Today, companies are faced with the challenge to develop and produce individualized products in the shortest possible time at very low cost in order to remain attractive under strong competitive pressure. For reasons of efficiency, products are therefore often developed in generations. Proven components are adopted in a new product generation and only some of the components are newly developed to meet new customer requirements. Many companies, therefore, have a large database of 3D CAD product models containing years of engineering experience. Nevertheless, it is often difficult to execute database queries to find which products or components already exist and could be reused or adapted for a new product generation or variant. As a result, many duplicates are created, which are associated with high effort and costs, and the risk of introducing design errors increases.
Therefore, the aim of this paper is to develop an automated approach for geometric similarity search that also takes company-specific features of components into account. Machine learning methods are capable of automatically extracting relevant geometric features by learning a suitable representation of the corresponding 3D object. ... mehrFor this purpose, an autoencoder is developed which is trained to extract class-specific feature vectors. To improve the representativeness of those vectors for the similarity search, the architecture and hyperparameters of the autoencoder are optimized based on several experiments. Considering a real use case with a data set from the field of mechanical engineering, it is shown that geometrically similar CAD models can be found very quickly using the learned representation, and that better results are obtained than with conventional methods based on meta information, e.g. volume and bounding box. On the one hand, the fast finding of similar models encourages the reuse of existing solutions. On the other hand, standardization and, thus, economy of scale is promoted.