Multi-scale GraphSAGE with class center balancing loss for rolling bearing fault diagnosis under extremely class imbalance

The imbalance between normal and fault data in the condition monitoring of rotating machinery often leads to models needing more focus on the information from the majority class. To this end, this work proposed a rolling bearing fault diagnosis method based on class center balancing loss (CCBL) and multi-scale GraphSAGE (MSGraphSAGE) to handle extreme class imbalance. First, a node-level pathgraph using frequency-domain signals enhances the model’s learning and generalization capabilities by associating signal features. Next, a multi-scale feature extractor is designed, employing DropEdge-based MSGraphSAGE in the first layer to improve the model’s feature extraction performance. Finally, a CCBL function is developed to reweight the class weights, reducing the weight loss assigned to the majority class to balance the class weights. Six imbalanced cases were designed on two bearing datasets, and the experimental results demonstrate the advantages of this method in highly imbalanced fault diagnosis tasks, validating the effectiveness and superiority of the proposed GNN model and class center balancing loss.