Toward Knowledge-Augmented Graph Neural Networks for Explainable Root Cause Analysis: A Multilevel Flow Modeling Approach

Safety-critical process systems face challenges in fault diagnosis due to limited labeled failure data and complex nonlinear coupling between components. Purely data-driven Prognostics and Health Management (PHM) methods can detect anomalies but often lack causal interpretability, limiting trust in safety-critical decision-making. This study explores a knowledge-infused framework that integrates Multilevel Flow Modeling (MFM), a functional modeling approach that encodes causal reasoning, into a Graph Neural Network (GNN) based diagnostic architecture. We construct a multi-relational graph representation where nodes represent system components and edges capture physical connections (mass/energy flow/control logic) and causal relationships (means-end relations). MFM knowledge is injected into the GNN through (1) differentiated graph structures reflecting physical and causal dependencies, and (2) causality-aware loss functions enforcing consistency with MFM propagation paths. Preliminary validation on a Minox deoxygenation system demonstrates that incorporating MFM-based functional constraints enhances Top-1 diagnostic accuracy by 10.0 % compared to standard physical-topology GNNs, achieving a Top-3 accuracy of 98.5 %. The framework's explainability is showcased through diagnostic heatmaps, providing a proof-of-concept for traceable decision support. This study explores the potential of bridging symbolic functional reasoning and graphbased deep learning, offering initial insights into enhancing the interpretability of safety-critical PHM.

Keywords: Prognostics and Health Management (PHM), Causal reasoning, Graph Neural Network (GNN), Knowledge-based modeling, Multilevel Flow Modelling (MFM), Artificial intelligence, Root Cause Analysis, Safety and reliability, Explainable Artificial Intelligence (XAI), Safety-Critical Systems.