Graph convolutional networks based contamination source identification across water distribution networks

Water distribution Networks (WDNs) are one of the most important infrastructures for modern society. Due to accidental or malicious reasons, water contamination incidents have been repeatedly reported all over the world, which not only disrupt the water supply but also endanger public health. To ensure the safety of WDNs, water quality sensors are deployed across the WDNs for real-time contamination detection and source identification. In the literature, various methods have been employed to improve the performance of contamination source identification (CSI) and recent studies show that there is a great potential to tackle the CSI problem by deep learning models. The success of deep learning based CSI methods often requires a large size of training samples being collected. In real-world situations, the number of contamination events occurring in a single WDN is rather small, especially for a newly built WDN. However, the existing CSI methods in the literature mostly focus on the study of training and applying models on the same WDNs and the knowledge of CSI gained from one WDN cannot be reused by a different WDN. To these ends, based on the application of graph convolutional networks, this paper provides a solution for cross-network CSI that can transfer the CSI knowledge learned from one WDN to a different WDN. Empirically, based on a benchmark WDN in the task of contamination source identification, we show that the proposed cross-network CSI method can achieve comparable accuracy even trained on a different WDN. (C) 2021 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

Automated summary

Water distribution networks are crucial infrastructure, and water contamination incidents pose threats to public health. Various methods and deep learning models have been employed for contamination source identification. This paper proposes a solution for cross-network CSI based on graph convolutional networks, showing comparable accuracy even when trained on a different WDN.