Deep Learning Architectures with Multiscale Attention for Acoustic Leak Detection in Digital Twin Water Networks

This chapter addresses the critical challenge of pipeline leak detection within Digital Twin (DT) frameworks for urban water management. While DTs enable proactive system monitoring, accurately classifying leak acoustics particularly for viscoelastic pipes like PVC remains hindered by the Continent of Confusion, where signatures from different failure modes overlap with background noise. We propose a novel hybrid Wavelet-Transformer deep learning model designed to overcome this limitation. The architecture integrates a trainable multi-scale convolutional front-end, functioning as a learned wavelet filterbank, with a Transformer back-end that models global, long-range dependencies in the acoustic signal. Benchmarked against CNN–LSTM, ResNet-Spectrogram, and Vision Transformer models on experimental hydraulic testbed data, the hybrid model demonstrates an accuracy of 77%. Quantitative metrics and qualitative t-SNE visualizations of the latent space reveal that our approach achieves optimal disentanglement of acoustically similar leak classes. The results indicate that combining adaptive multi-resolution analysis with self-attention most closely aligns with the underlying physics of leak generation and wave propagation. This integration provides a robust, physically interpretable solution for high-fidelity fault discrimination, advancing the capability of DT systems to enable precise, predictive maintenance and enhance the resilience of smart water infrastructure.