Abstract:

It is challenging to monitor and predict the health condition of a motor-bearing system in a physically interpretable manner. To address this issue, we develop a physics-informed machine learning (PIML) framework that incorporates empirical physics governing bearing health, characterized by lubricant film thickness and bearing surface roughness. The proposed PIML model consists of dual neural networks that characterize unknown dynamics of the lubricant film thickness and the surface roughness, respectively, along with a vibration model that incorporates random metal-to-metal impacts. Experimental data collected on a motor-bearing system under two different lubrication conditions are used to train the model and evaluate its performance. Results show that the trained model predicts four validation indicators within empirical ranges. Furthermore, the trained model also identified the onset of the starved-lubrication regime, in which bearing degradation accelerates. Finally, the effectiveness of the trained model is also verified in predicting the remaining useful life (RUL) of the bearing.