Multi-Branch Time-Frequency Spectrum Segmentation Approach for Gearbox Fault Diagnosis

In gearbox fault diagnosis research, fully utilizing the time-frequency information at meshing frequencies is crucial for enhancing the accuracy and robustness of fault diagnosis models. This study proposes a Multi-Branch Time-Frequency Spectrum Segmentation Network (TFSSN) model for gearbox fault diagnosis, employing deep learning to fully explore and integrate the time-frequency information from multiple meshing frequencies of multistage gearboxes. An adaptive time-frequency spectrum segmentation algorithm is proposed within this model to segment the time-frequency spectrum at various meshing frequencies of the gearbox. A multi-branch structure is then used to extract features from each segmented time-frequency spectrum. Finally, the designed Multi-Path Fully Connected Attention module fuses the features extracted from different time-frequency spectrum branches to accomplish fault diagnosis of multistage gearboxes. The performance of TFSSN was validated using gearbox datasets under various operating conditions. The experiments demonstrate that this method effectively extracts features strongly correlated with faults and significantly mitigates the overfitting issue in neural network models under conditions of sparse fault samples. Compared to conventional gearbox fault diagnosis methods, the proposed approach significantly improves the accuracy and robustness of fault diagnosis under conditions with limited fault samples. Specifically, the method achieves an accuracy of 0.982 and an F1 score of 0.99, demonstrating superior performance over existing methods.