数据驱动的热连轧板平直度分类预测  

Data-driven classification prediction of hot strip flatness

作　　者：李广焘 龚殿尧[2] 张殿华[2] 梁勋国[1] 陈驰 LI Guangtao;GONG Dianyao;ZHANG Dianhua;LIANG Xunguo;CHEN Chi(Headquarter of Steel and Iron,CISDI Group Co.,Ltd.,Chongqing 401122,China;The State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,Liaoning,China)

机构地区：[1]中冶赛迪工程技术股份有限公司钢铁事业本部,重庆401122 [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：《钢铁》2024年第7期83-93,共11页Iron and Steel

摘　　要：在热连轧带钢生产过程中,平直度是重要的板形质量指标。为提高热连轧带钢板形的控制水平,针对热连轧带钢平直度检测滞后与预测精度不高的问题,建立基于数据驱动的热连轧板平直度分类预测模型。将工艺参数、设备参数以及轧件参数中对带钢平直度影响显著的66个关键参数作为输入变量。按照平直度仪在轧制方向上的实际测量间距以及在带钢宽度方向上的测量通道宽度,将钢卷离散化为一系列条元,并生成不同条元带钢对应的平直度分类标签,研究实际以带钢宽度方向上中部条元的平直度分类标签作为输出变量。基于提出的数据预处理方法,获取用于构建板平直度分类预测模型的可靠且高质量的建模数据。采用深度置信网络(deep belief networks,DBN)构建基于数据驱动的热连轧板平直度分类预测模型。以交叉熵为评价指标,分析超参数对DBN模型性能的影响,确定模型的最佳超参数。在测试集样本类别分布不均衡的情况下,构建的DBN模型对大中浪平直度类别、无中浪平直度类别以及微中浪平直度类别的预测准确度分别达到了100%、96.87%与77.78%,平直度的错误预测类别主要为与其平直度值更接近的相邻平直度类别。该方法能对热连轧带钢平直度进行准确的分类预测,对提高板形预设定模型的计算精度与改善带钢板形具有重要作用。The flatness is an important quality indicator of the strip shape in hot strip rolling.In order to further improve strip shape control level,and in view of the problems of hot-rolled strip flatness on detection lag and low prediction accuracy,a data-driven classification model of the hot-rolled strip flatness is developed.The 66 key parameters that have a significant impact on the strip flatness among the process parameters,equipment parameters and strip parameters are used as input variables.According to the actual measured spacing of the flatness meter in the rolling direction and the measurement channel width along the strip width direction,the strip coils are dispersed into a series of strip units,and generating flatness classification labels corresponding to different strip units,the flatness classification labels of the middle strip units along the strip width direction are actually used as output variables in this study.The method of data preprocessing is proposed to obtain reliable and high-quality modeling data for building the flatness classification model.DBN algorithm is used to construct the data-driven classification model of the hot-rolled strip flatness.Using cross entropy as the evaluation index,the influence of hyperparameters on the performance of DBN model is analyzed to determine the best one.Although the sample categories in testing set are unbalanced,the accuracy rates of the three flatness categories,including the large-medium wave,the no-medium wave and micro-medium wave,are 100%,96.87% and 77.78% respectively,the wrong classification categories of flatness are mainly adjacent flatness categories,and their actual flatness values are very close.This method can accurately classify the finished hot-rolled strip flatness,and plays an important role in improving the calculation accuracy of the shape preset model and improving strip shape.

关 键 词：热连轧 带钢 板形预测 数据驱动建模 数据预处理 深度置信网络 

分 类 号：TG334.9[金属学及工艺—金属压力加工]