Improving Mechanical Fault Diagnosis Using Graph Neural Networks with Dynamic and Multiscale Features

Mechanical systems face a major drawback to fault diagnosis, as class imbalance greatly undermines it since minority class instances (critical faults) are underrepresented, resulting in biased predictions. This paper introduces a new Multiscale Receptive Fields and Dynamic Edge Weighting (MRS-GNN) framework, which fuses MRF and Dynamic Edge Weighting (DEW) on a GNN to improve classification performance in imbalanced datasets. Graph edge strengths are dynamically weighted with the learned node embeddings during training according to the DEW mechanism, and MRF allows the model to aggregate information from different neighborhood scopes for robust feature representation. In addition, a graph-specific oversampling algorithm, MR-SMOTE, was used to generate synthetic minority class nodes respecting and preserving the topology of the graph. The proposed model was evaluated through experiments on the 2009 PHM gearbox dataset and was found to have an accuracy of 92.1% and an AUC-ROC score of 0.95, better than traditional oversampling methods such as SMOTE, LR-SMOTE, and Graph-SMOTE. The results of an ablation study indicate that 3.7% and 2.4% accuracy drops occur in DEW and MRF removals, respectively, highlighting their importance. This study proposes a scalable and topology-preserving solution to the imbalanced fault diagnosis problem and makes substantial improvements compared to existing GNN-based methods.