Physics-Embedded Recurrent Graph Neural Network for Fault Diagnosis of Complex Systems

The improvement of fault diagnosis for complex equipment is an important step towards intelligent systems. Unlike component-level fault detection, system-level fault diagnosis presents new challenges, such as the integration of multiple sensors, the handling of massive datasets, and the necessity for robust system-level decision-making. To address these challenges, this paper proposes a novel Physics-embedded Recurrent Graph Neural Network (Pe-RGNN). By embedding physical knowledge about the interdependencies among sensors, the graph constructed of sensors are built through physical knowledge obtaining from high-fidelity model. With the sensory graph, the parameters are modeled through recurrent graph neural network to extract the temporal information. The proposed model has the advantage of simultaneously considering both spatial and temporal information. Experimental results from both simulated and real-world tests demonstrate the superior performance of the Pe-RGNN in early fault detection and fault identification accuracy when compared to traditional graph neural networks. These findings highlight the importance of incorporating both physical knowledge and time-series information in fault diagnosis models.