实验数据RBF神经网络模型中噪声的处理方法  

Strategies to the noise contained in experimental data in RBF neural network model

作　　者：李鸣鸣[1] 龚振邦[1] 欧阳航空[1] 孙麟治[1] 

机构地区：[1]上海大学精密机械研究所,上海200072

出　　处：《光学精密工程》2005年第z1期227-231,共5页Optics and Precision Engineering

基　　金：国家"863"计划资助项目(No.2002AA404450)

摘　　要：实验数据的非线性建模,是对各种仪器、设备的性能进行校正和补偿的基础。讨论了神经网络非线性建模时数据中的噪声成分造成的过拟合现象以及对模型精度的影响,针对RBF网络给出了2种提高建模精度的方法:建模数据预处理法和网络参数优化法。在数据预处理方法中,根据建模样本的特点,分别采用滑动平均法和灰色模型法对原始建模数据进行修正,并分析了它们的适用场合;对于后一种方法,选择径向基函数分布宽度和学习目标进行优化。以精密平台为例进行了实验,通过对其定位误差的测量、建模和预测,证明了上述各种方法的有效性,特别是后一种方法,可以得到非常高的建模精度。No-linear modeling to experimental data is a critical work before calibration or compensation to kinds of instruments and equipments.For neural networks modeling,noise contained in measuring data can cause over-fitting phenomena and then decrease the model's precision.As to radial basis function(RBF) networks,two strategies are introduced to cope with the defect.The first is pretreatment to original measuring data before modeling by means of the data's moving average and grey modeling.The second is to adjust the performance of RBF networks by choosing proper spread and training goal of radial basis function.Experiments are made on the precision stage by measuring,modeling and then predicting its positioning errors.The effectiveness is proved for all of these methods,especially for the last one.

关 键 词：数据建模 RBF网络 噪声 过拟合 

分 类 号：TN911.73[电子电信—通信与信息系统]