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

The safe and efficient operation of nuclear power systems largely relies on sensor networks that continuously collect and transmit monitoring data. However, due to the high sensitivity of the nuclear power field and strict privacy restrictions, data among different nuclear entities are typically not directly shareable, which poses challenges to constructing a global digital twin with strong generalization capability. Moreover, most existing digital twin approaches tend to treat sensor data as static, overlooking critical temporal patterns that could enhance fault prediction performance. To address these issues, this paper proposes an Edge Temporal Digital Twin Network (ETDTN) for cloud-edge collaborative, sensor-driven fault detection in nuclear power systems. ETDTN introduces a continuous variable temporal representation to fully exploit temporal information from sensors, incorporates a global representation module to alleviate the non-IID characteristics among different subsystems, and integrates a temporal attention mechanism based on graph neural networks in the latent space to strengthen temporal feature learning. Extensive experiments on real nuclear power datasets from 17 independent units demonstrate that ETDTN achieves significantly better fault detection performance than existing methods under non-sharing data scenarios, obtaining the best results in both accuracy and F1 score. The findings indicate that ETDTN not only effectively preserves data privacy through federated parameter aggregation but also captures latent temporal patterns, providing a powerful tool for sensor-driven fault detection and predictive maintenance in nuclear power systems.