Physics-constrained causal structure learning for root cause analysis in nuclear system monitoring

Root cause analysis (RCA) is important for identifying the initiating cause of faults and analyzing fault propagation paths in nuclear systems. However, robust RCA in nuclear accident monitoring remains challenging because it depends on reliable identification of causal relationships among monitoring variables, which is difficult under complex system coupling and limited-sample conditions. To address this issue, this paper proposes a physics-constrained causal structure learning method for RCA in nuclear accident monitoring. First, a simplified physical model of the nuclear system is constructed to extract a priori causal knowledge among key monitoring variables. Second, the causal relationships among different variables are learned by a score-based causal structure learning method under the guidance of the extracted physical constraints. In this process, physically infeasible edges are excluded and the admissible parent sets are restricted before score optimization. Finally, the root cause nodes of faults are identified based on the learned causal graph. The proposed method is validated using two representative accident scenarios generated from a nuclear simulation platform. The results show that the proposed method can improve root-node identification performance, reduce abnormal-edge misidentification, and enhance the robustness of causal structure learning under limited-sample conditions without introducing additional computational burden.