Probabilistic Models for Monitoring and Fault Diagnosis

Abstract

Reliably detecting and diagnosing faults is very important for autonomous systems. The problem is made difficult due to the large number of faults that can occur and the fact that most faults cannot be observed directly, but must be inferred from noisy sensor readings. Probabilistic models, such as Partially Observable Markov Decision Processes (POMDPs), are a natural representation for tracking the state of a stochastic system. To be useful for fault diagnosis, however, these models must be able to perform in real time and should be able to account for both anticipated and unanticipated faults. This paper presents some of our ongoing work in using POMDPs and particle filters for modeling and tracking faults in autonomous systems. We demonstrate how these methods can be used to detect, diagnose, and recover from faults, operating in real time on-board mobile robots.