Cross-regional leak detection in water distribution networks through domain transfer.

An effective leak detection system is critical for maintaining urban infrastructure operations. Machine learning-based leak detection systems represent a significant advancement, enhancing detection accuracy and efficiency. However, their performance is constrained by regional data characteristics, which limit model generalizability across domains and necessitate high-quality labeled datasets in target regions. To address these challenges, this study proposes a domain adaptation framework that leverages existing labeled data and adapts it to new regions. Experimental results demonstrate the effectiveness of the framework, achieving over 10 % higher accuracy than baseline models in cross-region transfer between Hong Kong and Dalian. Additionally, t-SNE visualizations illustrate the domain adaptation capabilities and leak detection performance of the framework. This work advances model generalizability and effectiveness while reducing development costs, thereby promoting the adoption of smart infrastructure management systems.