Advanced acoustic leak detection in water distribution networks using integrated generative model

Area of Science:

• Engineering
• Data Science
• Environmental Science

Background:

• Water distribution networks (WDNs) suffer substantial water loss from leaks, demanding effective detection methods.
• Current machine learning (ML) acoustic leak detection is hampered by limited and non-diverse datasets.

Purpose of the Study:

• To propose and evaluate an LSTM-GAN approach for generating synthetic acoustic leak signals.
• To enhance ML-based water leak detection in WDNs by augmenting limited datasets.

Main Methods:

• Collected acoustic signals from WDNs to train a Long Short-Term Memory-Generative Adversarial Network (LSTM-GAN) model.
• Generated synthetic leak signals using LSTM-GAN to augment the original dataset.
• Evaluated generative method validity using t-SNE and acoustic characteristics; compared LSTM leak detection models using original and augmented datasets.

Main Results:

• The LSTM-GAN model successfully generated high-quality, consistent synthetic acoustic signals indicative of leaks.
• Leak detection models trained on the augmented dataset showed improved performance compared to those using only original data.
• The proposed LSTM-GAN method outperformed other acoustic signal generation techniques in enhancing leak detection.

Conclusions:

• The LSTM-GAN approach effectively addresses data scarcity in ML-based acoustic leak detection for WDNs.
• Generated synthetic data significantly improves the robustness and performance of water leak detection models.
• This generative method offers an innovative solution for data-limited machine learning applications in infrastructure monitoring.