Distribution-based canonical correlation residual monitoring for industrial processes fault detection

Abstract

Abstract Reliable fault detection is essential for safe, efficient, and sustainable industrial processes. This paper proposes a Canonical Correlation Analysis (CCA)-based fault detection framework with distribution-based residual monitoring using Hellinger Distance (HD), Kullback–Leibler Distance (KLD), and Kantorovich Distance (KD). Specifically, the proposed approach evaluates residual distributions within a moving window and determines decision thresholds nonparametrically using Kernel Density Estimation (KDE), avoiding restrictive Gaussian assumptions. A key advantage is that the framework requires only fault-free data for model construction and threshold calibration, making it suitable when labeled fault data are scarce or unavailable. The framework detects faults that disrupt the CCA cross-correlation structure, including faults affecting variables in either the predictor or response variables. The proposed approach is evaluated on the Williams–Otto reactor and a Wastewater Treatment Plant (WWTP) under bias, intermittent, drift, and freezing faults. Additional experiments examine faults in CCA input variables, sensitivity to moving-window and KDE settings, and different fault magnitudes. The proposed approach is also compared with CCA-based Squared Prediction Error (CCA-SPE), Principal Component Analysis (PCA), and Support Vector Regression (SVR) monitoring under identical fault scenarios and evaluation conditions. The results show that distribution-based CCA indicators generally outperform conventional CCA-SPE and the considered PCA- and SVR-based strategies. Among them, CCA-KD achieves the best overall performance, with high detection and low false alarm rates. Overall, integrating CCA with distribution-based monitoring provides effective multivariate fault detection.