A novel initial weld position identification method for large complex components based on improved YOLOv5

Abstract

The initial weld position of large complex components has the characteristics of irregular shape and random spatial distribution. Traditional welding robots have technical bottlenecks such as low teaching programming efficiency and difficulty in offline programming in this application scenario. Therefore, at present, the method of "workers piling up" is used to weld large complex components. To overcome these problems, it is necessary and urgent to develop the intelligent initial weld position identification method for large components based on machine vision. Firstly, based on the theoretical background of YOLOv5 algorithm, an initial weld position recognition model for large complex components is established. Secondly, the model is optimized by changing the up-sampling method and fusing the Transformer Self-Attention mechanism. Finally, through the initial weld position detection experiment on the coco dataset, the experimental results show that the detection effect of this model is better than others and the original YOLOv5 model, and the detection speed of a single weld image is 11ms, and the average detection accuracy reaches 92.4%. Under the premise of ensuring the detection speed, the proposed method greatly improves the precision of initial weld position detection, and has better interference immunity and robustness.