Selection of manufacturing processes using graph neural networks
Marco Hussong, Patrick Ruediger-Flore, Matthias Klar, Marius Kloft, Jan C. Aurich
Abstract
The increasing complexity of modern manufacturing, driven by trends such as product customization and shorter product life cycles, presents significant challenges in process planning. Traditional methods for selecting manufacturing processes in industry rely on expert knowledge and manual intervention, which can be time-consuming and error-prone. Systems that can automate the selection of manufacturing processes become increasingly important. Current approaches for the selection of manufacturing processes focus on deep learning that convert the 3D CAD models to intermediate representations such as voxels, point clouds or dexels. However, this transformation can result in the loss of topological, geometrical, or Product and Manufacturing Information (PMI). To address these challenges, this paper proposes a neural network architecture MaProNet. MaProNet is a graph attention neural network (GAT) designed to capture topological and geometrical information through the analysis of Attributed Adjacency Graphs (AAG) and Mesh structures. MaProNet also incorporates a wide range of PMI information. • The paper proposes a neural network architecture MaProNet analyzing AAG, Meshes and PMI to predict manufacturing processes. • The paper extends the prediction of manufacturing processes to a semantic segmentation task on graphs. • The paper describes the creation of a new synthetical dataset for manufacturing process selection.