Research On Bearing Fault Diagnosis And Remaining Use Life Prediction Method Based On CNN And GAN

Rolling bearings play a particularly important role in industrial equipment,and are widely used in industrial devices such as generators,gearboxes,fans,gas turbines,etc.It was called the "joints" of industrial machinery systems.Generally speaking,the bearing working environment is relatively harsh,coupled with the need for long-term operation and heavy operating load,which promotes the bearing equipment to easily produce various types of failures during the service process and may threaten the safety of life and property in severe cases.It has important practical significance to monitor and maintain the state of bearing operation and service.However,in the actual production process,it is not easy to obtain the bearing fault state data sample,which will cause a serious imbalance of the sample data set.So it will reduce the accuracy of the corresponding diagnostic model trained and cause problems such as overfitting or under-fitting.In addition,for the prediction of the remaining service life of the bearing,on the one hand,it also has a large amount of service degradation data that is difficult to obtain.On the other hand,in the bearing life prediction regression model,the predicted degradation curve and the actual degradation curve are not well-fitted.The prediction model has poor robustness.In response to the above problems,this article mainly studies the following:(1)A balanced data set method based on Wasserstein Generative Adversarial Network(WGAN)and a diagnosis model based on Global Average Pooling Convolutional Neural Network(GAPCNN)are studied.This method first uses WGAN for adversarial training with a small amount of fault sample data to generates a large number of fault samples of this type.Then mixes the generated samples into the original few samples to achieve the balance of various fault samples.Next,the balanced data set is input into the diagnosis model,and the correct classification or diagnosis of various faults is realized through layer-by-layer adaptive extraction of features.The experimental results show that the proposed diagnosis method is superior to other algorithm models,and has strong generalization ability and robustness.(2)A method for predicting the remaining life of bearings based on AdCNN and a method for data enhancement based on CWGAN are studied.The prediction method realizes the remaining life prediction of the bearing by inputting the monitoring data into the one-dimensional convolutional neural network for confrontation training.The data enhancement method enhances the bearing degradation data set by introducing the Conditional Generative Adversarial Network(CGAN)into the Wasserstein distancebased generative adversarial network and solves the problem of less bearing degradation data.The experiment surface,AdCNN is more accurate in predicting the remaining life of the bearing.No matter what kind of prediction test index(MAE,MSE,RMSE,MAPE,SMAPE),AdCNN has less than other prediction methods.The data generated by CWGAN has the same characterization ability as the real data,but when the generated data accounts for more than 60%,its prediction accuracy will decrease.However,in this case,the prediction accuracy of this method is still higher than other methods.