Penetration identification of magnetic controlled Keyhole Tungsten inert gas horizontal welding based on OCR-SVM

Abstract

Keyhole Tungsten inert gas (K-TIG) welding can realize single-sided welding and double-sided forming. However, due to the influence of gravity, undercuts always occur in K-TIG horizontal welding. In order to expand the application scenarios of K-TIG and achieve automatic welding, a magnetic controlled K-TIG horizontal automatic welding system is proposed in this paper. A longitudinal magnetic field is used to weaken the influence of gravity and improve welding quality. The OCR (Object-Contextual Representations)-SVM (support vector machines) model is proposed to identify the welding penetration states during K-TIG horizontal welding, whose accuracy rate is 93%. In order to solve the problem of slow convergence and poor learning of difficult-to-learn classes, a loss function called Unified Focal loss was used, which achieves a mIoU (the mean of Intersection over Union) score of 91.48%.