Research On The Improvement Of Deep Transfer Learning Method In Fault Diagnosis Of Rolling Bearing Under Varying Working Conditions

The working environment of rolling bearing in actual industrial production is very complex and changeable.The vibration signal distributions of rolling bearings under different working conditions are very different.The labeled data under some working conditions are not very easy to be obtained.Therefore,it is of great significance to realize the intelligent fault diagnosis of rolling bearing under varying working conditions.For the above problems,this paper improves the existing deep transfer learning methods,an improved residual network deep subdomain adaptation method is proposed to realize rolling bearing fault diagnosis under varying working conditions.There are many problems in the aspect of data preprocessing,such as the insufficient number of samples in datasets and the limitation of one-dimensional data in fault diagnosis.For the problem of insufficient samples,a random data enhancement method is proposed to expand the datasets.Time-frequency analysis method is used to convert one-dimensional vibration signal into two-dimensional time-frequency image,which can enrich the information of samples,and reduce the impact of working condition changes on diagnosis result.For the receptive field in the shallow of residual network is not big and the correlations between channels is lack,and the model cannot extract features pertinently,the wide first-layer kernels mechanism and channel attention mechanism are introduced into the residual network,and the deep features of the input data can be extracted more purposefully.For the existing domain adaptation methods only align the global distribution of the samples,and the relationship between each subdomain is ignored in this process,the local maximum mean discrepancy is proposed to execute the operation of subdomain adaptation between the source domain and the target domain.Then,the distribution difference between each subdomain can be reduced and the feature aliasing can be avoided.The experimental results show that both the improved residual network and subdomain adaptation method can effectively improve the fault diagnosis accuracy of the model.Combining the improved residual network with the subdomain adaptation method,an end-to-end improved residual network deep subdomain adaptation model is constructed for fault diagnosis of rolling bearing under varying working conditions.Through experiments,short-time Fourier transform is selected as the time-frequency analysis method of data preprocessing.Comparative experiment of 8 transfer tasks are performed on CWRU dataset,and the average fault diagnosis accuracy of the proposed method is 98.2%.It is verified that the proposed method has good effect under varying working conditions.Combining the experiments of domain generalization and the validation experiments of MFPT dataset,the good generalization of the proposed method can be verified.