Research On Domain Adaptive Fault Diagnosis Of Rotating Machinery Based On Transfer Learning

Rotating machinery is the foundation of industrial development and an integral part of manufacturing.As one of the most important components in rotating machinery,the failure of rolling bearings can lead to industrial shutdowns and industrial accidents.Therefore,the health monitoring of rolling bearings is particularly important.Methods such as early analysis of failure frequency rely too much on the expertise of diagnosticians and are not time-sensitive.In recent years,the rapid development of deep learning has improved the ability to analyze fault data in the diagnosis process,but there are still problems of noise,variable working conditions and unbalanced samples leading to the decline of diagnostic accuracy.In view of the above problems,this paper uses transfer learning and adversarial generation methods to establish a fault diagnosis model based on deep neural network,and the main contents and innovation points are as follows:(1)In order to improve the ability of neural network to extract fault features,this paper proposes a bearing fault diagnosis method to improve the split attention network.Firstly,the continuous wavelet transformation method is used to convert the onedimensional vibration signal into a time-frequency image as the input of the network.Secondly,the channel attention mechanism and multi-branch structure are combined with the residual neural network,and the fixed convolution kernel is replaced by multiscale convolution kernels to increase the correlation between each channel and expand the receptive field of the convolution kernel.Experimental results show that the model can improve the ability of the network to extract fault features and improve the accuracy of fault diagnosis.(2)Due to the variable actual operating conditions of rolling bearings,there is a large difference in the distribution of source and target domain samples,and the diagnostic model trained with only single case samples has poor diagnostic effect in cross-condition diagnosis tasks.To solve this problem,this paper proposes a subdomain adaptive bearing fault diagnosis method based on improved split attention network.Firstly,the improved split attention network is used to extract the migrable fault characteristics;Secondly,the local maximum mean difference method is used to reduce the feature distribution difference between the source domain and the subdomains of the target domain in the feature distribution space,so as to avoid feature aliasing in the domain.The experimental results show that under unsupervised conditions,the model can achieve high diagnostic accuracy in cross-condition diagnosis tasks with only the training of a single working condition sample set.(3)The sample imbalance problem will lead to the lack of training of individual faults in the network,which will limit the migration diagnosis effect of the model.To solve this problem,this paper proposes a fault diagnosis method based on generative adversarial network.Firstly,the original JS divergence(Jensen Shannon divergence)is replaced with W distance(Wasserstein distance)to measure the difference between the generative distribution and the true distribution,so as to avoid the problem of gradient vanishing during the generator training process.Then,the gradient penalty method is used to optimize the network to improve the convergence speed and training stability of the network.Finally,the generative adversarial network is used to expand the unbalanced sample set of the source domain,and the expanded sample set is used as a new source domain for migration diagnosis tasks.Experimental results show that the model can be used to enrich the sample set to reduce the influence of unbalanced samples on the accuracy of final diagnosis.