Research On Classification Method Of Bearing Surface Defect Detection Based On Deep Learning

Bearing is widely used in the industrial field,and its quality is particularly important,which can directly affect the service life of the whole mechanical equipment.However,after the bearing assembly is completed,various types of defects will inevitably appear,such as stains,pits and scratches.This thesis studies the detection and classification method of bearing surface defects based on deep learning.Aiming at the problems that the size of bearing surface defects is small and there is little useful feature information,a convolution neural network based on convolution attention mechanism is designed to detect and classify four kinds of defects on bearing surface,such as stains,pits,scratches and scratches.This method has high classification accuracy and low delay.The specific research contents are as follows:(1)Analyze the types and characteristics of bearing surface defects,complete the selection of industrial cameras,lenses and light sources,build a digital image acquisition platform,and pre-process the collected image sample data to prepare a perfect dataset.(2)The application of neural network transfer learning in the classification of bearing defect detection is studied.With the development of deep learning,various deep learning neural networks train and learn in the field of image recognition on a large number of image datasets,and can migrate the network model with superior performance on the source domain dataset to the target domain with sample data scarcity through the transfer learning technology.The Res Net network model and the Sense Net network model are studied,and the residual block structure of Res Net can effectively avoid the gradient disappearance or gradient explosion of the convolutional neural network due to the deep network level in the backpropagation process,so that the performance of the overall network model is more superior,the error is controlled in a very small range,and the classification accuracy is high;Dense Net is improved on this basis,and the full connection module is designed to strengthen feature transmission and pay attention to feature reuse.Use fewer calculations to achieve higher performance.Experimental results show that on the bearing surface defect test set,the classification accuracy of Res Net is 96.75%,and the classification accuracy of Dense Net is 95.83%.(3)In view of the problem of small bearing surface defect size and little useful feature information,a convolutional neural network is designed,and a convolutional attention module and batch normalization layer are added to its infrastructure,and the effects of various activation functions and optimizers on the performance of the network model are compared and experimented,and the improved network model performs deeper feature extraction for the data set,paying more attention to the location and significance of the feature information,so that the extracted features are more refined,the feature reuse is strengthened,and the expressiveness of the model is enhanced.The experimental results show that in the case of less characteristic information in the bearing surface defect dataset and a relatively single classification object,the method studied in this paper achieves better results for the classification task than the deeper network model,the defect detection classification accuracy rate can reach 98.7%,and the average detection time of a single picture is about 9.7 ms,which meets the real-time detection needs.