Few-shot bearing fault diagnosis based on coordinate attention relation network
Bearing fault diagnosis is of great significance to ensure the normal operation of machinery.Nowa-days,fault diagnosis methods based on machine learning such as Alexnet,Resnet-18,Relation Network,Re-lation Network based on Channel Attention SENet(SERN)and Relation Network based on Mixed Atten-tion CBAM(CBRN)are extensively utilized.However,the performance of these methods can be seriously damaged by small samples and changing working conditions existing in practical engineering applications,and even the problem of overfitting can be resulted.In this paper,based on the coordinate attention network,a novel bearing fault diagnosis method is proposed for overcoming these problems.By embedding the coordi-nate information and generating coordinate attention,a coordinate attention relationship network is con-structed to solve the problem that the relational network model cannot be used to establish the long-distance dependence between feature map and fault feature location information,enhance the model's expression on fault features in the target region and reconstruct more discriminative fault sample features.Then,a feature embedding module is used to generate the feature vector of samples,by which the labeled samples and unla-beled samples can be spliced to generate the feature vector group.Finally,a relationship score module is used to measure the nonlinear distance of feature vector group,generate the relationship score and judge the class of unlabeled samples to achieve fault classification.Simulation results show that,comparing with the known methods,the proposed method has better classification performance on small sample bearing data sets.
国家科技期刊平台
NSTL
万方数据
