Fault Diagnosis of Gear Based on Multichannel Feature Fusion and DropKey-Vision Transformer

To solve the problem that it is single-channel vibration signals not being able to fully express fault feature information and diagnostic networks not being able to fully capture its information resulting in low diagnostic accuracy, a new gear fault diagnosis method is proposed. First, subtraction average-based optimizer (SABO) as an optimization algorithm is introduced to optimize the parameters of variational mode decomposition (VMD) quickly and with high quality to conduct signal preprocessing. Next, the noisy signals in each channel can be quickly and effectively processed to obtain clean 1-D and prominent vibration characteristics signals from multichannel. Then, multichannel information is fused to obtain image datasets for diagnosis based on symmetric dot pattern (SDP) to realize clear signals transformed into images. A diagnostic model is proposed based on DropKey added for vision transformer (DVit) to enhance the diagnostic network’s ability to comprehensively capture multichannel feature information. Finally, the proposed method is validated through three datasets from gear fault diagnosis experiments with the average accuracy in fault diagnosis reaching more than 99.5% whether it is the degree or type of fault diagnosis. The average accuracy has increased by at least 0.5% compared with before improvement, and it has increased about 2%–7% compared with other methods. The results with visualization form verify the effectiveness and superiority of the proposed method.