False Alarm Prevention Through Domain Knowledge-Driven Machine Learning: Leakage Detection in Water Distribution Networks

Abstract

Effective management of water distribution networks (WDNs) is critical for conserving water and reducing financial losses. This study addresses the problem of false alarms in WDNs' acoustic loggers, often triggered by electrical transformer noise. We propose a false alarm prevention framework that features a transformer noise-based frequency selection (TNFS) method, utilizing domain knowledge of the system. TNFS provides a uniform feature set without the iterative sampling or sample dependence typical of other methods. Its rapid processing and low computational needs make it exceptionally suitable for WDNs' acoustic loggers with constrained computing resources. In addition, we introduce the generative adversarial network-enhanced synthetic minority over-sampling technique (SMOTE), designed to tackle the mode collapse issue in generative adversarial networks (GANs). Our system, validated with real-world data spanning 17 months, dramatically reduces false alarms from electrical noise by 99.6%, highlighting the importance of domain-specific knowledge in the application of machine learning to industrial sensor networks.