Fault Detection and Identification for Advanced Vehicle Control Systems

Abstract

A preliminary design of a health monitoring system for automated vehicles is developed and tests using a high-fidelity nonlinear simulation are very encouraging. The approach is to fuse data from dissimilar instruments using modeled dynamic relationships and fault detection and identification filters. The filters are constructed so that the residual has static directional characteristics associated with the presence of a fault. These patterns may be blurred by sensor noise, disturbances, unmodeled dynamics and nonlinearities. A Bayesian neural network is designed to produce a probability that a fault has occured conditioned on the residuals. This is done by recognizing, in a stochastic sense, quasi-static fault patterns embedded in the residual.