Augmented geometry assurance digital twin with physics-based incremental learning

Abstract

This paper presents a novel digital twin framework employing batch incremental learning for geometry assurance. Addressing quality issues caused by part and process variation, the method evaluates three critical tasks: part matching, locator adjustments, and joining sequence. The proposed framework utilizes deep learning architectures, each trained on recursive simulation data. Employing incremental learning, the models adapt to new batch characteristics while maintaining predictive accuracy. A spot welded assembly demonstrated the proposed approach efficiency, achieving prediction accuracies with errors as low as 0.02 mm for part matching and 0.1 mm for locator adjustments.