摘要
针对传统最小二乘法辨识动力学模型精度不高的问题,结合深度学习方法,提出了一种基于人工神经网络的动力学参数辨识方法。使用线型整流单元(ReLU)作为神经网络的激活函数,使用RMSProp算法对神经网络权值进行迭代,使用Dropout方法防止过拟合。采用有限项傅里叶级数轨迹作为激励轨迹,对采集到的数据进行标准化处理及滤波处理。最后,对算法得到的模型进行比较验证。结果表明,本文提出的方法相对于传统方法有较高的精度,不需要对摩擦力进行建模,能够更好地应用于机器人模型控制系统。
The accuracy of traditional identification methods for dynamic parameters is limited.In order to solve this problem,a dynamic parameter identification method based on artificial neural network is proposed.Rectifier linear unit(ReLU)is used as the activation function of the neural network.RMSProp algorithm is used to iterate the weights of the neural network.Dropout method is used to prevent over-fitting.The finite-value Fourier series trajectory is used as the excitation trajectory,and the collected data is normalized and filtered.Finally,the torque calculated by the algorithm is compared and validated.The results show that the proposed method is more accurate than the traditional method and does not require modeling friction.The method can be better applied to the robot control system.
作者
杜其通
刘朝雨
闵剑
费燕琼
Du Qitong;Liu Zhaoyu;Min Jian;Fei Yanqiong(Research Institute of Robotics, Shanghai Jiao Tong University, Shanghai 200240)
出处
《高技术通讯》
EI
CAS
北大核心
2020年第5期495-500,共6页
Chinese High Technology Letters
基金
国家自然科学基金(51875335)
国家重点研发计划(2017YFD0700602)资助项目。
关键词
神经网络
参数辨识
动力学
neural network
parameter identification
dynamics
