Research And Implementation Of Defect Detection Method For Surface Of Varistor Based On Deep Learning

With the proposal of "made in China 2025",there are higher and higher requirements for the quality of industrial products in China,and the detection of product surface defects is an important part to ensure the quality of products.The varistor is a kind of resistance device,which plays an important role in lightning protection industry.And its surface defect detection methods can be divided into two categories:manual detection method and automatic detection method.The manual detection method has the disadvantages of low efficiency,poor reliability and high production cost,the automatic detection method can be realized by traditional machine learning and deep learning technology:the traditional machine learning based method has poor generalization ability and low accuracy,the deep learning based method is prone to overfitting problems when the training data set is small.In view of the above problems,this paper carries out the following research:(1)Construction of varistor surface images data set.In order to ensure the quality of the surface images data set of the varistor,this paper designs and builds the hardware environment of the surface defect detection device of the varistor for images acquisition.In order to collect automatically stable and clear surface images of varistor,the two-frame difference method is adopted to detect static target and improve collection efficiency.The images processing technology is used to automatically extract the effective information of thesurface images of the varistor,and then basic data is obtained.In order to obtain enough datas,the basic data set is expanded,and a total of 3350 varistor surface images are acquired,which provided sufficient and high-quality images data set for the surface defect detection model.(2)The model of varistor surface defect detection based on convolutional neural network was constructed.In order to solve the problem of poor generalization ability caused by manually defining data features in traditional machine learning methods.In this paper,the convolutional neural network is used to establish a varistor surface defect detection model.The specific method is that using the convolutional layers in the convolutional neural network automatically extract features from thel images data,remove redundant features by pooling layers,and then fuse effective features by full connection layers.Finally,the surface images of the varistor are classified,and then the defects of the varistor surface are judged.The experimental results show that the accuracy of the constructed model can reach 94%on the test set,the AUC value is 0.94,and the detection speed is 34 images per second.(3)An enhanced varistor images dataset model based on generative adversarial neural network was constructed.In order to further improve the accuracy,the AUC value of the surface defect detection model and the generalization ability of the model,this paper uses a generative adversarial neural network to enhance the varistor images data set.The specific method is that the convolutional layers of the generator and the discriminator of the generative anti-neural network learn the channel features of the varistor surface images,and then randomly combines the channel features to generate images similar to the varistor surface images structure distribution.Add the enhanced image datas to the data set,and train the convolutional neural network model again.The experimental results show that the surface defect detection model improves the accuracy by 4.3%on the test set,and the AUC value is increased by at least 2%.(4)An experimental system for detecting surface defects of varistor was designed and implemented.This paper designs the surface defect detection experimental system of varistor from the aspects of hardware system and software system,and realizes the function of images acquisition and defect detection.Experimental test results show that the system has good practicability and the accuracy rate can reach 96%at least,which meets the requirements of industrial production.