Fault detection: the effect of unknown distribution of residuals

Abstract

Residuals are typically used as indicators of normal (non-faulty) vs. abnormal (faulty) behavior in dynamic systems. The nonfaulty residuals are assumed to be Gaussian, zero-mean, uncorrelated, with a known variance. However, in many practical situations, the assumption of Gaussian-ness may not be valid. We propose a new type of fault detector which is essentially independent of the distribution of the residuals. This fault detector is based on an autoregressive modeling of the residual signal, augmented by a sample variance calculation. Usefulness of this new detector is demonstrated with the experimental fault data obtained at NASA Langley Research Center.