Edge Temporal Digital Twin Network for Sensor-Driven Fault Detection in Nuclear Power Systems.

Area of Science:

• Nuclear Engineering
• Data Science
• Artificial Intelligence

Background:

• Nuclear power systems rely on sensor networks for safe operation.
• Data sharing restrictions hinder the development of generalized digital twins.
• Existing digital twins often ignore temporal sensor data patterns crucial for fault prediction.

Purpose of the Study:

• To propose an Edge Temporal Digital Twin Network (ETDTN) for cloud-edge collaborative fault detection in nuclear power systems.
• To address challenges of data privacy and non-IID data in nuclear sensor networks.
• To improve fault prediction by incorporating temporal information from sensor data.

Main Methods:

• Developed an Edge Temporal Digital Twin Network (ETDTN) utilizing a continuous variable temporal representation.
• Incorporated a global representation module to handle non-IID data across subsystems.
• Integrated a temporal attention mechanism with graph neural networks for enhanced temporal feature learning.
• Employed federated parameter aggregation to ensure data privacy.

Main Results:

• ETDTN demonstrated significantly superior fault detection performance compared to existing methods in non-sharing data scenarios.
• Achieved state-of-the-art results in both accuracy and F1 score on real nuclear power datasets.
• Validated the effectiveness of temporal feature learning and privacy preservation.

Conclusions:

• ETDTN effectively preserves data privacy through federated learning.
• The network successfully captures latent temporal patterns in sensor data for improved fault detection.
• ETDTN offers a robust solution for sensor-driven fault detection and predictive maintenance in nuclear power systems.