Penalized spatial-temporal sensor fusion for detecting and localizing bursts in water distribution systems

In urban sustainability, the efficient management of water distribution systems (WDSs) plays a pivotal role in minimizing water loss and service disruptions. Detecting and localizing pipe bursts promptly is crucial for WDS maintainers to mitigate the impact of such incidents. Hydraulic measurements collected by spatially distributed sensors within a time interval often exhibit strong spatial–temporal (ST) correlation, which allows the development of data-driven burst detection methods. However, conventional approaches heavily rely on training data, which may be scarce or unavailable in real-world scenarios, making them less effective in detecting bursts. Moreover, these methods often overlook the intricate ST patterns inherent in the data and struggle to capture the burst characteristics embedded in hydraulic measurements. To address these limitations, our research introduces a novel unsupervised learning technique termed Penalized Spatial-Temporal Component Analysis (PSTCA). Unlike conventional methods, PSTCA does not depend on training data for burst detection. Instead, it leverages the inherent ST correlation in hydraulic measurements to decompose the data into distinct components: regular components, burst-induced anomaly components, and noise components, with each component regularized based on its unique ST characteristics. An optimization algorithm is developed to ensure accurate estimation of these components. The performance of the proposed method is evaluated with a simulation case study, demonstrating superior performance compared to benchmark methods, with a lower false alarm rate, higher burst detectability, and improved accuracy in burst localization and burst occurrence time estimation.