Contrastive learning method for leak detection in water distribution networks

Abstract

Detecting and mitigating leaks in water distribution networks are vital for water conservation. Machine-learning-based (ML) acoustic leak detection models were introduced as effective alternatives for leak management. However, ML model training requires sufficient labeled data, which hinders related development. To address this challenge, this study employed contrastive learning (CL) for leak detection using limited labeled signals. Experimental results indicate that flip-x and amplitude scaling are optimal combinations for contrastive learning. Besides, ablation and t-distributed stochastic neighbor embedding (t-SNE) results reveal that increasing the model depth does not always yield performance improvement, and five convolutional blocks are more suitable for the leak detection problem in this study. Comparison experiments demonstrate that contrastive learning outperforms supervised learning (SL) when trained with insufficient labeled data. The out-of-sample validation results indicate that the proposed leak detection model is robust and effective in unexplored pipelines. The proposed framework significantly advances ML-based leak detection research and supports sustainable water management practices.