Deep learning-based burst location with domain adaptation across different sensors in water distribution networks

Abstract

This study proposes a domain adaption method for pipe burst location based on deep learning. Multi-scale feature extractors are designed to extract pipe burst features, then three classifiers are trained by pipe burst features with different scales, and adversarial training is introduced during the edge domain fusion. Finally, the probability ranking of each pipeline is obtained according to the classification results of the three classifiers. In this study, a Net3 pipe network hydraulic model was used as an example to carry out related research. The pressure monitoring data of three sensors were used to train and test the model, and different scenarios of one, two and three sensors were considered at the same time. The results showed that the overall prediction accuracy of the three scenarios was over 90% when considering the five pipelines with the highest pipe burst probability.