基于Stacking集成学习的热轧带钢凸度诊断模型  

Crown diagnosis model of hot-rolled strip crown based on Stacking ensemble learning

作　　者：张殿华[1] 李贺 武文腾 霍光帆 孙杰[1] 彭文[1] ZHANG Dianhua;LI He;WU Wenteng;HUO Guangfan;SUN Jie;PENG Wen(State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,China;Hot Rolling Department,Shougang Co.Ltd.,Beijing 100041,China)

机构地区：[1]东北大学轧制技术及连轧自动化国家重点实验室,辽宁沈阳110819 [2]北京首钢股份有限公司热轧作业部,北京100041

出　　处：《中南大学学报（自然科学版）》2024年第10期3673-3682,共10页Journal of Central South University:Science and Technology

基　　金：中国五矿科技专项(2022ZXB03);国家自然科学基金资助项目(U21A20117);辽宁省人工智能重大科技专项(2023JH26-10100002)。

摘　　要：在热连轧生产过程中,凸度是重要的质量指标,过程数据的非平衡性限制了数据驱动模型的预测效果,为提高模型的预测精度,提出一种融合SMOTE和Stacking集成算法的热轧带钢凸度诊断模型。首先,采用SMOTE过采样方法处理凸度相关数据集,降低数据非平衡分布导致的影响;然后,构建以轻量级梯度提升机(LightGBM)、支持向量机(SVM)、K近邻(KNN)和随机森林(RF)为基学习器,逻辑回归(LR)为元学习器的Stacking集成模型,最后,使用某2160 mm热轧带钢实际生产数据进行模型验证。研究结果表明,诊断模型的准确率、少数类召回率、平衡F分数、几何平均值和ROC曲线下面积分别为0.9580、0.9595、0.9573、0.9589和0.9579,与XGBoost、LightGBM、KNN、SVM和随机森林模型对比,预测效果最优,证明了Stacking集成算法能够有效增强诊断模型的泛化能力,具有优良的诊断性能。In the hot rolling process,crown is an important quality index.The imbalanced nature of process data limits the prediction effect of data-driven models.The imbalance in process data restricts the predictive effect of data-driven models.In order to improve the prediction accuracy of the model,a hot rolled strip crown diagnosis model that combines SMOTE and Stacking ensemble algorithm was proposed.Firstly,the SMOTE over-sampling technique was employed to manipulate the crown-related datasets,aiming to mitigate the effects caused by the imbalanced distribution of data.Based on this,a Stacking ensemble model was constructed using LightGBM,SVM,KNN,and random forest as base learners and logistic regression(LR)as meta-learner.Finally,the model was validated using actual production data of a 2160 mm hot rolling strip.The results show that the diagnosis model's accuracy,minority class recall,balanced F score,geometric mean,and area under the ROC curve are 0.9580,0.9595,0.9573,0.9589,and 0.9579,respectively.The Stacking ensemble algorithm has the best prediction effect compared with XGBoost,LightGBM,KNN,SVM,and random forest models.It is proved that the Stacking ensemble algorithm can effectively enhance the generalization ability of the diagnosis model and has excellent diagnosis performance.

关 键 词：带钢凸度诊断 Stacking集成模型 非平衡数据 SMOTE 

分 类 号：TG335.5[金属学及工艺—金属压力加工]