A Probabilistic Digital Twin for Leak Localization in Water Distribution Networks Using Generative Deep Learning

Nikolaj Takata Mücke, Prerna Pandey, Shashi Jain, Sander M. Bohté, Cornelis W. Oosterlee

TL;DR

This paper introduces a novel probabilistic digital twin for water distribution networks that uses generative deep learning and Bayesian inference to pinpoint leaks with uncertainty quantification. This approach offers a faster, more accurate, and trustworthy method for leak detection compared to traditional techniques, crucial for maintaining critical infrastructure.

What Problem Does This Paper Address?

Water distribution networks are complex and vital infrastructure, but they are prone to leaks and blockages. Detecting and localizing these issues is challenging due to the sheer scale of these networks, the limited number of sensors, and the inherent noise in measurements. Existing methods often struggle to provide accurate leak locations with reliable confidence estimates, especially in real-time. This paper aims to bridge this gap by developing a framework that not only identifies leak locations but also quantifies the uncertainty associated with these predictions, making the system more robust and trustworthy.

Methods

The core of this work lies in combining two powerful techniques: Bayesian inference and generative deep learning.

  • Bayesian Inference: This statistical framework is used to model the leak localization problem as an inverse problem. It allows for the incorporation of prior knowledge about leak locations and, crucially, provides a posterior distribution that quantifies the uncertainty in the predicted leak location. However, traditional Bayesian inference can be computationally expensive.
  • Generative Deep Learning (Wasserstein Autoencoders - WAEs): To overcome the computational bottleneck of Bayesian inference, the authors employ a supervised Wasserstein Autoencoder (SupWAE). This deep learning model acts as a "stochastic digital twin" of the water distribution network. After training on simulated data, the decoder of the SupWAE can rapidly generate plausible pressure heads and flow rates for a given leak location and time. This generative model effectively replaces the computationally intensive forward model in the Bayesian inference process, enabling faster likelihood computations.

The framework uses the decoder of the trained SupWAE to approximate the likelihood function within the Bayesian inference framework. This allows for efficient computation of the posterior distribution of leak locations, even with noisy sensor data and uncertain parameters. The authors also explore two specific neural network architectures: Residual Neural Networks (ResNets) and Transformers, to demonstrate the flexibility of their approach.

Datasets

The authors generated synthetic data for training and evaluation using the EPANET hydraulic simulation software. They created three distinct water distribution network (WDN) test cases:

  • WDN1 (Hanoi network): A smaller network.
  • WDN2 (Net3): A medium-sized network.
  • WDN3 (Modena): A larger, more complex network.

For each WDN, they simulated various scenarios by varying:

  • The number and placement of sensors.
  • The level of noise in sensor readings and demand.
  • The leak location and size.

The generated data included pressure head and flow rate measurements, along with the true leak location. The Epanet .inp files for these networks are available in a GitHub repository.

Key Insights and Ideas

  • Generative Models as Surrogate Twins: The paper demonstrates that generative deep learning models, specifically WAEs, can serve as effective "stochastic digital twins." They can learn the complex, high-dimensional distributions of network states (pressure, flow) and provide rapid, probabilistic outputs, significantly speeding up computationally intensive tasks like Bayesian inference.
  • Uncertainty Quantification is Key: The probabilistic nature of the Bayesian approach, combined with the generative model, allows for robust uncertainty quantification. This is crucial in real-world applications, as it tells users when they can trust a prediction and when more data is needed. The entropy of the posterior distribution is used as a direct measure of this uncertainty.
  • Transformer Architectures Shine: For smaller to medium-sized WDNs, transformer architectures showed superior performance, likely due to their ability to effectively model long-range dependencies and interactions within the network.
  • Prior Knowledge Matters: Incorporating prior knowledge about likely leak locations significantly improved the accuracy and reduced the topological distance of leak predictions across all test cases.
  • Flexibility in Sensor Configurations: A single trained model can handle different sensor configurations, which is a significant advantage over traditional classification methods that require retraining for each new setup.

Key Results

The proposed framework demonstrated strong performance across three test cases with increasing complexity:

  • Accuracy: Achieved accuracies upwards of 83% for the simplest case, 72% for the second, and 42% for the most complex case, depending on noise and sensor configurations.
  • Topological Distance (ATD): The ATD ranged from 0.3 to 3 for the simplest case, 0.75 to 4 for the second, and 1.5 to 10 for the most complex case.
  • Speed: Computation times ranged from 0.1 to 13 seconds, depending on the neural network size, enabling near real-time leak localization.
  • Comparison to Baseline: The proposed Bayesian framework consistently outperformed a conventional classification approach in terms of both ATD and accuracy, especially when using prior information.
  • Uncertainty Estimation: The framework provided more informative uncertainty estimates (higher entropy ratios for incorrect predictions) compared to the baseline classifier.

Limitations and Future Work

  • Training Stage Cost: While the online inference stage is fast, the offline training of the generative model can be computationally intensive.
  • Posterior Computation Efficiency: The paper notes that more efficient integration methods than Monte Carlo simulation could further speed up posterior computation.
  • Physics Integration: The authors suggest incorporating the physical constraints of the problem (e.g., mass and momentum conservation) directly into the neural network training to improve accuracy and robustness.
  • Model Architectures: The performance varied between ResNet and Transformer architectures across different WDNs, suggesting that the choice of architecture is important and may require investigation for specific network types.

The paper concludes that generative deep learning, integrated with Bayesian inference, offers a powerful approach for creating probabilistic digital twins for critical infrastructure like water distribution networks, paving the way for more intelligent and reliable leak detection systems.

Summary updated

BibTeX
@inproceedings{mucke2023probabilistic,
  author = {Nikolaj Takata Mücke and Prerna Pandey and Shashi Jain and Sander M. Bohté and Cornelis W. Oosterlee},
  title = {{A Probabilistic Digital Twin for Leak Localization in Water Distribution Networks Using Generative Deep Learning}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=xMCumEiFNS},
}
Figure