Adaptive Subband Downsampling for Leak Noise Extraction in Water Distribution Networks

In urban water distribution network leak detection, acoustic methods offer greater flexibility and accuracy than flow-based and pressure-based approaches. However, leak noise is a persistent, broadband stochastic signal that is severely contaminated by ambient noise, making accurate detection challenging. This paper proposes an adaptive subband downsampling and filtering method that leverages the cumulative energy of the power spectral density (PSD) to identify the main frequency band. The input signal is decomposed into three subbands: the main band is processed at the full sampling rate, and the high- and low-frequency bands undergo downsampling and upsampling. This scheme preserves the integrity of leak noise while reducing computational load. Compared with wavelet denoising (WD) and empirical mode decomposition (EMD), the proposed method is adaptive, requires no frequent parameter tuning, and overcomes the reconstruction errors, distortions, and the need for a separate desired signal inherent in conventional subband adaptive filter (SAF) approaches. Field experimental results demonstrated that the proposed method significantly enhances leak localization precision. Compared with the raw input signal, the proposed method reduces the time delay estimation (TDE) error from 0.52 to 0.11 ms. Additionally, it increases the spectral contrast by 5.33 dB. Whereas traditional methods like WD suffer from severe phase distortion, the proposed approach effectively suppresses noise while preserving phase integrity.