Canadian Journal of Chemical Engineering, Vol.96, No.2, 426-433, 2018
ONLINE INCIPIENT FAULT DIAGNOSIS BASED ON KULLBACK-LEIBLER DIVERGENCE AND RECURSIVE PRINCIPLE COMPONENT ANALYSIS
Fault detection and isolation (FDI) methods based on the principal component analysis (PCA) model have achieved a large number of theoretical studies and applications, especially for complex and highly dimensional processes. However, the Hotelling's T-2, that is the most common used statistical distance, can fail in detecting small shifts such as a sensor incipient fault with low fault-to-noise ratio (FNR). Although an incipient fault develops slowly, it cannot be ignored and is necessary to be detected early enough to avoid more serious consequences. In this study, a realistic online diagnosis method for incipient faults with low FNR is presented. Based on probability distribution measure, the Kullback-Leibler divergence (KLD) is utilized to compare the probability density of each of the latent scores to a reference one. Under the hypothesis of Gaussian distribution, dynamic changes of KLDs are computed via the mean and variance of score vectors, which can be updated online utilizing the recursive principal component analysis (RPCA). From simulations, it is shown that the proposed approach can detect, isolate, and estimate the sensor incipient fault of the multivariate AR system successfully.