Exploring deep learning algorithms for enhanced yield in manufacturing industry: data augmentation, knowledge distillation, and novel category discovery

Abstract

Due to various constraints, research on the application and operation of deep learning methodologies in the manufacturing industry has remained limited compared to other fields. Consequently, we believe that deep learning can significantly contribute to productivity and yield improvements in this sector. To address these challenges, we have explored cost-effective image-based data augmentation methods, self-knowledge distillation for efficient learning, and unsupervised continuous learning for discovering new categories in environments where vast amounts of unlabeled data are generated. These methods have been validated on both the various benchmark datasets and the datasets collected from actual industrial settings, demonstrating performance improvements compared to existing methodologies, respectively. We expect that ongoing research tailored to the characteristics of the manufacturing industry will continue, and we hope this paper will be a valuable contribution to such endeavors.