Joseph, Kiran ORCID: https://orcid.org/0000-0003-4992-3933 (2025) Development of Software and Hardware-Based Methods for the Detection of Leaks and Bursts in the Water Supply Systems for Developing an Intelligent Water Network. PhD thesis, Victoria University.

Abstract

Centralised water supply systems have been utilised globally for more than a century; however, ageing infrastructure, population growth, climate variability, and increasing urbanisation have intensified operational pressures on water utilities. Globally, Non-Revenue Water (NRW) primarily due to physical losses from leaks and bursts is estimated to account for approximately 20–30% of total water supplied in many systems, with significantly higher losses reported in ageing networks. These losses result in substantial economic costs associated with water production, energy consumption, infrastructure repair, and service disruptions, while also contributing to environmental degradation and reduced water security. As of late 2024 and early 2025, global NRW is estimated at 346 million cubic meters per day, equivalent to a loss of approximately USD $39 billion annually. This research was undertaken to develop and validate software- and hardware-based methodologies for leak and burst detection as part of a broader framework for establishing an Intelligent Water Network (IWN). This thesis integrates two complementary studies to develop a comprehensive, scalable, and field-validated framework for intelligent monitoring of pressurised water distribution systems. The first study focuses on real-time leak and burst detection using operational Supervisory Control and Data Acquisition (SCADA) data from an Australian pumping main and controlled measurements from a purpose-built experimental testbed. The second study proposes an optimisation-based methodology for identifying the most hydraulically informative sensor locations within a large-scale international water network. Together, these studies establish a unified technical foundation to support the transition toward future IWNs. The leak and burst detection study was conducted on the 6.24-km, 450-mm Sunbury Pumping Main in Victoria, Australia, which conveys water from the Loemans Road Pump Station to the Shepherds Lane Tank. The system operates under variable pump-speed control, with SCADA continuously logging pressure, flow, and pump-speed measurements. A deterministic hydraulic logic model was developed to analyse pressure, flow, and pump-speed behaviour across start-up, steady-state, and shutdown conditions. Adaptive pressure thresholds were applied to distinguish minor leaks, major leaks, and bursts, while the Darcy–Weisbach formulation enabled accurate localisation of burst events along the pipeline alignment. Machine Learning analysis was also applied to the Sunbury Pumping Main to identify abnormal hydraulic behaviour in the absence of labelled leak events. Five unsupervised anomaly-detection algorithms including Local Outlier Factor (LOF), Isolation Forest, One-Class Support Vector Machine (One-Class SVM), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and the K-Means Clustering Algorithm (K-Means) were trained on normalised SCADA telemetry. LOF demonstrated the highest anomaly-detection performance after class balancing, achieving an F1-score of approximately 0.69 and a Receiver Operating Characteristic – Area Under the Curve (ROC–AUC) of 0.72, owing to its ability to capture subtle, localised deviations in pressure, flow, and pump-speed interactions. The final LOF model was integrated into a real-time web-based monitoring platform for automated leak and burst alerting. To provide controlled experimental validation, a 111-m long, 110-mm polyethylene (PE) prototype pipeline was constructed in Melton, Victoria, connected to an existing 100-mm supply line. Two flow meters and two pressure sensors were installed at hydraulically informative locations. Controlled leak scenarios were introduced using 20-mm and 25-mm lateral outlets, representing realistic small-to-moderate leak conditions commonly associated with service connections, fittings, and joint failures in distribution systems. Fourteen supervised Machine Learning classifiers were trained on high-resolution hydraulic datasets representing no-leak, minor-leak, and major-leak scenarios. Decision Tree, k-Nearest Neighbours (KNN), and Random Forest algorithms achieved perfect classification accuracy and F1-scores, demonstrating excellent capability under controlled conditions. The second investigation develops a robust Optimal Sensor Placement (OSP) framework applied to the Hanoi Water Distribution Network (WDN), a widely used benchmark model in hydraulic research. Unlike the Sunbury trunk main and Melton prototype system, the Hanoi network represents a meshed distribution system, enabling evaluation of sensor placement strategies under complex nodal interactions. Leak scenarios were simulated using EPANET by introducing emitter-based leak representations at 32 nodes with equivalent orifice diameters of 10, 50, 100, and 150 mm. These diameters represent a spectrum of leak severities, from small joint-related losses to larger structural failures. Pressure deviations at all nodes were filtered using a 0.5-m sensitivity threshold to construct a hydraulic sensitivity matrix representing nodal responsiveness to leaks. Four optimisation algorithms including Greedy, Sensitivity-based Top-K, D-Optimal, and a Hybrid Greedy→D-Optimal strategy were benchmarked against a Random baseline. The deterministic algorithms achieved consistently high (>96%) to perfect detection coverage, with the Hybrid approach providing the best balance between detection performance, redundancy reduction, and sensor stability. Sensitivity-evolution analysis identified Nodes 29, 28, 27, 15, and 30 as the most consistently informative locations across all leak magnitudes. Additional robustness, runtime, and threshold-perturbation analyses confirmed the scalability and reliability of the proposed OSP framework for utility-scale implementation. Collectively, the integrated methodologies improve fault detectability, reduce false alarms, enhance localisation accuracy, and optimise sensor deployment strategies across both trunk main and distribution network contexts. The outcomes offer significant benefits to water utilities through reduced non-revenue water, lower operational costs, improved infrastructure resilience, and enhanced service reliability for communities. Environmentally, the framework supports sustainable water-resource management by minimising unnecessary abstraction and treatment energy. From a research perspective, this thesis advances the integration of hydraulic theory with data-driven analytics, contributing a scalable and practically deployable foundation for next-generation Intelligent Water Networks.

Search Google Scholar

Repository staff login