The Role of Graph Neural Networks in Enhancing Knowledge Inference Capabilities

Abstract

Graph Neural Networks (GNNs) provide a powerful framework for knowledge inference over relational data, offering significant advantages for Artificial Intelligence-empowered Intelligent Engineering Systems (IES). Grounded in systems engineering, cyber-physical systems (CPS) theory, and information-theoretic principles of structured message passing, this review examines how GNNs enhance inference capabilities to drive IES evolution. GNNs enable four key transitions: from standardized to personalized manufacturing through relational configuration reasoning; from reactive to proactive prediction via multi-hop inference on system graphs; from manual to autonomous control by learning dynamics on interaction topologies; and from isolated to integrated ecosystems through cross-domain dependency modeling. Applications are illustrated in fault diagnosis and predictive maintenance, process optimization and quality control (knowledge graphs for interdependent parameter inference), and human-machine collaboration (interaction graphs for operational efficiency). While GNNs deliver superior relational generalization, challenges remain in graph construction quality, scalability, interpretability, and physics integration. Future directions include hybrid neuro-symbolic and physics-informed architectures. GNNs thus constitute a pivotal advance for robust, structure-aware knowledge inference in complex engineering systems.