A fast-running failure prognostic algorithm based on a non-homogeneous markov chain

Abstract

Typically, model-based prognostic algorithms estimate the remaining-useful-life distribution by characterizing the probable system trajectories described through state-space models. Unfortunately, as the state dimension increases or the prognostic horizon enlarges, the computational time of such algorithms augments considerably, complicating their real-time execution. To overcome this difficulty, this work proposes a paradigm change in model-based prognostic algorithms; instead of tracking the state-space trajectories, a Fast-Running Markov Chain-based Prognostic Algorithm (FRMC-PA) is proposed, capable of estimating the time-of-failure probability mass function directly. FRMC-PA is based on a two-state non-homogeneous discrete-time Markov chain, where state 0 describes the operative situation and state 1 represents a catastrophic failure event. FRMC-PA is composed of two stages: i) offline stage, which leverages historical data to train a regression model that maps the system variables to the transition probabilities of the binary-stochastic process; ii) online stage, which combines the regression model built previously and real-time observations to estimate the system s remaining useful life. This method is validated using the case study of battery discharge. Results show that FRMC-PA can transfer most of the computational cost to the offline stage, achieving an online computational-time reduction of 99% compared with a Monte-Carlo-based prognostic, without significantly sacrificing the prognostic accuracy.