Node Classification in CAD Data: A Graph Learning Approach for Machining Feature Localization

Pattern recognition is becoming increasingly important in many sectors, such as the manufacturing industry. With rising customer demands and advancing technology, product data, often represented as 3D meshes, is becoming highly complex. Therefore, intelligent approaches are needed to recognize and localize features in 3D meshes to determine complex manufacturing processes. However, previous approaches solely work on voxel-, pixel-or. STEP-based formats and not directly on mesh-based formats such as STL data. To overcome these issues, we present a configurable data generator and customize graph neural network architectures for the problem of localizing machining features directly in 3D meshes. After comparing the graph neural networks against each other, the best resulting network is compared to an existing state-of-The-Art approach, i.e., SsdNet. Our results indicate that the best-performing graph neural network exhibits superior classification accuracies on meshes with few machining features compared to SsdNet with an average F1 score of 80%. While the accuracy slightly falls behind SsdNet with more features, the graph neural network shows substantially better runtime performance results, which are, on average, 70 times faster.