A Novel Method for Detecting the Leakage Duration of Water Pipe Network Based on the Anchor Box Mechanism

In recent years, automated leakage detection has played an increasingly important role in the safe operation of water distribution systems. To a large extent, the physical localization of leaks depends on the leakage detection results of the temporal signals collected from the water distribution network. In view of the high noise characteristics of pipeline signal acquisition and the inability of traditional temporal anomaly detection methods to accurately detect leakage duration, this paper proposes a leakage duration detection method based on an anchor box mechanism for water distribution networks. Firstly, in order to remove the noise interference in the pressure signal of the water distribution network, the VMD decomposition method is used to denoise the collected pressure data. Secondly, in order to capture the temporal correlation of leakage signals at different scales, causal convolutions with different dilation rates are employed to extract features from the denoised pressure data. Then, self-attention mechanism is utilized to further enhance the extracted features. Thirdly, to reduce the number of anchor boxes, a sliding window mechanism based on attention scores is employed to generate candidate leakage duration. Finally, the leakage duration is obtained through the position regression and category prediction. Experiments are conducted on two publicly available datasets, LeakDB and BattLeDIM, and the results show that the achieved Accuracy (ACC) is ${8 6 . 1 9 \%}$ and $83.41 \%$ respectively, and the Intersection over Union (IoU) is 0.76 and 0.74 respectively. Compared with the existing advanced methods, although the proposed method adopts a lightweight network structure, it can also obtain higher precision performance, and can be used for rapid diagnosis of leakage in high leakage areas.