基于双向权值调整算法的神经网络非线性系统辨识

使用神经网络建模是非线性系统辨识的一个重要方法。为克服传统BP算法训练多层前向神经网络进行系统辨识中存在的一些问题,本文提出一种使用双向权值调整学习算法训练单隐层前向神经网络进行非线性系统辨识的方法。此辨识方法使用结构简单的单隐层前向神经网络,在正向阶段由Moore-Penrose广义逆确定输出权值,反向阶段则按误差梯度下降原则对隐层权值进行调整。算法能在正向和反向两个过程对网络的权值做出调整,具有较快的学习速度,并且能在一定程度上保证神经网络的泛化能力。通过基准辨识仿真实验验证,基于此方法的非线性系统辨识具有建模结构简单、训练速度快且辨识精度高的特点。

Nonlinear System Identification Using BPWA Based Single Hidden-layer Feedforward Neural Networks

Modeling with neural networks is one of important methods for nonlinear system identification.To avoid shortcomings of the traditional BP algorithm and modeling,a new method for nonlinear system identification is presented,which uses the Bi-Phases Weights Adjusting(BPWA) Algorithm to train single hidden-layer feedforward neural networks.The structure of the single hidden-layer feedforward neural network is simple.The learning algorithm is able to adjust the weights during both the forward and backward phases,which has the advantages of fast convergence speed and improved generalization.The method is tested over several benchmark problems.The simulation results demonstrate that the proposed method performs well for nonlinear system identification.