A probabilistic approach to residual processing for vehicle fault detection

This paper presents a probabilistic method for processing and analyzing residuals for the purpose of fault detection. The method incorporates residuals from multiple models using a hybrid dynamic Bayesian network in order to yield a low-cost, complete, diagnostic system. Continuous residuals are used as evidence directly in the network, and this paper discusses options for representing their probability distributions. The Bayesian network is used to model the temporal behavior of the faults, and the assumptions necessary to do this are analyzed. The diagnostic method is demonstrated on a car's handling system and experimental results are presented.