A Diagnostic Framework for Harmonic Drives Based on Dynamic Graph Data Augmentation and Adaptive Knowledge Distillation for Graphs

The operational state of harmonic drives demonstrates nonlinear and nonstationary characteristics, which pose challenges for traditional methods to extract features. Graph neural networks (GNNs) have shown significant potential in harmonic drive fault diagnosis owing to their ability to capture high-order correlations among nodes and adapt to dynamic changes. However, the industrial application of GNNs is limited by two major constraints: 1) the scarcity of fault samples and 2) the high-computational cost. To mitigate these constraints, this article proposes a diagnostic framework based on dynamic graph data augmentation and adaptive knowledge distillation for graphs (DGDA-AKDG). Specifically, DGDA employs a multiview feature extraction module and a weighted feature fusion module for graph data augmentation to mitigate the data imbalance problem. AKDG utilizes a policy network and a routing feature fusion mechanism to determine the distillation path, thereby optimizing the distillation location in the GNN model. Ablation and comparative experiments conducted on bearing datasets and industrial robot harmonic drive datasets indicate that DGDA improves dataset accuracy by 2.91%–12.60% after balancing, while AKDG reduces distillation accuracy loss by a factor of 3.54–14.51. These results demonstrate the superiority of DGDA-AKDG in handling heterogeneous distillation under data imbalance conditions, thereby further expanding the application of GNNs in intelligent manufacturing.