基于集成学习的连退带钢质量在线预报方法  

Online Prediction of Steel Strip Quality in Continuous Annealing Based on Ensemble Learning

作　　者：王显鹏[1] 黄灿明[1] 徐子睿[1] 王丹敬 

机构地区：[1]东北大学工业与系统工程研究所,沈阳110819

出　　处：《控制工程》2017年第3期481-486,共6页Control Engineering of China

基　　金：国家自然科学基金资助项目(61573086)

摘　　要：钢铁企业连续退火生产过程由于工艺复杂并且包含众多相互影响的过程变量,导致带钢产品质量波动较大,难以进行在线测量。针对该问题,提出了一种基于混合集成学习的带钢产品质量在线预报方法。该方法以Bagging为基础,引入了Adaboost中对误差较大样本的重点学习策略,以提高混合集成学习模型的精度并改进泛化能力;在子学习机的训练中提出了基于动态加权的最小二乘支持向量机方法,以改进子学习机的鲁棒性。基于实际连退生产过程数据的测试结果表明,所提出的混合集成学习方法在预测精度和泛化能力上均要优于Bagging和Adaboost等传统的集成学习建模方法。To handle the frequent large fluctuation of steel online way, an online quality prediction method of steel strip quality which is hard to be measured in an strip based on a hybrid ensemble learning was developed. This method adopted the main framework of Bagging and incorporated the emphasis learning on samples with large errors used in Adaboost, so as to improve the precision and generalization ability. Besides, a dynamic weight least square support vector machine （LSSVM） was proposed to improve the robustness of the individual learning machine. Computational results based on practical samples of continuous annealing illustrated that the proposed hybrid ensemble learning method was superior to the canonical Bagging and Adaboost algorithms in both the precision and the generalization ability.

关 键 词：连续退火 质量在线预报 集成学习 最小二乘支持向量机 

分 类 号：TP27[自动化与计算机技术—检测技术与自动化装置]