摘要
以MATLAB的神经网络工具箱为平台,依据实例系统的试验数据,分析神经网络的隐层数、神经元数、初始权值、初始阈值和样本的选择方法及输入个数对系统辨识精度的影响。在充分考虑各参数之间的交互作用对系统辨识精度影响的前提下,提出了循环嵌套编程的训练方法,获得了使收敛精度及泛化效果达到最优时各项参数的取值。利用神经网络在单输入系统辨识方面的优点,提出将多输入系统转化为单输入系统进行系统辨识的方法,并进行了对比分析。基于理论分析的结果,对实例系统进行了辨识,取得了满意的效果。
Influences of artificial neural network＇s （ANN） parameters including number of hidden layers and nodes, initial weights, initial biases and the determination methods of the training set to the accuracy of system identification, based on the experiment data of the exemplified system, are analyzed. Based on MATLAB toolbox, nest of programming method, which considers the interaction among the parameters and can help gain optimal parameters of ANN, is presented. In addition, a method that converts a multi-input system to a one-input system and then establishes its model using ANN is also presented. According to these methods, model of the exemplified system is established and satisfactory result is obtained.
出处
《机械工程学报》
EI
CAS
CSCD
北大核心
2006年第7期217-221,226,共6页
Journal of Mechanical Engineering
基金
上海汽车工业总公司(0308)
上海市教委曙光计划(02SGU8)资助项目。
关键词
神经网络
参数
系统辨识精度
影响分析
确定方法
Artificial neural network Parameters System identification accuracy Influence and analysis Determination method
