Artificial Intelligence (AI) is increasingly deployed in safety-critical infrastructures to support anomaly detection, fault diagnosis, and predictive maintenance. However, purely data-driven models often operate as black boxes, limiting their transparency, causal interpretability, and practical suitability for risk-informed decision-making in safety-critical systems. This work proposes a knowledge-augmented Graph Neural Network (GNN) framework integrated with Multilevel Flow Modeling (MFM) to enable explainable and causally consistent fault diagnosis. MFM, a qualitative functional reasoning method, is used to encode system goals, physical flows (mass and energy), control logic, and means-end relations into a structured multi-relational graph. This graph representation is then mapped into a multi-domain GNN architecture, where distinct adjacency matrices represent mass flow, energy flow, control logic, and functional relations. Domain knowledge is embedded at three levels: (1) structured graph construction derived from MFM; (2) knowledge-guided feature engineering; and (3) a semantic regularization loss function that enforces alignment between diagnosed fault types and their corresponding physical domains. This design prevents “right-for-the-wrong-reasons” predictions and enhances causal consistency between model attention and first-principle constraints. The proposed framework bridges functional safety modeling and graph-based deep learning, offering a transparent diagnostic mechanism that aligns AI reasoning with physical laws and system objectives. The framework establishes a pathway toward trustworthy AI-assisted decision-making in safety-critical infrastructures.