- 発表日 2023/07/10
- 20th Annual International CAD Conference（国際会議，Mexico City）
- DOIコード 10.14733/cadconfP.2023.254-259
Graph Neural Network-Based Finite Element Feature Recognition from B-rep Model
Finite element (FE) analysis requires mesh generation as a preprocess, which partitions a Boundary Representation (B-Rep) computer-aided design (CAD) model into FE meshes. Because mesh quality significantly affects analysis accuracy, manufacturers specify company-internal mesh generation rules for some types of FE features on CAD models, such as bosses and ribs, including free-form surfaces. However, at present, the recognition of the FE features from CAD models relies heavily on human eyes and hands, making it time-consuming, and error-prone. Therefore, a reliable, and versatile FE feature recognition method from CAD models is strongly required for efficient high-quality mesh generation. Recently, several deep-neural-network (DNN)-based methods have been proposed for feature recognition from CAD models. They have an advantage in that they do not require algorithm design specific to each feature type, unlike classical methods. However, most DNN-based methods approximate input CAD model geometries with voxels or point clouds as an input of DNNs, causing discretization loss in model resolution or an increase in the data size, which results in more significant memory consumption or longer training time. To resolve these issues, other DNN-based feature recognition methods have been proposed in recent years that use “graphs” as an input of the DNN, which take advantage of high compatibility with standard B-Rep CAD models. Nevertheless, those methods also have issues. First, the recognition significantly depends on the model poses. Second, the recognition method targeted the machining features or geometric modeling procedures and was not tested with FE features that included free-form surfaces. In this study, we propose an FE feature recognition method from a B-Rep CAD model using a graph neural network (GNN), which has a recognition ability invariant to model rotation or translation. The proposed method comprises a graph construction method with descriptors invariant to the translation or rotation, a neural network structure used for feature recognition, and data augmentation (DA) techniques for robust recognition. We tested our method with the original dataset of FE features, including bosses and ribs, and compared its performance with that of an existing method using GNN, similar to ours.
関連記事
