Research On Defect Detection Method Of PCB Plug-In Solder Joints Based On CNN

Printed Circuit Board(PCB)is the core component of intelligent electronic products,and defect detection is the basic requirement of quality control in the production process.With the rapid implementation of "China’s Intelligent Manufacturing 2025",my country’s manufacturing industry has begun to transform to intelligence.As a key technology for intelligent production,machine vision has been widely used in workpiece size measurement,robot identification and sorting,and defect identification.Therefore,it is of great significance and value to study and realize the detection method of PCB plug-in solder joint defect based on machine vision.This paper studies the problem that the existing PCB solder joint defect detection algorithm cannot meet the high-precision detection,low false alarm rate and high-speed detection at the same time.The YOLOv3 detection algorithm that can quickly detect is adopted.The research contents of this paper are as follows:First of all,this paper considers the optimization of the prior frame of the original network,the backbone network and detection network,and the loss function.Using the method of ordered probability density weighting,the highest priority inspection frame is obtained;using the attention network to optimize the original,while eliminating redundant parameters in the network,the network is more focused on the detection of solder joint defects,which greatly reduces the In the case of network missed detection and error detection;the generalized Intersection over Union(GIo U)is used to solve the problem of non-learning under extreme conditions of the model,which speeds up the model convergence and improves the detection accuracy.Secondly,this paper considers further optimization on the original feature pyramid structure to improve the performance of dense small object detection.In the original pyramid upsampling process,more upper-layer semantic information is explored and fused through multiple atrous convolutions to reduce the semantic difference between feature layers;channel attention is used to filter out important feature information channels and eliminate redundant parameters.Finally,the system is applied to the actual industry,and the results show that the developed deep learning-based PCB plug-in defect solder joint detection system can achieve a correct rate of 97.42%,and the detection speed fully meets the production speed requirements of the factory’s assembly line.It has certain reference and application value for the detection of solder joint defects in industrial assembly lines.