Enhanced Underground Pipeline Leakage Detection via Deep Denoising of Mel Spectrograms and Efficient CNN Classification

Detecting underground water supply pipeline leakage poses significant challenges in complex urban environments due to the limited accuracy of conventional methods. This study aims to address these challenges and develop a more reliable leakage detection approach. To this end, a novel signal processing framework integrating advanced deep denoising techniques with efficient classification algorithms is proposed: first, an improved Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) model is used to denoise Mel spectrograms via multi-scale feature fusion, enhancing faint leakage features; subsequently, a modified Convolutional Neural Network (CNN)—incorporating lightweight bottleneck structures, and the Mish activation function (to improve feature representation and gradient flow)—accurately classifies processed signals, with this architectural refinement boosting detection accuracy while reducing computational resource consumption. Experimental results on real pipeline datasets demonstrate that the proposed method significantly outperforms standard Convolutional Neural Network approaches across key metrics (accuracy, precision, recall, and F1-score), presenting an innovative solution for underground pipeline leakage detection and facilitating the evolution of related technologies from conventional modes toward smarter, more precise, and highly generalizable solutions.