改进的RBF神经网络在非线性系统中的应用

在RBF神经网络的各种学习算法中,最近邻聚类算法学习时间短、计算量小,不需要事先确定隐单元的个数,完成聚类所得到的网络是最优的,并且可以在线学习,是一种自适应聚类学习算法,非常适合非线性实时系统的应用。但常规最近邻聚类算法在实时性要求较高的系统预测中学习时间相对较长。针对这一问题,提出了系统离线学习时采用减聚类算法,在线学习时采用改进的最近邻聚类算法,并变步长修正聚类半径和限制学习样本数。在函数拟合实验中,这种改进算法明显缩短了RBF神经网络的学习时间,在钢包精炼炉电极系统的在线辨识中的成功应用进一步表明对最近邻聚类算法的改进是有效的。

Application of an Improved RBFNN in Nonlinear System

The nearest neighbor-clustering algorithm has a short training time,less work to calculate and the number of hidden units is not to be determinated in advance in the various RBFNN learning algorithms,the network is optimization after clustering and can be trained on-line,it is an adaptive clustering algorithm for nonlinear real-time system.But it needs a longer training time in a real-time system which require high precision.Aimed at this problem,an improved algorithm was presented by using subtractive clustering algorithm in off-line learning and nearest neighbor-clustering algorithm in on-line learning,the clustering radius was modified by using variable step,the example number was limited in on-line identification.This algorithm cuts training time of RBFNN a lot in function fitting.The application in on-line identification of ladle furnace electrode system indicates the efficiency of the improved algorithm.