Research On Visual Detection Algorithms For Rolling Bearing Rings Lack Of Material And Cage Riveting Defect

Rolling bearings are commonly used parts of rotating machinery,and their manufacturing defects will seriously affect subsequent assembly and service performance.At present,many small and medium-sized enterprises in the manufacturing of deep groove ball bearings assume that the components(inner and outer rings,rolling elements and cage)are intact,and for the quality of the bearing ring and cage after the sleeve is riveted,only manual visual inspection is used.This inspection method is likely to cause missed inspections for defects generated in the assembly process of the bearing.Machine vision is widely used in industrial fields as a non-contact detection method.Based on this,the detection algorithm research is carried out for the lack of material of rolling bearing ring and the defect of cage riveting.The specific research methods are as follows:The bearing image online collection device is designed.The device includes an automatic feeding mechanism,an image acquisition mechanism,an automatic turning and an automatic unloading mechanism.The device can realize the process of loading,image collection,and unloading of the bearing.The setting of the automatic turning mechanism enables the device to collect images of the upper and lower surfaces of the bearing at the same station.Aiming at the problem that the collected bearing image contains the flipped cylinder,an algorithm for bearing positioning and background separation is designed.First,perform filtering operations on the collected images,use peak signal-to-noise ratio(PSNR)and mean square error(MSE)to evaluate the effects of classic filtering algorithms,and select bilateral filtering methods based on the evaluation results.Then,the morphological reconstruction technology is used to remove the bearing surface details,and only the contour information of the image is retained.Edge detection is performed on the morphologically reconstructed image.The edge image is detected by the Hough transform to obtain the radius and center of the bearing outer ring.Finally,the image of the bearing is separated from the background image of the acquisition device.Aiming at the detection problem of lack of material in bearing rings,an online recognition method of manufacturing defects based on connected domain detection and convolutional neural network(CNN)is proposed.For the connected domain detection method,the bearing normalized expansion algorithm is first used to expand the bearing from the annular area to the rectangular area.The inner and outer ring areas of the bearing are determined by the horizontal projection method,and the connected areas are detected in this area to remove the small-area connected areas to avoid noise interference.This method can realize the accurate marking of the lack of material defect area.This method is used to identify the bearing samples collected in this article,and the accuracy rate is 83%.For the CNN detection method,the bearing shortage defect area is small and unevenly distributed,and the image feature extraction is difficult.CNN can automatically learn the sample feature.Use image processing technology to expand the sample,build a CNN-based bearing lack defect discrimination model,perform 300 epochs of training on the model,and finally the model has an accuracy of 100% on the training set and 97% on the test set.For rolling bearing cage riveting defects,a defect detection algorithm based on the AdaBoost algorithm is proposed,and the bearing cage region positioning and segmentation algorithm is designed;the orientation gradient histogram(HOG)feature and Gabor wavelet feature of the cage region image are extracted as feature matrix.For the problem of significant data imbalance,the Synthetic Minority Oversampling Technology(SMOTE)algorithm is used to balance the sample size.In order to reduce the dimension of the feature matrix,the Principal Component Analysis(PCA)method is used to reduce the dimension of the feature matrix.The AdaBoost integration strategy uses k-nearest neighbors,Naive Bayes method,and linear discriminant analysis as weak classifiers for ensemble learning of the two types of feature matrices,and compares the detection results of the two types of features combined with AdaBoost.The results show that the accuracy rate of HOG-AdaBoost is 95.83%,the detection effect of Gabor-AdaBoost model is better,and the accuracy rate is as high as 100%,and the Precision,Recall,Specificity and F-value evaluation indicators are better than those of the HOG-AdaBoost model.