• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2025, Vol. 61 ›› Issue (7): 301-314.doi: 10.3901/JME.2025.07.301

• 机器人及机构学 • 上一篇    下一篇

扫码分享

无人自行车运动控制器重构及认知学习优化

黄用华, 梁子彦, 庄未, 杨海洋   

  1. 桂林电子科技大学机电工程学院 桂林 541004
  • 收稿日期:2024-04-26 修回日期:2024-10-26 发布日期:2025-05-12
  • 作者简介:黄用华,男,1977年出生,博士,研究员,硕士研究生导师。主要研究方向为机器人动力学与运动控制技术。E-mail:huangyonghuaxj@sina.com
    庄未(通信作者),女,1977年出生,博士,研究员,硕士研究生导师。主要研究方向为机器人动力学及运动控制技术。E-mail:zhuangweibupt@sohu.com
  • 基金资助:
    国家自然科学基金(52165001,51865005,51765011)和广西自然科学基金(2018GXNSFAA281297,2018GXNSFAA281301)资助项目。

Reconstruction and Cognitive Learning Optimization of Unmanned Bicycle Motion Controller

HUANG Yonghua, LIANG Ziyan, ZHUANG Wei, YANG Haiyang   

  1. School of Mechanical and Electrical Engineering, Guilin University of Electronic Technology, Guilin 541004
  • Received:2024-04-26 Revised:2024-10-26 Published:2025-05-12

摘要: 针对无人自行车平衡控制器结构及参数未知及其控制系统泛化能力不足的问题,提出一种用于重构未知控制器并提升控制系统动态性能的认知学习系统。该认知学习系统主要由两部分构成:其一是利用高斯过程回归(GPR)的理论方法,以无人自行车车体侧向倾角、车体侧向倾角速度、车把转角、车把转角速度为输入,以车把控制力矩为输出,构建概率模型表征系统控制器输入输出之间的映射关系;其二是结合反应式认知学习优化策略,引入感知模块、认知模块和执行模块,以所建立的概率模型输出作为无人自行车认知学习控制器的训练初值,对其进行控制性能的迭代优化。数值仿真和样机实验结果表明:高斯过程回归方法能够较好地重构未知控制器,认知学习算法可以有效地改善控制系统的动态性能;两者综合使无人自行车可以在相对平整水泥路面实现稳定的平衡行走,其侧向倾角在[‒0.042 rad, 0.044 rad]范围内波动,且相比于GPR控制器,优化后的认知学习系统使车体侧向倾角及车把转角的振幅分别减小了约55.7%和57.1%,并且对湿滑路面、坡面、草坪的运行场景表现出较好的泛化性能。

关键词: 无人自行车, 平衡运动, 控制器重构, 认知学习, 高斯过程回归

Abstract: Aiming at the unknown structure and parameters of the balance controller of unmanned bicycle and the lack of generalization ability of the control system, a cognitive learning system for reconstructing the unknown controller and improving the dynamic performance of the control system was proposed. The cognitive learning system is mainly composed of two parts: One is to use Gaussian process regression (GPR) theory to establish a probability model to represent the mapping relationship between the input and output of the controller, taking the lateral inclination Angle, lateral inclination speed, handlebar Angle and handlebar Angle speed as inputs and the control torque of the handlebar as outputs. Second, combining the reactive cognitive learning optimization strategy, the perception module, cognitive module and executive module are introduced, and the output of the established probability model is taken as the initial training value of the cognitive learning controller of the unmanned bicycle, and the control performance of the controller is iteratively optimized. The results of numerical simulation and prototype experiment show that the Gaussian process regression method can reconstruct the unknown controller well, and the cognitive learning algorithm can effectively improve the dynamic performance of the control system. The combination of the two enables the unmanned bicycle to achieve stable and balanced walking on a relatively flat cement road surface, and its lateral inclination Angle fluctuates within the range of [-0.042rad, 0.044rad]. Compared with the GPR controller, the optimized cognitive learning system reduces the lateral inclination Angle of the vehicle body and the amplitude of the handlebar Angle by about 55.7% and 57.1%, respectively. It also shows good generalization performance for the running scenarios of wet road, slope, lawn.

Key words: unmanned bicycle, balance exercise, controller reconfiguration, cognitive learning, Gaussian process regression

中图分类号: