基于神经网络的时变非线性系统迭代学习辨识  

NEURAL NETWORK BASED ITERATIVE LEARNING IDENTIFICATION OF TIME-VARYING NONLINEAR SYSTEMS

作　　者：孙明轩[1] 何海港[1] 孔颖[1] 

机构地区：[1]浙江工业大学信息工程学院,杭州310023

出　　处：《系统科学与数学》2013年第6期671-684,共14页Journal of Systems Science and Mathematical Sciences

基　　金：国家自然科学基金(60874041;61174034)资助项目

摘　　要：时变神经网络结构可简单地取为常规神经网络连接形式,但连接权却是时变的.如何确定时变权是应用时变神经网络时的难题.迭代学习方法是一种合理的选择,它不同于将时变连接权展成Taylor级数,通过训练多项式系数的处理方法.而且,后者的处理方式不可避免地存在截断误差.对于有限区间连续时变非线性系统的神经网络建模与辨识,借助于重复运行过程,以迭代学习算法调整权值,进行网络训练.不计逼近误差,提出的学习算法能够使得辨识误差在整个区间上渐近收敛于零.为处理非零但有界的逼近误差,采用带死区的迭代学习算法.逼近误差界值已知时,文中证明带死区修正的迭代学习算法使得辨识误差在整个区间上渐近收敛于由死区界定的邻域内.对于逼近误差界值未知的情形也进行了讨论.The architecture of a class of time-varying neural networks can be deter-mined by simply adopting that of the conventional neural networks, while the weights are allowed to vary with time. The challenge lies how to select the weights, when ap-plying a time-varying neural network. The conventional treatment is to use Taylor＇s series expansion for the time-varying weights, and the existing training algorithms are directly applicable for the coefficients, which are time-invariant. However, trun- cation errors exist, which may cause deterioration in performance. In this paper, we use the iterative learning methodology for training time-varying neural networks, and the neural networks are proposed for modeling and identification of continuous-time time-varying nonlinear systems. The zero identification error is achieved over the entire interval, if the used neural network is perfect in approximation. In the case of non-zero approximation error, iterative learning algorithms with dead zone mod-ification are proposed for updating the time-varying weights, and the identification error is ensured to converge to the bound, which is proportional to the approximation error.

关 键 词：辨识 迭代学习 时变系统 时变神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP181[自动化与计算机技术—控制科学与工程]