Aus Wissenschaft und Forschung 18 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 Detection of wear in sliding bearings with the use of machine learning techniques Martin Winnertz, Georg Jacobs, Thao Baszenski, Mattheüs Lucassen, Benjamin Lehmann* Eingereicht: 01.09.2023 Nach Begutachtung angenommen: 03.01.2024 Dieser Beitrag wurde im Rahmen der 64. Tribologie-Fachtagung 2023 der Gesellschaft für Tribologie (GfT) eingereicht. Die Zustandsüberwachung von Maschinenelementen ist im Rahmen der Digitalisierung im Maschinenbau von grundlegender Bedeutung für die Wirtschaftlichkeit und Ressourceneffizienz. Mithilfe der Zustandsüberwachung können kritische Betriebszustände sowie sich anbahnende Schäden frühzeitig detektiert werden. Durch eine Erkennung und Differenzierung von Schäden können Maschinen bedarfsgerecht gewartet werden, sodass die Wartungsintervalle verlängert und Kosten eingespart werden können. In dem vorliegenden Beitrag wird Mischreibung mit Hilfe der Messung der thermoelektrischen Spannung detektiert. Eine messbare Veränderung der thermoelektrischen Spannung tritt bereits durch den Kontakt von Gleitlager und Welle auf, sodass vor dem Eintreten erster Schäden am Gleitlager eine Veränderung des Messsignals vorliegt. Die Detektion der Mischreibung erfolgt mit Hilfe eines Machine Learning Modells, da dies eine zuverlässige Identifizierung von Mischreibung in Messdaten verschiedener Gleitlagersysteme ermöglicht, auch wenn für diese keine Trainingsdaten vorliegen. Für die Generierung von Trainings-, Testsowie Validierungsdaten werden Experimente an einem Gleitla- Kurzfassung gerprüfstand durchgeführt. In den Experimenten werden Lager im Mischreibungsbetrieb unter einer definierten spezifischen Pressung und Gleitgeschwindigkeit untersucht. Am Prüfstand ist ein Sensor appliziert, mit welchem die thermoelektrische Spannung zwischen Lager und Welle gemessen wird. Für die Bewertung einer Übertragung des Machine Learning Modells auf Messdaten von anderen Gleitlagersystemen werden an einem Getriebeprüfstand Validierungsdaten erzeugt. Für die Detektion von Mischreibung in verschiedenen Gleitlagersystemen wird ein überwachter Machine Learning Algorithmus entwickelt, welcher Messdaten der thermoelektrischen Spannung bezüglich des Reibungszustandes klassifiziert. Mit dem entwickelten Machine Learning Modell kann Fluid- und Mischreibung in Gleitlagersystemen zuverlässig differenziert werden und somit kritische Betriebsbedingungen von Maschinen auf Basis der Zustandsüberwachung vermieden werden. Schlüsselwörter Condition Monitoring, Gleitlager, Machine Learning, thermoelektrische Spannung In the context of digitalization in mechanical engineering, condition monitoring of machine elements is of fundamental importance for economic efficiency and resource efficiency. With the help of condition monitoring, critical operating states and incipient damage can be detected at an early stage. By detecting and differentiating damage, machines can be maintained as needed, so that maintenance intervals can be extended and costs saved. In this paper, mixed friction is detected by measuring the thermoelectric voltage. Contact between the sliding Abstract bearing and the shaft causes a measurable change in the thermoelectric voltage, so that a change in the measurement signal is noticeable before the first damage occurs. The detection of mixed friction is carried out using a machine learning model, which can reliably identify mixed friction in measurement data from different sliding bearing systems, even if no training data is available. For the generation of training, test as well as validation data, experiments are conducted on a sliding bearing testbench. In the experiments, bearings are examined TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 18 1 Introduction Unexpected bearing failures are considered one of the major causes of machine downtime. For example, 60 % of wind turbine downtime is caused by gearbox damage, 67 % of which is due to bearing failure [1]. In combustion engines of ship propulsion systems, main bearing failure with subsequent crankshaft damage is the largest cost factor for repair [2]. These two examples illustrate the influence of bearing damage on the economy and reliability of machinery. The use of condition monitoring (CM) systems to detect damage at an early stage and minimise unplanned downtime is one way of counteracting the influence of bearing damages on the economy and reliability of machinery [3]. CM systems are already used, for example, to monitor roller bearings, where a change in the raceway of the rolling elements is detected by analysing the acceleration [4, 5]. This method is not transferable to sliding bearings due to the lack of roll over events. Common methods for CM of sliding bearings include analysis of bearing temperature, thermoelectric voltage, particles in the lubricant, frictional torque, shaft orbit, acceleration and acoustic emissions (AE) [6]. In practice, the monitoring of bearing temperature, acceleration, frictional torque and shaft orbit is made difficult by the often non-existent but necessary access to the sliding bearing, or by the excessive space required for the sensors [6, 7]. In contrast, the monitoring of the AE, the thermoelectric voltage and the particles in the lubricant does not require the sensors to be positioned directly on the sliding bearing [6, 8, 9]. A disadvantage of AE is the demanding and complex data processing due to the large amount of data recorded during AE measurements [10]. The disadvantage of using sensors to monitor particles in the lubricant is that it takes additional efforts to associate the particles with a machine component (gear, bearing, etc.) [11]. In [9] the influence of mixed friction on the thermoelectric voltage was investigated. The BEAROMOS sensor from Schaller Automation [12] was used in [9] to determine the thermoelectric voltage. It was shown that the thermoelectric voltage is suitable for monitoring sliding bearings with respect to potentially damaging operating conditions. Thermoelectric voltage sensors make use of the Seebeck effect, which describes the formation of a voltage between two different metallic bodies in contact at different temperatures [13]. Equation (1) shows the relationship between the voltage U t , the temperature difference dT and the material and temperature dependent Seebeck coefficients Q A and Q B of the two materials. If the materials and the temperatures are identical, there is no thermoelectric voltage [14]. (1) In sliding bearing systems consisting of a sliding bearing, a shaft and a lubricant, there is a thermoelectric voltage between the sliding bearing and the shaft. If the two bodies are completely separated by the lubricant, there is a constant, system-dependent thermoelectric voltage level. If the lubricant film is not high enough to separate the bearing and shaft, the two bodies come into contact and mixed friction occurs. The high frictional power in mixed friction between the shaft and the bearing causes both bodies to heat up. Due to the different geometries and heat capacities of the shaft and the bearing, different temperatures are reached in the bodies. As a result of the temperature difference and the contact of the two bodies, a thermoelectric voltage gradient occurs. However, this thermoelectric voltage gradient can be influenced either by particles in the lubricant supply of the sliding bearing or by other contacts in the system, such as in roller bearings. The dependence on other machine elements such as gears is unknown. In addition, the behaviour of the = ( ( ) ( ) ) Aus Wissenschaft und Forschung 19 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 * Martin Winnertz, M.Sc. Univ.-Prof. Dr.-Ing. Georg Jacobs Orcid-ID: https: / / orcid.org/ 0000-0002-7564-288X Thao Baszenski, M.Sc. Orcid-ID: https: / / orcid.org/ 0000-0002-0978-0563 Mattheüs Lucassen, M.Sc. Orcid-ID: https: / / orcid.org/ 0000-0001-6072-5650 Benjamin Lehmann, M.Sc. Orcid-ID: https: / / orcid.org/ 0000-0002-8328-2503 RWTH Aachen, Institut für Maschinenelemente und Systementwicklung, Schinkelstraße 10, 52062 Aachen in mixed friction operation under a defined specific pressure and sliding velocity. A sensor is applied to the testbench to measure the thermoelectric voltage between bearing and shaft. To evaluate the transferability of the machine learning model to measured data from other bearing systems, validation data will be generated on a gearbox testbench. A supervised machine learning algorithm is developed to detect mixed friction in various sliding bearing systems by classifying thermoelectric voltage measurement data with respect to the friction condition. The developed machine learning model can be used to reliably discriminate between fluid and mixed friction in sliding bearing systems, thus avoiding critical operating conditions of machines based on condition monitoring. Keywords Condition Monitoring, Sliding Bearings, Machine Learning, thermoelectric Voltage TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 19 found on the transfer of a model trained on thermoelectric voltage data for the detection of mixed friction in sliding bearings to another sliding bearing system. However, it is known that thermoelectric voltage is suitable for monitoring sliding bearings with respect to potentially damaging operating states. In addition, a change in the thermoelectric voltage can already be measured when contact occurs between the bearing and the shaft. This has the advantage that potentially harmful operating conditions can be detected at a very early stage. For this reason, a system for detecting friction states based on thermoelectric voltage is developed in the following investigations. As the suitability of SVC for detecting friction states has already been demonstrated on the basis of vibration and AE data, an SVC is also used in this work. The SVC is trained, tested and validated with data from experiments on a component testbench. To investigate the transfer of the SVC to another system, an experiment is performed on a gearbox testbench. 2 Materials and testbenches The testbenches and materials used to generate data for the ML model are presented below. Experiments to generate training, test and validation data are carried out on a component testbench. In addition, a test to generate validation data to investigate the transfer of the trained ML model to unknown sliding bearing systems. The BEAROMOS sensor from Schaller Automation was used to determine the thermoelectric voltage. 2.1 Component testbench Training, test and validation data is generated on the sliding bearing component testbench shown in Figure 1. This testbench allows sliding bearings to be tested at different sliding velocities and static loads. The bearing is mounted torsionally stiff and pulled against the shaft by an electric cylinder. The cylinder and the sliding bearing are connected by a tilting parallelogram, the so-called friction scale. The frictional torque between the bearing and the shaft is measured using the friction scale, a force sensor and the known lever arm between the force sensor and the bearing’s axis of rotation. The radial force applied to the bearing, the speed, the temperature of the be- Aus Wissenschaft und Forschung 20 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 magnitude and gradient of the voltage change for different systems is unknown. [9] To evaluate the measured data of the thermoelectric voltage, the use of machine learning (ML) methods is suitable, since the influences of machine elements and different systems on the thermoelectric voltage cannot be described. ML is a subdomain of artificial intelligence that recognises patterns in data to make decisions without being explicitly programmed to do so [15]. Compared to conventional data analysis, machine learning has the advantage of being able to recognise complex correlations that are difficult or impossible for humans to identify [16]. One of the widespread applications of ML is the classification of measurement data. In [17], Osisanwo compares seven different data classification algorithms and shows that the supervised support vector machine classifier (SVC) achieves the highest number of correct classifications. The results are applicable to use cases with numerical input data. In [18], Cervantes points out the widespread use of SVC. For example, SVC is used in text categorisation, image classification, face recognition and many other applications. In addition, he shows the increase of publications and book chapters with the search term Support Vector Machine (SVM) from 23 in 2000 to about 2600 in 2017. Jack and Nandi use an SVC in [19] to detect damage to roller bearings based on vibration measurements. In [20], Goyal also uses an SVC based on vibration measurements to detect damage to roller bearings and to classify different types of damage. Moosavian and Mokhtari both use SVC to classify mixed friction in sliding bearing systems, with Moosavian using vibration data and Mokhtari using AE data. It is shown that the SVC, together with suitable sensor data, reliably detects friction conditions in sliding bearings. The method has not been applied to new sliding bearing systems [21, 22]. The quality of classification algorithms such as SVC is evaluated on the basis of the three criteria recall, precision and F1 - score (equations (2) - (4)). Recall indicates the probability that a positive object is correctly identified as positive. Precision is the ratio of the correctly positive class to all results classified as positive. The F1 - score combines recall and precision and forms the harmonic mean. [15] (2) (3) (4) Despite extensive searches in established scientific databases and journals, no relevant literature was found on the detection of friction states in sliding bearings based on thermoelectric voltage. Similarly, no literature was = + = + 1 = 2 + Figure 1: Sliding bearing component testbench TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 20 aring in the load zone and the thermoelectric voltage between the bearing and the shaft are also recorded. The thermoelectric voltage measurement circuit is implemented between the shaft and the friction scale. The contact to the rotating shaft is made by sliding contacts. 2.2 Gearbox testbench To investigate the system transferability of the ML model, a test is carried out on the gearbox testbench shown in Figure 2. This gearbox testbench is used in current and past research projects to investigate the behaviour of sliding bearings in planetary gearboxes. The testbench abstracts a planetary gearbox with three shafts on which helical gears are mounted. The sliding bearing is shrunk onto the pin of the planet gear. Lubricant is supplied to the plain bearing through holes drilled in the pin. The design of the testbench allows the sliding bearings to be tested under loads applied to planetary bearings in planetary gearboxes. The control of the testbench records the bearing temperature in the load zones, the speed and torque on the planet as well as the thermoelectric voltage between the sliding bearing and the planet gear. 2.3 Bearing The test bearings on the component testbench are made of the material CuSn12Ni2-C, with a diameter of 30 mm and a width of 15 mm. The bearing clearance is 1.67 ‰. The roughness is Ra 0.6 µm. On the gearbox testbench, a bearing of the material CuSn12Ni2-C is also used. The bearing has a diameter of 100 mm and a width of 112 mm. The bearing clearance is 1 ‰ and the roughness is Ra 1.7 µm. 2.4 Sleeve A hardened steel sleeve (100Cr6) is used as a counter body on the component testbench. The roughness is Ra 0.1 µm. The counter body on the gearbox testbench is the inner surface of the planetary gear. A polished 18CrNiMo7-7 sleeve is pressed into the gear to adjust the bearing clearance. The surface roughness of the sleeve is Ra 0.6 µm. 2.5 Lubricant On the component testbench, the bearing is lubricated with the additivefree mineral oil FVA2. This lubricant has a density of 850 kg/ m 3 and a kinematic viscosity of 32 mm 2 / s at 40 °C and 5.35 mm 2 / s at 100 °C. The lubricant was chosen to maintain comparability with previous research projects. In the gearbox testbench, Castrol Optigear Synthetic CT 320 is used, which is a synthetic lubricant. This has a kinematic viscosity of 275.3 mm 2 / s at 40 °C and 31.6 mm 2 / s at 100 °C. The lubricant is used on this testbench because it is commonly used in wind turbine gearboxes. 3 Method The procedure for developing an SVC to detect mixed friction in sliding bearing systems is presented below. To this end, tests will be carried out on the testbenches described in chapter 2.1 and 2.2, the recorded data will be preprocessed to positively influence the quality and efficiency of the SVC and finally the SVC will be trained, tested and validated. 3.1 Experiments Four tests with a total running time of 145 h were performed on the component testbench to generate training, test and validation data. During the tests, the specific pressure was kept constant and the sliding velocity was reduced in steps as shown in Figure 3. In this test pro- Aus Wissenschaft und Forschung 21 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 Figure 2: Gearbox testbench Sliding velocity [m/ s] Time Specific pressure [MPa] Sliding velocity [m/ s] Specific pressure [MPa] Specific pressure [MPa] Sliding velocity [m/ s] Experiment 1 2 3 4 2 - 0.1 2 - 0.1 3 - 0.1 4 - 0.1 3 4 6 8 Figure 3: Specific pressure and sliding velocity of the tests on the component testbench TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 21 SVC. This processing includes filtering, labelling and normalising the data, as well as calculating a feature and splitting the data into training, test and validation data. The thermoelectric voltage measurement data is filtered to suppress outliers and statistical variations in the measurements. The filtering is done using the Exponential Weighted Moving Average (EWMA) [23] according to equation (5), where ŷ t is the filtered data and y t is the unfiltered data at time t. α is the weighting factor and is chosen experimentally to be α = 0.0003. (5) With the selected weighting factor, new data is considered with a weighting of 0.03 % and old data is considered with a weighting of 99.97 %. This leads to y t having only a small influence on the value of ŷ t . However, if there is a large change in y t (y t << y t-1 or y t >> y t-1 ), this change will have a large influence on ŷ t . Figure 5 shows the application of EWMA to thermoelectric voltage measurement data. The raw thermoelectric voltage signal is shown in blue, the EWMA filtered signal is shown in green and the sliding velocity is shown in yellow. The specific pressure is constant at 8 MPa. It can be seen that small outliers, such as at t = 114000 s, have a very limited effect on the filtered signal. Despite the constant specific pressure and sliding velocity, the thermoelectric voltage slowly increases from t = 112000 s to about 76 µV from t = 114000 s. This increase in the thermoelectric voltage is due either to the running-in of the sliding bearing or to the formation of a wear groove. During the running-in of bearings, existing roughness is worn away by mixed friction and the bearing surface is smoothed. After the running-in process is complete, the sliding bearing and shaft were again completely separated by an oil film. If wear occurs on the sliding bearing, a wear groove is formed. Once the groove is formed, the sliding bearing can switch back to fluid = + ( 1 ) Aus Wissenschaft und Forschung 22 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 cedure, the bearing was initially operated in fluid friction at high sliding velocities and increasingly in mixed friction as the sliding velocity decreased. On the gearbox testbench, a test was carried out to investigate the transfer of the trained ML model to unknown sliding bearing systems. Figure 4 shows the sliding velocity and specific pressure during the test. The operating points have been selected so that mixed friction occurs at a specific pressure of 8 MPa and fluid friction at a specific pressure of 2 MPa. The sliding velocity is kept constant at 0.22 m/ s. The friction condition at the operating points was determined in advance by an EHD calculation. This test sequence was selected to specifically test the bearing in mixed friction and fluid friction conditions. This is not possible with a test sequence as shown in Figure 3. 3.2 Data preprocessing The measurement data recorded in the experiments are processed before training, testing and validating the 108000 110000 112000 114000 116000 [s] 120000 55 60 65 70 [μV] 80 Thermoelectric voltage 0.0 0.1 0.2 0.3 0.4 [m/ s] 0.6 Time Sliding velocity Thermoelectric voltage EWMA Sliding velocity Figure 5: Thermoelectric voltage, EWMA of thermoelectric voltage and sliding velocity from a component testbench experiment Sliding velocity [m/ s] Time Specific pressure [MPa] Sliding velocity [m/ s] Specific pressure [MPa] Specific pressure [MPa] Sliding velocity [m/ s] Experiment 1 0.22 2 & 8 Figure 4: Specific pressure and sliding velocity of the test on the gearbox testbench TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 22 friction operation. Both the running-in and the formation of a wear groove can occur again when the sliding velocity or the specific pressure is changed. As SVC is one of the supervised ML methods, the thermoelectric voltage data must be labelled. A distinction is made between fluid and mixed friction. This is done on the basis of the measured frictional torque, bearing temperature, thermoelectric voltage data and the results of EHD calculations. As the stability and convergence speed of SVC is strongly dependent on the range of values of the input and the range of the feature data, these are normalised [24]. This is done by centering the data around zero and scaling to the range [-1,1]. The data is centred by subtracting the value of the thermoelectric voltage at full fluid friction from the measured data. The scaling of the data is done with a MaxAbsScaler according to equation (6) [25]. This is done by dividing the data y by the absolute maximum value of y. The arithmetic mean from equation (7) is used as the thermoelectric voltage feature. (6) (7) The final step of the data preprocessing is to divide the measurement data generated on the component testbench into training, test and validation data. The measurement data from three tests are mixed and divided into training and test data with a ratio of 80: 20. The data from the fourth test is used for validation. The measurement data resulting from the test on the gearbox testbench are used exclusively for the validation of the system transferability. 3.3 Training of SVC An SVC with an experimentally determined fifthdegree polynomial kernel is used to classify mixed and fluid friction. In addition, the misclassification of mixed friction is weighted nine times more severely than that of fluid friction. = max | | = 1 This is necessary because in the training, test and validation data, the fraction of data with fluid friction is much larger than with mixed friction. Without the misclassification weighting, the SVC would tend to always predict the more common class [26]. To evaluate the accuracy of the model, recall, precision and F1 - score are calculated according to equations (2) - (4). The results after training are shown in Figure 6. The letters M and F are used as abbreviations for fluid and mixed friction in the following figures. For the detection of fluid friction, the F1 - score is 95 % and for mixed friction it is 90 %. The F1 - scores for fluid and mixed friction show a very good value of at least 90 %. The recall for mixed friction is already good at 82 %, but some mixed friction conditions are not recognised as such, so potentially harmful conditions are not detected. 4 Validation Once the model has been trained, it is validated using new data from the component testbench and the gearbox testbench, which is unknown to the model. The results of the validation on the component testbench are shown in Figure 7. Fluid friction is marked with a value of zero and mixed friction with a value of one. The thermoelect- Aus Wissenschaft und Forschung 23 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 0 5 10 15 20 25 [h] 35 -1.0 -0.5 0.0 [-] 1.0 normalized thermoelectric Voltage -1.0 -0.5 0.0 [-] 1.0 Time Class Normalized thermoelectric Voltage True Class Predicted Class Figure 7: Normalised thermoelectric voltage of the validation test on the component testbench with true and predicted classes Figure 6: Confusion matrix and scores of the SVC based on the test data TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 23 Figure 8 shows the confusion matrix as well as the recall, precision, and F1 - score for the validation on the component testbench. The F1 - score for fluid friction is slightly higher at 97 % compared to the F1 - score achieved during training. The F1 - score for mixed friction remains the same as the training score at 90 %. The values for the recall are identical. Slight deviations of a maximum of three percent occur in the Precision. In summary, fluid friction is reliably detected. Some short mixed friction states are not detected and the mixed friction states in the classification last shorter than in reality. The results of the gearbox testbench validation are shown in Figure 9. It shows that the start of mixed friction states is always detected slightly too late and the end too early. This leads to mixed friction being misclassified as fluid friction. Conversely, fluid friction is never misclassified as mixed friction. As a result, any condition with fluid friction is detected. Figure 10 shows the confusion matrix as well as the recall, precision, and F1 - score for validation on the gearbox testbench. The results of the F1 - score with 97 % for fluid and mixed friction show slightly better values than in the training and validation on the component testbench. This difference can be explained by comparing the detection of the end of mixed friction states of the validation tests on the component and gearbox testbenches in Figure 7 and Figure 9. It can be seen that the end is detected much more accurately on the gearbox testbench than on the component testbench. The difference is due to the fact that on the gearbox testbench, the transition from mixed friction to fluid friction operating conditions occurs at a dis- Aus Wissenschaft und Forschung 24 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 ric voltage normalized by the MaxAbsScaler is shown in blue, the true class is shown in green and the predicted class is shown in yellow. It can be seen that the end of mixed friction is always detected too early, e.g. at t = 31.5 h, so that the duration of the predicted mixed friction states is always somewhat shorter than in reality. In addition, short mixed friction states such as t = 15 h are not detected. 0 500 1000 1500 2000 [s] 3000 -1.0 -0.5 0.0 [-] 1.0 normalized thermoelectric Voltage -1.0 -0.5 0.0 [-] 1.0 Time Class Normalized thermoelectric Voltage True Class Predicted Class Figure 9: Normalised thermoelectric voltage of the validation test on the gearbox testbench with true and predicted classes Figure 8: Confusion matrix and scores of the SVC based on the validation data of the component testbench Figure 10: Confusion matrix and scores of the SVC based on the validation data of the gearbox testbench TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 24 crete point in time by reducing the specific pressure. (see Figure 4). In contrast, on the component testbench, this transition between friction states takes place when the bearings run in. As this is not a discrete but a continuous and long term process, the boundary between mixed and liquid friction cannot be clearly defined by the SVC and misclassifications occur in the component testbench data. Overall, the validation results on the gearbox testbench are slightly better than those on the component testbench. This is due to the different transition from mixed to fluid friction in the test sequences on each bench. 5 Summary and outlook An SVC with a fifth-degree polynomial kernel and a misclassification weighting of nine to one has been developed to detect mixed friction in sliding bearing systems based on thermoelectric voltage. Experiments were performed on a component testbench and a gearbox testbench to generate training, test, and validation data. On the component testbench, four sliding bearings were tested under static specific pressure at a stepwise decreasing sliding velocity. The gearbox testbench consisted of three parallel shafts, abstracting a planetary gear. The sliding bearing was used to support the planet and was tested at constant sliding velocity and varying specific pressure. The recorded thermoelectric voltage data were filtered with EWMA, labelled, and then centered and scaled with a MaxAbsScaler. The arithmetic mean was calculated as the feature and input of the SVC. Test data from the component testbench was used to train, test and validate the SVC. Data from the gearbox testbench experiment was used to validate the transferability of the friction condition detection to sliding bearings in other technical systems. The results show that fluid friction is detected with a recall of 100 % on both testbenches. Mixed friction is detected with a recall of 82 % on the component testbench and 93 % on the gearbox testbench. The precision for fluid friction detection is 94 % on the component testbench and 95 % on the gearbox testbench. Mixed friction is detected with 100 % precision on both testbenches. The F1 - score for fluid friction detection is 97 % for both testbenches. For mixed friction, the F1 - score is 90 % on the component testbench and 97 % on the gearbox testbench. The results of the investigations show that the detection of mixed friction on the basis of measured thermoelectric voltage data is possible and that the trained SVC can be transferred to other plain bearing systems. In the future, the aim is to reliably classify short mixed friction states. In addition, the identified mixed friction conditions are further investigated to predict the remaining life of the bearing. Acknowledgement The results of this work have been generated as part of a cooperation project with Schaller Automation, funded by the German Federal Ministry of Economics and Climate Protection. Project Partner This publication has been produced in cooperation with Schaller Automation. References [1] K. Scott, D. Infield, N. Barltrop, J. Coultate, and A. Shahaj, “Effects of Extreme and Transient Loads on Wind Turbine Drive Trains,” 2012, doi: 10.2514/ 6.2012-1293. [2] The Swedish Club, “Main engine damage,” 1998. [3] A. Jablonski, Condition Monitoring Algorithms in MAT- LAB ® , 1st ed. Cham: Springer International Publishing; Imprint Springer, 2021. [4] E. Hart et al., “A review of wind turbine main bearings: design, operation, modelling, damage mechanisms and fault detection,” Wind Energ. Sci., vol. 5, no. 1, pp. 105- 124, 2020, doi: 10.5194/ wes-5-105-2020. [5] F. P. García Márquez, A. M. Tobias, J. M. Pinar Pérez, and M. Papaelias, “Condition monitoring of wind turbines: Techniques and methods,” Renewable Energy, vol. 46, pp. 169-178, 2012, doi: 10.1016/ j.renene.2012.03.003. [6] J. Deckers, “Entwicklung einer Low-Cost Körperschallsensorik zur Überwachung des Verschleißverhaltens von wälz- oder gleitgelagerten Kreiselpumpen kleiner Leistung,” Dissertation, Gerhard-Mercator-Universität Duisburg, Duisburg, 2001. [7] W. Zhou, T. G. Habetler, and R. G. Harley, “Bearing Condition Monitoring Methods for Electric Machines: A General Review,” 2007. [8] Z. Nazarchuk, Acoustic Emission: Methodology and Application. Cham: Springer International Publishing, 2017. [9] W. Yu, G. Jacobs, C. Sous, F. König, and H. Brünnet, “Transient condition monitoring of sliding bearings subjected to mixed friction conditions,” Göttingen, Sep. 25 2019. [10] C. Grosse and M. Ohtsu, Acoustic Emission Testing. Berlin, Heidelberg: Springer Berlin Heidelberg, 2008. [11] M. Dickerhof, “Potentiale der Schallemissionsanalyse zur Überwachung und Diagnose tribologischer Systeme.,” Institut für Produktentwicklung, Karlsruher Institut für Technologie, Karlsruhe, 2011. [12] H. Brünnet, P. Kyrtatos, C. Barro, and M. Flemmer, “Combination of physical and virtual sensors for the condition-based monitoring of large engine sliding bearings,” 2022. [13] W. Demtröder, Experimentalphysik 2: Elektrizität und Optik, 5th ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009. [Online]. Available: http: / / nbn-resolving. org/ urn: nbn: de: bsz: 31-epflicht-1616872 Aus Wissenschaft und Forschung 25 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 25 nosis system for bearings,” J Intell Manuf, vol. 31, no. 5, pp. 1275-1289, 2020, doi: 10.1007/ s10845-019-01511-x. [21] A. Moosavian, H. Ahmadi, A. Tabatabaeefar, and B. Sakhaei, “An Appropriate Procedure for Detection of Journal-Bearing Fault using Power Spectral Density, K-Near est Neighbour and Support Vector Machine,” International Journal on Smart Sensing and Intelligent Systems, vol. 5, 2012. [22] N. Mokhtari, J. G. Pelham, S. Nowoisky, J.-L. Bote-Garcia, and C. Gühmann, “Friction and Wear Monitoring Methods for Journal Bearings of Geared Turbofans Based on Acoustic Emission Signals and Machine Learning,” Lubricants, vol. 8, no. 3, p. 29, 2020, doi: 10.3390/ lubricants8030029. [23] J. S. Hunter, “The Exponentially Weighted Moving Average,” Journal of Quality Technology, vol. 18, no. 4, pp. 203-210, 1986, doi: 10.1080/ 00224065.1986.11979014. [24] A. B. A. Graf and S. Borer, “Normalization in Support Vector Machines,” Pattern Recognition, vol. 23, 2001. [25] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825-2830, 2011. [26] T. Imam, K. M. Ting, and J. Kamruzzaman, “z-SVM: An SVM for Improved Classification of Imbalanced Data,” Advances in Artificial Intelligence, vol. 19, 2006. Aus Wissenschaft und Forschung 26 Tribologie + Schmierungstechnik · 70. Jahrgang · 6/ 2023 DOI 10.24053/ TuS-2023-0036 [14] R. Pelster, R. Pieper, and I. Hüttl, “Thermospannung - Viel genutzt und fast immer falsch erklärt! ,” Physik und Didaktik in Schule und Hochschule, pp. 10-22, 2005. [15] A. Géron, Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow: Concepts, tools, and techniques to build intelligent systems. Beijing, Boston, Farnham, Sebastopol, Tokyo: O'Reilly, 2019. [16] T. Hastie, R. Tibshirani, and J. Friedman, The elements of statistical learning: Data mining, inference, and prediction: Springer, 2009. [17] Osisanwo F.Y., Akinsola J.E.T., Awodele O., Hinmikaiye J. O., Olakanmi O., Akinjobi J., “Supervised Machine Learning Algorithms: Classification and Comparison,” International Journal of Computer Trends and Technology, vol. 48, 2017. [18] J. Cervantes, F. Garcia-Lamont, L. Rodríguez-Mazahua, and A. Lopez, “A comprehensive survey on support vector machine classification: Applications, challenges and trends,” Neurocomputing, vol. 408, pp. 189-215, 2020, doi: 10.1016/ j.neucom.2019.10.118. [19] L. B. Jack and A. K. Nandi, “Fault Detection using Support Vector Machines and Artificial Neural Networks, Augmented by Genetic Algorithms,” Mechanical Systems and Signal Processing, vol. 16, 2-3, pp. 373-390, 2002, doi: 10.1006/ mssp.2001.1454. [20] D. Goyal, A. Choudhary, B. S. Pabla, and S. S. Dhami, “Support vector machines based non-contact fault diag- Linguistik \ Litera Slawistik \ Skand \ Gesundheit \ R schaft \ Linguisti schaft \ Slawistik \ Sport \ Gesund wissenschaft \ Li wissenschaft \ S philologie \ Spo Fremdsprachend \ VWL \ Maschin schaften \ Sozio Bauwesen \ Frem TuS_6_2023.qxp_TuS_6_2023 01.02.24 14: 18 Seite 26