Time-Transformer for acoustic leak detection in water distribution network

Abstract

Accurate leak detection for water distribution networks (WDNs) is a critical task to minimize water loss and ensure efficient infrastructure management. Machine learning (ML) algorithms have demonstrated significant potential in establishing effective acoustic leak detection systems. However, the utilization of time-series models, specifically designed to handle sequential signals, in the field of water leak detection remains relatively unexplored, and there is a lack of research discussing their applicability in this context. Therefore, this study introduces a novel approach for precise leak detection in WDNs using a Time-Transformer model, which effectively captures long-range dependencies through self-attention mechanisms, enabling it to outperform other time-series models. This study conducted field experiments on WDNs in Hong Kong to demonstrate the superior performance of the proposed approach in accurately detecting leaks. The model structure is optimized through parametric experiments. Besides, leak detection and t-SNE results highlight the model's significant potential to enhance leak detection in WDNs compared to 1D-CNN and CNN–LSTM. The proposed Transformer-based model shows significant potential in advancing leak detection in WDNs, improving accuracy and precision, and supporting efficient water management.