Search for a command to run...
- 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.