Degradation modeling under time-varying operating conditions: Inference and prognosis with particle filter

Abstract

The health state estimation and remaining useful life prediction have always been challenging when a mechanical system works under time-varying operating conditions (TVOC). TVOC can not only influence the amplitude of degradation growth, but also have an impact on how the baseline degradation process is observed. In this paper, we propose a state-space model to estimate the health state of a single-unit system under continuous TVOC. Compared to most existing works where only one discrete operating condition is considered, our model accounts for multi-dimensional and continuous TVOC. The baseline (hidden) degradation is described by stochastic processes or general path models, and the observations are subject to TVOC-related observation bias and noise. Given the raw observations and the TVOC sequences, a particle filter (PF) is used to estimate the hidden degradation process. When the model parameters are unknown, Monte Carlo Expectation-Maximization is used for parameter estimation. As a result, the health state at any time and any TVOC can be estimated. The proposed method is applied to a real choke valve erosion dataset, showing a significant improvement in the health state estimation compared to traditional methods.