数据清洗对热轧微合金钢性能预报模型的改进  

Data cleaning for improvements on mechanical property prediction model of hot rolling micro-alloyed steel

作　　者：李维刚 王肖[1] 杨威 赵云涛 LI Wei-gang;WANG Xiao;YANG Wei;ZHAO Yun-tao(Engineering Research Center for Metallurgical Automation and Detecting Technology of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;National-provincial Joint Engineering Research Center of High Temperature Materials and Lining Technology,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China)

机构地区：[1]武汉科技大学冶金自动化与检测技术教育部工程研究中心,湖北武汉430081 [2]武汉科技大学高温材料与炉衬技术国家地方联合工程研究中心,湖北武汉430081

出　　处：《钢铁研究学报》2019年第10期920-927,共8页Journal of Iron and Steel Research

基　　金：国家自然科学基金资助项目(51774219)

摘　　要：热轧生产过程实测数据具有噪音大、信噪比低等特点,运用合适的方法对异常数据进行清洗将有助于提高钢材力学性能预报模型的精度。基于带钢热连轧过程数据的分布特点,采用孤立森林算法对热轧过程异常数据进行清洗,提高了性能预报模型的预测精度。首先,基于收集到的大量热轧微合金钢生产过程数据,采用孤立森林算法计算原始数据集中每条数据记录的异常分值;接着结合异常分值排序与力学性能建模实验,确定异常数据记录的个数;最后,基于清洗后的数据集合,运用融合数据与机理的建模方法建立力学性能预报模型,并对抗拉强度和屈服强度进行预测。预测实践表明,抗拉强度和屈服强度预报的平均绝对百分误差分别为2.50%和3.42%,且分别有93.13%和86.30%的数据预测值和实测值绝对误差在±6%之内;采用孤立森林算法对热轧生产过程异常数据进行清洗,可显著提高热轧带钢力学性能预报模型的精度。The process data of hot-rolled strip has the characteristics of high noise and low signal-to-noise ratio.The accuracy of the mechanical property prediction model could be improved through reasonable methods to clean the abnormal data.Based on the distribution characteristics of the process data of hot-rolled strip,the isolation forest was used to clean the abnormal data,which improves the accuracy of the prediction mode.Firstly,according to a large amount of collected production process data of hot rolled micro-alloyed steel,the anomaly score of each data record in the original data set was calculated using the isolation forest.Then,combining the anomaly score order and the mechanical property modeling experiment,the number of abnormal data was determined.Finally,based on the cleaned dataset,the mechanical property prediction models were built by the modeling method of combining data and mechanism,and the tensile strength and yield strength were predicted.Predictive results show that the mean absolute percentage errors for tensile strength and yield strength are 2.50%and 3.42%,respectively.Besides,the amounts of data with a relative error within±6%for tensile strength and yield strength account for 93.13%and 86.30%,respectively.Hence,the prediction accuracy of mechanical property prediction model can be improved significantly by the isolation forest to clean the abnormal data in hot rolling process.

关 键 词：热轧带钢 孤立森林 数据清洗 力学性能预报 

分 类 号：TG3[金属学及工艺—金属压力加工]