Research Of Weld Feature Extraction Methods Based On Deep Learning

As a key part of weld automatic tracking system,weld feature extraction is an important guarantee of welding quality and efficiency.In the process of welding,due to the interference of strong noise such as arc,smoke,splash and so on,the traditional single or specific geometric feature extraction algorithms can not fully adapt to the complex welding environment.In order to improve the self adaptability and anti-interference ability of weld feature extraction,it is necessary to study the intelligent weld feature extraction algorithm of multi-level feature extraction with extensive learning ability.Therefore,based on the analysis of the characteristics of the weld image and the current methods of weld feature extraction,this paper combines the deep learning technology with the extraction process of the weld feature,designs different extraction schemes of the weld feature from the two angles,extracting the fringe feature of the weld and directly obtaining the key points position of the weld groove,and verified the effectiveness of the proposed method through experiments.The research content of this paper includes:(1)The characteristics of the weld image are analyzed in depth,and the related principles of the convolution neural network are studied,and its strong feature extraction ability and feature expression ability are used to design the weld stripe feature segmentation scheme and the weld feature point extraction scheme based on the deep learning.(2)In order to improve the anti-interference ability of laser stripe segmentation of weld,the end-to-end pixel level classification of welding image is established by fully convolutional networks,so as to segment the laser stripe and interference background of welding.At the same time,the high-level and low-level feature fusion is used to supplement the edge detail information and improve the segmentation accuracy of welding stripe.The experimental results show that the method based on the fully convolutional networks can achieve the accurate segmentation of the laser stripe features of the weld,and compared with the gray gravity algorithm and Steger algorithm,the method in this paper can fuse the features of the image of weld at multiple levels,and has deep learning ability and noise suppression ability.(3)In order to improve the field welding efficiency,aiming at the situation that only the key position of weld groove needs to be extracted from the disturbed weld image,a weld feature points extraction network based on multi-layer convolution is designed.By introducing prior boxes,the weld image is transferred from the global to the local to improve the detection accuracy.The experimental results show that the network model designed in this paper has high positioning accuracy,the average error is 0.238mm,the maximum error is 0.702mm,and the storage size of network parameters is only 2.63MB,which can basically meet the requirements of automatic weld tracking system,and has good anti-interference ability and generalization ability.