An unsupervised acoustic denoising model for water leakage detection

The global issue of water loss due to leakage in water distribution networks (WDNs) is considerable. Acoustic methods are preferred for leak detection because they are noninvasive, efficient, and cost-effective. However, distinguishing leaks from background noise remains a major challenge due to the reliance on predetermined thresholds in conventional methods and the substantial dependence of mainstream supervised deep learning approaches on the quality of training data. To overcome these obstacles, this study proposes an unsupervised acoustic denoising model (UADM), designed specifically for identifying and reducing noise to enhance leak detection accuracy within a WDN. This model uses an encoder-decoder architecture and incorporates domain-specific loss functions to guide the denoising process. Tests with publicly available data sets show that the proposed UADM significantly enhances the distinction between leak and nonleak signals. The improvements in accuracy, recall, F1 score, and precision were notable, with average increases of 8.1%, 14.5%, 8.0%, and 6.4%, respectively. The proposed UADM offers a stable and efficient tool for stakeholders involved in WDN management. By enhancing the antinoise capability of acoustic leak detection systems, the UADM model contributes to the proactive identification and mitigation of water leaks, thereby minimizing water loss and associated financial costs.