基于正交最小二乘的傅立叶神经网络结构选取方法  

Structure Selection for Fourier Neural Network Based on Orthogonal Least Square Optimization

作　　者：段超霞[1] 田学民 

机构地区：[1]国家知识产权局专利局专利审查协作广东中心,广州510530 [2]中国石油大学信息与控制工程学院,山东青岛266580

出　　处：《石油化工自动化》2012年第6期19-23,共5页Automation in Petro-chemical Industry

基　　金：国家自然科学基金资助(51104175);山东省自然科学基金资助(ZR2011FM014)

摘　　要：针对傅立叶神经网络基频选取及隐含层神经元数目难以确定这一问题,提出一种基于正交最小二乘的傅立叶神经网络结构选择方法。该方法依据傅立叶神经网络激活函数是一系列不同函数的特点,给定一个较小的基频,把一系列关于基频整数倍的正余弦函数作为可供选择的激活函数集,然后由正交最小二乘算法选择对网络性能贡献显著的函数作为激活函数,并简化网络结构。以聚合反应器为对象的建模仿真研究表明,正交最小二乘算法可以更有效地确定傅立叶神经网络结构。Contrapose to the problem of difficult determination of the base frequency selection and neural hidden nodes number of Fourier Neural Network （FoNN）, the Orthogonal Least Square （OLS） method is used to optimize FoNN. According to the characteristic that the activation function in FoNN is different, it is regarded as a problem of activation function selection. Given a small base frequency, a series of sin/cos functions regarding integral multiple of base frequency are treated as a set of function, and then OLS algorithm is employed to choose distinctly contributing function subset as activation function. The simulation study of polymerization reactor shows that the proposed method is effective.

关 键 词：傅立叶神经网络正交最小二乘算法激活函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]