Unsupervised Water Pipe Network Leakage Detection Method Based on Discrete-Time Dynamic Graph Network

Compared to the large number of non-leakage signals during normal water flow in pipelines, leakage samples are scarce and difficult to collect, resulting in leakage detection being a small sample problem. In addition, training models is difficult to reflect dynamic adaptability, and an increase in time span will inevitably cause inference delays. Therefore, This paper proposes an unsupervised leakage detection method for water pipe networks using a discrete-time dynamic graph network. First, the time graph is sliced into snapshots, where node embeddings and flow distributions are learned from the geometric and temporal relationships of measurement points. Spatial and temporal attention networks enhance node features and graph embeddings, while conditional flow estimates the positive sample density of these features. Second, a recurrent neural network dynamically adjusts the graph convolutional network’s parameters, enabling adaptive modeling. Finally, leakage detection is performed by computing the KL divergence between the test sequence and the trained positive sample density. To validate the effectiveness of our method, we conducted experiments on the LeakDB Net1 dataset and Hanoi dataset. On the Net1 dataset, TPR, FPR, and AUC were 95.3%, $4.8 \%$, and $93.9 \%$, respectively, while on the Hanoi dataset, they were $92.7 \%, 9.5 \%$, and $90.1 \%$, respectively. The experimental results fully demonstrate the effectiveness of the proposed method in detecting water pipe network leakage.