A Transformer-Based Approach to Leakage Detection in Water Distribution Networks

Area of Science:

• Engineering
• Computer Science
• Environmental Science

Background:

• Efficient leakage detection in water distribution networks (WDNs) is vital for water supply safety and urban operations.
• Traditional machine learning methods like CNNs and Autoencoders struggle with long-term dependencies in pressure data.
• Reliance on local pressure information limits the effectiveness of existing methods.

Purpose of the Study:

• To propose and evaluate a novel transformer-based model for enhanced leakage detection in WDNs.
• To address the limitations of existing methods in capturing long-term temporal dependencies in pressure data.
• To improve the accuracy and reliability of leakage detection systems.

Main Methods:

• Development of a transformer-based model utilizing an attention mechanism.
• Incorporation of correlations between historical and same-day pressure data to capture long-term dependencies.
• Application of pressure data normalization and concatenation of position embeddings to prevent feature misleading.

Main Results:

• The transformer-based model demonstrated superior performance in leakage detection accuracy and F1-score.
• Experimental results on simulated WDN datasets confirmed significant outperformance over traditional CNN methods.
• The model effectively learned data distributions and accounted for complex correlations in pressure series.

Conclusions:

• The proposed transformer-based model offers a significant advancement in WDN leakage detection.
• The attention mechanism effectively captures long-term dependencies, leading to improved detection performance.
• This approach enhances the reliability of municipal water supply and urban operational efficiency.