Short-Term Cross-Sectional Time-Series Wear Prediction by Deep Learning Approaches

Abstract

Wear is one of the major causes that affect the performance and reliability of tribo-systems. To mitigate its adverse effects, it is necessary to monitor the wear progress so that preventive maintenance can be timely scheduled. An online visual ferrograph (OLVF) apparatus is used to obtain online measurements of wear particle quantities, and monitor the wearing of a four-ball tribometer under different lubrication conditions, and several popular deep learning algorithms are evaluated for their effectiveness in providing maintenance decisions. The obtained data are converted to the cross-sectional time series (CSTS), for its effectiveness in representing the variation trends of multiple variables, and the data are used as the input to the deep learning algorithms. Experimental results indicate that the CSTS together with the bidirectional long short-term memory (Bi-LSTM) architecture outperforms other tested settings in terms of the mean-squared error (MSE). Increased prediction accuracy is observed for tribological pairs with a stochastically changing coefficient of friction.