LLM-Inspired Vision Transformer Framework for Intelligent Quality Recognition in Aerosol Jet Printing

作者：Haining Zhang, Joon Phil Choi, Xingchen Liu, Nannan Liang, Yangwen Yu, Yongrae Kim, Dongwoon Shin, Seung Ki Moon, Yong-Jin Yoon · 发表于：Engineering · 年份：2026 · DOI：10.1016/j.eng.2026.03.021 · 被引用次数：1 · 研究领域：Aerosol Filtration and Electrostatic Precipitation、Computer Graphics and Visualization Techniques、Fire Detection and Safety Systems

Aerosol jet printing (AJP) is an advanced direct-ink writing technique within additive manufacturing (AM) that enables the high-resolution fabrication of micro-scale electronic components through the controlled deposition of aerosolized functional inks. Despite its potential, AJP faces significant challenges in maintaining consistent printing quality due to dynamic process variability and material heterogeneity. Existing approaches are often insufficient to address these challenges. In particular, traditional data-driven models relying on numerical inputs inherently omit critical visual morphology features, while conventional supervised vision methods struggle to generalize under practical variability. To address this, we propose an intelligent quality recognition framework that adapts large language model (LLM)-inspired Transformer architectures to manufacturing vision tasks. Specifically, Vision Transformers (ViT), which translate the self-attention mechanisms of LLMs into computer vision, are employed to achieve robust and generalizable defect detection in AJP. The proposed approach leverages a dataset composition of approximately 46 000 unlabeled images for self-supervised pre-training to enhance feature representation, followed by fine-tuning on 9369 labeled images generated through randomized printing experiments. Afterwards, comparative studies demonstrate the effectiveness of the proposed system, achieving classification accuracy of 98.1% with high robustness. By leve...