Scholarlink

<p>In basic oxygen furnace (BOF) steelmaking process, the splashing may affect the stability of operation and threaten production safety. An innovative deep network with dual-attention mechanism (AM) and transfer learning is firstly proposed to realize automatic recognition of splash degree. Due to the insufficient amount and non-equilibrium distribution of splash images, the pre-trained residual network is used as a preliminary feature extractor in combination with transfer learning to reduce the image data required for network training. The dual AM is designed to extract the features in parallel on the spatial and channel dimensions, which further improves the ability of the network to capture important features. Besides, Focal Loss function is applied to assign a higher weight to the insufficient splash categories, which is used to solve the problem of non-equilibrium distribution of image samples. Finally, compared with the existing networks and ablation study tests, the advantages and applicability of the proposed network are verified by using the images of a true BOF steelmaking process. Experimental results show that recognition accuracy, precision rate, recall rate, and F1-Score of the proposed network reach 95.54, 93.53, 95.54, and 94.52&#xa0;pct, respectively. Accurate splash recognition results provide a guarantee for the design of intelligent control strategies to avoid splashing.</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

Deep-Transfer-Learning Network for Recognizing Splash of BOF Steelmaking Process with Non-equilibrium Samples

  • 期刊论文

  • 发表日期：

    2025年3月31日

查看全文
  • Shijian Dong,
  • Jun Liu,
  • Xiaoqing Jiang,
  • Xingtong Li,
  • Wei Gong,
  • Huimin Xu

摘要

In basic oxygen furnace (BOF) steelmaking process, the splashing may affect the stability of operation and threaten production safety. An innovative deep network with dual-attention mechanism (AM) and transfer learning is firstly proposed to realize automatic recognition of splash degree. Due to the insufficient amount and non-equilibrium distribution of splash images, the pre-trained residual network is used as a preliminary feature extractor in combination with transfer learning to reduce the image data required for network training. The dual AM is designed to extract the features in parallel on the spatial and channel dimensions, which further improves the ability of the network to capture important features. Besides, Focal Loss function is applied to assign a higher weight to the insufficient splash categories, which is used to solve the problem of non-equilibrium distribution of image samples. Finally, compared with the existing networks and ablation study tests, the advantages and applicability of the proposed network are verified by using the images of a true BOF steelmaking process. Experimental results show that recognition accuracy, precision rate, recall rate, and F1-Score of the proposed network reach 95.54, 93.53, 95.54, and 94.52 pct, respectively. Accurate splash recognition results provide a guarantee for the design of intelligent control strategies to avoid splashing.