Scholarlink

<p>Precisely controlling the end-point carbon content of liquid steel is significant for interstitial-free (IF) steel in the RH refining. The models for predicting the end-point carbon content in RH refining based on machine learning algorithms, SHAP interpretation framework, and metallurgical mechanism were established in the present study. The Noise-Based Data Augmentation method was used to expand the samples to 1000 sets of data for training the machine learning models. Randomly divided the 1000 samples in an 8:2 ratio, with 800 groups in the training set and 200 groups in the test set. The dimension of influencing factors was decreased based on the Principal Component Analysis method to mitigate the over-fitting risk in training machine learning models. The prediction performance of five models established, respectively, by Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), Random Forest, Linear Regression, and Support Vector Machine algorithms was compared. The Data Augmentation-GBDT (DA-GBDT) model achieves the optimal performance, as evidenced by the determination coefficient of 0.995 in the test set, mean square error (MSE) of 0.0021, and root mean square error (RMSE) of 0.0028. The DA-GBDT model is significantly superior to other four models in terms of hit rate within the error interval (0, 0.005&#xa0;pct). Among the twelve influencing factors, the decarburization time, initial carbon content, and the time to reach the ultimate vacuum are the key factors among the influencing factors in the order of decarburization of IF liquid steel on the basis of the SHAP interpretation framework and metallurgical mechanism analysis.</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

Prediction of End-Point Carbon Content of IF Steel in RH Refining Based on Machine Learning and Shapley Additive Explanations

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Yifan Meng,
  • Chengbin Shi,
  • Zhiguang Li,
  • Huai Zhang

摘要

Precisely controlling the end-point carbon content of liquid steel is significant for interstitial-free (IF) steel in the RH refining. The models for predicting the end-point carbon content in RH refining based on machine learning algorithms, SHAP interpretation framework, and metallurgical mechanism were established in the present study. The Noise-Based Data Augmentation method was used to expand the samples to 1000 sets of data for training the machine learning models. Randomly divided the 1000 samples in an 8:2 ratio, with 800 groups in the training set and 200 groups in the test set. The dimension of influencing factors was decreased based on the Principal Component Analysis method to mitigate the over-fitting risk in training machine learning models. The prediction performance of five models established, respectively, by Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), Random Forest, Linear Regression, and Support Vector Machine algorithms was compared. The Data Augmentation-GBDT (DA-GBDT) model achieves the optimal performance, as evidenced by the determination coefficient of 0.995 in the test set, mean square error (MSE) of 0.0021, and root mean square error (RMSE) of 0.0028. The DA-GBDT model is significantly superior to other four models in terms of hit rate within the error interval (0, 0.005 pct). Among the twelve influencing factors, the decarburization time, initial carbon content, and the time to reach the ultimate vacuum are the key factors among the influencing factors in the order of decarburization of IF liquid steel on the basis of the SHAP interpretation framework and metallurgical mechanism analysis.