GSPO-LAD: A Graph Neural Network-Based Sensor Placement and Anomaly Detection in Water Distribution Systems

AI-based sensor placement for anomaly detection in water distribution systems (WDS) enhances system reliability and efficiency. Advanced WDS are attractive targets for malicious actors who may gain control over these systems to disrupt or manipulate water services for personal gain. Monitoring various points in the network is crucial for effective anomaly detection. However, increasing the number of sensors or placing them inaccurately can raise computational costs, increase management overhead, and degrade detection performance. In this study, we address sensor placement optimization (SPO) and anomaly detection (AD) in WDS to improve performance while reducing computational and management overhead by focusing on critical sensor locations. We use the BATADAL dataset, which includes key WDS components such as pump units, pressure junctions, and tanks. Our implementation identifies the most critical pump unit locations for monitoring flow rates, allowing us to focus only on essential points to lower computational costs and management overhead in anomaly detection. For SPO, we implement Graph Neural Networks (GNN) to extract relationships between pump units and other components, identifying the most significant pump units for anomaly prediction. GNN-based SPO (GSPO) selects critical locations, and we then use flow rate data from these selected sensors to train the AD model. We employ a Long Short-Term Memory (LSTM)-based AD approach (LAD) to capture sequential patterns in the sensor data. The LAD approach achieved significant results with an accuracy of 0.925 and an F1-score of 0.75. To further enhance model performance over time, we adopt a continuous learning approach. We fine-tune the LSTM model monthly with new incoming data while retaining previous knowledge using a weight interpolation approach, ensuring the model remains up-to-date and continues to improve in detecting anomalies.