Explainable Machine-Learning Leak Identification Framework for Water Distribution Networks

Leak identification in urban water distribution networks (WDNs) is a critical task with potential consequences. Acoustic-based convolutional neural networks (CNNs) have been widely employed for leak identification, leveraging its advantages with fast processing and advanced pattern recognition capabilities. However, CNN exhibits a deficiency in interpretability, which hinders a comprehensive understanding of its decision-making processes. This limitation may reduce their reliability and impede further improvements. To address this issue, this study proposes an explainable CNN-based water leak identification model that incorporates gradient-weighted class activation mapping (Grad-CAM) for improved interpretability. For data collection, acoustic signals were collected from WDNs in Hong Kong. Besides, variational mode decomposition (VMD) was employed to denoise the signals. Three CNN-based models [residual network (ResNet), visual geometry group (VGG), and AlexNet] were developed and assessed based on their leak identification performance. After training, Grad-CAM analysis was employed on the above models to reveal their working mechanism for better interpretability. The results demonstrate that, among the developed models, AlexNet and VGG-19 demonstrated superior identification performance on testing data sets and out-of-sample testing. The Grad-CAM visualization results revealed that VGG-19 and Grad-CAM's attention regions effectively captured essential signal components. Furthermore, a comparison of Grad-CAM results for different leakage conditions indicated that the model primarily relied on the frequency components of 750-1,250 Hz to classify signals as leaks, while frequency components below 500 Hz were the primary focus for noleak signals. These insights shed light on the decision-making process of the model and highlighted the specific frequency bandwidth associated with different identification models and leakage scenarios. Overall, the proposed method offers an innovative approach to enhance and explain acoustic CNN-based leak identification models, potentially advancing their effectiveness and interpretability.