Scholarlink

<p>This study presents a machine learning approach to predict the slag eye area in gas-stirred steelmaking ladles. The data from physical modeling experiments and real industrial measurements to improve prediction accuracy and reliability are combined, and several regression models were tested, with Artificial Neural Networks (ANNs) showing the best performance. A single-layer ANN with around 300 nodes achieved the highest accuracy, with an <i>R</i><sup>2</sup> value close to 0.97. SHAP (Shapley Additive Explanations) analysis was used to understand how different features influence slag eye formation. The study revealed substantial differences between the physical modeling and industrial data. Ladle diameter, molten bath height, <i>etc</i>. were the most influential features in predicting the slag eye size. By combining the data from physical models and industrial trials, the ANN model avoided the limitations of generalizing encountered when relying only on physical modeling results. A fivefold cross-validation strategy confirmed the robustness of the ANN model tuned using Bayesian optimization across different test samples. This study highlights the advantages of blending physical modeling and industrial data in making the predictive models robust and illustrates the potential of machine learning in forecasting slag eye behavior and optimizing process efficiency in secondary steelmaking.</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

A Robust Machine Learning Framework for Predicting Slag Eye Formation in Industrial Steelmaking Ladles

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Somenath Mukherjee,
  • M. A. Sankar,
  • Vishnu Teja Mantripragada

摘要

This study presents a machine learning approach to predict the slag eye area in gas-stirred steelmaking ladles. The data from physical modeling experiments and real industrial measurements to improve prediction accuracy and reliability are combined, and several regression models were tested, with Artificial Neural Networks (ANNs) showing the best performance. A single-layer ANN with around 300 nodes achieved the highest accuracy, with an R2 value close to 0.97. SHAP (Shapley Additive Explanations) analysis was used to understand how different features influence slag eye formation. The study revealed substantial differences between the physical modeling and industrial data. Ladle diameter, molten bath height, etc. were the most influential features in predicting the slag eye size. By combining the data from physical models and industrial trials, the ANN model avoided the limitations of generalizing encountered when relying only on physical modeling results. A fivefold cross-validation strategy confirmed the robustness of the ANN model tuned using Bayesian optimization across different test samples. This study highlights the advantages of blending physical modeling and industrial data in making the predictive models robust and illustrates the potential of machine learning in forecasting slag eye behavior and optimizing process efficiency in secondary steelmaking.