Geometric deep learning as an enabler for data consistency and interoperability in manufacturing

Abstract

Skilled labor shortages and the growing trend for customized products are increasing the complexity of manufacturing systems. Automation is often proposed to address these challenges, but industries operating under the engineer-to-order, lot-size-one production model often face significant limitations due to the lack of relevant data. This study investigates an approach for the extraction of assembly-relevant information, using only vendor-independent STEP files, and the integration and validation of these information in an exemplary industrial use case. The study shows that different postprocessing approaches of the same segmentation mask can result in significant differences regarding the data quality. This approach improves data quality and facilitates data transferability to components not listed in leading ECAD databases, suggesting broader potential for generalization across different components and use cases. In addition, an end-to-end inference pipeline without proprietary formats ensures high data integrity while approximating the surface of the underlying topology, making it suitable for small and medium-sized companies with limited computing resources. Furthermore, the pipeline presented in this study achieves improved accuracies through enhanced post-segmentation calculation approaches that successfully overcome the typical domain gap between data detected solely on virtual models and their physical application. The study not only achieves the accuracy required for full automation, but also introduces the Spherical Boundary Score (SBS), a metric for evaluating the quality of assembly-relevant information and its application in real-world scenarios.