Scholarlink

<p>Weld seam tracking in reflective metal welding environments remains a challenging task due to intense arc glare, spatter, and smoke, which severely degrade image quality. Conventional laser-based vision systems often fail under such conditions, leading to inaccurate seam localization. This study presents a lightweight denoising autoencoder (L-DAE) specifically designed to enhance weld seam visibility while enabling real-time deployment on embedded hardware. The proposed architecture incorporates adaptive spatial and channel attention modules to suppress localized noise patterns while preserving fine-grained laser line features. A dedicated dataset of 38,000 noisy–clean image pairs was collected from a Cartesian robotic welding platform, covering diverse part orientations and reflective galvanized steel surfaces. Experimental evaluations show that the L-DAE achieves 31.8 dB PSNR, 98.2% IoU, and 99.5% F1-score, outperforming baseline CNN and GAN models. The model delivers 30 FPS on an NVIDIA RTX 3060 GPU and maintains ± 0.2 mm tracking accuracy across varying orientations. On the NVIDIA Jetson NX, the network achieves real-time performance with only marginal accuracy degradation, confirming its suitability for industrial deployment where computing resources are limited. These results demonstrate that the proposed L-DAE offers a robust and computationally efficient solution for weld seam tracking in highly reflective environments. Future work will address cross-material generalization and integration with multi-modal sensing for enhanced stability in dynamic manufacturing settings.</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

Lightweight denoising autoencoder for real-time weld seam tracking on reflective metals

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Nurettin Gökhan Adar,
  • Ali Yildiz

摘要

Weld seam tracking in reflective metal welding environments remains a challenging task due to intense arc glare, spatter, and smoke, which severely degrade image quality. Conventional laser-based vision systems often fail under such conditions, leading to inaccurate seam localization. This study presents a lightweight denoising autoencoder (L-DAE) specifically designed to enhance weld seam visibility while enabling real-time deployment on embedded hardware. The proposed architecture incorporates adaptive spatial and channel attention modules to suppress localized noise patterns while preserving fine-grained laser line features. A dedicated dataset of 38,000 noisy–clean image pairs was collected from a Cartesian robotic welding platform, covering diverse part orientations and reflective galvanized steel surfaces. Experimental evaluations show that the L-DAE achieves 31.8 dB PSNR, 98.2% IoU, and 99.5% F1-score, outperforming baseline CNN and GAN models. The model delivers 30 FPS on an NVIDIA RTX 3060 GPU and maintains ± 0.2 mm tracking accuracy across varying orientations. On the NVIDIA Jetson NX, the network achieves real-time performance with only marginal accuracy degradation, confirming its suitability for industrial deployment where computing resources are limited. These results demonstrate that the proposed L-DAE offers a robust and computationally efficient solution for weld seam tracking in highly reflective environments. Future work will address cross-material generalization and integration with multi-modal sensing for enhanced stability in dynamic manufacturing settings.