知识与数据融合驱动的无取向硅钢力学性能预报Stacking集成模型研究

电工钢 ›› 2025, Vol. 7 ›› Issue (5): 63-.

知识与数据融合驱动的无取向硅钢力学性能预报Stacking集成模型研究

黄望芽^1,2,康　琦²,黄若麟³

1.宝钢股份有限公司硅钢事业部,上海 201999;
2.同济大学电子与信息工程学院,上海 200092; 3.新疆大学计算机科学与技术学院,新疆乌鲁木齐 830046

出版日期:2025-10-28 发布日期:2025-10-21

Knowledge⁃data fusion driven stacking ensemble model for mechanical property prediction of cold⁃rolled non⁃oriented electrical steel strip

HUANG Wangya¹^,², KANG Qi², HUANG Ruolin³

1.Silicon Steel Division,Baoshan Iron&Steel Co.,Ltd.,Shanghai 201999,China;
2.College of Electronics and Information Engineering,Tongji University,Shanghai 200092,China; 3.School of Computer Science and Technology,Xinjiang University,Urumqi 830046,China

Online:2025-10-28 Published:2025-10-21

摘要/Abstract

摘要： 以新能源汽车驱动电机用无取向硅钢为研究对象,针对冷轧带钢力学性能预测难题,创新性地提出“HALL⁃PETCH方程+数据挖掘”的双驱动建模策略。通过整合材料学机理与工业大数据分析,系统构建了高精度的力学性能预测模型。研究首先从300余项生产特征(涵盖化学成分、工艺参数及磁性能数据)中,通过BOOTSTRAP随机森林法进行预测变量筛选,采用了JMP软件进行主成分筛选,主成分可解释度累计达到80.5 %;在主成分筛选的基础上,采用随机森林法、KNN、神经网络和逐步回归等多种机器学习算法进行建模对比分析。实验结果表明,单一模型中随机森林法表现最优(R²=0.987),而通过STACKING集成策略(以随机森林和KNN为基模型,GBDT为元模型)构建的融合模型预测性能显著提升,最终模型的R²达到0.998,RMSE为1.986。该模型已成功部署于制造执行系统(MES),实现了基于在线铁损数据的力学性能实时预测,为生产工艺动态调整提供了可靠依据。

关键词: 力学性能预测, 机器学习, Stacking集成模型, 特征工程

Abstract: This study develops a hybrid modeling approach combining the Hall⁃Petch equation with data mining techniques to predict mechanical properties of non⁃oriented electrical steel for new energy vehicle motors. By integrating materials science principles with industrial big data analytics, a high⁃precision prediction system was established. From over 300 production parameters (including chemical composition, process variables, and magnetic properties), key features were selected via bootstrap random forest, followed by principal component analysis (80.5 % cumulative variance) using JMP software. Comparative analysis of multiple algorithms (random forest, KNN, neural networks, and stepwise regression) revealed random forest as the best standalone model (R²=0.987). The Stacking ensemble model, combining random forest and KNN as base learners with GBDT meta⁃learner, achieved superior performance (R²=0.998, RMSE=1.986). Deployed in manufacturing execution systems (MES), this model enables real⁃time mechanical property prediction using online iron loss data, supporting dynamic process optimization.

Key words: mechanical property prediction, machine learning, stacking ensemble model, feature engineering

黄望芽, 康　琦, 黄若麟.

知识与数据融合驱动的无取向硅钢力学性能预报Stacking集成模型研究 [J]. 电工钢, 2025, 7(5): 63-.

HUANG Wangya, KANG Qi, HUANG Ruolin.

Knowledge⁃data fusion driven stacking ensemble model for mechanical property prediction of cold⁃rolled non⁃oriented electrical steel strip [J]. Electrical Steel, 2025, 7(5): 63-.