Leak detection in water supply networks using two-stage temporal segmentation and incremental learning for non-stationary acoustic signals

• Emphasis on integrating deep learning with non-stationary acoustic signal analysis • Two-stage segmentation reduces fluctuating noise and enhances signal features • Mel-spectrograms enhance low frequencies, where key information in signals resides • Incremental learning continuously integrates filtered new data for better detection Acoustic detection is a primary method for identifying leaks in urban water supply networks. However, acoustic signals within pipelines are highly susceptible to dynamic interference noise. This complicates the differentiation between leak and non-leak signals. To address this challenge, this paper presents a temporal segmentation-based approach for processing acoustic signals. Specifically, the two-stage temporal segmentation approach, which applies long-term segments to isolate non-stationary characteristics and short-term segments for capturing quasi-stationary features in acoustic signals, is introduced. We then applied the CNN model to recognize the Mel spectrogram features of the two-stage segmented signals and compared its performance with other models. Results indicate that this approach enhances both the accuracy and stability of leak detection, with the model achieving an average detection accuracy of 95%. Moreover, the model is designed as an adaptive and continuous learning model, integrating its detection outcomes and newly labeled data segments into its training dataset. In practical applications, this continuous learning capability enables the model to improve its detection efficacy over time as data volume expands.