Interpretable deep learning for acoustic leak detection in water distribution systems.

Area of Science:

• Water resource management
• Machine learning applications in engineering
• Acoustic signal processing

Background:

• Leak detection is vital for water system safety and conservation.
• Current machine learning models for leak detection lack interpretability, hindering practical adoption.
• Transparency and credibility are essential for real-world leak detection systems.

Purpose of the Study:

• To develop and evaluate a Multi-channel Convolution Neural Network (MCNN) for enhanced leak detection.
• To compare the MCNN model's performance against the Frequency Convolutional Neural Network (FCNN).
• To improve the interpretability of machine learning models in leak detection using visualization techniques.

Main Methods:

• Implementation of a Multi-channel Convolution Neural Network (MCNN) model.
• Comparative analysis with the Frequency Convolutional Neural Network (FCNN) using experimental and field data.
• Application of Multi-channel Gradient-weighted Class Activation Mapping (MGrad-CAM) for model interpretability.
• Utilizing clustering methods to analyze factors influencing acoustic leak signals.

Main Results:

• The MCNN model significantly outperformed the FCNN on both laboratory and real-world datasets.
• Achieved a high accuracy rate of 95.4% for leak detection in real-field scenarios.
• MGrad-CAM successfully visualized critical signatures in acoustic signals, enhancing model interpretability.
• Identified that higher pressure, closer proximity, and increased leak flow rate correlate with increased acoustic signal bandwidth.
• Confirmed the importance of high-frequency components for accurate leak detection.

Conclusions:

• The MCNN model offers a more accurate and interpretable solution for water system leak detection.
• MGrad-CAM provides valuable insights into the decision-making process of deep learning models for leak detection.
• Understanding the acoustic signal generation mechanism aids in developing more effective leak detection strategies.