Research On Classification Method Of Surface Defects Of Bearing Raceway Based On Deep Learning

Bearings as supporting components are widely used in aerospace and machinery manufacturing in many heavy industries.The smoothness of the bearing raceway surface determines the vibration fluctuation and life of the bearing to a large extent.Therefore,in the manufacture of bearings,it is particularly important to study the classification of defects on the bearing raceway surface.This paper uses two batches of bearing track surface data for classification research,which comes from a well-known bearing manufacturer in my country.Compared with the mainstream image classification data set,the difference between the defect of the two batches of data used in this experiment is small,and it is difficult to accurately classify the defects of different classes.The first batch of bearing groove surface defect data sets have the characteristics of rich details but not prominent features.According to the traditional feature extraction methods for defect classification,there are disadvantages of modeling difficulties and low classification accuracy.This paper proposes an improved residual network(Residual Network,ResNet)to achieve high-precision classification of surface defects of bearing grooves.The convolutional neural network is used as the basic model architecture,the residual block is used as the main feature calculation method,and the Inception module is integrated into the deep network for feature dimensionality reduction and splicing to obtain more detailed image features.At the same time,batch normalization(BN)is introduced in the convolutional neural network(CNN)feature calculation to perform data regularization processing to accelerate the model convergence.Compared with the first batch of data,the second batch of data has more subtle differences between classes.Then have dense texture,and the differences are almost indistinguishable by the naked eye.In the feature calculation,the spatial structure is easily lost.In this paper,the network is further improved,and the High Resolution Net(HRNet)module is incorporated into the original residual network to parallel and interactively calculate feature maps of different resolutions to retain low-dimensional image information,and calculate the features between different dimensions.In comparison,images with subtle differences are easier to accurately classify.The method is tested on the two batches of bearing channel surface defect data sets,compared with the classic image classification model Lenet5 and ResNet,the results show that this method is exceedingly good and convergent on the high-precision classification problem of the bearing channel surface defect.Among them,the accuracy of the improving residual network can reach 98.84% in the first batch of data.It can reach 91.08% in the second batch of data,and the residual network integrated into the HRNET can reach 99.46% in the first batch of data.Batch data can reach 98.17%。At the same time,this paper develops a defect classification system to complete the pretreatment of real data,and the verification of the above model is implemented.