Acoustic signal-based leakage detection for water distribution networks with multispectrogram inputs

ABSTRACT

Accurate leak detection is fundamental to ensuring the operation of water distribution networks (WDNs). However, most existing acoustic-based methods are rarely applied in real WDNs because they often capture only incomplete or single-type features, resulting in low leak detection accuracy. This study proposes a multistage, multidimensional spectrogram-based leak detection method using real acoustic signal data, which includes the following features: (1) reconstruction of original acoustic signals using the Iterative Amplitude Adapted Fourier Transform to enhance the size of the original dataset; (2) conversion of acoustic signals into multidimensional spectrograms that integrate Mel, linear, and logarithmic spectrograms as inputs; and (3) adoption of a two-dimensional residual convolutional neural network (Res2D-CNN) model for leak detection. Compared with inputs with nonreconstructed original data and single spectrogram feature, as well as against multiple methods such as SVM, one-dimensional residual CNN (Res1D-CNN), CNN, the hybrid CNN and Informer (CNN-Informer), the hybrid CNN and random forests (CNN-RF), and SqueezeNet, the proposed method achieves superior performance across all evaluation metrics, including accuracy, precision, F1-score, specificity, and sensitivity, with each exceeding 98%. Therefore, this framework is demonstrated to be an effective approach for detecting leaks.