A Comparison of Fully-Linear Deep Learning Methods for Pipe Burst Localization in Water Distribution Networks

Pipe bursts cause a considerable loss of treated water, increase the risks of environmental contamination and are a health hazard for the end-user as they can create a passage for contaminants to enter water distribution networks (WDN). Identifying pipe burst locations will help water service providers promptly repair them. Many methods have been proposed to locate pipe bursts. But none have proved to produce results accurate enough for water service providers to heavily rely on. Therefore, this paper aims to develop a fully-linear ResNet to accurately locate pipe bursts using pressure sensors. The performance of the model was compared to the fully-linear DenseNet and ResNet methods. The proposed FL-ResNet had a mean accuracy of 98.23% which is comparable to the 99% achieved by FL-DenseNet and it required significantly less training time compared to the FL-DenseNet. As part of future work, transfer learning could be explored, by training the model on simulated data and then on real-world data.