Scholarlink

<p>Selective laser melting (SLM) is widely used to fabricate high-precision metal components, yet microdefects formed along individual laser scan lines can degrade mechanical reliability and drive costly rework and scrap. We address this engineering problem with MobileViT-SLM, a lightweight CNN–Transformer hybrid explicitly designed for edge-deployable quality monitoring. Using optical microscopy, 22 high resolution images were segmented into two datasets totaling 9,284 scan line samples, reflecting the small-data conditions typical of production settings. The proposed model couples local convolutional features with global self-attention to achieve robust binary classification of scan-line quality. During 5-fold cross-validation, MobileViT-SLM achieves up to 99.12% accuracy and AUC = 0.994 while maintaining a compact footprint ( 1.9&#xa0;M parameters). Deployment tests demonstrate real-time inference on embedded hardware (2.4 ms/sample on Jetson Nano) with &lt; 2% false-alarm rate of 2% and stable operation over &gt; 8&#xa0;h, validating readiness for on-machine use. SHAP-based explanations expose the image regions driving decisions, supporting operator trust and process traceability. By enabling early fine-grained detection and closed-loop parameter feedback during printing, MobileViT-SLM functions as a drop-in module for SLM quality control that can shorten inspection cycles and reduce downstream rework in metal additive manufacturing.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MobileViT-SLM: real-time edge-deployable CNN–transformer hybrid for fine-grained scan line defect classification in additive manufacturing

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Nhu-Quynh Tran,
  • Hoa-Cuc Nguyen,
  • Bich-Ngoc Mach,
  • Nghi N. Nguyen,
  • Thanh Q. Nguyen

摘要

Selective laser melting (SLM) is widely used to fabricate high-precision metal components, yet microdefects formed along individual laser scan lines can degrade mechanical reliability and drive costly rework and scrap. We address this engineering problem with MobileViT-SLM, a lightweight CNN–Transformer hybrid explicitly designed for edge-deployable quality monitoring. Using optical microscopy, 22 high resolution images were segmented into two datasets totaling 9,284 scan line samples, reflecting the small-data conditions typical of production settings. The proposed model couples local convolutional features with global self-attention to achieve robust binary classification of scan-line quality. During 5-fold cross-validation, MobileViT-SLM achieves up to 99.12% accuracy and AUC = 0.994 while maintaining a compact footprint ( 1.9 M parameters). Deployment tests demonstrate real-time inference on embedded hardware (2.4 ms/sample on Jetson Nano) with < 2% false-alarm rate of 2% and stable operation over > 8 h, validating readiness for on-machine use. SHAP-based explanations expose the image regions driving decisions, supporting operator trust and process traceability. By enabling early fine-grained detection and closed-loop parameter feedback during printing, MobileViT-SLM functions as a drop-in module for SLM quality control that can shorten inspection cycles and reduce downstream rework in metal additive manufacturing.

Graphical Abstract