Research On Intelligent Diagnosis Method Of Bearing Fault Based On Transfer Learning

Bearings are an important component in industrial equipment,and their health is related to the efficiency of industrial equipment.Bearing fault diagnosis has become a necessary task for factory operations.Nowadays,with the continuous application of machine learning and deep learning methods to various fields,it also promotes the development of data-driven bearing fault diagnosis methods.However,this diagnosis method is based on the condition of collecting a large amount of bearing fault data with markers.In practical application scenarios,the marked data already requires a lot of manpower to obtain,and it takes a long time to collect the bearing fault data to a small amount.Therefore,the diagnostic effect of this method is not ideal in practical application scenarios.In view of the above problems,this thesis uses the method of transfer learning to transfer the knowledge of labeled bearing data to unlabeled bearing data.And learn from the idea of GAN to add a domain discriminator to the transfer learning model to improve the generalization ability of the model.In this thesis,the transfer learning is researched and improved by focusing on the problems of large differences in cross-domain data distribution,small amount of labeled sample data,and low diagnostic accuracy when carrying out transfer learning for bearing fault diagnosis across operating conditions and equipment.The main work is as follows:(1)The distribution difference between the two sets of bearing signal data across working conditions is not conducive to knowledge transfer,and in practical application scenarios,the amount of bearing fault signal data is small and the data lacks diversity.This thesis considers enhancing the signal data before feature extraction,and increases the diversity and complexity of the data by rotating and shifting the signal points.In addition,multi-layer domain adaptation is used to assign different weights to the results extracted by each convolutional layer,so as to reduce the adverse effects of the features extracted in the deep network that are not suitable for migration on model training.Finally,by performing multiple sets of cross-working-condition migration tasks on three sets of public bearing datasets,it can be verified from the experimental results that the method proposed in this chapter can effectively improve the two problems aimed at.(2)Thinking about fault diagnosis across working conditions,and cross-equipment migration diagnosis is also a common phenomenon.Therefore,we also carried out research on the improvement of migration learning method for the data of different types of bearings.Considering that cross-equipment bearing data will have different sampling frequencies,the data is resampled during data processing to ensure frequency alignment.Then,for the problem that the label space of different types of bearings may be inconsistent,the single domain discriminator is divided into multiple sub-domain discriminators,starting from each type of data,and shortening the distance between the cross-domain data under each type of label.Experiments on the transfer task between bearings,the accuracy of the algorithm proposed in this chapter is about 4% higher than the other four algorithms,which proves the effectiveness of the method.