Research On Classification Method Of Bearing Quality Grade Based On Imbalanced Data Augmentation

Rolling bearings are important basic components in modern mechanical equipment.The quality of bearings will directly affect the performance,running accuracy,life and security of mechanical moving parts.Therefore,it is necessary to test the quality of the bearing before it is put into use.The vibration signal analysis method is often used for bearing quality inspection.However,in the manufacturing process of bearings,most of the samples are high-quality grade samples,and only a few samples are low-quality grade samples due to component processing errors or assembly errors resulting in reduced manufacturing accuracy.The imbalanced distribution of bearing samples with different quality grades will make it difficult for traditional classification models to learn strong robust generalization characteristics from a few samples,and make the prediction results tend to favor the majority samples,thus affecting the robustness and generalization of the classification model.Based on this,a classification method of bearing quality grade based on imbalanced data augmentation is proposed.Aiming at the problem that the lack of low-precision grade samples(minority class samples)when the bearing quality grade classification model is constructed leads to the low prediction generalization ability of the model for bearing samples of different precision grades.Random oversampling(ROS),synthetic minority oversampling technique(SMOTE),borderline synthetic minority oversampling technique(Borderline-SMOTE),adaptive synthetic sampling(ADASYN)and integrated sampling technique(SMOTE Tomek)were used to enhance the minority-class samples respectively.The support vector machine(SVM)model based on whale optimization algorithm(WOA)optimization was established.The experimental results show that the model established by the SMOTE Tomek method combined with WOA-SVM was the best.Compared with the model established by the original imbalanced data,the F-measure of the test set is increased from 0.595 to 0.933,and the accuracy was increased from 0.714 to 0.933.The SMOTE Tomek method effectively improves the prediction ability of the bearing quality grade classification model and enhances the robustness of the model.In order to solve the modeling problem of data imbalance from the level of classification algorithm,Ada Boost,Bagging,Random Forest(RF)and SMOTE Tomek-Bagging fusion methods were used to establish the bearing quality classification model.The fusion method of SMOTE Tomek-Bagging was to enhance the minority-class samples through the SMOTE Tomek method at the sampling level,combined with the repeated downsampling of the Bagging algorithm,so that the model can be fully trained,thereby improving the model recognition ability.The experimental results show that only from the classification algorithm level,the model results have over-fitting phenomenon,which cannot effectively solve the problem caused by imbalanced data,the fusion model of SMOTE Tomek-Bagging can solve the model over-fitting problem well.And the test set F-measure and accuracy of the model both reached 0.983.Aiming at the problem of overfitting in imbalanced data modeling.A data enhancement method was proposed to convert one-dimensional vibration signals into two-dimensional vibration images and adaptively generate high-quality virtual samples through a deep convolutional generative adversarial network(DCGAN)method,combined with convolutional neural network(CNN)to achieve the extraction and modeling of deep-level features of samples.The experimental results show that compared the DCGAN-CNN model with the CNN model established by the original data,the F value of the test set was improved from 0.595 to 0.914,and the accuracy rate was improved from 0.714 to 0.917.Therefore,the DCGAN-CNN model further improved its generalization and robustness under the premise of ensuring the classification accuracy,and provided a new modeling strategy for the analysis of imbalanced data.