Gear fault diagnosis based on SCResNeSt50 and transfer learning for small samples
LIU Jie
GUO Zefeng
YANG Na
Abstract：[Objective]There are still shortcomings in the transfer learning of gear fault diagnosis between different devices in existing research,especially under small sample conditions,and the diagnostic accuracy still needs to be improved.Therefore,a gear fault diagnosis method for small samples was proposed in this paper,which combined self-calibrated convolution integrated split attention network SCResNeSt50 and transfer learning strategy.[Methods]Firstly,the continuous wavelet transform was used to perform time-frequency analysis on the gear signal,generating a time-frequency graph as the model input.Secondly,based on the residual split attention network(ResNeSt)structure,a split attention mechanism and self-calibrated convolution were integrated to improve the linear processing mode of traditional convolutional neural networks for time-frequency graphs.The self-calibrated convolution was used to replace the conventional convolution in the ResNeSt module to achieve adaptive response calibration and multi-scale feature encoding,thereby expanding the receptive field and enhancing the ability to characterize fault features.Finally,a transfer learning strategy was adopted to fine tune the classifier parameters of the pre-trained model in the source domain,and the feature extraction layer was frozen,in order to achieve effective adaptation of the target task while preserving the general knowledge and feature representation of the source model and improving the accuracy of gear fault diagnosis under small sample conditions.[Results]Experiments were conducted on the gearbox dataset of Southeast University and the gear dataset of the University of Connecticut to verify the effectiveness of the method.The experiment included two scenarios:transfer learning under variable operating conditions and cross-dataset transfer learning.The proposed method was compared and analyzed with existing fault diagnosis methods.The results show that in the experiment of transfer learning under variable operating conditions,the diagnostic accuracy of the target domain reaches 98.7%and 98.9%.In the migration experiment from the dataset of Southeast University to that of the University of Connecticut,when the sample size of each gear state in the target domain training set is 25,20,16,12,8,and 6,the diagnostic accuracy reaches 98.1%,98.1%,97.8%,97.5%,96.5%,and 93.1%,respectively.[Conclusions]This method achieves better diagnostic accuracy than other methods in multiple experiments,which indicates that the improved self-calibrated convolution effectively enhances the characterization ability of gear fault features,and the transfer learning strategy significantly enhances the reliability of fault diagnosis under small sample conditions.This study provides a feasible solution to gear fault diagnosis under small sample conditions,promoting the development of intelligent fault diagnosis technology.
Keywords：small samplefault diagnosisgearself-calibrated convolutionsplit attention mechanismtransfer learningsource domaintarget domain
Publication Date：2026-01-25
Online Publishing Date：2026-03-17(First online date of this platform, not the publication date of the document)
Pages：10( 110-119 )