机器人逆动力学神经网络建模.PDFVIP

12 2 Vol112 N o12 2008 3 ELE CTR IC MACH INES AND CONTRO L M ar. 2008 栾楠 ( , 200240) : 针对神经网络逆动力学模型用于前馈控制的网络泛化能力问题, 通过对实际机器人的 真实验, 分析了前馈神经网络学习的机理, 对神经网络训练过程中的数据采样环节进行了改进, 提 出了在状态空间域中进行数据筛选和处理的神经网络学习方法通过二自由度机器人运动 真实 验表明, 该方法提高了模型泛化能力, 有利于实时动力学前馈控制方法的实用化; 与传统 PID控制 相比, 该神经网络模型进行动力学前馈控制能大幅度减少动态误差, 改善了系统稳定性 : 神经网络; 动力学逆模型; 前馈控制; 泛化能力 : TP242 : A : 100 - 449X( 2008) 02- 01 4- 05 Study on robot inverse dynam ics modeling w ith neural network LUAN Nan ( School ofM echan ical Eng ineering, Shanghai Jiaotong U inversity, Shanghai 200240, Ch ina) Abs tract: Considering the generalization ab ility of the IDM ( inverse dynam ics model) by neural network w hich is often used in dynam ics feedforw ard control, w ith p lenty of smi u lation expermi ents on a robo,t the characteristics of neural network learn ing are d iscussed. An mi proving method for neural netw ork training w as presented. The mi provement ami d at sample data obtaining and disposing, and themethodw as based on data filtering and processing on state space. The effectiveness of them ethod w as confirmed by smi u la- tion expermi ents on a 2 DOF ( degree of freedom ) robo.t The expermi ents show that the generalization a- b ility of the IDM is mi provedw ith this method. It is of benefit to utility of real-tmi e dynam ics feedforw ard contro.l Comparing w ith PID control, the dynam ics feedforw ard control using such IDM can reduce the dynam ic error greatly and mi prove the stab ility of the system.