Physics-informed prognostic model for rolling element bearing using auxiliary particle filter with resample-move and inline oil debris data

Abstract

Rolling element bearings are susceptible to spalling, a major failure mechanism due to rolling contact fatigue. It poses failure risks in rotating machinery like wind turbines and aircraft engines. The data-driven method for remaining useful life (RUL) prediction has shown significant potential in intelligent prognostics when data availability is limited, and model interpretability is critical. This paper proposes a novel framework for RUL prediction of rolling element bearings by integrating inline oil debris data from Gastops Ltd.’s MetalSCAN sensors with an auxiliary particle filter (PF) enhanced by a Resample-Move (APF-RM) strategy. The proposed data-model fusion approach incorporates physics-informed knowledge of the spalling process to improve interpretability and robustness under limited data conditions. In recent years, the PF has gained popularity in Bayesian state estimation. However, conventional PFs suffer from inherent challenges, notably particle degeneracy and sample impoverishment. The proposed APF-RM mitigates these challenges by 1) leveraging an APF to improve the efficiency of the sampling distribution and 2) applying an RM technique to increase the diversity of particles in both state and parameter spaces. Experimental results show that the proposed framework outperforms the conventional PF, achieving improved prediction accuracy and more robust inference performance. Furthermore, the framework shows potential for extension to other bearing types and sizes, particularly in applications with limited experimental data.