Machining Feature Recognition Method Based on Improved Mesh Neural Network

• Published in: Iranian journal of science and technology. Transactions of mechanical engineering 2023-12, Vol.47 (4), p.2045-2058
• Main Authors: Jia, Jia-Le, Zhang, Sheng-Wen, Cao, You-Ren, Qi, Xiao-Long, WeZhu
• Format: Article
• Language: English
• Subjects: Engineering; Mechanical Engineering; Research Paper
• ISSN: 2228-6187; 2364-1835
• DOI: 10.1007/s40997-023-00610-8

Machining feature recognition, which bridges computer-aided design (CAD) and computer-aided process planning (CAPP) systems, is the key technology to achieve the integration of CAD, CAPP, and computer-aided manufacturing (CAM). To overcome the low recognition efficiency and intelligence level of traditional CAPP systems and effectively recognize the machining features, this paper proposes a novel machining feature recognition method for mechanical parts based on Mesh-Faster RCNN, which combines the original MeshCNN model with the Faster RCNN model. Based on the improved MeshNet mesh recognition framework, a convolutional neural network is constructed for the automatic machining feature recognition of machining surface mesh data. By collecting the mesh data set of the machining feature surface in the CAD model, a triangulated mesh data sample library suitable for network learning is constructed, and the optimal neural network model is obtained by sample training. Then, the machining features of the parts are quickly and accurately obtained and converted into triangular mesh data based on model-based definition (MBD) technology. On this basis, combined with the triangular mesh data processing algorithm, the processed machining feature data are imported into the optimal neural network model to complete the feature recognition process. The results of experiments show that a recognizer of machining features can be obtained by the network training of the CNN and can automatically recognize 24 types of machining features with over 99.6% accuracy. Furthermore, this method requires a small amount of feature mesh data and has a simple data processing process. It also has good robustness and a good recognition effect for intersecting features.