Fault Detection and Remaining Useful Life Estimation Using Switching Kalman Filters

Abstract

The use of condition monitoring (CM) data in degradation modeling for fault detection and remaining useful life (RUL) estimation have been growing with increasing use of health and usage monitoring systems. Most degradation modeling methods requires fault detection thresholds to be established. When the CM measure exceeds the detection threshold, RUL prediction is then performed using a time-invariant dynamical model to represent the degradation path to the failure threshold. Such approaches have some limitations as detection thresholds can vary widely between individual units and a single dynamical model may not adequately describe a degradation path that evolves from slow to accelerated wear. As such, most degradation modeling studies only focuses on segments of their CM data that behaves close to the assumed dynamical model. In this paper, the use of Switching Kalman Filters (SKF) is explored for both fault detection and remaining useful life prediction under a single framework. The SKF uses multiple dynamical models describing different degradation processes from which the most probable model is inferred using Bayesian estimation. The most probable model is then used for accurate prediction of RUL. The proposed SKF approach is demonstrated to track different evolving degradation path using simulated data. It is also applied onto a gearbox bearing dataset from the AH64D helicopter to illustrate its application in a practice.