遗传算法-模糊聚类动态模糊神经网络辨识

针对非线性系统辨识特点,在剖析具有递归环节的T-S模糊神经网络结构的同时,提出了一种新型的3步设计优化方案,即非线性区域的线性划分、离线训练和在线辨识.将融合了模糊c-mean聚类(GA-FCM)(称为双群体并行聚类)算法引入到RTSFNN中,对非线性系统的输入输出空间进行聚类(线性划分),再在每个线性区域上建立ARMAX模型;利用GA实数编码,同时优化前件隶属函数的中心和宽度、递归增益及后件参数;在线时利用FCM在线分析输入数据特征,确定是否对现有划分进行改动,并利用GA迭代一定代数优化其他参数,直到误差满足要求为止.通过对非线性动态系统的辨识仿真,验证了所提出方法在训练时的寻优速度、训练误差及校验误差指标上均有很大优势.

Dynamic fuzzy neural network identification using combination of genetic algorithm and fuzzy clustering

In view of the characteristics of identification in nonlinear systems,this paper gives an analysis of the structure of a recurrent T-S fuzzy neural network(RTSFNN).Then a three-staged optimization method was proposed,i.e.,linear division of the non-linear region,off-line training,and online identification.In the first stage,an improved fuzzy c-mean clustering algorithm,or double groups GA-FCM parallel clustering algorithm,is introduced into RTSFNN to cluster the input and output spaces of a nonlinear system,then an auto-regressive moving average exogenous(ARMAX) model in each linear region is established.In the second stage,using real-number codes in GA algorithm optimizes the center,the width and recursive gain of antecedent membership function and the consequent parameters are optimized simultaneously.In the third stage,an improved fuzzy c-mean clustering method is employed to analyze the characteristics of input data online,so as to determine whether the existing division should be modified;then using GA optimizes the parameters until they satisfy the requested error.By simulating the process of identification in a non-linear dynamic system,the authors confirmed the superiority of their method in convergence rate,training errors and verification errors.