Efficient On-line Fault Diagnosis for Non-Linear Systems

Fault diagnosis is a critical task for autonomous operation of systems such as spacecraft and planetary rovers, and must often be performed on-board. Unfortunately, these systems frequently also have relatively little computational power to devote to diagnosis. For this reason, algorithms for these applications must be extremely efficient, and preferably anytime. In this paper we introduce the Gaussian particle filter (GPF), an efficient variant on the particle filtering algorithm for non-linear hybrid systems. Each particle samples a discrete mode and approximates the continuous variables by a multivariate Gaussian that is updated at each time-step using an unscented Kalman filter. The algorithm is closely related to Rao-Blackwellized Particle Filtering and equally efficient, but is more broadly applicable. We show that given the same computation time GPF performs diagnosis with a significantly lower rate of incorrect diagnoses and with a much lower error on the continuous parameters. We also use the GPF to diagnose data from the K-9 rover at NASA Ames Research Center.