Research on weld start point detection and localization using deep learning techniques

Abstract

In industrial automated welding, precise positioning of the weld start point is crucial for ensuring weld stability and quality. To address this challenge, this paper proposes an improved weld start point recognition algorithm based on the Real-Time Detection Transformer (RT-DETR), enabling the welding robot to adaptively recognize and initially position the weld start point. Images of the weld start point are captured using a vision system based on the D435i camera, and a dataset is constructed for deep learning training. The algorithm is based on the RT-DETR model, utilizing a Residual Network architecture with 18 layers (RT-DETR-R18), integrating the Context Guided Block (CGBlock) from the Context Guided Network (CGNet) with the BasicBlock of Residual Network 18 (ResNet18) to form a new architecture, Context Guided Residual Network 18 (CG-ResNet18), enhancing the model's global perception and accuracy. Additionally, a Large Separable Kernel Attention (LSKA) mechanism is incorporated into the backbone network to improve applicability and robustness. The hybrid encoder of RT-DETR is further optimized using Group Separable Convolution (GSConv) and Efficient Cross Stage Partial Network Block (VoV-GSCSP) modules, enhancing feature fusion and reducing computational load. Experimental results show that the improved model achieves a mean Average Precision (mAP) of 99.4 %, a 5 % increase from the original, with a 32 % reduction in parameters and a 31 % reduction in computational load. The detection error of the weld start point is below 0.4 mm, meeting the real-time detection and positioning requirements of the welding robot.