Integrating machine learning models and acoustic signals for water pipeline leakage localization

Water leakage detection and localization in the water distribution networks are fundamental in limiting the loss of water, maintaining efficiency in its operations, and the sustainable management of the resource. Other conventional methods, like visual surveillance and pressure threshold monitoring, are inadequate as they tend to respond slowly and lack spatial resolution. In this paper, an experimental architecture combining acoustic sensing with the sophisticated machine learning regression algorithm is presented and effective leak localization is realized. High sensitivity hydrophones record leak induced acoustic signals and the Time Difference of propagation between sensor pairs. The essential characteristics of the acoustic cues are obtained in the form of Fast Fourier Transform (FFT) and frequency domain results, which provide powerful spectral representations to be used as models. A comparative evaluation is conducted using standard regression error metrics. With a root mean square error (RMSE) of 0.0551 m and mean absolute error (MAE) of 0.0369 m, the KNN regression model achieved the lowest localization error among all evaluated models. The proposed framework is non-invasive, cost-efficient, and scalable, providing a practical solution for real-time detection and localization in complex pipeline infrastructures.