基于UKF的自组织模糊神经网络训练算法  

Training self-organizing fuzzy neural networks with unscented Kalman filter

作　　者：李庆良[1] 雷虎民[1] 徐小来[1] 

机构地区：[1]空军工程大学导弹学院,陕西三原713800

出　　处：《系统工程与电子技术》2010年第5期1029-1033,共5页Systems Engineering and Electronics

基　　金：总装备部武器装备预研基金(9140A0405040TJ133201);航天科技创新基金(CASC0209)资助课题

摘　　要：如何生成最优的模糊规则数及模糊规则的自动生成和修剪是模糊神经网络训练算法研究的重点,针对这一问题,提出了基于无迹卡尔曼滤波(unscented Kalman filter,UKF)的自组织模糊神经网络的训练算法。分析了模糊神经网络的非线性动力系统表示,并用递推最小二乘法(recursive least square,RLS)和UKF分别学习线性和非线性的参数,给出了模糊规则生成的准则和参数更新的策略;然后,用误差下降率方法作为模糊规则修剪的策略,删除作用不大的规则。通过典型的函数逼近和系统辨识实例,表明所提算法得到的模糊神经网络的结构更为紧凑,泛化性能更佳。Much of the current research interest in neuro-fuzzy hybrid systems is focused on how to generate an optimal number of fuzzy rules in a neuro-fuzzy system and investigate the automated methods of adding and pruning fuzzy rules.To deal with this problem,a self-organising fuzzy networks training algorithm based on unscented Kalman filter（UKF） is presented.Firstly,a non-linear dynamical system expression of fuzzy networks is analyzed,and RLS and UKF are used to learn linear and non-linear parameters respectively.Secondly,guidelines of how to generate a new rule and update parameters are presented.Then,the method of the error descending rate is used as fuzzy rule pruning strategy,so that the rule which plays an unimportant role in the system is deleted.Finally the typical experiments of function approximation and system identification indicate that the fuzzy network obtained by the proposed algorithm has a more tighten structure and better generalization than other algorithms.

关 键 词：无迹卡尔曼滤波 自组织神经网络 T-S模型 系统辨识 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]