| Rolling bearings are the main components of rotating equipment.In order to ensure reliable and economical operation of rolling bearings,intelligent fault diagnosis technology is essential.Due to the non-linear and non-stationary characteristics of rolling bearing vibration signals,whether representative fault features can be extracted has a direct impact on the results of fault diagnosis.In addition,when faced with the problem of few fault samples and insufficient label information in the data collected by the actual industry,the performance of the traditional deep learning method may be reduced or even invalid.Transfer learning can solve such problems by learning knowledge of one domain to solve new and related tasks in another domain.In response to the above problems,this article focuses on feature extraction and cross-domain migration learning to study bearing failures.The main contents are as follows:(1)In order to extract effective features from the signal,a method combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)and energy entropy is proposed.On the bearing data set,it is verified that the proposed method can accurately and effectively decompose the signal and obtain effective fault characteristics,which lays the foundation for subsequent diagnosis and analysis of rolling bearing faults.(2)The improved Softmax function combined with LSTM is used to identify the CEEMDAN energy entropy feature vector extracted from the bearing vibration signal.Center loss(Center Loss)introducing Softmax loss can reduce the distance between classes and further improve the classification accuracy.The experimental results on the CWRU bearing data set are visualized by T-SNE.The visualization results on the CWRU bearing data set show that after introducing Center Loss,each sample is as close as possible to the center of the sample,making it easier to classify.(3)In the cross-domain fault diagnosis,the generalization ability of the migration model is poor when the target domain marked data is small.Therefore,a weakly supervised transfer learning model based on pseudo-labels is proposed.The target domain data set with pseudo tags is constructed to help the tagged source domain data train the model together.Wasserstein metric is used to calculate the distribution difference between two domains.Through iterative learning and 1D-CNN learning to transfer features,domain adaptation is realized.Experiments on CWRU,IMS and MPC data sets prove that the model has better migration effect,and higher diagnostic accuracy can be obtained when the target domain has no sample tags or few sample tags. |
