基于神经网络的超声波电机模型辨识  

Identification of Ultrasonic Motors Based on Neural Networks

作　　者：段梦凡[1] 谭永红[2] 

机构地区：[1]桂林电子科技大学智能系统与工业控制研究室,广西桂林541004 [2]上海师范大学信息与机电工程学院,上海201814

出　　处：《控制工程》2010年第5期607-610,690,共5页Control Engineering of China

基　　金：上海师范大学重点学科资助项目(DZL811);上海师范大学前瞻性项目(DYL200809);上海教委科研创新重点资助项目(09ZZ141);国家自然科学基金资助项目(60572055)

摘　　要：超声波电机存在着死区、迟滞等复杂的非线性特性。采用传统的系统辨识方法难以直接对该系统进行辨识,因此,根据超声波电机的静态、动态特性,提出了一种改进的BP神经网络以建立关于该电机的一种新的模型。通过引入迟滞算子构造扩张输入空间,将迟滞的多值映射转换为一一映射。提出了变斜率与带死区的神经元,以便于描述电机的死区特性。在训练神经网络时引入了广义梯度,以近似非光滑点处的梯度。最后给出了相应的实验结果,训练、泛化结果证明该建模方法是有效的。To the problems that the ultrasonic motors has the complex nonlinearities,such as dead zone and hysteresis,and the traditional identification methods are hard to be used to identify such systems directly,a modified back-propagation neural-network is proposed based on the static and dynamic characteristics of the ultrasonic motor.The ultrasonic motor model is established.By introducing a hysteretic operator to construct an expanded input space,the multi-valued mapping of hysteresis is transformed into a one-to-one mapping.The neuron with varying slope and dead zone is proposed to describe the feature of the dead zone in the motors.For the training of the proposed neural network,the generalized gradient is applied to approximate the gradient at the non-smooth points.The experimental results of the training and the corresponding model validation show the effectiveness of the proposed modeling method.

关 键 词：BP神经网络 超声波电机 辨识 非线性 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]