Research Progress of Weld Tracking Image Processing Technology Based on Deep Learning Theory

Abstract

In this paper, a convolutional neural network is used to localize the weld seam feature points with noise interference in complex welding environments. A priori frames are introduced into the feature point extraction network, combined with position prediction and confidence prediction, to improve the accuracy and anti-interference ability of the weld tracking system. To improve welding efficiency by utilizing the continuity of weld tracking, the weld tracking network is designed based on the twin structure. The weld detection network designs the first frame to locate the key position of the bevel and inputs into the weld tracking network as a template, and the weld tracking network completes the automatic tracking of the subsequent welds. At the same time, the network introduces a hybrid domain attention mechanism, which makes full use of the weld feature channel dependence and spatial location relationship and puts more attention near the inflection point of the weld laser line to ensure the accuracy of weld tracking. The research results show that the extraction error of weld seam feature points based on the convolutional neural network is within 17, which is much lower than that of the grayscale center of gravity method and Steger's algorithm. In the weld tracking experiments under the workpiece tilting state, the average value of the absolute error of the tracking trajectory in the X-axis direction is not more than 0.7 mm, and the maximum value is less than 1.15 mm. The absolute tracking error in the Z-axis direction is relatively low, with an average of 0.638 mm and a maximum of 1.573 mm. Therefore, the weld-tracking image processing technique proposed in this paper has strong anti-noise interference capabilities and high localization accuracy. And high accuracy in localization.