Explainable deep learning-based model development for leakage prediction and severity assessment of branched water distribution systems

Given that leakage in water distribution systems (WDSs) usually causes significant operational inefficiencies and resource depletion, effective detection and classification frameworks are necessary to address this issue. Nevertheless, current techniques that involve hardware and software-based methods frequently demonstrate insufficient accuracy in classifying leak types and assessing severity under real-world conditions. Therefore, this study classified leaks and assessed their severity levels within a WDS by examining the performance of three deep learning (DL) models: (i) long short-term memory (LSTM), (ii) recurrent neural network (RNN), and (iii) gated recurrent unit (GRU). These DL models were trained utilizing a multi-sensor dataset that included dynamic pressure, vibration, and acoustic signals collected from laboratory-scale pipe networks during controlled leak scenarios. The modeling process then involved analyzing five leakage types: (i) orifice leak (OL), (ii) gasket leak (GL), (iii) circumferential crack (CC), (iv) longitudinal crack (LC), and (v) non-leak (NL). Subsequently, severity levels were categorized into minor, moderate, and severe based on variations in flow rate. This study also evaluated the impact of outlier removal on the accuracy of the model. Consequently, the DL models demonstrated satisfactory performance, surpassing RNN while offering an optimal balance between performance and computational efficiency. Notably, the OL and GL classes experienced significant impacts due to data scarcity after outlier removal, resulting in decreased recall and F1-scores. The reintegration of outliers then improved classification in all models. Although classification accuracy also surpassed 85% for severe cases in severity assessment, this value declined for minor leak detection owing to signal overlap with NL scenarios. Overall, the reliability of DL models in WDS diagnostics was confirmed, emphasizing the importance of preprocessing strategies to preserve data diversity and model robustness in real-world applications.