Scholarlink

<p>As blast furnace smelting processes become increasingly complex, traditional anomaly detection methods relying solely on physical variables have shown significant limitations. At the same time, newly commissioned blast furnaces generally suffer from insufficient historical data, making it difficult to directly deploy stable and reliable anomaly detection systems. To address these issues, this paper proposes a bilinear collaborative transfer learning (BCTL) method for blast furnace anomaly detection. This method employs a bilinear mapping mechanism, overcoming the shortcomings of traditional vector-based feature transfer. Subspace matching is performed directly in the two-dimensional matrix space, preserving key structural information of the image to the greatest extent. Within the collaborative learning framework, physical variables and image data are fused in a multimodal manner to achieve feature alignment and enhance anomaly recognition accuracy. In addition, the model introduces a cross-domain key sample weighting strategy, which effectively improves its adaptability to new operating conditions and new equipment domains. Experimental results show that BCTL has excellent subspace matching and cross-domain migration capabilities. The detection rate for uneven pulverized coal injection reached 94.15%. The detection rate for localized overheating faults in the furnace body reached 97.00%. The experimental results fully validated the practical value and application prospects of BCTL in the diagnosis of anomalies in industrial blast furnaces.</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

Blast furnace multimodal anomaly diagnosis method based on bilinear collaborative transfer

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Xin Sha,
  • Xinye Song,
  • Rong Sun,
  • Zheng Wang,
  • Shengxuan Wang

摘要

As blast furnace smelting processes become increasingly complex, traditional anomaly detection methods relying solely on physical variables have shown significant limitations. At the same time, newly commissioned blast furnaces generally suffer from insufficient historical data, making it difficult to directly deploy stable and reliable anomaly detection systems. To address these issues, this paper proposes a bilinear collaborative transfer learning (BCTL) method for blast furnace anomaly detection. This method employs a bilinear mapping mechanism, overcoming the shortcomings of traditional vector-based feature transfer. Subspace matching is performed directly in the two-dimensional matrix space, preserving key structural information of the image to the greatest extent. Within the collaborative learning framework, physical variables and image data are fused in a multimodal manner to achieve feature alignment and enhance anomaly recognition accuracy. In addition, the model introduces a cross-domain key sample weighting strategy, which effectively improves its adaptability to new operating conditions and new equipment domains. Experimental results show that BCTL has excellent subspace matching and cross-domain migration capabilities. The detection rate for uneven pulverized coal injection reached 94.15%. The detection rate for localized overheating faults in the furnace body reached 97.00%. The experimental results fully validated the practical value and application prospects of BCTL in the diagnosis of anomalies in industrial blast furnaces.