基于正交法的神经网络结构研究和应用  

Study and Application for Structure of Neural Networks Based on Orthogonal Method

作　　者：邵信光[1] 杨慧中[1] 

机构地区：[1]江南大学自动化研究所,江苏无锡214036

出　　处：《控制工程》2004年第S2期58-60,共3页Control Engineering of China

摘　　要：总结了正交法的应用研究,提出用正交化方法来确定前馈神经网络的结构,包括隐层数、节点数以及网络训练步数。并将谈网络用于描述聚丙烯腈生产过程软洲量混合模型的中间参数k1,k2,k3,P与现场操作条件的非线性关系,从而实现质量指标的在线估计。仿真结果证明了该方法的有效性。Based on summarizing of the application of onhogonal method, an onhogonal method for deciding the number of hidden layers, neurons and the training step of neural networks is presented. Further, the networks are applied to describe the nonlinear relations of k1, k2, k3, P and onsite operating conditions in a soft measurement hybrid model of polyacrylonitrile. Then the quality performances are estimated online. The results of simulation show that this method is effective.

关 键 词：丙烯腈 聚合 神经网络 正交化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]