Abstract: Aiming at the challenge in decomposing three-dimensional geometric models of computer numerical control (CNC) machine tools into basic geometric primitives, a primitive feature recognition algorithm is proposed. Firstly, a supervised convolutional neural networks (CNN) is employed to identify the geometric primitive category of each triangular face. Simultaneously, a multi-scale feature fusion-based surface division algorithm is employed to segment the model's outer surface into four feature types, namely cuboid planes, cylindrical cap planes, cylindrical surfaces, and spherical surfaces. Subsequently, a hybrid correction algorithm is introduced to determine feature surface categories by integrating divided surfaces and their constituent triangles' classifications, correct triangular face categories based on surface features, merge the feature surfaces accordingly, and effectively update the initial results from both the recognition network and the division algorithm. Experimental results demonstrate that the algorithm achieves 98.97% average accuracy on basic primitives and over 90% on practical machine models.