International Colloquium Tribology
ict
expert verlag Tübingen
125
2022
231
Opportunities and Applications for Artificial Intelligence in Sealing Technology
125
2022
Matthias Baumann
Lukas Merkle
Frank Bauer
ict2310471
23rd International Colloquium Tribology - January 2022 471 Opportunities and Applications for Artificial Intelligence in Sealing Technology Matthias Baumann University of Stuttgart, Institute of Machine Components, Stuttgart, Germany Corresponding author: matthias.baumann@ima.uni-stuttgart.de Lukas Merkle University of Stuttgart, Institute of Machine Components, Stuttgart, Germany Frank Bauer University of Stuttgart, Institute of Machine Components, Stuttgart, Germany 1. Introduction Sealing technology is an expert science in which the experience and the expertise of the players are important in many respects. In the case of failure analysis, it is often the small things that reveal the causes of failure and the damage mechanisms. In many places, a great deal of effort is therefore being expended to optimize the methods used for the analysis of sealing systems, [1]. Specialized equipment has even been developed, for example to produce the best possible images of the components of a sealing system or to measure relevant parameters user independently [2-4]. Despite considerable effort in failure analysis and the use of standardized evaluation matrices (see Figure 1), interlaboratory tests with several participants regularly show scattered evaluation results. This is a delicate matter in standardized seal test procedures, which potentially aim for the release of an oil for a given application. Of course, the inherent scatter of tribological tests contributes to these results, [5]. However, when the measurement data of all participants are considered together as a team, there is usually a clear unification of the evaluation results. This is ultimately due to the different levels of knowledge and the individual subjective evaluation of the experts performing the tests. One way of standardizing the evaluation of tribological tests and reducing the influence of the single user is to use artificial intelligence methods. Assistance systems that are trained with previously coordinated and appropriately classified image data can preserve and transfer expert knowledge. Furthermore, they can help to keep the track about the sometimes hundreds of involved images and analysis results. 2. Failure Analysis Methodology for Rotary Shaft Seals Standardized test specifications for evaluating e.g. the oil-elastomer compatibility of newly developed oils with rotary shaft seals made of standard elastomers have become established in the industry, [6]. Systematic procedures form the basis for the evaluation and, associated with this, the release of the tested oils for the respective application. In addition to measuring the wear width, the radial load and the inner diameter of the tested seals, such approvals also include a visual inspection of the elastomeric sealing edges of the rotary shaft seals. To ensure consistent results over a longer period of time and a higher number of tests, an evaluation matrix was developed at the institute. The so-called IMA-MARS (Matrix for the Advanced Rating of Seals, ) has been iteratively improved ever since then, Figure 1. Methodically, the evaluation includes an examination of the sealing edges in the uncleaned and cleaned condition. 472 23rd International Colloquium Tribology - January 2022 Opportunities and Applications for Artificial Intelligence in Sealing Technology Figure 1: Matrix for the Advanced Rating of Seals - MARS [1] Based on 4 groups (Mechanical Failure, Thermal Attack, Oil Carbon Deposits, Chemical Attack), the condition of the elastomer sealing edge is systematically rated in terms of the sealing function. But however, a systematic approach alone does not lead to fully consistent results. Carried out individually, there are usually always certain deviations in the ratings due to the individual subjective influence of the user. Therefore, comparison tables are used and the ratings are carried out in large expert teams at joint meetings. All this represents a not inconsiderable effort. In order to work efficiently and consistently, good preparation and coordination is required. This has led to the development of intelligent assistance software at the institute, which supports the raters in their work. 3. Image Classification Part of the evaluation criteria, which are determined in the course of the visual analysis of sealing edges described above, are determined on uncleaned images of the sealing edge and the other part on cleaned images of the sealing edge, both taken e.g. with the IMA Sealobserver [7]. The following rating can be done conveniently if the rater is specifically shown the correct image data in each case. To make this possible in an automated way, Convolutional Neural Networks (CNNs) from the field of machine learning can be used. These are particularly well suited for classifying images and can thus distinguish undeclared image data. Such a CNN is roughly composed of an image input layer, a combination of different convolutional and pooling layers and a classification layer. It converts an image file into a classification vector via previously trained computing operations. We trained a CNN with the image data of sealing edges, obtained from past research projects of the institute. The image data were previously divided manually (supervised machine learning) into the classes: uncleaned, undamaged and damaged, whereas damaged was mainly focused on the group thermal attack to the sealing edge. A split of the data into 80 % training data and 20 % validation data led to a validation accuracy of 99 % after the training procedure. The resulting classification network was transferred into a graphical user interface and now allows a targeted display of the image data. The tool is used in rater meetings, which have since become much more efficient and less time consuming. Figure 2: Graphical user interface for the rating seals according to the IMA-MARS 4. Image Regression In addition to classification problems, CNNs can also be used to calculate regressions. This is of interest when image data must be converted into parameters, as is also required here in the case of the individual evaluation criteria. The otherwise manually performed evaluation is is therefore automated. The rater receives an evaluation proposal from an intelligent assistance system and can use this as a basis for checking the image data and, if necessary, taking corrective action. The problem here is that the evaluation criteria presented above are based on a global view of a large number of images. Accordingly, a complete derivation of all evaluation criteria cannot be made on the basis of a single image, which is the way CNNs work. Several images have to be linked to get a result. For this purpose, a Long Short-Term Memory (LSTM) [8] network was implemented. This 23rd International Colloquium Tribology - January 2022 473 Opportunities and Applications for Artificial Intelligence in Sealing Technology network architecture belongs to the Recurrent Neural Networks (RNN) and was developed for the analysis of data sequences such as videos or speech. Series of images can be processed sequentially, whereby information about arbitrary intervals is stored in a LSTM cell. Through this procedure it is possible to transform a set of image data into a regression of the evaluation criteria. An LSTM network was trained with the previously mentioned image data. The evaluation required for this using the IMA MARS was previously carried out in an expert team meeting. The network was integrated into the graphical user interface and now supports the user in the evaluation of the image data. However, an interlaboratory test among several users is still pending. The aim is to investigate whether the user-dependent variation in the evaluation can be reduced by means of such assistance systems and whether consistent results are obtained overall. 5. Conclusion The field of artificial intelligence is growing at a tremendous rate. The tools that have already been developed enable a wide range of applications in almost all technical areas. There are also many possible applications in sealing technology. Convolutional Neural Networks can be used for classification of image data in damage analysis to make the evaluation process more comfortable for the user. LSTM networks also permit the creation of systemic evaluation matrices from image data sets. These can serve as a basis guess for a user, they preserve expert knowledge and are intended to make results user-independent and even more consistent in the future. Both examples show the benefits of the application of machine learning methods for application within the sealing technology. Many other applications are further conceivable, as for example artefact detection on sealing counterfaces, classification of manufacturing methods and much more. References [1] Bauer, F.: Federvorgespannte-Elastomer-Radial- Wellendichtungen, Wiesbaden: Springer Fachmedien Wiesbaden, 2021, - ISBN 978-3-658-32921-1. [2] Baumann, M.; Bauer, F.: Moderne visuelle Untersuchungsmethoden für die Verschleißanalyse am Beispiel Radial-Wellendichtring. 20th International Sealing Conference (ISC), Stuttgart, 10.-11. Oktober 2018; Fluidtechnik; Frankfurt am Main: Fachverband Fluidtechnik im VDMA e.V, 2018, S. 93-104 - ISBN 978-3-8163-0727-3. [3] Schollmayer, T.; Burkhart, C.; Kassem, W.; Thielen, S.; Sauer, B.: Verschleißanalyse an Radialwellendichtringen und weiteren Maschinenelementen mittels Laserprofilometrie. Tribologie-Fachtagung. 2021, 62 , 70/ 1-70/ 10. [4] Fehrenbacher, C.; Hörl, L.; Bauer, F.; Haas, W.: Description of the Pumping Rate of Shaft Counterfaces in the Sealing System Radial Lip Seal Using the 3D Parameters of ISO 25178. Tribology Online. 2016, 11 (2), S. 69-74. [5] Bauer, F.: Tribologie, Wiesbaden: Springer Fachmedien Wiesbaden, 2021, - ISBN 978-3-658-32919-8. [6] Hüttinger, A.; Hermes, J.; Wöppermann, M.; Prem, E.: Neues Prüfverfahren für dynamische Dichtungen von Getriebemotoren. Erschienen in: Jahrbuch Dichtungstechnik 2016, Friedrich Berger; Sandra Kiefer; Mannheim: ISGATEC GmbH, 2015 - ISBN 978-3-9811509-9-5. [7] Universität Stuttgart, Institut für Maschinenelemente: IMA-Sealobserver. Universität Stuttgart, Institut für Maschinenelemente (Hrsg.), IMA- TechSheet, V1, #102070, Stuttgart, URL: https: / / www.ima.uni-stuttgart.de/ dokumente/ forschung/ dichtungstechnik/ einrichtungen/ 102070.pdf. [8] Hochreiter, S.; Schmidhuber, J.: Long short-term memory. Neural computation. 1997, 9 (8), S. 1735- 1780.
