A water pipe leakage detection method based on Multi-view Embedding Distance Feature Fusion Network.

Leak detection is crucial for the healthy operation of water pipelines and efficient water resource conservation. However, traditional acoustic signal-based leak detection methods for water pipelines often capture only shallow or incomplete features, resulting in suboptimal accuracy in identifying leaks of varying severity. To address this limitation, we propose a water pipeline leak detection method based on a Multi-View Embedding Distance Feature Fusion network (MDFF) to comprehensively integrate leakage information from multiple perspectives. This method first designs encoders based on the characteristics of pipeline vibration signals from different views to extract deep features as embedding representations. Subsequently, an attention prototype network is proposed to extract multi-view distance features, which are fused using a Dempster-Shafer (DS) fusion module. Finally, the fused results are utilized to classify normal pipeline conditions and leakage states with varying severity. The proposed approach was tested on actual pipelines and compared with single-view methods and other multi-view approaches. The experimental results demonstrate that, in the face of normal pipeline conditions and three levels of leakage, the MDFF method achieves an accuracy of 99.69 % even with limited sample data. Consequently, the proposed model significantly enhances the accuracy of identifying different leak severity levels, possesses strong generalization capabilities, and exhibits superior detection performance.