Water contamination anomaly detection is a challenging task due to heterogeneous sensor data and stringent privacy constraints. In this paper, we propose a federated approach to the detection of water contamination anomalies that addresses both technical and organizational challenges in water monitoring utilizing open-source water quality data from the city of Austin. The model applies various data-related techniques to build a high-performance contamination detection system by using long-short-term memory (LSTM) networks. This approach ensures data-knowledge distillation from various water quality datasets without centralized environmental monitoring data, creating a paradigm for data utility and privacy protection as a critical infrastructure process. Experimental results show that our FedAvg-based federated learning approach achieves a test-set MSE of 0.07—substantially lower than the best traditional LSTM algorithm with a MSE of 0.13, while preserving data privacy.
