Research On Fault Diagnosis Method Of Gearbox Bearing Based On Two-Dimensional Convolution

In the face of the high complexity and integration of modern mechanical systems,the fault diagnosis of bearings needs to introduce higher dimensional feature data.In order to solve the problem of low feature dimension,this thesis uses two-dimensional convolution features to do the following research.Firstly,the ADAM algorithm is improved,and the convolution structure and input dimension are improved.Lenet-5 model was selected by transfer learning with accuracy and rate as indexes.A new dynamic learning rate method is proposed,and the exponential attenuation formula is introduced as a new learning rate attenuation method to realize the fast convergence of the gearbox bearing fault diagnosis model.This method is proved to be effective.The residual block idea is introduced to construct the convolution module,and the results of the bearing experiments show that the proposed external parallel module can effectively reduce the parameters.In order to facilitate subjects to cross quickly,two-dimensional convolution is used instead of onedimensional convolution,and the experimental results show that the method has high accuracy and low loss.Secondly,aiming at the problem of few sample data of gearbox bearings,a data expansion method of“Three plus one” mode is proposed.This paper makes full use of the two methods of up-sampling and defining loss function to solve the data imbalance,combines the non-overlapping overlapping sampling with image enhancement,and further strengthens the data by using smote method to fine-tune the loss function.The experimental results show that this method can make up the defect of unbalanced data.In response to the lack of interpretability of convolutional neural network,a visual analysis of the learning results of convolutional neural network was carried out using gradients.The results of the gradient-weighted class-activation diagram are in good agreement with the theoretical analysis,which verifies the effectiveness of the learning position.The two-way proof of feature extraction using unsupervised learning and supervised learning proves the correctness of feature extraction.Finally,the 4DCEK method is proposed to further enhance the generalization of gearbox bearing fault diagnosis,and to further avoid the influence of small sample data,on the basis of improved channel attention,the mechanism of dual attention was put forward.Multi-group convolutional networks extract feature data of different scales,then fuse the traditional energy kurtosis index to improve the generalization.Experimental results show that this method has strong generalization performance.In order to solve the problem of abnormal diagnosis caused by different engineering problems by using the same bearing fault diagnosis model,the network structure of diagnosing gearbox bearing fault is studied in this paper.The experimental results show that the proposed method has certain engineering guidance significance.