Study On Bearing Fault Diagnosis Based On Deep Transfer Learning

The structure of mechanical equipment is becoming more and more complex and automated,and the requirements for equipment safety and reliability are getting higher and higher.The detection and diagnosis of mechanical faults is particularly important.Data-driven methods based on vibration signal analysis have been extensively studied.However,most intelligent fault diagnosis technologies are based on the same distribution of training data and testing data,and the influence of complex working conditions in practice makes the diagnosis effect of the theoretical method greatly affected.The introduction of transfer learning has provided new ideas for solving bearing fault diagnosis problems under different distributed data.This paper mainly studies the cross-domain bearing fault diagnosis method based on deep transfer learning.First of all,considering that the complex and changeable working conditions in the actual operation of rotating machinery often lead to deviations in data distribution,which reduced the recognition accuracy of a single recognition model.In this paper,domain adaptation is introduced into the process of bearing fault diagnosis.Based on deep learning and transfer learning,a transfer diagnosis network with synchronous adaptation of feature space domain and label probability distribution is proposed.In this network,one-dimensional dense convolution network and attention mechanism are integrated to realize the automatic extraction of complex fault features;domain adaptation processing constrains the network to learn domain invariant features by jointly minimizing the difference of feature probability distribution and label probability distribution;finally,the rolling bearing fault under variable conditions is identified with high accuracy.Secondly,to address the problems of less labeled samples and insufficient resource utilization of multiple similar data sets,this paper proposes a bearing fault diagnosis method based on multi-source domain adaptation.This method is based on the feature extraction of convolutional neural network,and comprehensively utilizes the extracted multiple sets of source domain data information.First extract the domain invariant information between different source domain data and target domain data,and construct each source domain classifier,and then realize the complementary utilization of multiple source domain information through the weighted combination of source domain classifiers,and finally complete the target domain sample class recognition.The experimental results of cross-machine diagnosis under multiple data sets prove the effectiveness of this method.The method proposed in this paper was verified on multiple data sets such as the Western Reserve Bearing Data Set and IMS Bearing Data Set.The important parameters of the proposed method and its influence were deeply analyzed and the proposed methods are compared with the common methods.The experimental results show the effectiveness and better performance of the research method in this paper.