Abstract
Water Distribution Networks (WDNs) are vital infrastructures, andcontamination poses serious public health risks. Harmful substances caninteract with disinfectants like chlorine, making chlorine monitoring essentialfor detecting contaminants. However, chlorine sensors often become unreliableand require frequent calibration. This study introduces the Dual-ThresholdAnomaly and Drift Detection (AD&DD) method, an unsupervised approach combininga dual-threshold drift detection mechanism with an LSTM-based VariationalAutoencoder(LSTM-VAE) for real-time contamination detection. Tested on tworealistic WDNs, AD&DD effectively identifies anomalies with sensor offsets asconcept drift, and outperforms other methods. A proposed decentralizedarchitecture enables accurate contamination detection and localization bydeploying AD&DD on selected nodes.