Research On Fault Diagnosis Method Of Rolling Bearing Based On Migration Learning Of Deep Stack Self Encoder

The efficiency of modern industrial production depends on the efficiency of equipment.As the most frequently used component in mechanical equipment,if the rolling bearing can judge its running state and even predict the existence of faults in advance,it will be of extraordinary significance to improve the level of mechanical management.The traditional rolling bearing fault diagnosis technology depends on manual experience and can not meet the diagnosis needs of big data.Deep learning technology can get rid of the dependence on signal processing and manual extraction,and has become the most mainstream application technology.However,the problem of ignoring the original features in extracting deep features by neural network and the influence of small sample data set lead to the poor adaptive ability of existing networks.In order to solve these problems,this paper proposes an embedded stack sparse automatic encoder model to improve the performance of network feature extraction,and verifies the impact of different kernel functions on the transfer learning method to improve small sample data.As the basis of subsequent analysis,this paper studies the influence of network parameters on the performance of neural network.Through the study of network parameters,regularization parameters and optimization parameters,the most appropriate parameter combination of neural network is selected.Aiming at the situation that the existing deep stacked autoencoder(DSAE)does not consider the original features in the network structure and training process,and the complementarity between deep features and original features is not ideal,an embedded stacked group sparse autoencoder ensemble(ESGSAE)is proposed.Firstly,by evaluating the effectiveness of hidden layer output,the constraint of original information is introduced in the layered training process to enhance the complementarity between deep features and original features.Secondly,an embedding unit is designed in the input structure of each self encoder,which embeds the original information into each coded output,integrates the original features and the output of the current layer,and constructs the feature representation in a higher hidden layer.Finally,the group sparsity constraint is used to eliminate redundancy and improve the degree of simplification.The experimental results show that the proposed ESGSAE model has better feature extraction ability and classification effect than DSAE model.In the production situation of real life,there are more unlabeled data and less labeled data,or at least sample data.Migration learning can be applied to the field of small sample data,learning labeled data close to the target data,migrating the learned data labels to supplement the data of the data set;Or use the whole or part of the trained model to train the target domain data.Kernel function is the key parameter in the accounting method.Based on the transfer component analysis method in the feature-based transfer learning method,this paper makes a comparative analysis of the impact of different kernel functions on the algorithm.The results show that the combined kernel with the ability to fuse multi-source information performs well under various migration conditions,and has better classification accuracy compared with other representative kernel functions.