Bayesian fault diagnosis: Common approaches and challenges

Abstract

In this paper we describe a Bayesian approach to fault diagnosis based on Markov chain Monte Carlo algorithms. These approaches are largely applied to hybrid diagnosis problems in which the system being diagnosed is modelled with a mixture of discrete and continuous state variables. We describe the probabilistic hybrid automaton model typically used, and an algorithm based on particle filtering that can be applied to these models. Diagnosis provides some particular challenges for Monte Carlo approaches, including large dimensional state spaces, and low probability transitions in the Markov chain. We discuss these and some proposed solutions to them. Finally, we examine some open challenges for the Bayesian approach.