Res50GCN-JEGP: a novel gearbox fault diagnosis method based on optimized graph structure training

Abstract The planetary gearbox is widely used in fields such as automobiles, aerospace, machinery, power, and metallurgy. To address issues like insufficient comprehensiveness and correlation of feature extraction, low computational efficiency, and poor classification performance in the identification of gear fault signals using graph neural networks, this paper proposes an improved model, Res50GCN-JEGP, based on the combination of the residual network ResNet50 and the graph convolutional network (GCN), with an embedded joint enhanced graph pooling (JEGP) layer for dimensionality reduction and data fusion in deep learning. Specifically, the paper first plots the signals in the time domain and splits them into multiple images using a sliding window, followed by feature extraction from all input images using the ResNet50 model. Then, semi-supervised learning training is performed within the GCN framework, and the JEGP layer is inserted to achieve a comprehensive diagnosis of multi-class signal samples by integrating different pooling strategies. Finally, a detailed analysis of the two test set samples is conducted, with the sparse confusion matrix of predicted categories and the visualization of predicted labels being plotted, followed by the calculation of four evaluation metrics. The proposed model achieves a classification accuracy of 96.66% and 96.933%, respectively. Further experiments reveal that the removal of the JEGP module leads to significant performance degradation. Additionally, when replacing the feature extractor with other CNNs (such as CNN, MobileNet, ZFNet, and ResNet34), the model accuracy is lower than the proposed method, validating the universality of the JEGP module and the suitability of ResNet50’s feature representation in this framework.