基于聚类特征选择的热轧过程带钢头部厚度预测  

Strip head thickness prediction during hot rolling process based on clustering feature selection

作　　者：武凯 武文腾 谢松雨 孙彦广[1] 彭文[2] 孙杰[2] WU Kai;WU Wenteng;XIE Songyu;SUN Yanguang;PENG Wen;SUN Jie(Automation Research and Design Institute of Metallurgical Industry,Beijing 100071,China;State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,Liaoning,China)

机构地区：[1]冶金自动化研究设计院有限公司,北京100071 [2]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

出　　处：《中国冶金》2024年第3期131-138,共8页China Metallurgy

基　　金：国家重点研发计划资助项目(2022YFB3304800)。

摘　　要：厚度是热轧产品关键质量指标之一,带钢头部厚度预测精度直接影响自动厚度控制(AGC)的控制效果,进而影响产品质量和成材率。热轧生产过程复杂多变,大量冗余工艺特征严重影响厚度预测建模效果。为提高预测模型精度,采用层次聚类和互信息相结合的方法进行特征选择,分别基于深度神经网络(DNN)、极端梯度提升(XGBoost)、支持向量机回归(SVR)以及梯度提升决策树(GBDT)建立带钢头部厚度预测模型,通过平均绝对误差(E_(MA))、均方误差(E_(MS))、最大百分比误差(E_(MAP))以及决定系数(R^(2))对模型的泛化能力进行评估。结果表明,在所建的预测模型中,DNN预测模型具有比其他模型更优的预测精度,测试集数据的E_(MA)、E_(MS)、E_(MAP)和R^(2)分别为0.015 4、0.000 3、0.004 4、0.992 1,并有97.15%的数据预测偏差小于0.03 mm,最大偏差小于0.04 mm。最后采用机器学习模型解释方法SHAP进行特征分析,得到相关工艺参数对带钢头部厚度的影响程度,实现了热轧带钢头部厚度的高精度预测。Thickness is one of the key quality indicators for hot-rolled products. The prediction accuracy of strip head thickness directly affects the control effect of automatic gauge control(AGC), which in turn affects product quality and yield. Due to the complexity and variability of the hot rolling production process, excessive redundancy features seriously affect the effect of thickness prediction modeling. To enhance the accuracy of the prediction model, a method combining hierarchical clustering and mutual information was adopted for feature selection. Strip head thickness prediction models were established using deep neural networks(DNN), extreme gradient boosting(XGBoost), support vector regression(SVR), and gradient boosting decision trees(GBDT). Evaluation of the models′ generalization capability was performed using metrics such as mean absolute error(E_(MA)), mean squared error(E_(MS)), maximum absolute percentage error(E_(MAP)) and coefficient of determination(R^(2)). The results show that the DNN model demonstrates superior precision compared to the others models constructed. The test dataset exhibits E_(MA), E_(MS), E_(MAP) and R^(2) values of 0.015 4, 0.000 3, 0.004 4, and 0.992 1, respectively, with 97.15% of data having a prediction error less than 0.03 mm and a maximum deviation under 0.04 mm. Finally, SHAP method is used for feature analysis, and the influence of related process parameters on strip head thickness is obtained, the high precision prediction of strip head thickness is realized.

关 键 词：热轧 精轧机 聚类特征选择 深度学习 头部厚度预测 

分 类 号：TG335.11[金属学及工艺—金属压力加工]