DyLoFViT: A Novel Approach for Real‐Time Metal 3D Printing Surface Quality Classification

ABSTRACT Real‐time surface quality monitoring of selective laser melting (SLM) is critical for defect identification and process stability but is still plagued by the trade‐off between high precision and efficiency. We propose DyLoFViT, a light‐weight vision transformer specifically optimized for industrial SLM defect classification. DyLoFViT makes four significant architectural contributions: (1) dynamic positional encoding that retains spatial coherence and facilitates translation invariance, (2) dynamic sparse gating that adaptively filters the most informative tokens, (3) Nyström‐based sparse self‐attention that estimates long‐range dependencies with lower computation, and (4) frequency‐domain compression with discrete cosine transform that filters high‐frequency noise with preservation of low‐frequency structure vital for defect classification. Experiments on an industrial test set of 1320 high‐resolution images of four melt quality classes show that DyLoFViT achieves a Top‐1 accuracy of 96.1% and an F1‐score of 95.6% with merely 13.2M parameters and 2.9 GFLOPs. Compared with the prior CNN and transformer architectures, DyLoFViT achieves superior precision with an order of magnitude lower inference latency (0.3 ms/image), making it extremely well‐suited for real‐time, edge‐device deployment in additive manufacturing. DyLoFViT hence sets a new standard for efficient, accurate in situ metal 3D printing defect classification.