Incipient Gradual Fault Detection via Transformed Component and Dissimilarity Analysis

Abstract

This paper proposes a novel method named recursive transformed component dissimilarity analysis (RTCDA) combining dissimilarity analysis algorithm and traditional sliding window technique for detecting incipient gradual faults. Firstly, orthogonal transformed components (TCs) corresponding to a new set of data in the sliding window are obtained using a recursive algorithm based on rank-one modification. Then, to quantitatively estimate the distribution difference of TCs, the dissimilarity index between TCs of the new dataset and that of referenced dataset is calculated. The distribution of TCs changes more dramatically than that of original data after a small quantitative bias in the original data. Compared with original data, TCs are more sensitive to tiny quantitative variation of dataset. Finally, case studies on a numerical example and a practical industrial fed-batch penicillin fermentation process are carried out to evaluate the performance of RTCDA method for incipient gradual fault detection.