Research And Development Of Abnormal Detection Algorithm In Electronic Information Product Manufacturing

With the widespread use of surface mount technology in the production of electronic information products,the density of components on the PCB(printed circuit board)is getting higher and higher,and the size of the components is getting smaller and smaller.The production line is equipped with AOI(Automatic Optical Inspection)equipment.Carry out timely quality inspections on products.Problems such as easy omissions reduce the efficiency of the production line.Here,machine learning technology based on deep learning is used to develop a variety of component-type bad recognition algorithms to replace the current manual re-judgment to improve the accuracy of the re-judgment and the efficiency of the production line.This thesis develops an automatic recognition algorithm for AOI review.First,use feature visualization technology to classify images,train multi-classifiers based on deep networks for bad detection? then study bad recognition algorithms based on template matching and propose a fast clustering algorithm that automatically generates templates?finally,use twin networks to learn A better similarity measure is used for bad identifica-tion.The specific contributions of this article are as follows:1.Developed a bad detection algorithm based on deep classification network.First,use the t SNE-based dimensionality reduction algorithm to visualize image features to realize the classification of images,and then train a multi-class deep network for bad detection.2.Propose a fast bottom-up hierarchical clustering algorithm for template generation.The bottom-up clustering algorithm has the shortcoming of slow execution speed.This paper develops a fast bottom-up clustering algorithm based on the idea of divide and conquer,which greatly improves the efficiency of template generation.3.Developed a bad detection algorithm based on template matching.The image segmentation algorithm is used to cut out the component area from the image to match the template image,and compare a variety of image features and similarity measures.4.Developed a bad detection algorithm based on metric learning.The twin network is used to learn the similarity between images.The experimental results show that the twin network can better describe the similarity between images than traditional methods.