A two-stage leak monitoring framework for water distribution networks based on acoustic signals

[Display omitted] Timely monitoring of pipeline leaks is essential to avoid secondary contamination of drinking water and economic losses and to ensure the sustainability of water distribution networks (WDNs). However, it is prominent that the field acoustic signal sensors have high false alarm rates and unclear leak management priorities. To address the above issues, a two-stage leak monitoring framework based on acoustic signals is proposed in this work by drawing on meteorological domain knowledge, including two stages of leak detection and leak assessment. (1) In the leak detection stage, the acoustic signals are converted to Mel spectrograms, and the Residual Convolutional and Transformer networks (RCT-Net) are coupled to achieve local and global extraction of spectral features. Then, the pipeline state (leak and no-leak) is determined after calculating the leak probability of each pipeline by nonlinear feature fusion. (2) In the leak assessment stage, the pre-trained RCT-Net in the first stage is retrained and fine-tuned to achieve different leak level predictions (Light, Medium, and Severe), and improve the fault tolerance of field applications. Taking acoustic signals of real WDNs as an example, comparing the deep learning models of CNN architecture and Transformer, the framework has the best recognition performance and robustness, with more than 97% of all evaluation metrics on the leak detection and level assessment, including AUC up to 99.7%. The proposed monitoring framework can provide comprehensive leak monitoring information, which provides a new perspective on leak management prioritization, reducing water losses, and guiding water utilities toward sustainable development.