Masayuki Tamura

A study on the number of principal components and sensitivity of fault detection using PCA

Selection of the number of principal components (PC) in the fault detection method using principal component analysis (PCA) is considered. In this paper, we focus on the relationship between the sensitivity of fault detection and the number of PCs to retain. Consideration of the signal-to-noise ratio of fault detection (Fault SNR) is proposed. The Fault SNR shows different dependency 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 a priori data is available, that is, operation data measured during faulty conditions, optimization of the number of PCs for the process fault is also possible. In a case where a priori information of the fault is not available, monitoring multiple models with various numbers of PCs in parallel is considered the next best strategy.

A Study on the Selection of Model Dimensions and Sensitivity of PCA-Based Fault Detection

Abstract A method of determining model dimensions (number of principal components) which maximize the sensitivity of fault detection was studied. In this paper it is shown that the sensitivity of PCA-based fault detection generally depends on the number of principal components. Most of the existing methods give only one value as a recommended number of components, and so sometimes the sensitivity is poor for certain kinds of faults. Among existing methods, although the Variance of Reconstruction Error (VRE) criterion gives a recommended model dimension which depends on the kind of fault, it is not intended to maximize sensitivity. This paper presents a new method of determining the model dimension which maximizes the sensitivity of PCA-based fault detection

Fault Diagnosis by Parity Relations Designed Through Partial Least Squares

Abstract In this paper, a new design method of parity relations for fault diagnosis by partial least squares regressions is presented. The method has an advantage that parity relations can be designed based on the vast amounts of correlated data without significant knowledge of devices. The method has an advantage of easy isolation of sensor faults.