Research On Key Technology Of Surface Defect Classification Of Bearing Outer Ring Based On Deep Learning

Bearing is an important component of rotating machinery.Due to the uncertainties in some production process,there may be some defects like scratches,chips etc.on the surface of its outer ring which shortens the service life of the machine and impacts the stability for the operation of the machine.Bearing will be inspected and classified before the delivery.The major method for the inspection is manual sampling inspection.This method is time and effort consuming.Besides,not all the parts are inspected.Recently,a visual technology is developed for the inspection by machines.Based on the advantages of fastness,stableness,accuracy,and nondestructive testing,it is applicable to the flow production line for the inspection on the bearing defects.By the integration of the vigorous development for deep learning technology and its outstanding performance on the classification of features,this article designs a device to collect images for the surface of outer ring of bearing and develops an algorithm for the classification of defects on the surface of bearing based on deep learning.The specific research content includes the following aspects:(1)The design of the device for image collection.Through the analysis on the structural features of bearings to be inspected and the requirement of inspection,complete the design of transmission belt,test platform,mechanical hand and the internal circuit of the bearing,complete the type selection for devices of camera,light source,analyze the characteristics of motion for the respective devices in accordance with the flow for the image collection of the device,design the control program for the movement of the device.(2)Based on the technology of deep learning and the theory of automatic encoder,design the automatic encoder of convolution for the improvement,complete the dimensionality reduction of sample data.Design the class network,complete the classification of defects in accordance with the characteristics after the dimensionality reduction of the automatic encoder.Optimize the performance of neural network model by the adjustment for the parameters of network weights and the hyperparameters during the training.Analyze the influence of parameters like dimensions of output,batch size and Dropout etc.of the automatic encoder on the performance of class network.Compared with the traditional method of dimensionality reduction,the dimensionality reduction effect of PCA(Principal Component Analysis)better helps the automatic encoder to effectively extract the features that are helpful to improve the accuracy of classification;compared with the classification effect of traditional classification method such as SVM(Support Vector Machine),the classification accuracy of the classifier designed in this article is higher with stronger generalization ability.(3)Based on the theory of transfer learning,design a class network which extracts the defects of bearing in the network of InceptionV3 based on the classic features.Conduct a research on the applicability of transfer learning on the classification of bearing defects in the condition of small samples.It is discovered in the experiment that the network can effectively extract the features of defects and obtain a good accuracy for classification.(4)Based on the comparison for the class network of automatic encoder and the class network based on transfer learning on the consumption of hardware resources and the time required for classification,the class network of automatic encoder is more advantageous and more suitable for the complicated environment of a factory under the requirements of real time classification for the system.