Graph Neural Networks for Modelling Structural and Functional Dependencies in Smart Cyber Physical Systems

Smart Cyber-Physical Systems (SCPSs) exist at the interface of physical phenomena together with computational intelligence, which requires modelling the structural interconnections and the functional dependences. The connection is established among and between the various elements with precision and in real-time scenarios. The traditional methods of modelling frequently miss the temporal dynamics and heterogeneous relationships that are characteristic of SCPSs. This paper proposes the use of a Graph Neural Network (GNN) in SCPSs (GNN-SCPS) methodology to design the structural and functional relations that exist between sensors, actuators, control units, and communication interfaces in SCPSs. The system is represented by a time-varying multi-relational graph, where nodes represent entities within the system and edges reflect dynamic dependencies, whether physical or cyber. The model proposed integrates message-passing GNN layers, referred to as temporal gating mechanisms and attention-based aggregation, to learn robust representations of node behaviors. The deployment of the model is operated under variable operational conditions and fault conditions. The SCPS artificial environment was designed to generate graph sequences with injected anomalies that simulate reality-related scenarios in industry. Experimental findings indicate that our approach outperforms fault localization, dependency inference, and anomaly detection compared to classical graph models and existing state-of-the-art mechanisms. The framework is also characterised by interpretability, as the mechanism of interconnection between diverse system parts can be recorded. This study constructs a database-driven, scalable study of modelling and monitoring SCPSs based on spatio-temporal graph deep learning.