Mechanism-Aligned Nuclear Power Plant Accident Diagnosis via Physically Guided Concepts and Evidence Paths

Accident diagnosis in nuclear power plants (NPPs) should provide mechanism-aligned evidence that can be reviewed by operators and safety engineers, rather than only a high-confidence accident label. Existing data-driven methods achieve strong classification performance but often express explanations as attention maps, anomalous nodes, prototypes, or causal links separately, making it difficult to obtain a unified diagnostic evidence chain. To address this limitation, we propose a Concept-Constrained Physical Graph (CCPG) framework that formulates accident diagnosis as structured evidence generation. CCPG groups multivariate transient signals into operator-readable physical nodes, extracts node-wise temporal features, and propagates them over a mechanism-guided graph. It then couples a concept bottleneck with implicit latent features, prototype learning, and edge/stage supervision to predict the accident class and a reviewable evidence package. On the evaluated NPPAD five-class simulated benchmark, CCPG achieved saturated clean-set classification (1.000 accuracy) and high paired-challenge accuracy (0.996) while providing concept, affected-node, edge-template, stage-order, and prototype evidence. Additional analyses, including Transformer baselines, feature-restricted and early-window stress protocols, calibration, statistical testing, open-set detection, and scalability profiling, further characterize the robustness, reliability, and applicability of the proposed framework.