Optimal Selection of Components in Fault Detection Based on Principal Component Analysis

Abstract

Selection of the optimal number of principal components (PCs) in fault detection using principal components analysis (PCA) is considered in this paper. The focus is on the relationship between the sensitivity to a particular fault and the number of PCs retained. The selection method that is based on signal to noise ratio of the fault detection (known as fault SNR) is compared to the cumulative percent variance (CPV) and the Scree methods of determining the optimal number of PCs to be retained for fault detection based on PCA. The SNR fault detection method shows different dependencies on the number of PCs for different kinds of faults. The number of PCs that gives the maximum sensitivity is easily determined for sensor faults by examining the fault SNR. If apriori data is available as operational data that has been measured during faulty conditions, then optimization of the number of PCs for the process fault is possible. The methods are applied to a thermal system.