<p>The accuracy of the recognition on weld seam between impeller hub and fan blade are still low due to the texture characteristic of welded workpieces and lighting, thus the recognition on weld seam based on point clouds needs to be improved. Besides, the exiting recognition methods are generally performed based on the point clouds of weld seam of flat surface, which are not suitable for the curved fan blade, therefore, we propose an automatic recognition method on the weld seam to improve the accuracy. Firstly, the original point clouds are denoised and compressed by Gaussian filter and voxel filter, respectively, and the point cloud model is divided into two types using region growth method. Next, the shape of the welding surface was recognized by principal components analysis (PCA) and normal distribution theory, and the position of the weld seam was obtained by least square method, furthermore, the welding path was determined by graph algorithm and bubble sorting method. The welding experimental results showed that the weld seam between the fan blade with different shapes and impeller hub was recognized successfully, and the offset error between the weld seam position obtained using our algorithm and the actual position was less than 1 mm, which improved accuracy of the recognition on weld seam. Besides, the weld seam meets the requirements of the welding technology based on the weld seam inspection.</p>