Leak Signal Collection and Classification in Water Pipelines Using Resource-Constrained Devices

Water leakage poses a critical threat to water distribution systems. This research addresses leak detection by investigating the correlation between acoustic signatures and leak severity. A resource-constrained system (utilizing LoRa with SF12 and coding rate 4/5) is developed for large-scale deployment, emphasizing affordability and low maintenance. Convolutional Neural Networks (CNNs) are utilized, and experiments with Spectrogram, Mel Spectrogram, and MFCCs features provide crucial insights. While Spectrogram and Mel Spectrogram enable highly accurate classification, MFCCs offer a compact representation (up to 90% reduction) and significantly faster transmission times (over 10 times faster). This size reduction comes at a cost in accuracy, but in the 2-class model, MFCCs remain within a 10% accuracy difference from Mel Spectrograms, making them well-suited for initial leak screening. This system has the potential to optimize water management by proactively identifying leaks and enabling targeted maintenance for a more sustainable network.