Construction of a Deep Learning-Driven Reliability Assessment Model for Nuclear Fuel Cycle Facilities

Abstract

Reliability assessment of nuclear fuel cycle facilities is crucial for the safe operation of nuclear energy. However, traditional methods suffer from drawbacks such as difficulty in fusing multi-source heterogeneous data, insufficient capture of equipment topology correlations, and inadequate learning of dynamic failure features. To address these issues, this paper proposes a Dynamic Attention Graph Convolutional Neural Network (DAGCN) for facility reliability assessment. The model constructs a three-layer collaborative module: a multi-source data preprocessing module that uses sliding window Z-score normalization to eliminate dimensions and combines multimodal attention to highlight core indicators; a dynamic attention graph convolutional module that constructs a condition-dependent equipment topology graph to accurately locate key failure nodes and capture associated failure features; and a temporal coding module that learns the progressive evolution of operating conditions through bidirectional LSTM. Experiments are conducted using real data (100,000 data points) from a nuclear power plant from 2019 to 2023 and Monte Carlo simulation extreme data (20,000 data points), comparing six types of traditional, machine learning, and deep learning models. The results show that DAGCN achieves an accuracy of 98.2% and a recall of 97.8%, representing improvements of 3.5 and 4.2 percentage points respectively compared to the optimal baseline model GAT. The MAE is 1.23 and RMSE is 1.56, representing reductions of 22.1% and 19.3% compared to GAT. Even under extreme operating conditions, the accuracy remains above 95.8%, demonstrating significantly better stability than other models. This research provides technical support for the intelligent upgrading of reliability assessment for nuclear fuel cycle facilities.