Application of neural network models integrating network topology and mechanistic knowledge for burst localization in...

Accurate pipe burst localization is critical to the emergency response of water distribution networks (WDNs) because it directly affects the timeliness of repair responses and the operational safety of water supply systems. Traditional machine learning-based pipe burst localization methods typically adopt the tabular input format for monitoring data (TIFMD), which may struggle to effectively characterize the propagation characteristics of pipe burst signals in the network. To address this limitation, this study proposes a burst localization framework based on network topology and mechanistic knowledge (BLF-NTMK). The framework explicitly maps the network structure and monitoring data through a topological matrix, enabling the convolutional neural network to learn the spatial propagation patterns of pipe burst features. Meanwhile, a difference enhancement method based on an online hydraulic model and a range constraint method based on District Metered Area zoning are introduced to construct a mechanistic knowledge-driven feature enhancement layer, which improves the model's ability to identify pipe burst signals and the interpretability of localization results from a mechanistic perspective. Results demonstrate that BLF-NTMK significantly improves localization accuracy, achieving >62.97% reduction in mean squared error compared with baseline TIFMD-CNN and outperforming an advanced FL-DenseNet model under the TIFMD paradigm, despite employing a substantially lighter CNN backbone. The framework also exhibits strong robustness to hyperparameter variations and burst leakage intensity changes, and maintains stable localization performance once an appropriate topology matrix resolution is adopted. Moreover, BLF-NTMK maintains high localization performance under limited training data, achieving satisfactory accuracy with only 10% of the full dataset, indicating excellent few-shot learning capability. Validation on a real-world water distribution network in F-City further confirms the practicality of the proposed framework, which can retain acceptable localization accuracy even under high sensor failure ratios, demonstrating its strong engineering applicability in actual large-scale and complex water distribution systems. These findings highlight the effectiveness of topology-aware and mechanism-informed data representations for reliable pipe burst localization in real-world WDNs.