Interpretability study of a typical fault diagnosis model for nuclear power plant primary circuit based on a graph neural network

Weak interpretability has become a huge obstacle for the practical application of artificial intelligence diagnosis models in the nuclear field. In order to solve the above problem, this study proposes a fault diagnosis method of Graph Neural Networks (GNNs) combined with fault causal directed graph. The method summarizes the typical fault causal directed graphs of nuclear power plant through the system physical structure and expert knowledge, and combines it with the spatial inductive framework of GNNs to achieve qualitative interpretable diagnosis. Furthermore, this study analyses the feature representation weights of various types of sensor nodes in the fault diagnosis process based on the self-attention mechanism, which is used to elucidate the decision-making process of the model's fault diagnosis and to achieve quantitative interpretability analysis. The proposed model is validated by simulation data from the simulator of Fuqing No.1 pressurised water reactor nuclear power plant. The results show that the proposed model is able to diagnose the fault types effectively, and the decision-making process of the model is logical and interpretable. Therefore, this study opens up a new technical approach with both accuracy and interpretability in the field of nuclear power plant fault diagnosis. By improving the interpretability of the intelligent diagnosis model, it effectively promotes the application of artificial intelligence technology in the fault diagnosis field of the nuclear industry, and provides a new enlightenment for the application of complex system fault diagnosis in other fields.