Abstract: In robot arc welding of multi-variety and small-batch steel structural components, due to the low efficiency of manual teaching and offline programming, the image processing method is susceptible to the interference of the environment and the surface material of the workpiece to be welded, and the deep learning methods are difficult to create datasets for training of multi-variety and small-batch workpieces, which usually requires a large amount of computing costs. This paper proposes an automatic recognition method that integrates the prior knowledge of CAD models and three-dimensional point clouds. Based on the Open CASCADE platform, a double-constraint contact surface discrimination mechanism is proposed. Furthermore, a systematic process of "Boolean intersection-boundary traversal-parametric characterization" is designed to automatically extract and quantify the weld intersection line parameters from the assembly. Finally, innovatively, the extracted welds are mapped to the actual workpiece point cloud through point cloud registration to complete the identification and extraction of welds. The experimental results show that this method can effectively overcome the interference of unstructured environments and accurately and stably identify common weld seams, which can be referred to in the automated welding of small-batch customized workpieces.