Predictive maintenance of journal bearings: Remaining Useful Lifetime Prediction using Isolation Forest and Temporal Convolutional Networks

Journal bearings are machine elements that are widely used in equipment ranging from vehicles to industrial machines. Their reliable operation is in many cases critical and journal bearing failure may cause subsequent high costs. This is particularly true for bearing seizure, which is a catastrophical failure of a heavily stressed journal bearing where extremely large amounts of heat are being generated. This motivates to study bearing failure and different ways to predict it from actual operation. In this work, two approaches to predict the remaining useful lifetime of journal bearings are presented. The fi rst attempt tries to classify the lubrication conditions - the occurrence of mixed lubrication is the required precursor to seizure. For this task, an isolation forest model was established on the basis of simulation data. The results obtained with the isolation forest showed good classifi cation performance, although not accurate enough for real-world applications. The second attempt goes significantly beyond the fi rst approach and utilizes measured data of seized bearings to predict the actual remaining load to failure. For this approach a Temporal Convolutional Network (Temporal Convolutional Network (TCN)) was trained using a dataset consisting of several experimental seizure tests where bearings are tested in different ways until failure. The results with the TCN showed very good Remaining Useful Lifetime (RUL) prediction accuracy, being able to predict the load to failure with a very small Root Mean Squared Error (RMSE) error of only about 10%.