Abstract: Objective To address the issues of severe point cloud noise interference and insufficient feature extraction accuracy in the automated welding of circular plug welds, this study aims to develop a high-precision robotic welding system guided by 3D point cloud data. A composite point cloud processing workflow was proposed by integrating normal filtering, Euclidean clustering, and an improved RANSAC algorithm. Methods This method was utilized to segment and fit high-noise point clouds for extracting the 3D geometric features of circular hole welds, guiding the robot to accomplish automated welding trajectory planning through a hand-eye vision system. Results Experimental results indicate that under complex working conditions with noise exceeding 20%, the center positioning error is ≤0.05 mm and the radius measurement error is ≤0.084 mm. Compared to the standard RANSAC algorithm, the positioning accuracy is improved by 84.6%, with the entire processing workflow completed within 545 ms. The system’s positioning repeatability reaches ±0.1 mm, increasing welding efficiency by approximately 30% compared to manual operations, and achieving uniform weld bead formation on both flat plates and circular ring workpieces. Conclusion The proposed system balances exceptional algorithmic robustness and industrial real-time performance, providing an efficient and reliable automated solution for the precise feature extraction and automated welding of circular hole welds in high-noise environments.