Improving the leak detection efficiency in water distribution networks using noise loggers

Area of Science:

• Environmental Engineering
• Signal Processing
• Machine Learning

Background:

• Water leakage in distribution networks poses significant challenges, with laboratory-based detection models often lacking practical efficiency due to environmental complexities.
• Real-world water distribution networks (WDNs) present unique challenges like ambient noise and irregular usage, hindering the effectiveness of traditional leak detection methods.
• Developing robust leak detection techniques for actual WDNs is crucial for water management and infrastructure integrity.

Purpose of the Study:

• To develop and evaluate machine learning (ML)-based leak detection models for real water distribution networks (WDNs).
• To investigate the impact of signal de-noising and feature extraction on the performance of ML-based leak detection.
• To compare the efficacy of different ML algorithms for leak detection in both metal and non-metal WDNs.

Main Methods:

• Wireless sensors were deployed to capture acoustic signals from real WDNs.
• Discrete wavelet transform was utilized for de-noising the acquired acoustic signals.
• Seventeen features were extracted from raw and de-noised signals using linear prediction for ML model development.
• Performance evaluation involved Decision Tree (DT), Support Vector Machine (SVM), Artificial Neural Network (ANN), and K-Nearest Neighbor (K-NN) algorithms.

Main Results:

• ML models utilizing features from de-noised signals demonstrated superior classification accuracy compared to those using raw signals.
• For de-noised signals, DT, SVM, and ANN algorithms achieved 100% accuracy, precision, and recall in detecting leaks in both metal and non-metal WDNs.
• The study provided a comprehensive comparison of ML model performance across different WDN types, feature sets, and algorithms.

Conclusions:

• Machine learning, particularly when applied to de-noised acoustic signals, offers a highly effective solution for leak detection in real-world WDNs.
• The developed ML models, especially DT, SVM, and ANN, show exceptional performance, achieving perfect scores for accuracy, precision, and recall.
• This research highlights the importance of signal processing techniques in enhancing the practical applicability of ML for water infrastructure monitoring.