一种快速在线小波神经网络辨识算法

A RAPID LEARNING AND STEPWISE UPDATING ALGORITHM FOR MULTIRESOLUTION NEURAL NETWORK

本文以小波理论的多分辨率分析为基础,分析了各种小波网络的频率特性,指出了由尺度函数和小波函数组成的多分辨率小波网络的优点.针对多分辨率小波网络,本文提出了一种在线辨识算法.该算法,当增加样本和新神经元时,在线修正网络权值,同时通过正交化在线优选神经元,达到优化网络结构和网络权值的目的,训练速度快,辨识精度高,仿真结果表明了该算法的有效性.

In this paper, the frequency features of various wavelet neural networks are analyzed using multiresolution analysis theory, and the advantage of multiresolution neural networks (MRNN) composed of the scaling and wavelet functions is pointed out. Then a new identification method for multiresolution neural network is proposed to find the optimal weights of the network. The stepwise updating algorithm updates the weights instantly when a new pattern or a new neural node is added, and selects the optimal neural network by the orthogonal method on-line. The simulation results indicate that the proposed algorithm is very effective.