鲁棒的递推核学习建模方法在高炉过程的应用  

Robust recursive kernel learning modeling method with application to blast furnace

作　　者：喻海清[1] 刘毅[2] 陈坤[1] 纪俊[1] 李平[1] 

机构地区：[1]浙江大学工业控制技术国家重点实验室,浙江杭州310027 [2]浙江工业大学化工机械设计研究所,浙江杭州310032

出　　处：《浙江大学学报（工学版）》2012年第4期705-711,共7页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(61004136)

摘　　要：针对现场采集时间序列数据中的离群点显著影响时变非线性工业过程在线模型性能这一问题,提出鲁棒的递推最小二乘支持向量机软测量建模方法.在模型训练阶段,采用支持向量聚类(SVC)排除离群点,建立有效的数据区域.将SVC用于递推过程前向学习阶段,并引入更有效的增删节点准则,在快速递推的前提下提高了模型的推广能力.将该方法应用于工业高炉过程铁水的硅质量分数预测,通过试验连续预测566炉高炉铁水硅质量分数,命中率高达81%,预测均方根误差为0.054 7,表明了较其他方法有更好的鲁棒性与精度.A robust recursive least squares support vector regression（R-RLSSVR） soft-sensing modeling method was proposed in order to overcome the embarrassment that outliers in the time series have a remarkable negative impact on the online modeling of the time-varying and nonlinear industrial processes.During the learning stage,the support vector clustering（SVC） approach was adopted to detect and remove the outliers and then to obtain the valid data area.Moreover,an improved recursive learning strategy,with nodes online growing and pruning,was developed to enhance the model generalization with a similar computation load.Through an example,the proposed method which was applied to predict the silicon mass fraction in hot metal of blast furnace showed a good performance,with 81% percentage of target hitting and 0.054 7 of the root mean square error of prediction when the size of predicted sample set was 566.These criteria are better than alternative methods,implying that the R-RLSSVR based modeling method is more robust and precise.

关 键 词：时间序列 离群点 支持向量聚类(SVC) 递推辨识 最小二乘支持向量机回归 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] TQ02[自动化与计算机技术—计算机科学与技术]