| Rolling bearing is the main component of rotating equipment,fault diagnosis of rolling bearing can ensure its reliable operation.In the practical application,the rolling bearing is in the normal state most of the time,and the number of samples collected in the normal state is much larger than that in the fault state,thus,there is a phenomenon of unbalanced data collection.However,the fault type samples with low frequency and relatively small number of samples are focuses of bearing fault diagnosis,so how to achieve accurate diagnosis of various types of faults in the case of data imbalance also has important research significance.Intelligent diagnosis technology based on deep learning provides strong support for solving this problem,In addition,Feature fusion makes the fused features more comprehensive and representative by extracting useful information from different angles,so as to improve the diagnosis accuracy.However,for the fault diagnosis of rolling bearing under unbalanced data,the traditional deep learning methods mainly analyze the signal in time domain and frequency domain,or use only one neural network to extract signal features,which will lead to the problems of low accuracy of rolling bearing fault diagnosis and poor performance of a few types of diagnosis under unbalanced data.Therefore,from the perspective of feature fusion,this thesis proposes an improved method to improve the accuracy of fault diagnosis,The main research work is as follows:(1)A diagnosis method based on over sampling and time-frequency feature fusion is proposed.The smote borderline algorithm is used to expand the minority samples,dual networks are used to extract the frequency domain and time domain features of the balanced signal respectively,The fusion features are used to classify the faults,which solves the problem of low accuracy of bearing fault diagnosis under unbalanced data.The experimental results on the CWRU bearing data set show that the average diagnostic accuracy is 94.14%,95.74% and 97.43%respectively by using the features of time-frequency fusion under three different unbalance ratio data sets,Compared with single domain and single network analysis,the diagnostic accuracy has been improved.(2)This thesis presents a diagnosis method based on generative adversarial networks(GAN)and cross layer features fusion.Gan is introduced to enhance the data,and combined with the global and periodic characteristics of the signal to realize the diagnosis,which improves the diagnosis accuracy of a few types of faults.The experimental results on XJTU-SY bearing data set show that the average diagnosis accuracy of the model is 95.25% for four kinds of minority faults when the data is unbalanced,which improves the diagnosis accuracy under the condition of unbalanced fault data.(3)In order to improve the ability of feature selection,attention mechanism is introduced into the constructed network model to highlight the key features that contribute a lot.After feature extraction by deep learning network,the attention mechanism is used to suppress the unimportant features with small contribution,the problem of low accuracy of a few kinds of diagnosis is further solved.Experimental results show that compared with the model without attention mechanism,the average diagnostic accuracy of the proposed method for minority classes is improved by 10.5%. |
