Water Supply Network Leakage Detection Method Based on Fusing Spatiotemporal Features and Attention Mechanism​

Urban water supply networks (WSNs) are essential infrastructure but remain vulnerable to leakage caused by pipe aging, construction defects, and external stresses. Conventional detection methods often underperform in noisy or data-sparse environments. This study develops an attention-based spatiotemporal fusion framework for robust leak detection in WSNs. Pressure-driven simulations combined with an improved K-means algorithm enable optimized sensor placement and capture hydraulic dynamics. A CNN-LSTM-Attention model is then employed to integrate spatial–temporal representations and enhance feature focusing. The unified framework strengthens the identification of weak leakage signals under complex hydraulic conditions. Field validation in a mountainous county in southeastern China achieved 99.4% detection accuracy and a 99.6% F1 score, outperforming SVM, LSTM, and CNN benchmarks. The proposed method provides a reliable approach for intelligent leakage diagnosis and supports the development of resilient, data-driven smart water systems.