24th International Colloquium Tribology - January 2024 161 Detection of Critical Operation in Porous Journal Bearings Using Machine Learning Josef Prost, Guido Boidi, Georg Vorlaufer, Markus Varga * AC2T research GmbH, Wiener Neustadt, Austria * Corresponding author: markus.varga@ac2t.at 1. Introduction Many tribosystems are designed for long-lasting operation. Therefore, a condition monitoring strategy is necessary to timely identify critical operation and to avoid catastrophic failure. This work introduces a semi-supervised Machine Learning (ML) approach for the classification of operational states on the example of porous journal bearings [1]. Time and frequency feature sets allow to combine data from multiple sensors in order to gain new insights from tribological data by training more robust and accurate models. With this approach, ML algorithms can be trained to assess the operational state, indicating the current health status, as well as to predict the remaining useful lifetime of machine elements. 2. Experimental The presented model was developed using data generated during accelerated lifetime tests of porous journal bearings. The test rig allows to run up to five tests simultaneously and was equipped with sensors measuring, e.g., friction torque, sample holder displacement, and acceleration, see Fig. 1. Ironand bronze-based bearing materials were investigated at room temperature, 100°C, and 160°C. The bearings run on a ø8 mm X90CrMoV18 shaft and were impregnated with apolyglycol-based lubricant prior to the tests. A specific bearing load of 3 MPa was applied at all experiments, and they were run with a triangular profile from +20 to -20 rpm with 0.1-Hz frequency. Especially the constantly changing rotation direction leads to mixed lubrication conditions and accelerated wear, which allowed for lifetime tests of the bearings. Tests were run until a preset critical temperature was exceeded after several 10 to 300 hours. Tests running longer than 300 hours were stopped manually. 3. Machine Learning To achieve higher robustness and classification accuracy, data from these sensors were combined by calculating time and frequency feature sets from 60 second time windows of hourly recorded high-speed data. Each of these time windows was assigned to one of four operational regimes: ‘Run-in’, ‘Steady’, ‘Vibration’, and ‘Critical’. The labelling procedure was assisted by multivariate statistics, i.e., principal component analysis (PCA) and clustering. The most relevant features were selected by an automated iterative feature selection algorithm to further enhance the performance of the model. In a comparative study, six different “classical” ML classifiers were investigated. Hereby, analgorithm based on aggregated decision trees (“Extra-Trees”) proved best performing and was selected for training the model. Figure 1: Detailed view of one sample stage of the accelerated lifetime test rig with instrumentation. 4. Results and Discussion Exemplarily, acceleration data of two running conditions are presented in Fig. 2. The top image depicts the stable running condition, with low vibrations even at the change points of the running direction. In the bottom picture, the acceleration signal of critical running behavior shortly before the failure is shown, where at low velocities very high accelerations can be measured in both running directions. Figure 2: Accelerometer signal from different running conditions during one test run. 162 24th International Colloquium Tribology - January 2024 Detection of Critical Operation in Porous Journal Bearings Using Machine Learning Figure 3: Classification of operational states of an iron-based porous journal bearing operated at 160°C. Each colored dot represents a 60 second time window. The grey line indicates the mean coefficient of friction for each time window. An example of the trained model applied to a test run is shown in Fig. 3. The trained model was able to identify the correct operational state with an overall accuracy of 0.98. Thereby, critical operation was detected up to 50 hours before the stopping criterion of the test rig was reached. A similar model may be integrated into a predictive maintenance strategy, as it can be implemented into an online monitoring tool for identifying critical operation in real time. Impending failures can be detected hours before an alarm raised by a conventional threshold based system. This allows the operator to undertake timely countermeasures before the machinery would suffer catastrophic damage. 5. Acknowledgments This work was funded by the Austrian COMET-Program (Project K2 In Tribology1, no. 872176) and carried out at the Austrian Excellence Center for Tribology (AC2T research GmbH). Part of the work was also funded by the Interreg AT- CZ 279 project ‘‘AI-based Predictive Maintenance’’. References [1] J. Prost, G. Boidi, A.M. Puhwein, M. Varga, G. Vorlaufer. Classification of operational states in porous journal bearings using a semi-supervised multi-sensor Machine Learning approach. Trib. Int. 184 (2023) 108464.