Enhancing GNN based fault diagnosis through implicit multi-sensor relation mining in limited sample conditions

Abstract In a rotating mechanical system, the physical connections among its components provide a natural topological structure for the graph neural network (GNN). The graph model constructed based on these physical connections can transform the operation data of each component into a structured representation, thereby assisting in the extraction and identification of fault features. However, the dynamic correlations among multi-sensor data contain rich information that goes beyond physical connections. These correlations reflect the internal coupling mechanisms and the evolution patterns of faults in the mechanical system. Thus, this paper proposes a small-sample fault diagnosis method based on the mining of implicit relationships among multiple sensors. This method deeply analyses the potential connections among the data of multiple sensors, effectively enhancing the feature extraction efficiency of the GNN and significantly improving the accuracy and generalization ability of fault diagnosis. To verify the reliability and practicality of this method, this paper conducts experiments on the PHM2024 dataset and the Southeast University Bearing Dataset. The results show that this method improves the ability to extract fault features by mining the implicit relationships among multiple sensors. Especially in scenarios with scarce small-sample data and noisy interference, this method still maintains a high diagnostic accuracy.