Scholarlink

The accuracy and reliability of welded components are essential in industrial manufacturing, where ensuring high-quality welds is essential for maintaining production standards. This study investigates the application of advanced object detection models-YOLO12, RT-DETR, and Faster R-CNN to classify friction stir welding (FSW) results of aluminum alloys such as AA5083 and AA6061 as successful, semi-successful, or unsuccessful. We evaluated these models using both original and preprocessed images, incorporating several techniques to improve image quality. The results were analyzed to determine the impact of preprocessing on detection accuracy. Our findings showed that preprocessing significantly improved the performance of the models. Also, we evaluated the inference times to identify the best-performing model for this application, balancing speed and accuracy. The results highlight the potential of these models to improve welding quality inspection, providing a foundation for further improvement and potential real-time application in industrial environments.

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

Comparative Performance of Convolutional Neural Networks and Vision Transformers for Quality Assurance of a Welding Process

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Paula Arcano-Bea,
  • Agustín García-Fischer,
  • Manuel Rubiños,
  • Pablo Fariñas,
  • Francisco Zayas-Gato,
  • José Luis Calvo-Rolle

摘要

The accuracy and reliability of welded components are essential in industrial manufacturing, where ensuring high-quality welds is essential for maintaining production standards. This study investigates the application of advanced object detection models-YOLO12, RT-DETR, and Faster R-CNN to classify friction stir welding (FSW) results of aluminum alloys such as AA5083 and AA6061 as successful, semi-successful, or unsuccessful. We evaluated these models using both original and preprocessed images, incorporating several techniques to improve image quality. The results were analyzed to determine the impact of preprocessing on detection accuracy. Our findings showed that preprocessing significantly improved the performance of the models. Also, we evaluated the inference times to identify the best-performing model for this application, balancing speed and accuracy. The results highlight the potential of these models to improve welding quality inspection, providing a foundation for further improvement and potential real-time application in industrial environments.