【作者】 杜春燕；

【导师】 吴爱国；

【作者基本信息】 天津大学 ， 控制理论与控制工程， 2005， 硕士

【摘要】 本文对递归型神经网络在机械手逆模学习控制中的应用进行了研究。传统的基于前馈型神经网络的机械手逆模学习控制中,将机械手的动力学逆模型看作非线性静态映射,利用前馈型神经网络对非线性函数的逼近能力对其进行辨识,所需被测量较多,网络结构复杂,计算量大。如果将机械手的逆模型看作动态系统,并利用递归型神经网络对动力学系统的逼近能力对其进行辨识,则可以减少被测量,简化网络结构。本文使用递归型二阶神经网络对机械手逆动力学模型进行逼近,设计了基于递归型二阶神经网络的机械手逆模学习控制系统,对该系统进行了仿真,并与基于前馈型神经网络的系统进行了比较,以明确递归型神经网络对机械手逆模型的逼近能力,以及将递归型神经网络应用于机械手逆模学习控制的可行性和优越性。仿真实验表明,本文所设计的递归型二阶神经网络进行离线训练后对机械手逆动力模型有较好的逼近能力,学习效率要优于前馈型神经网络,因此更适用于机械手的在线学习控制;在轨迹控制效果方面,与基于前馈型神经网络的机械手逆模学习控制相比控制精度相差不大;在输出控制量的品质方面,基于递归型二阶神经网络的控制器输出力矩连续稳定,没有强烈的跳动,而基于前馈型神经网络的控制器输出力矩连续性较差,有强烈的突变和跳动;在控制器计算的复杂程度方面,递归型二阶神经网络的结构更简单,所需输入信息更少;在系统模型发生变化时,基于递归型神经网络的控制系统更加稳定。更多还原

【Abstract】 A type of recurrent neural network named second order recurrent neural networkis applied to the inverse model learning control of manipulators. In the traditionalinverse model learning control method, the model of the manipulator is seen as anonlinear static function and approximated by feed forward neural networks. Thedisadvantages of this method are that much of information of the controlled systemhas to be measured, the size of the networks is large and the computation burden isheavy. In this paper the inverse model of the manipulator is seen as a dynamic systemand approximated by recurrent neural networks, so as to downsize the networks,reduce the requisite information and lessen the computation burden. An inverse modellearning control system of manipulators based on second order recurrent neuralnetworks is proposed and simulated to confirm the feasibility and advantages of theapplication of the recurrent neural networks. Simulation experiments show that comparing with the method based on feedforward neural network, the advantages of the method presented in this paper are thatthe recurrent neural network is more efficient in offline approximation of the inversemodel of the manipulator, which implies its online control performance will be better;the control precision is not of much difference;the control signal is continuouswithout jerkiness;the size of the network is smaller and the computation burden islighter;when the controlled object model is changing, the control system is morestable.更多还原