Particle Filter and Its Variants for Degradation State Estimation and Remaining Useful Life Prediction

Hengyangcan Gu,Hassan Mahmoud,Raj Lourd Arun,Jie Liu,Xinyi Ma

DOI: https://doi.org/10.1109/phm58589.2023.00055

2023-01-01

Abstract:State estimation and future state prediction is the main objective of prognostics and health management (PHM) frameworks. The health state can sometimes be either inaccessible or complicated to be measured directly under operating conditions. The degradation process estimation is usually described as a nonlinear or non-Gaussian online tracking problem, and the uncertainty due to multi-source variability makes the estimation challenging. Inference through Particle Filter (PF) is one of the Bayesian paradigms to allow the estimation possible. The motivation is to compute the posterior of a Markov process's hidden health state with noisy and partial observations. Particle Filter is a sequential Monte Carlo method. It uses particle representation to recursively update the posterior of a stochastic process when new observations become available. However, the generic algorithm, also known as Sequential Importance Sampling with Resampling (SISR), suffers the well-known drawbacks of particle degeneration and impoverishment. With rising expectations, Regularized Particle Filter (RPF) and Auxiliary Particle Filter (APF) are proposed to alleviate the problems. In a recent development, Regularized Auxiliary Particle Filter (RAPF) reportedly performs better than the other variants. In this work, we reviewed and assessed the framework and performance of each PF variant using the benchmark model and a case study of lithium batteries to estimate the health state and remaining useful life. The results reflect the previous research work, indicating that RAPF gives the best estimation among all methods in both experiments.