I. Introduction
The problem of detecting and isolating (identifying or classifying) a fault (or a target, or an anomaly) in the presence of random noise and deterministic uncertainty often appears in engineering applications (see, for instance, [1]–[5]): network monitoring, navigation integrity monitoring, nondestructive quality control, continuous process monitoring, image processing, etc. It is assumed that the vector of observations represents either an additive sum of a fault, an unknown nuisance deterministic vector and a Gaussian noise, or just a sum of an unknown nuisance vector and an additive Gaussian noise. If it is present, a fault (or a target, or an anomaly) has to be identified (or classified) as one of several predefined types.