Research On Ball Bearing Raceway Defect Transitive Transfer Diagnosis Method Based On Small Sample Data

In the practical application of data-driven bearing fault diagnosis,it is hoped that the method has a certain degree of cross-domain and cross-working condition diagnosis generalization adaptability.This is because the available labeled data is often not from the equipment to be diagnosed,and the data itself can not cover all kinds of working conditions.In this context,transfer learning methods have been widely studied in the field of fault diagnosis.However,limited by the performance of existing transfer learning methods,due to the large span of working conditions and the difference of vibration characteristics caused by the difference of equipment structure,the objectively huge differences between the source domain and target domain data provide a challenge for the effectiveness of transfer diagnosis.In response to this problem,this paper proposes a cross-working condition data supplement method based on Cycle-GAN and dynamics model,which can use dynamic simulation data to approximate the missing part of existing small sample data in the condition through cross-domain mapping,and further use it for transfer learning diagnosis in the target domain,to complete the transmission process of the actual signal diagnosis from the simulation data through the experimental data set.First,for the research needs,the bearing dynamic model and vibration test platforms are used to obtain sufficient vibration time-domain signals of bearing faults under multi-working conditions,which are used as the data basis for subsequent research.At the same time,at this stage,according to the characteristics of the time-frequency map generation,band-pass filtering,threshold noise reduction,logarithmic mapping and other methods are combined to highlight the useful features in the data samples,so that the model can learn more clear fault feature information.After obtaining the data set that meets the experimental requirements,take the experimental data of the limited working condition as the target domain,the simulation data corresponding to the working condition as the source domain,and use the working condition as the benchmark to constrain the data correspondence between the two data sets.The Cycle-GAN model is used to learn the feature mapping from simulation to experiment.When the model training effect is good,the corresponding fault diagnosis model training set can be generated for the detected data under any working condition within a certain range based on the relatively easy-to-obtain simulation data by the generator in the corresponding direction.In this paper,the effectiveness of the transfer method is tested.First,build a discriminant model for bearing fault diagnosis based on the Res-net model,test its basic performance through the obtained data set,and study the performance degradation of the model and the transfer learning method based on the adaptive probability distribution in cross-domain diagnosis tasks under corresponding working condition and cross-domain-and-working conditions diagnosis tasks.