Design Of Visual Inspection System For Surface Defects Of Large Marine Engine Forgings Based On Deep Learning

With the continuous development of intelligent manufacturing,quality inspection standard are getting higher and higher.Due to the large size and weight of large marine engine forgings,Large uneven temperature changes and large difference in plastic fluidity of metals,easier to form defects,defect detection on its internal and surface was an important prerequisite to ensure the safe and reliable operation of diesel engines.Research on the surface defect detection of large forgings.Due to the various shapes and sizes of the surface defects of the forgings,and the random positions of the defects.Rely on manual visual inspection to detect,low degree of stability,easy to cause missed detection and false detection.can not to meet the automatic detection requirements of high speed,high precision and "100% detection".Use marine engine connecting rod as experiment and test object,aiming at the shortcomings of traditional manual inspection methods.A visual inspection system for surface defects based on the optimized YOLOv4 method was proposed.The binocular vision and deep learning technology were applied to the defect inspection system of forgings,which can realize the tasks of defect detection,classification and depth measurement.Through adopting two industrial cameras to collect images of forging surface defects at different locations,and the optimized model was used to analyze and detect the image,so as to obtain the information of the defect category.Calibrate the binocular camera to establish the mapping relationship from the two-dimensional pixel coordinate system to the three-dimensional space coordinate system,Realize the depth information extraction of pit defects.The main research work is as follows:1.By comparing the mainstream target detection algorithms in single stage and two stages,yolov4 algorithm with better detection accuracy and speed was selected.According to the defect detection requirements,the feature extraction network of yprov4 CSP darknet53 was optimized,and the CBAM attention module was inserted.At the same time,k-means++ cluster optimization was used to optimize the prior frame selection strategy to generate a priori frame suitable for surface defect detection.The experimental results show that the optimized model improves the accuracy of defect detection,reduces the rate of false detection and missed inspection,and the target detection frame is closer to the real defect.2.The binocular vision model and measurement principle were studied.The calibration method of Zhang Zhengyou camera is used to calibrate the binocular camera,and the internal and external parameters of the camera are obtained.The calibration experiment is completed by Opencv program,and the calibration results are compared with MATLAB.Then,the pit area identified after defect detection was compared with the results of different feature points extraction algorithm,and finally,surf algorithm was used to extract the feature,and the polar constraint was added to stereo matching.Then,according to the matching point pairs,the depth information was extracted,and the depth difference of matching points was calculated to get the depth of the pit.The comparison between the manual measurement results and the analysis of the causes of the errors were made.3.According to the defect detection process and requirements,the system hardware selection was carried out,and the defect visual detection system was built.Considering the industrial application,the visual human-computer interaction interface was developed,and the functions of image acquisition,camera calibration,defect detection and depth measurement were designed and realized.