data and 20 % validation data. 3 Results This chapter discusses the results of the friction state classification. It has already been shown that the AE signal is speed and temperature dependent [6]. Thus, it is assumed that the classification of the friction condition can be improved if the bearing temperature and the sliding speed are included as input features in the classification algorithm. This approach does not conflict with the wind turbine application, since the necessary measurements are also available in the real application. Table 2 shows the achieved classification accuracy for all three algorithms with different input features. The given accuracy is the ratio of the number of overall correct classifications to the number of overall datapoints. The highest Science and Research 19 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 ( ) = 0 , = 0 1 , 0 < 15 2 , 15 < 45 3 , > 45 Equation 5 sliding speed used as input features are shown in Figure 10. It can be observed for all three algorithms that the friction states “fluid friction” and “severe mixed” friction can be detected relatively accurate (values above 90 %). “Mild mixed friction” is more challenging to detect (accuracy values of around 70 to 80 %) due to the very weak change in AE signal characteristics (see also Figure 4). In summary all three algorithms performed with Science and Research 20 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 accuracy is achieved by the GP classifier using AE data, bearing temperature and sliding speed as input features. This corresponds to the findings presented in [14], where the AE signals have been proven to be highly speed dependent. Overall, it should be noted that the accuracy of the classification in this work increases with the number of input features. The confusion matrices for all three classification algorithms with the AE signals, bearing, temperature and Support vector machine classifier Gaussian process classifier Multi-layer perceptron classifier Input: Accuracy [%] Accuracy [%] Accuracy [%] AE 81,6 85,6 80,7 AE + T 84,1 88,1 82,7 AE + v 86,1 90,4 87,0 AE + T + v 89,8 92,6 90,9 Table 2: Classification performance of the three implemented algorithms based on the AE measurements and different additional feature inputs Figure 10: Confusion matrices for all three algorithms with the AE signals, bearing temperature and sliding speed as input features Figure 11: (a) Planetary journal bearing test gearbox (b) AE sensor mounted to the front face of the planetary pin (a) (b) an acceptable accuracy. Since SVM and GP surpass the neural network in terms of simplicity and training effort, both approaches are favorable for the application in a CMS. a. Transfer of the method to the gearbox application The previously presented measurements result from experiments on a component test rig for radial journal bearings. To show the transferability of the presented method from component level to planetary journal bearings in wind turbine gearboxes further experiments were conducted. Figure 11 (a) shows a test rig for planetary journal bearings which allows for the examination of the bearing behavior under the influence of operating conditions typical for planetary bearings in wind turbines. The test rig consists of two parallel shafts driving a planetary gear between them. The planetary gear rotates around a fixed axis (planetary pin) with a planetary journal bearing. In contrast to the previously investigated component test rig, the planetary bearing test rig considers gear influences and planetary gear tilting. Thus, the behavior of the test bearing is more realistic to the real WT application. The AE Sensor in this test setup is mounted to the front face of the planetary pin (see also Figure 11 (b)). In this case the distance from the sensor to the center of the journal bearing is 180 mm, which is closer to the application in a gearbox in the field. There the accessibility is limited so that it is expected, that the sensor needs to be place at a significant distance to the load zone. In order to ensure comparability with the previous experiments, tests are also carried out on the planetary journal bearing test rig with constant sliding speed and different pressure levels. The bearing and gear material and lubricant are identical to those on the aforementioned component test rig. The post-processing of the measurement data was done identically to the previous experiments. The bearing temperature is measured directly under the sliding surface in the bearing sleeve. Figure 12 (a) shows an example for the conducted test runs. Both AE signal features behave similarly to the previous experiments. In addition, an amplitude modulation of the signals is noticeable that can be explained through the gear mesh influence. It is assumed that the overall effectiveness of the signal features for the classification of the friction condition is not affected by this influence and that the transfer of the signal processing developed on the component test rig is transferable to a gear stage application. In total five different sliding speeds have been examined between v = 0.1…0.3 m/ s. All tested operating points are within the range of typical load cases for planetary journal bearings in wind turbines. The results from the radial journal bearing test rig show that the gaussian process classifier achieves the best classification performance (Table 2). Figure 12 (b) shows the confusion matrix resulting from a GP classifier trained to differentiate between the same four friction states using the AE data from the planetary journal bearing. The labels for this dataset are created using an EHD model from the planetary journal bearing test gearbox with an equal parameterization to the previously shown model. The classification results show that the presented method is transferable to the planetary journal bearing application. The overall classification accuracy for the friction state “mild mixed friction” has even increased by 11 %. Science and Research 21 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 Figure 12: (a) AE measurement results on the planetary journal bearing test gearbox (b) Friction state classification with a GP classifier achieved with the AE-measurement from the planetary journal bearing test (a) (b) density with special attention to a low-noise turbine operation. In: Conference for Wind Power Drives 2019 Conference Proceedings [3] Hase, A., Mishina, H. u. Wada, M.: Correlation between features of acoustic emission signals and mechanical wear mechanisms. Wear (2021) [4] Strablegg, C., Summer, F., Renhart, P. u. Grün, F.: Prediction of Friction Power via Machine Learning of Acoustic Emissions from a Ring-on-Disc Rotary Tribometer. Lubricants 11 (2023) 2, S. 37 [5] König, F., Marheineke, J., Jacobs, G., Sous, C., Zuo, M. J. u. Tian, Z.: Data-driven wear monitoring for sliding bearings using acoustic emission signals and long shortterm memory neural networks. Wear 476 (2021), S. 203616 [6] Mokhtari, N., Guhmann, C. u. Nowoisky, S.: Approach for the Degradation of Hydrodynamic Journal Bearings based on Acoustic Emission Feature Change. In: IEEE International Conference on Prognostics and Health Management. 2018 [7] König, F., Jacobs, G., Stratmann, A. u. Cornel, D.: Fault detection for sliding bearings using acoustic emission signals and machine learning methods. IOP Conference Series: Materials Science and Engineering 1097 (2021) 1, S. 12013 [8] König, F., Sous, C., Ouald Chaib, A. u. Jacobs, G.: Machine learning based anomaly detection and classification of acoustic emission events for wear monitoring in sliding bearing systems. Tribology International 155 (2021), S. 106811 [9] Mokthari, N. u. Glühmann, C.: Classification of journal bearing friction states based on acoustic emission signals. Technisches Messen (2018) [10] König, F., Sous, C. u. Jacobs, G.: Numerical prediction of the frictional losses in sliding bearings during start-stop operation. Friction (2020) [11] Greenwood, J. A. u. Tripp, J. H.: The Contact of Two Nominally Flat Rough Surfaces. Proceedings of the Institution of Mechanical Engineers 185 (1970) 1, S. 625-633 [12] AVL: EXCITE Power Unit User Manual [13] Rasmussen, C. E. u. Williams, C. K. I.: Gaussian Processes for Machine Learning. The MIT Press 2005 [14] Mokhtari, N., Pelham, J. G., Nowoisky, S., Bote-Garcia, J.-L. u. Gühmann, C.: Friction and Wear Monitoring Methods for Journal Bearings of Geared Turbofans Based on Acoustic Emission Signals and Machine Learning. Lubricants 8 (2020) 3, S. 29 Science and Research 22 Tribologie + Schmierungstechnik · volume 71 · issue 2/ 2024 DOI 10.24053/ TuS-2024-0008 4 Summary and outlook For the condition monitoring of journal bearings, a reliable detection of mixed friction is essential. This work presents an approach for a friction state detection based on AE measurements and supervised machine learning methods. The experiments for establishing the method are performed on a test bench for radially loaded journal bearings. The collected data is fed into three different classification algorithms, of which the GP proves to have the best classification accuracy. Finally, experiments on a planetary journal bearing test rig demonstrated that the method developed in this work can be applied to planetary journal bearings. As shown above the simulative distinction between different friction states strongly depends on the accuracy of the model’s parameterization. In this work the stochastic contact model according to Greenwood and Tripp is used and its parameters (e.g. elastic factor K) are fitted to represent the performed experiments. For a field application of the presented method in a real wind turbine environment the precision of the friction state labels should be further improved. The methods presented in this work will be extended in the future to feature additional metrics such as surface acoustic wave measurements from the bearing’s sliding surface and particle counting in the oil supply lines. In future work the classification software will be extended by unsupervised learning methods to enable an anomaly detection for mixed friction occurring during special events like dry-running or particle contamination. In a further step, the CMS measurement technology will be transferred to a WT gearbox with planetary journal bearings and tested on a system test bench to finally demonstrate the transferability. This will bring the CMS methods closer to the application in the field. Acknowledgement This research was funded by the German Federal Ministry of Economic Affairs and Climate Action. Literature [1] Umweltbundesamt: Erneuerbare Energien in Deutschland. Daten zur Entwicklung im Jahr 2022 (2023) [2] Lubenow, K., Schuhmann, F. u. Schemmert, S.: Requirements for wind turbine gearboxes with increased torque