Study On Multi-scale And Multi Channel Feature Learning And Transffr Of Deep Convolutionional Neural Networks Towards Bearing Fault Diagnosis

Modern large-scale mechanical equipment is becoming more and more complex,intelligent and precise.As an important part of most mechanical equipment,bearing is widely used in the equipment transmission system.With the development of the system’s increasingly extreme efficiency and the systematization of functional modules,the connection of each part of the equipment is closer.Once the mechanical failure occurs,the loss and influence will be equivalent Therefore,accurate bearing fault diagnosis plays a key role in mechanical health management.In recent years,the introduction of deep learning method provides new ideas and methods for fault diagnosis and prediction.This paper mainly studies the fault diagnosis based on deep learning of residual dense network and transfer learning.Firstly,aiming at the problem that it is difficult to obtain comprehensive information from the original bearing data and a large number of features extracted by deep learning model have different importance for bearing fault diagnosis,a bearing fault diagnosis method based on multi domain information fusion and improved residual dense network is proposed.By multi domain transformation and multi domain information fusion,this method can mine rich fault information from the original signal,use the improved residual dense network to adaptively extract robust features from the signal,and use Convolutional Block Attention Module to weight the extracted fault features,and finally use softmax to output the prediction results of the network.This method can further improve the accuracy of fault diagnosis under the given sample number.Secondly,in view of the difference of receptive fields of different convolution check features and the lack of a large number of labeled data,this paper proposes a bearing fault diagnosis method based on multi-scale convolution network and transfer learning.The method adaptively extracts the rich and complementary features of the source domain and the target domain through multi-scale convolution network,and supplements the target domain data with the help of transfer learning.Finally,labeled samples of target domain are used to test the classification performance of the network.This method overcomes the problem of only a small number of labeled data samples in the target domain,and realizes intelligent fault diagnosis with high accuracy in the target domain.The method proposed in this paper is verified on the bearing data set of Case Western Reserve University and the data set of XJSY rolling bearing.Through a large number of comparative experiments,the influence factors of the key model such as the number of convolution layers,the number and size of convolution kernels are analyzed,and compared with common methods.The experimental results show the effectiveness of the research method in this paper.