Physics-informed Lightweight Temporal Convolution Networks for Fault Prognostics Associated to Bearing Stiffness Degradation

Abstract

This paper proposes hybrid methods using physics-informed (PI) lightweight Temporal Convolution Neural Network (PITCN) for bearings’ remaining useful life (RUL) prediction under stiffness degradation. It includes three PI hybrid models: a) PI Feature model (PIFM) — constructing physics-informed health indicator (PIHI) to augment the feature space, b) PI Layer model (PILM)—encoding the physics governing equations in a hidden layer, and c) PI Layer Based Loss model (PILLM)—designing PI conflict loss, taking into account the difference before and after integration of the physics input-output relations involved module to the loss function. We simulated 200 different bearing stiffness degradations, using their discrete monitored vibration signals to verify the effectiveness of the proposed method. We also investigate their inference process through feature heat map analysis to interpret how the models melt physics knowledge to assist in capturing the degradation trend. The physics knowledge considered in this paper is the dynamic relationship between vibration amplitude and stiffness in a damped forced vibration model. The results show that all three PITCN models effectively capture degradation-related trend information and perform better than the vanilla lightweight TCN. Furthermore, the visualization of the feature channels highlights the important role of physics information in model training. Channels containing physics information demonstrate higher correlation with results as they significantly dominate the heat map compared to other channels.