Enhancing Acoustic Leak Detection Under Environmental Interference in Smart Water Networks Using Corruption-Aware Restoration Denoising.

Leak detection in drinking water distribution networks is essential for reducing water loss and preventing infrastructure degradation. To support large-scale monitoring, acoustic IoT sensors are increasingly being deployed to collect routine noise measurements for leak detection, and machine-learning-based classification enables these data to be interpreted automatically at scale. However, when these classifiers are trained or deployed under field conditions, their practical performance remains strongly affected by environmental interference, despite substantial existing research on increasingly sophisticated classifier architectures. This makes data quality a critical bottleneck for reliable leak detection and establishes denoising as a necessary pre-processing step. Existing denoising approaches can suppress noise, but they often struggle when corruption is intermittent, localised and structured, making it difficult to preserve the continuity of leak-related signatures and support reliable downstream classification. In this study, we addressed this problem by proposing a machine-learning-based corruption-aware restoration denoising framework (CARD) that first localises corrupted regions in acoustic spectrograms and then reconstructs the underlying time-frequency structures using targeted restoration. Across denoising and downstream leak-classification evaluations, CARD outperformed the evaluated baselines, achieving higher leak detection sensitivity with a lower false alarm rate. These results indicate that improving spectrogram data quality by accurately identifying and restoring corrupted regions can improve the robustness of acoustic leak detection and support more reliable acoustic IoT monitoring in noisy operational environments.