Assessing Sensor Integrity for Nuclear Waste Monitoring Using Graph Neural Networks.

Area of Science:

• Geological Engineering
• Data Science
• Sensor Networks

Background:

• Long-term monitoring of deep geological repositories for radioactive waste is crucial for safety.
• Sensor networks used for monitoring degrade over time due to environmental factors.
• Assessing sensor integrity and data consistency is vital for facility oversight.

Purpose of the Study:

• To develop and validate a Graph Neural Network (GNN) model for assessing data integrity in sensor networks.
• To leverage experimental data from a real-world underground research laboratory for GNN training.
• To evaluate the GNN's performance in detecting faulty sensors under realistic conditions.

Main Methods:

• Utilized experimental data from Andra's Underground Research Laboratory (URL) emulating thermal loading of a high-level waste demonstrator cell.
• Developed a modified GraphSAGE GNN model incorporating Graph Net elements to process temperature field data (current and past).
• Trained the GNN to classify individual sensors as faulty or non-faulty, even with up to 50% sensor failure.

Main Results:

• The GNN model successfully identified faulty sensors using real experimental data from a deep geological environment.
• The model demonstrated high effectiveness in detecting sensor failures, even when a significant proportion of sensors were compromised.
• Comparison against standard classification methods confirmed the GNN's superior performance in data integrity assessment.

Conclusions:

• GNNs are highly suitable for detecting faulty sensors in complex monitoring networks within deep geological repositories.
• The study validates the use of real experimental data and advanced GNN architectures for robust sensor network monitoring.
• This approach enhances the reliability of long-term monitoring essential for the safety and management of radioactive waste facilities.