Machine Learning-Based Digital Twin for Water Distribution Network Anomaly Detection and Localization

Prerna Pandey, Nikolaj T. Mücke, Shashi Jain, Parthasarathy Ramachandran, Sander M. Bohté, Cornelis W. Oosterlee2024 · DOI

TL;DR

This paper introduces a Digital Twin (DTwin) approach for real-time leak detection and localization in water distribution networks (WDNs) using machine learning. They developed both single-stage and two-stage models, demonstrating high accuracy in identifying leaks on a real-time dataset from the Indian Institute of Science campus.

What Problem Does This Paper Address?

Water distribution networks are critical infrastructure, and leaks can lead to significant water loss, economic damage, and potential contamination. Traditional methods for leak detection are often slow and imprecise. The integration of IoT sensors has enabled real-time monitoring, but effectively analyzing this data to pinpoint leaks remains a challenge. This paper aims to address this by developing a data-driven Digital Twin that can accurately detect and localize leaks in real-time.

Methods

The authors propose a Digital Twin (DTwin) framework that leverages IoT-enabled pressure and flow sensors. They explore two main machine learning approaches:

  • Single-Stage Model: This approach directly uses sensor data to classify anomalies as leaks or no-leaks. They tested Logistic Regression and Random Forest for this task.
  • Two-Stage Model: This model first uses Linear Regression to predict pressure differences between sensor pairs under normal (no-leak) conditions. In the second stage, it analyzes the residuals (differences between predicted and actual pressure differences) to classify leaks and estimate their locations using Multinomial Logistic Regression and Random Forest.

Datasets

The study utilized a real-time dataset collected from a treated wastewater supply line at the Indian Institute of Science (IISc), Bangalore. This network is 8.974 km long and comprises 90 nodes, 93 pipes, and 2 reservoirs. For experimental purposes, three controlled leaks were introduced at different pipe segments. The data was collected using nine IoT-enabled flow and pressure sensors, with readings taken every minute. The dataset was divided into training (70%) and testing (30%) sets.

Key Insights and Ideas

  • Digital Twins for WDNs: The paper highlights the potential of Digital Twins, augmented with real-time sensor data and ML, to revolutionize WDN management by enabling proactive leak detection.
  • Two-Stage Model Effectiveness: The two-stage approach, particularly when combined with Linear Regression and Multinomial Logistic Regression, demonstrated superior performance, achieving 100% accuracy in leak detection and localization.
  • Sensor Placement Matters (with limited sensors): While using all nine sensors yielded high accuracy across models, the study revealed that when using a limited subset of sensors (e.g., three), both the model choice and the spatial distribution of sensors significantly impact performance.
  • Random Forest Robustness: The single-stage Random Forest model showed robustness, maintaining good accuracy even with limited and potentially less optimal sensor combinations.

Key Results

  • High Accuracy with Full Sensor Data: When all nine sensors were utilized, both single-stage models (Logistic Regression and Random Forest) achieved 100% accuracy. The two-stage model combining Linear Regression and Multinomial Logistic Regression also achieved 100% accuracy.
  • Two-Stage Model Performance: The two-stage approach using Linear Regression and Random Forest achieved a slightly lower accuracy of 77.5% due to some misclassifications.
  • Limited Sensor Performance: When using only three sensors, the single-stage Random Forest and Logistic Regression models achieved over 70% accuracy in a significant number of sensor combinations (57 and 43, respectively). The two-stage models showed varying performance based on sensor combinations. The best accuracy for all models was achieved with the most scattered sensor combination (locations 1, 2, and 9).

Limitations and Future Work

  • Limited Leak Locations: The study tested only three controlled leak locations. While these were introduced in separate pipe segments, a broader range of leak scenarios and locations would provide a more comprehensive evaluation.
  • Sensor Placement Sensitivity: The paper notes that with a limited number of sensors, performance is sensitive to placement. Future work could focus on optimizing sensor placement strategies to maximize detection accuracy.
  • Real-World Complexity: While tested on a real-time dataset, the controlled leaks might not fully replicate the complexity of spontaneous leaks in a live network, which can be influenced by various environmental factors and operational changes.
  • Future Directions: The authors suggest further exploration into optimizing sensor placement and investigating more advanced ML techniques for enhanced leak localization and prediction.

Summary updated

Abstract

This paper presents the development of a Digital Twin (DTwin) to detect and localize the leaks in water distribution networks (WDNs), using single-stage and two-stage data-driven models. In the single-stage model, we test the anomalies in the dataset using Logistic Regression and Random Forest. In the two-stage model, a linear regression model predicts pressure differences between sensor pairs in the first stage. Based on this, we compute the distribution of residuals. In the second stage, changes in the residual distribution are classified using Multinomial Logistic Regression and Random Forest models to compute possible leak locations’ posterior probabilities. We have tested these models on a real-time dataset.
BibTeX
@misc{pandey2024machine,
  author = {Prerna Pandey and Nikolaj T. Mücke and Shashi Jain and Parthasarathy Ramachandran and Sander M. Bohté and Cornelis W. Oosterlee},
  title = {{Machine Learning-Based Digital Twin for Water Distribution Network Anomaly Detection and Localization}},
  year = {2024},
  doi = {10.3390/engproc2024069201},
}
Figure