Intelligent diagnosis of water supply pipe leakage based on multi-modal information fusion and improved heterogeneous temporal graph neural network

Leakage detection and diagnosis of water supply pipe are crucial for ensuring urban water safety and reducing waste of water resources. Nowadays, pipe leakage is mainly detected using single-modal information (such as images, sound, etc.) combined with various numerical models or algorithms. However, there are many factors that affect pipe leakage in real environment. It is difficult for single-modal data to reflect the true information of pipe leakage accurately, and the collected data often have the problem of asymmetric information or partial missing. To address these challenges, an improved heterogeneous graph neural network model (p-HTGNN) with multi-modal information fusion is proposed, and the leakage monitoring experiment of the water supply pipe is carried out. The experimental and analytical results show that p-HTGNN achieves an F-score of 88.5 % for the classification of leakage defects and an F-score of 86.5 % for the diagnosis of leakage risk level. The recognition accuracy for different features all exceeds 95 %, with overall performance superior to other traditional detection algorithms. This work provides a novel method for accurately reflecting the actual situation of water supply pipe leakage and for carrying out leakage diagnosis intelligently and efficiently. • A novel intelligent diagnosis method of water supply pipe leakage based on multi-modal information fusion and improved heterogeneous temporal graph neural network(p-HTGNN). • Process and integrate the multi-modal information obtained from multiple sensors. • The LSTM module is added to improve the performance of the HTGNN model. • Ablation studies and model comparison experiments were carried out to verify the superiority of the p-HTGNN.