Enhancing Fault Detection with Clustering and Covariance Analysis

Fault detection plays an important role in identifying abnormalities in high-cost, large-scale industrial processes. Clustering in combination with dimensionality reduction is a common practice in data analysis and anomaly detection but is not well explored in the field of industrial fault detection. In this paper, we apply correlation clustering before and after dimensionality reduction to enhance fault detection on the Tennessee Eastman Process. The reduction techniques employed are principal component analysis (PCA) and dynamic inner principal component analysis (DiPCA). This paper also introduces a novel index (delta(2)(S)) for monitoring the covariance of principal and residual components. Adding the novel index increases fault detection rates of PCA by more than 20% on 5 faults, and clustering before PCA increases fault detection by over 10% on 5 faults. This paper demonstrates two points: our clustering method can considerably increase fault detection rates, and the novel delta(2)(S) index can boost performance on noisy faults, which were previously considered to be difficult to detect. Copyright (c) 2022 The Authors.This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)