Deep Learning Defect Detection System Of SMT Solder Joint Based On Artificial Intelligence

As one of the most important steps in the Surface Mounted Technology(SMT)process for electronics packaging,to a large extent soldering determines the performance of electronic products.The quality of solder joints directly affects the output of the entire process flow.In recent years,some semi-automatic solder joint defect detection systems have used machine vision methods to replace human eyes,which to a certain extent solve the issue of unreliable human visions that are greatly affected by environmental and subjective factors,however the whole inspection process of those detection systems cannot be visualized and fully automated due to the deep level of defective solder joints cannot be extracted features,therefore it has been unable to meet the increasing density and accuracy requirements of the SMT industry year by year.The Printed Circuit Board(PCB)manufacturing industry in our country has gradually expanded in scale,ranking first in the world in terms of total volume,and at the same time showing a sustained growth trend of prosperity.However,there is still a big gap between our country and the most developed countries regarding the advanced manufacturing of PCB.Based on the research status of domestic and foreign,this thesis proposes a real-time SMT solder joint defect detection system based on deep learning for the current problems of low precision,slow speed and high cost solder joint defect detection technology.The detection system is based on computer vision technology and for the first time adopts the YOLOv3 algorithm for single-stage target detections in solder joint defect detection system.After building and training the network model using the Paddle Paddle deep learning framework,the collected image information of the target board can be sent to the solidified model for real time prediction.In addition,on the basis of optimized selection of electronic components and detailed analysis of software function,a stage motion control system and a human-computer interaction system are included in the system for image processing,completing a cost-effective,high-precision real-time solder joint defect detection system.After 1000 iterations of training,the loss function of the model dropped from the initial value of 10352 to 8.26,and there was no further significant decline.Using dummy data set,the m AP value reached 81.97%,and the processing speed was 200ms/frame.Experiments show that the model can fit the data well and can correctly detect most of the defective solder joints.