3D CAD model dynamic clustering based on inertial feature encoder

Abstract

The number of three-dimensional (3D) computer-aided design (CAD) models of mechanical parts in cyber manufacturing has experienced explosive growth. Classified CAD model shape knowledge based on induction is conducive to model retrieval, design reuse, and machining reuse. However, 3D CAD feature extraction primarily utilizes projected views, point clouds, voxels, and meshes for dimensionality reduction. Nonetheless, complex processes and high computational costs impede effective shape analysis. Traditional distance measures in data spaces or shallow linear embedded spaces are susceptible to errors when assessing similarity in data clusters. Furthermore, as the size of the database increases, data distribution may change in dynamic clustering, leading to data drift. This paper proposes an automatic unsupervised learning shape classification method based on deep embedding for 3D mechanical part CAD models. First, an inertial feature descriptor that effectively represents shape characteristics was established to extract the multidimensional moment of inertia of the 3D CAD model. Second, the inertial feature data space was nonlinearly mapped to a low-dimensional feature space, and the clustering accuracy was improved through the joint training of the encoder and clustering layers. Simultaneously, we revealed the influence of eps and min samples of Density-based Spatial Clustering of Applications with Noise (DBSCAN) algorithm on the clustering distribution of the CAD models. Third, adding new data can effectively achieve dynamic clustering based on the original clustering results. This paper explains the potential problems of fuzzy clustering boundaries that may arise from adding new data. Experimental data showed that the silhouette coefficient calculated by the proposed method is 0.78, and the normalized mutual information is 0.82, which has an excellent automatic classification effect.