Explainable simulation-to-field transfer learning for transient-based leak detection in water distribution networks

Area of Science:

• Engineering
• Artificial Intelligence
• Environmental Science

Background:

• Water distribution networks (WDNs) face challenges in detecting leaks due to limited field data.
• Transient-based methods offer potential but suffer from a simulation-to-reality gap.

Purpose of the Study:

• To develop a robust data-driven framework for transient-based leak detection in WDNs.
• To improve leak classification and localization accuracy using advanced AI and transfer learning.
• To interpret AI model predictions for better understanding of leak detection mechanisms.

Main Methods:

• Generated large simulation datasets for training artificial neural networks (ANNs) including CNN, LSTM, GRU, and Transformer models.
• Applied transfer learning to fine-tune models using limited experimental data, addressing the simulation-to-reality gap.
• Utilized Integrated Gradients for model interpretability to analyze feature importance.

Main Results:

• Transfer learning significantly enhanced leaking-pipe classification accuracy (up to 62.50%) and reduced leak-location mean absolute error (MAE) (down to 6.57 m).
• Models trained on simulations alone performed poorly on experimental data, highlighting the necessity of transfer learning.
• Interpretability analysis revealed models focus on later-time transients and fine-tuning refines feature reliance.

Conclusions:

• The proposed framework effectively leverages simulations and limited experimental data for reliable transient-based leak detection in WDNs.
• Transfer learning is crucial for adapting AI models to real-world conditions in WDN leak detection.
• Model interpretability tools aid in understanding and improving the robustness of leak detection strategies.