Methodology on Establishing Multivariate Failure Thresholds for Improved Remaining Useful Life Prediction in PHM

Abstract

Prognostic approaches commonly try to predict the Remaining Useful Life (RUL) based on machine health status by either directly establish a mapping or setting up a failure threshold to determine the End-of-Life (EoL). On the one hand, determining a failure threshold is crucial but subjective for most reported cases. Machine operation risks, which are intuitional but difficult to quantify, can be used to bridge the gap between prediction and determining a multivariate failure threshold. On the other hand, historical machine life information is rarely considered together with the condition indicators for such prognostic tasks. Building multivariate failure thresholds based on quantifiable operation risks for prognostic tasks is the general topic that is rarely studied due to the following challenges: 1) How to quantify operation risks under multiple variables? 2) How to determine the multivariate failure thresholds? 3) How to make reliable extrapolations of future conditions? To address these questions, as the extension of our previous work (Jia, Li, Wang, Li, & Lee, 2020), this paper proposes 1) a Gaussian Copula model-based risk quantification method to determine multivariate failure thresholds, and 2) a Similarity enhanced Blackwellized Particle Filter (RBPF) to predict future system conditions. Two examples of establishing tri-variate and bi-variate failure thresholds are given. The proposed methodology is validated on the aero-engine RUL prediction task based on the C-MAPSS dataset from the PHM society data competition 2008. The result suggests that the proposed methodology has better explainability and practicability with comparable prediction capability.