| In the strip rolling process,usually the tail of the front rolled strip is welded to the head of the back rolled strip to achieve continuous rolling process,if the welding quality is not qualified,it will cause weld cracking during the rolling process,and even malicious accidents such as broken strip,causing major property damage.Therefore,before proceeding to the next process,the strip weld quality must be checked.Currently,the common method used in industry is to do the cupping test on the weld seam,which is generally manually operated,labor-intensive and inefficient.This paper presents an automatic inspection system for cupping test of strip weld based on machine vision.The two key technologies of weld seam positioning and cupping test result detection were studied.Firstly,the convolutional layer and pooling layer of convolutional neural network are extended,and the working principles of two target detection algorithms based on anchor boxes and anchor-free boxes are discussed.Among them,the target detection algorithm based on anchor-free frame provides a design idea for the later study of weld positioning,and the depth-separable convolution provides a lightweight theoretical basis for the later study of weld cupping effect classification.Secondly,in view of the problem that angle information cannot be obtained when using traditional target detection for weld positioning,referring to the Center Net algorithm,the rotation target detection is realized by adding the regression of the rotation angle on the basis of the original network,and the data set is designed according to the input parameters of the network.In addition,to further improve the accuracy and robustness of the model,the standard convolution structure at downsampling is replaced with deformable convolution,and a pyramid segmentation attention module is introduced at upsampling.The experimental results show that the use of deformable convolution can improve the recall rate of the model by 3.00%,the introduction of the pyramid segmentation attention module alone can improve the accuracy rate by 1.01%,and the simultaneous use of the two modules can improve the F value of the original model by 3.02%.The stability experiments prove that the optimized network has better robustness.Thirdly,by analyzing the characteristic information of the collected weld cupping effect map,the samples were classified into qualified and unqualified samples according to the cracking status.Expand the dataset to some extent using mirroring,rotation,adjusting brightness and contrast,and adding Gaussian and pretzel noise.For the problem that the number of samples in the expanded dataset is still too small,a training idea based on migration learning is used.The experimental results show that only the classification part of the lightweight network can be retrained to obtain high accuracy,and the accuracy of all tested algorithms reaches more than 94%.Heat map visualization of the middle layer of the trained model by Grad-CAM algorithm verifies the authenticity of the model classification basis.Finally,a vision system was built to verify the above weld seam localization algorithm and the classification algorithm of the cup-and-bump test results,and part of the software design and debugging were completed,and the software operation results verified the feasibility of the detection algorithm. |
