摘要
针对聚丙烯熔融指数软测量建模问题,提出了一种基于最小化最大绝对预测误差的组合神经网络软测量建模方法,并将该方法应用于聚丙烯熔融指数软测量研究中。通过建立多个不同结构的BP神经网络模型,并合理组合各个模型,可显著改善单一神经网络模型的泛化能力。鉴于合适的组合权重对取得良好预测性能是至关重要的,因此提出将最小化最大绝对预测误差作为控制预测精度的目标,来选择合适的组合权重。聚丙烯熔融指数软测量研究结果表明:通过与各个单一神经网络模型的预测精度比较,采用该方法建立的聚丙烯熔融指数组合神经网络软测量模型具有更佳的预测精度和鲁棒性。
To build an accurate soft-sensor model of polypropylene melt index,a new method based on minimum maximum was studied.Single neural network model generalization capability could be significantly improved by using a number of different structures of BP neural network model and reasonable stacked neural network model.Proper determination of the stacking weights was essential for good performance,so determination of appropriate weights for combining individual networks using minimum maximum absolute prediction error was proposed.The results of melt index prediction demonstrated significant improvement in model accuracy and robustness,as compared to using a number of different structures neural network model.
出处
《科技通报》
北大核心
2011年第3期403-407,共5页
Bulletin of Science and Technology
基金
国家高技术研究发展计划项目(2006AA04Z178)
关键词
聚丙烯
熔融指数
软测量
组合神经网络
最大绝对预测误差
polypropylene
melt index
soft sensor
stacked neural networks
maximum absolute prediction error