数据挖掘在卷取温度控制中的应用  

Application of data mining in coiling temperature control

作　　者：张敬哲 陈冬 张瑞[1] 李振垒[1] 袁国[1] ZHANG Jingzhe;CHEN Dong;ZHANG Rui;LI Zhenlei;YUAN Guo(State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,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

出　　处：《轧钢》2023年第2期105-110,共6页Steel Rolling

摘　　要：卷取温度的控制精度直接影响热轧钢材的组织性能,为改善外界工况条件变化对卷取温控制造成的准确率低、温度波动大和换规格轧制首卷钢温度命中难等共性问题,结合轧制数据特点与异常值检测、特征降维、数据均衡化等多种智能算法,改善了数据存在异常、冗余信息及分布不均等问题,提高了数据可靠程度;通过调节优化方法改进传统BP神经网络模型,建立了高精度卷取温度预测模型。结果表明:孤立森林算法可以准确并高效地筛查出异常值,使数据更加稳定、有效;在对化学成分信息进行数据降维时结合先验知识采用MIRF算法,成功减少了冗余信息与噪音数据所造成的空间复杂度高、误差大等问题,避免了关键信息的缺失;基于Adam优化方法的4层BP神经网络建立的卷取温度预测模型具有很强的泛用性及可靠性,该模型预测值与实际值线性拟合良好,R^(2)达0.990 6,绝大多数预测值的绝对误差在10℃以内。应用数据挖掘方法在卷取温度控制中优势明显,为进一步以数据驱动研究温度控制提出了一种新的思路。The control accuracy of the coiling temperature directly affects the microstructure and properties of the hot rolled steel.In order to improve the common problems of coiling temperature control caused by the change of external working conditions,such as low accuracy,large temperature fluctuation,and difficulty in rolling the first coil of steel for changing specifications,combined with the characteristics of rolling data,the application of various intelligent algorithms such as outlier detection,feature dimensionality reduction and data equalization,the problems of data anomalies,redundant information and uneven distribution were improved,finally improved data reliability.And by adjusting the original optimization method to improve the traditional BP neural network model,a high-precision coiling temperature prediction model was established.The results show that the isolation forest algorithm can accurately and efficiently screen out outliers,making the data more stable and effective.When reducing the dimensionality of chemical composition information,the MIRF algorithm is used in combination with prior knowledge,which successfully reduces the problems of high space complexity and large errors caused by redundant information and noise data,and avoids the loss of key information.Based on the Adam optimization method,the coiling temperature prediction model established by the 4-layer BP neural network has strong versatility and reliability,the linear fit between the predicted value of the model and the actual value is good,the R^(2) score is 0.9906,and the absolute error of most predicted values is within 10℃.The application of data mining method has obvious advantages in coiling temperature control,and a new idea is proposed for further data-driven research on temperature control.

关 键 词：卷取温度 数据挖掘 数据预处理 神经网络 

分 类 号：TG335.11[金属学及工艺—金属压力加工]