- This study presents a novel, high-precision pipeline leak detection and localisation method that integrates the Transient Reflection Method (TRM) with Mel-Frequency Cepstral Coefficients (MFCC) and a lightweight Artificial Neural Network (ANN) for leak size estimation. Unlike conventional time-domain or FFT-based approaches, the proposed method uses MFCC to characterise the spectral signature of transient pressure wave reflections caused by leak-induced impedance discontinuities, rather than relying on leak-generated frequency shifts. Laboratory experiments on a 152-m Medium-Density Polyethene (MDPE) pipe system achieved an average localisation error of ±1.98 m and 96.5% leak detection sensitivity for leaks as small as 1 mm. The ANN regression model demonstrated high predictive reliability, producing a mean absolute error (MAE) of 0.15 mm for leak size estimation. Field validation of a 220-m buried MDPE pipeline yielded comparable performance, maintaining a localisation error of ±2.12 m and 94.1% detection sensitivity, confirming practical scalability under real municipal operating conditions. A structured preprocessing pipeline including signal normalisation and MFCC feature vectorisation proved essential for model stability, improving leak size MAE by more than 12% post-normalisation while preserving reflection-based spectral patterns. The system requires only single-point hydrant access, uses minimal portable hardware, and avoids the computational overhead typical of deep convolutional models. These results confirm a cost-effective, portable, and scalable solution for early leak detection and quantification in pressurised water pipeline systems, offering high spatial resolution and strong robustness for field deployment.
Ghazali et al. (Sun,) studied this question.