Probability-Relevant Incipient Fault Detection and Diagnosis Methodology With Applications to Electric Drive Systems

By dealing with the crowding problem caused by incipient faults, this brief will develop a new fault detection and diagnosis (FDD) scheme called probability-relevant principal component analysis from the probability view point. The proposed methodology cooperates with Kullback-Leibler divergence from the information field and Bayesian inference from the machine learning domain. Compared with the standard FDD methods under the framework of multivariate statistical analysis, this new FDD scheme is more sensitive to faults under an acceptable false alarm ratio, especially to incipient faults; moreover, it is more accurate in diagnosing faults with the aid of improved fault detectability. The effectiveness of the proposed FDD method is illustrated by mathematical analysis and geometric descriptions, and validated via a numerical example and a real experimental setup on the electric drive system of a high-speed train.