基于神经网络的音圈电机迟滞特性建模  

Modeling of Hysteresis for Voice Coil Motor Based on Neural Network

作　　者：赵景波[1] 薛琨[1] 张磊[1] 刘慧敏[1] 

机构地区：[1]青岛理工大学,青岛266520

出　　处：《系统仿真学报》2014年第7期1503-1510,共8页Journal of System Simulation

基　　金：山东省自然科学基金(ZR2013FM014);山东省高等学校科技计划项目(J12LN37);泰山学者海外特聘专家项目(C2010-T005)

摘　　要：阐述了由音圈电机驱动的定位系统,并为该系统设计了微位移检测电路。对音圈电机施加35Hz的正弦波电压时,绘制的电机输出位移曲线为迟滞环,并且采集了0~35V的任意三角波驱动电压下的输出位移数据,作为训练样本。设计了径向基函数网络的迟滞辨识模型。网络的激励函数采用高斯核函数,提出加入最近邻规则的混合型K—均值聚类算法,基宽度由平均法确定,解决了RBF中心的初始化和基宽度由经验公式确定的问题。经验证,权值的修正采用改进的BP算法。仿真训练结果表明,RBF网络迟滞辨识模型平均误差为0.115μm,误差最大值为0.323μm。当辨识音圈电机的迟滞特性时,改进的RBF网络学习速度和精度都优于BP网络。The voice coil motor（VCM） positioning system was constructed and the micro-displacement detection circuit was designed. When 35 Hz sine wave voltage was applied to the VCM, the motor output displacement curve shows the hysteresis loop. The experiment collected the output displacement data under 0~35V random triangular wave voltage, as input pattern. The hysteresis model was designed by using radial basis function network. Some basis functions were analyzed and then Gaussian kernel function was chosen as basic function of hidden layer. Formally, the center’s initialization and the base width were determined by the experience formular; the K-means clustering algorithm and nearest neighbor rules combining to strike the center of the base function was proposed, the width of the basic function was determined by the average method. The weights connected the hidden layer to the output layer are modified by the improved BP algorithm. The simulation results show that the average error is 0.115μm, maximum error is0.323μm. Finally, the speed and accuracy of RBF network learning is better than BP network, and it is more suitable for hysteresis approximation of the voice coil motor.

关 键 词：BP神经网络 RBF神经网络 音圈电机 迟滞 仿真 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]