Research On Bearing Fault Diagnosis Method Under Unbalanced Sample Condition

As a key component in the mechanical structure,rolling bearings have a direct impact on the operation of the entire mechanism.Real-time monitoring of the bearing status is particularly important.At this stage,there are still many intelligent diagnosis methods that need to construct feature sets through artificial prior knowledge,which cannot achieve "endto-end" intelligent fault diagnosis,and is mostly limited by the demand for a large number of balanced training samples.On the other hand,the lack of failure signals is an unavoidable problem in bearing fault diagnosis.Research often uses function simulation signals,laboratory simulation fault signals and other signals to replace it.However,in the actual production process,some faults are often difficult to obtain through simulation experiments,and the function simulation signal is quite different from the real vibration signal.At present,in the research of bearing fault diagnosis methods,the starting point is to improve the fault diagnosis model to obtain better diagnosis accuracy under the sample equilibrium scenario,and there are few methods to improve the model fault diagnosis ability under the sample imbalance scenario.With the development of integrated learning and deep learning technology,it is a new research challenge to build a rolling bearing fault diagnosis model suitable for sample imbalance scenarios.In response to the above problems,this paper is based on convolutional neural networks,combined with random forests,generative adversarial networks and other methods.According to different scenarios,from the perspectives of optimized models and amplified samples,the study of rolling bearing fault diagnosis under unbalanced samples is carried out.Optimized the model angle to build a CNNRF model,and further improved the diagnostic ability to build a Bagging-MCNN model;amplify the sample perspective to build a DCGAN model,and build an EEMD-DCWGAN model to enhance model stability.The main research results of the thesis are as follows:(1)A fault diagnosis method for rolling bearings based on CNN-RF is proposed.This method uses deep learning non-linear fitting capabilities to intelligently extract original sample features,reducing the dependence on manual and expert knowledge.Then build a random forest diagnosis model based on CART decision tree,which improves the sensitivity of the model to minority samples.Combining the advantages of convolutional neural networks and random forests makes the model more adaptable to scenarios with unbalanced samples.Finally,it is verified through examples.(2)A fault diagnosis method for rolling bearings based on Bagging-MCNN is proposed.This method uses the class reorganization method to increase the sampling probability of minority samples,and further increases the "emphasis" of the diagnostic model on minority samples.Then,in order to solve the problem of insufficient diagnosis ability in the CNN-RF rolling bearing fault diagnosis model,the Bagging method was introduced to construct a multichannel convolutional neural network model through an integrated method,which effectively improved the diagnosis accuracy of the model and the stability of the diagnosis results.And verify through examples.(3)A fault diagnosis method of rolling bearing based on DCGAN is proposed.From the perspective of amplified samples,this method further explores fault diagnosis methods under extreme sample imbalance scenarios.First,in order to eliminate the influence of the initial phase on the generation countermeasure network,the fast Fourier transform method is used to transform the bearing vibration signal into a frequency domain signal.At the same time,the convolutional neural network is used to construct the generator and discriminator,which improves the sample generation ability of the ordinary GAN model.Through the verification of examples,the model has stronger generalization for the unbalanced sample scenario.(4)A fault diagnosis method of rolling bearing based on EEMD-DCWGAN is proposed.This method uses EEMD to decompose the original vibration signal into multiple eigenmode functions,and constructs a correlation matrix to fuse the obtained eigenmode functions,so as to highlight the local characteristics of the original sample.Secondly,Wasserstein distance is introduced as a new loss function to improve the stability of the model generated samples.And verify through examples.