【作者】 王新泽； 王贵东； 刘金琨；

【Author】 Xinze Wang;Guidong Wang;Jinkun Liu;School of Automation Science and Electrical Engineering,Beihang University;China Academy of Aerospace Aerodynamics;

【机构】 北京航空航天大学自动化科学与电气工程学院； 中国航天空气动力技术研究院；

【摘要】 在仪器测量、机器人控制和数控机床等领域,理论模型和实际对象之间一般都存在偏差,由于引起误差的因素众多且影响规律复杂,通过完备的动力学模型来描述实际的误差模型往往是难以实现的。随着计算机技术和算法的进步,神经网络在误差模型的建立中应用逐渐广泛,传统的BP神经网络等学习算法大多采用梯度下降方法,会出现训练速度慢,容易陷入局部最优,对学习速率敏感,容易过拟合等问题。本文在极限学习机神经网络的基础上,提出了一种基于高斯基函数特征映射的模糊极限学习机神经网络,既可以避免初始随机权值的不确定性带来的影响,又具有很好的泛化性能,可以避免严重的过拟合现象。通过所提出的神经网络设计了对应的误差补偿算法,针对具体的实验对象进行实验,补偿后实际对象和实验模型的误差范围和均方误差大幅缩小,实现了良好的补偿效果。更多还原

【Abstract】 There are generally deviations between the theoretical model and the actual object in the fields of instrument measurement,robot control,and CNC machine tools.Due to the numerous factors that cause errors and the complex rules of influence,it is often difficult to describe the actual error model through a complete dynamic model.With the progress of computer technology and algorithms,neural networks are gradually widely used to establish error models.Traditional learning algorithms such as BP neural networks mostly use gradient descent methods,which will lead to problems such as slow training speed,easy falling into local optimization,sensitivity to learning rate,and easy to overfit.On the basis of the extreme learning machine neural network,this paper proposes a fuzzy extreme learning machine neural network based on Gaussian basis function feature mapping,which can not only avoid the impact of the uncertainty of initial random weights but also has good generalization performance,and can avoid severe overfitting phenomenon.A corresponding error compensation algorithm was designed through the proposed neural network,and experiments were conducted on specific experimental objects.After compensation,the error range and mean square error of the actual object and experimental model were significantly reduced,achieving good compensation effects.更多还原