Learning localized features in 3D CAD models for manufacturability analysis of drilled holes

We present a novel feature identification framework to recognize difficult-to-manufacture drilled holes in a complex CAD geometry using deep learning. Deep learning algorithms have been successfully used in object recognition, video analytics, image segmentation, etc. Specifically, 3D Convolutional Neural Networks (3D-CNNs) have been used for object recognition from 3D voxel data based on the external shape of an object. On the other hand, manufacturability of a component depends on local features more than the external shape. Learning these local features from a boundary representation (B-Rep) CAD model is challenging due to lack of volumetric information. In this paper, we learn local features from a voxelized representation of a CAD model and classify its manufacturability. Further, to enable effective learning of localized features, we augment the voxel data with surface normals of the object boundary. We train a 3D-CNN with this augmented data to identify local features and classify the manufacturability. However, this classification does not provide information about the source of non-manufacturability in a complex component. Therefore, we have developed a 3D-CNN based gradient-weighted class activation mapping (3D-GradCAM) method that can provide visual explanations of the local geometric features of interest within an object. Using 3D-GradCAM, our framework can identify difficult-to-manufacture features, which allows a designer to modify the component based on its manufacturability and thus improve the design process. We extend this framework to identify difficult-to-manufacture features in a realistic CAD model with multiple drilled holes, which can ultimately enable development of a real-time manufacturability decision support system.