Gearbox Fault Diagnosis Based on ResNet50-MSVEA-GATv2 Model

Addressing the issue of incomplete feature extraction during the recognition of different types of fault signals in the sun gear of a planetary gearbox and the insufficiently close node relationships in the graph neural structure operations, this paper embeds a Multi-scale Vein Expansion Aggregation (MSVEA) module into the joint ResNet50-GATv2 model, which combines a ResNet50 and a GATv2. This integration is aimed at achieving accurate classification of gearbox fault signals. Specifically, first, the ResNet50 model is used to extract features from the sub-time domain graphs segmented by sliding windows, which are then used as node information in the GATv2 model. Secondly, the MSVEA model is used to retrieve multiple strongly correlated single-line graph structures and attention algorithms are employed to redistribute the weights of node relationships. Finally, experiments and data analysis are conducted, sparsity confusion matrices of predicted categories are plotted, and four evaluation metrics for test samples are calculated. The experimental results show that the ResNet50-MSVEA-GATv2 classification model proposed in this paper outperforms other experimental groups in terms of classification results, with an accuracy rate of $\mathbf{9 6. 4 5 \%}$. This provides a new approach for applying graph neural networks in the field of mechanical fault diagnosis.