Extracting Dissimilarity of Slow Feature Analysis between Normal and Different Faults for Monitoring Process Status and Fault Diagnosis

Abstract

A process monitoring method based on the dissimilarity (DISSIM) of slow feature (SF) analysis is proposed for effective fault detection in the process industry. The useful information from this method is mainly contained in low-frequency data signals. The sensitive slow features (SSFs) of a single fault status with maximum dissimilarity between normal data SFs and fault data SFs are initially selected by a dissimilarity analysis and are used to construct monitoring statistics and obtain the control limits of the corresponding fault status. The most probable fault status for online monitoring is selected by addressing the smallest Euclidean distance between the SSFs of the online data and the corresponding SFs of the fault data. Subsequently, the SSFs are redefined according to the corresponding fault status, and the statistics and corresponding control limits are compared to detect faults. A simulation of the Tennessee Eastman process demonstrates that the proposed method outperforms conventional methods.