Explainable simulation-to-field transfer learning for transient-based leak detection in water distribution networks

This paper deals with a data-driven framework for transient-based leak detection in water distribution networks (WDNs) under limited field data. Leak detection are characterized as three supervised tasks: (i) leaking-pipe classification, (ii) leak-location regression, and (iii) leak-size regression. Large partially calibrated simulation datasets are generated to produce transient pressure responses, but models trained solely on these simulations perform poorly when applied to measurement data, reflecting a pronounced simulation-to-reality gap. To address this problem, we train several advanced artificial neural network (ANN) architectures, including Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), Gated Recurrent Units (GRU), and Transformer models, on simulated transients contaminated with measurement-based noise. Subsequently, we apply transfer learning to fine-tune the pre-trained models using experimental records and evaluate the adaptation through leave-one-out cross-validation. The transfer-learning tests include both the original 17-scenario experimental dataset and an expanded 40-scenario dataset collected under a different PRV set pressure. Across architectures, transfer learning consistently improves leaking-pipe classification and leak-location regression relative to direct simulation-to-experiment testing; in the expanded 40-scenario evaluation, the best after-transfer classification accuracy reaches 62.50%, compared with a best direct-test accuracy of 30.00%, while the best leak-location MAE decreases from 18.44 m in direct testing to 6.57 m after fine-tuning. To interpret the learned models, we use Integrated Gradients technique to identify the sensor-time regions that drive the predictions and to compare their behavior before and after transfer learning. The attribution maps show that the networks rely primarily on later-time, multiply scattered transients rather than solely on the first-half characteristic wave period, i.e., the time required for the wave to travel from the transient source to the furthest boundaries and back, and that fine-tuning (via transfer learning) suppresses spurious time features while preserving physically meaningful features. Overall, the proposed framework provides a robust leak-detection strategy that leverages simulations, limited experimental data, and interpretability tools to improve the reliability of transient-based defect detection in practical WDN applications.