Leak and Burst Detection in Water Distribution Network Using Logic- and Machine Learning-Based Approaches

Kiran Joseph, Jyoti Shetty, Ashok K. Sharma, Rudi van Staden, P. L. P. Wasantha, Sharna Small, Nathan Bennett

Urban water systems worldwide are confronted with the dual challenges of dwindling water resources and deteriorating infrastructure, emphasising the critical need to minimise water losses from leakage. Conventional methods for leak and burst detection often prove inadequate, leading to prolonged leak durations and heightened maintenance costs. This study investigates the efficacy of logic- and machine learning-based approaches in early leak detection and precise location identification within water distribution networks. By integrating hardware and software technologies, including sensor technology, data analysis, and study on the logic-based and machine learning algorithms, innovative solutions are proposed to optimise water distribution efficiency and minimise losses. In this research, we focus on a case study area in the Sunbury region of Victoria, Australia, evaluating a pumping main equipped with Supervisory Control and Data Acquisition (SCADA) sensor technology. We extract hydraulic characteristics from SCADA data and develop logic-based algorithms for leak and burst detection, alongside state-of-the-art machine learning techniques. These methodologies are applied to historical data initially and will be subsequently extended to live data, enabling the real-time detection of leaks and bursts. The findings underscore the complementary nature of logic-based and machine learning approaches. While logic-based algorithms excel in capturing straightforward anomalies based on predefined conditions, they may struggle with complex or evolving patterns. Machine learning algorithms enhance detection by learning from historical data, adapting to changing conditions, and capturing intricate patterns and outliers. The comparative analysis of machine learning models highlights the superiority of the local outlier factor (LOF) in anomaly detection, leading to its selection as the final model. Furthermore, a web-based platform has been developed for leak and burst detection using a selected machine learning model. The success of machine learning models over traditional logic-based approaches underscores the effectiveness of data-driven, probabilistic methods in handling complex data patterns and variations. Leveraging statistical and probabilistic techniques, machine learning models offer adaptability and superior performance in scenarios with intricate or dynamic relationships between variables. The findings demonstrate that the proposed methodology can significantly enhance the early detection of leaks and bursts, thereby minimising water loss and associated economic costs. The implications of this study are profound for the scientific community and stakeholders, as it provides a scalable and efficient solution for water pipeline monitoring. Implementing this approach can lead to more proactive maintenance strategies, ultimately contributing to the sustainability and resilience of urban water infrastructure systems.