Scholarlink

<p>The quality of welding joints is critical to the reliability of metal connections, driving the need for advanced inspection technologies in the welding industry. Current methods often rely on manual feature extraction or shallow algorithms, limiting their real-time performance and ability to predict nugget size and strength accurately. To address these challenges, we propose TRANSInception, a novel deep learning algorithm that integrates Convolutional Neural Network (CNN) and multi-head self-attention mechanisms for end-to-end real-time quality inspection of spot welding. The multi-scale one-dimensional convolutional blocks extract detailed sequence features, while the multi-head self-attention mechanism captures long-term dependencies. Experimental results demonstrate the superiority of our approach: using dynamic resistance signals, the model achieves an average error of 0.219 mm and a relative root mean square error of 4.95&#xa0;% in nugget diameter prediction. For industrial deployment, the model is optimized using TensorRT for accelerated inference and integrated into a user interface system for real-time quality monitoring. The system is further deployed on both laptops and edge artificial intelligence devices, showcasing its versatility and efficiency in real-world applications. This work bridges the gap between advanced AI techniques and practical engineering solutions, providing a robust framework for real-time welding quality inspection in industrial 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

Spot welding quality detection system based on convolutional neural network and self-attention mechanism

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Rui Wang,
  • Ruichen Mi,
  • Zhonglin Gao,
  • Hao Xu,
  • Weipeng Liu,
  • Tao Liang

摘要

The quality of welding joints is critical to the reliability of metal connections, driving the need for advanced inspection technologies in the welding industry. Current methods often rely on manual feature extraction or shallow algorithms, limiting their real-time performance and ability to predict nugget size and strength accurately. To address these challenges, we propose TRANSInception, a novel deep learning algorithm that integrates Convolutional Neural Network (CNN) and multi-head self-attention mechanisms for end-to-end real-time quality inspection of spot welding. The multi-scale one-dimensional convolutional blocks extract detailed sequence features, while the multi-head self-attention mechanism captures long-term dependencies. Experimental results demonstrate the superiority of our approach: using dynamic resistance signals, the model achieves an average error of 0.219 mm and a relative root mean square error of 4.95 % in nugget diameter prediction. For industrial deployment, the model is optimized using TensorRT for accelerated inference and integrated into a user interface system for real-time quality monitoring. The system is further deployed on both laptops and edge artificial intelligence devices, showcasing its versatility and efficiency in real-world applications. This work bridges the gap between advanced AI techniques and practical engineering solutions, providing a robust framework for real-time welding quality inspection in industrial settings.