Sparse dissimilarity analysis based on distribution dissimilarity decomposition for online diagnosis of incipient faults

Abstract

For fault detection and diagnosis, the conventional multivariate statistical process control (MSPC) methods in general quantify the distance between the new sample and the modeling samples. They, however, do not check the changes of data distribution as long as monitoring statistics stay inside normal region enclosed by control limit, which, are not sensitive to incipient changes. In the present work, a novel fault diagnosis method is developed which can isolate the incipient abnormal variables that change the data distribution structure and does not need any prior knowledge of historical fault process. First, the distribution dissimilarity is decomposed deeply and significant dissimilarity is extracted to integrate the critical difference of variable covariance structure between the reference normal operation distribution and the actual distribution resulting from incipient process disturbances. Second, a sparse dissimilarity (SDISSIM) algorithm is proposed to isolate abnormal variables associated with changes of distribution structure. It shows that fault diagnosis based on distribution dissimilarity analysis can be formulated as a regression-type optimization problem. Sparse coefficients are obtained with only a small fraction of variables' coefficients nonzeros, pointing to abnormal variables. As illustrations, SDISSIM is applied to both simulated and real industrial process data with encouraging results to figure out the slight distortions.