Research And Application Of Solder Joint Defect Detection Method

The quality of the welded joints of the steel plate directly affects the reliability of the steel plate,while the traditional manual inspection method takes a long time to inspect,which affects the production capacity of the steel plate to a certain extent.In recent years,with the rapid development of machine learning,good results have been achieved in image processing,target detection and other research fields,and have been widely used.At the same time,the continuous reduction of computing costs has also made it possible for inspection methods based on machine vision to land in industrial production scenarios.Therefore,the introduction of machine visionbased methods into steel plate solder joint quality inspection has great industrial application value and research significance.This paper focuses on the quality defects of surface solder joints in the production and processing of steel plates.First,a comparative test of classic algorithms is carried out,and then an improved algorithm is proposed.Finally,a set of steel plate solder joint defects detection system based on machine vision is designed and implemented,details as follows:(1)Using classic welding joint defect detection algorithms based on traditional machine learning and deep learning to conduct comparative experiments on steel plate welding joints.Experimental results show that both types of algorithms have high accuracy rates,and the former has better classification capabilities,but the feature extraction method based on image processing has become a bottleneck for improving the accuracy of the overall model.The latter has stronger feature extraction capabilities,but its average F1-Score is slightly lower than algorithms based on traditional machine learning.(2)An improved algorithm is proposed,which effectively combines the powerful feature extraction capabilities of deep learning algorithms with the accurate classification capabilities of traditional machine learning algorithms.The overall model is divided into two steps for training,firstly training the feature extraction ability of the neural network,and secondly training the classification ability of the traditional machine learning classifier.Continuously optimize the accuracy of the overall model in the iterative training process of the model.Experimental results show that the improved algorithm has achieved significant improvements in all indicators,with a recall rate of 97%,which is of great significance for further industrial application research.(3)Designed and implemented a machine vision-based solder joint defect detection system suitable for actual industrial production scenarios.According to the overall process of solder joint defect detection,the system is modularized and implemented,including: window cutting module,solder joint positioning module,image preprocessing module,data enhancement module,and defect detection module.It has achieved the function of automatic,fast,accurate and efficient detection of welding spot defects of steel plate.