Welding seam detection and location: Deep learning network-based approach. (April 2023)

• Record Type: Journal Article
• Authors: Wang, Jianyong; Mu, Chunyang; Mu, Song; Zhu, Rui; Yu, Hua
• Is Part Of: International journal of pressure vessels and piping. Volume 202(2023)
• Publication Date: 2023-04
• Subjects: Welding seam defect -- Deep learning network -- Detection -- Faster R–CNN -- FPN; Pressure vessels -- Periodicals; Pipe -- Periodicals; Récipients sous pression -- Périodiques; Tuyaux -- Périodiques; Pipe; Pressure vessels; Periodicals; 681.76041
• DOI: 10.1016/j.ijpvp.2023.104893
• Languages: English
• ISSNs: 0308-0161

Abstract

Abstract: There are many kinds of surface defects, large difference in size and strong randomness in the process of material welding, which makes it difficult for some conventional target detection algorithms to quickly and accurately locate and identify. This paper works out an improved Faster R–CNN network model. By adding FPN pyramid structure, variable convolution network and background suppression function to the traditional Faster R–CNN network, the path between the lower layer and the upper layer is greatly shortened, and the location information is better preserved. The simulation test shows that the model above improves the detection accuracy of the network, shortens the detection time, and reduces the loss value. The mAP is increased by 0.243, and the detection time of a single picture is shortened by 0.019s. The detection accuracy of five weld defects is significantly improved, and the accuracy of weld deviation is improved by 0.234, which is the one that improved the most significantly. It shows that the improved model is effective and has good convergence, and can meet the requirements of the detection of weld defects in the actual production process. Highlights: This paper works out an improved Faster R-CNN network model is proposed to locate and identify the surface defects of weld rapidly and accurately. The accuracy has been significantly improved, and has good convergence. The improved model can meet the detection of weld defects in the actual production process.