Scholarlink

<p>During the continuous casting production of 82B steel, a significant number of surface crack defects are observed in the billets. Early prediction of such defects facilitates timely adjustments of continuous casting process parameters. For the collected dataset comprising 315 sets of 160&#xa0;×&#xa0;160 billet samples, preprocessing is performed using box plots, the empirical mode decomposition (EMD) algorithm, and probability density distribution methods, resulting in the selection of 172 sets of data. Key factors influencing billet surface crack defects are identified through the Pearson correlation coefficient and the random forest model. The reliability and stability of different deep learning models in predicting surface crack defects are evaluated using metrics such as precision, confusion matrix, and receiver operating characteristic (ROC) curves. When the training-to-test set ratio is set at 8:2, the gated recurrent unit (GRU) model achieves a surface crack defect prediction accuracy of 97.12&#xa0;pct, providing valuable guidance for on-site continuous casting production.</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

Real-Time Surface Crack Defects Prediction of 82B Steel Billet Based on Multi-model Fusion

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Baorong Wang,
  • Yongkun Yang,
  • Zhiheng Yu,
  • Weian Wang,
  • Silong Zhang,
  • Zhibin Geng,
  • Shuan Wang,
  • Xiaoming Li

摘要

During the continuous casting production of 82B steel, a significant number of surface crack defects are observed in the billets. Early prediction of such defects facilitates timely adjustments of continuous casting process parameters. For the collected dataset comprising 315 sets of 160 × 160 billet samples, preprocessing is performed using box plots, the empirical mode decomposition (EMD) algorithm, and probability density distribution methods, resulting in the selection of 172 sets of data. Key factors influencing billet surface crack defects are identified through the Pearson correlation coefficient and the random forest model. The reliability and stability of different deep learning models in predicting surface crack defects are evaluated using metrics such as precision, confusion matrix, and receiver operating characteristic (ROC) curves. When the training-to-test set ratio is set at 8:2, the gated recurrent unit (GRU) model achieves a surface crack defect prediction accuracy of 97.12 pct, providing valuable guidance for on-site continuous casting production.