Label-free anomaly detection in wastewater treatment plants via enhanced independent component analysis

Abstract Effective fault detection in wastewater treatment plants (WWTPs) is crucial for maintaining operational efficiency and preventing costly failures. This paper presents a semi-supervised fault detection framework that requires only fault-free data for training. The proposed method integrates Independent Component Analysis (ICA) for extracting statistically independent latent features from multivariate process data, combined with the Kolmogorov–Smirnov (KS) test for detecting distributional changes in residuals via sample-wise comparison. To ensure flexible and reliable thresholding, Kernel Density Estimation (KDE) is employed. The ICA–KS approach is evaluated using benchmark WWTP data across various fault types, including bias, drift, intermittent faults, freezing faults, and magnitudes, as well as simultaneous faults. Experimental results show that the method consistently outperforms traditional PCA- and ICA-based strategies, offering high accuracy and good sensitivity to weak and evolving faults.