Confident learning-based Gaussian mixture model for leakage detection in water distribution networks

Area of Science:

• Environmental Engineering
• Data Science
• Water Resource Management

Background:

• Water distribution systems face challenges with leakage, leading to water waste and potential contamination.
• Data-driven approaches using SCADA data (flow and pressure) are emerging for real-time leak detection.
• Lack of labeled leakage data often necessitates unsupervised methods with strong assumptions.

Purpose of the Study:

• To develop a data-driven framework for accurate leakage detection in water distribution systems.
• To address the challenge of limited labeled leakage data by leveraging historical repair records.
• To infer normal pressure characteristics and identify anomalies indicative of leaks.

Main Methods:

• Proposed a hybrid framework combining Confident Learning (CL) for label cleaning with Gaussian Mixture Model (GMM) for unsupervised anomaly detection.
• Utilized historical pressure and flow data from Supervisory Control and Data Acquisition (SCADA) systems.
• Validated the methodology using both synthetic and real-world measured data from a water distribution network.

Main Results:

• The GMM-based approach demonstrated superior identification of leak features from pressure data compared to four other unsupervised methods.
• In a real-world K city water distribution system (91 pressure sensors), the framework achieved an average true positive rate of 0.78 and a false positive rate of 0.11.
• The methodology effectively infers normal pressure characteristics and identifies leakage patterns.

Conclusions:

• The proposed framework offers a promising, data-driven tool for real-time leakage detection in large-scale water distribution networks.
• Combining label cleaning with unsupervised methods enhances the accuracy and reliability of leak detection.
• This approach reduces reliance on hardware and improves water resource management by minimizing waste and pollution risks.