Unsupervised anomaly detection in water distribution SCADA system using Gaussian mixture models with GMM-lime explainability

Reliable fault detection in existing benchmark water distribution systems (WDS) dataset is complicated by severe class imbalance, high sensor dimensionality, and significant distribution shift between historical fault signatures and faults observed during operation. This paper presents a three-stage framework that addresses each challenge in turn. Six supervised classifiers are evaluated on a 43-sensor WDS dataset and shown to perform poorly on fault-sensitive metrics, with the best F1 score of only 0.41 (Random Forest). A diagnostic analysis using Mahalanobis distance shows a 30-fold gap between training and test fault distributions, explaining why supervised training fails. Recursive feature elimination (RFE) driven by a Random Forest then reduces the sensor set from 43 to 15, retaining the sensors with the highest joint discriminative value. A Gaussian mixture model (GMM) trained exclusively on normal data subsequently achieves F1 = 0.533, AUC-ROC = 0.853, and AUPR = 0.582, outperforming all supervised baselines. For interpretability, the GMM-LIME (local interpretable model-agnostic explanations) framework replaces standard Gaussian perturbation with GMM-structured sampling, improving average Jaccard similarity from 0.544 to 0.684 for fault instances across five repeated runs. The proposed system offers a practical, interpretable route to assist human operators in WDS SCADA fault monitoring where labelled fault data are scarce and non-representative.