中国机械工程 ›› 2026, Vol. 37 ›› Issue (10): 2577-2585.DOI: 10.3969/j.issn.1004-132X.2026.10.020

• 机械基础工程 • 上一篇    下一篇

物理内嵌高斯过程模型及其在智能汽车路径跟踪控制器参数整定中的应用

赵克刚1, 陈浩强1, 梁志豪2, 叶杰3(), 王炽军4   

  1. 1.华南理工大学机械与汽车工程学院, 广州, 510641
    2.广东电网有限责任公司佛山供电局, 佛山, 528010
    3.佛山大学机电工程与自动化学院, 佛山, 528231
    4.南方(韶关)智能网联新能源汽车试验检测中心有限公司, 韶关, 511200
  • 收稿日期:2025-09-01 出版日期:2026-10-25 发布日期:2026-11-10
  • 通讯作者: 叶杰
  • 作者简介:赵克刚,男,1977年生,副教授。研究方向为先进传动系统的设计与控制、混合动力汽车的能量管理策略、智能网联汽车的控制与规划
    叶杰*(通信作者),男,1991年生,讲师。研究方向为自动驾驶规划与控制。E-mail: yejie@fosu.edu.cn。
  • 基金资助:
    国家自然科学基金(52572474);2023年度广东省特支计划领军人才项目

Physics-embedded Gaussian Process Model and Its Application in Parameter Tuning of Intelligent Vehicle Path Tracking Controllers

ZHAO Kegang1, CHEN Haoqiang1, LIANG Zhihao2, YE Jie3(), WANG Chijun4   

  1. 1.School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou,510641
    2.Foshan Power Supply Bureau of Guangdong Power Grid Co. ,Ltd. ,Foshan,Guangdong,528010
    3.School of Mechatronic Engineering and Automation,Foshan University,Foshan,Guangdong,528231
    4.South (Shaoguan) Intelligent Connected and New Energy Vehicle Testing & Inspection Center Co. ,Ltd. ,Shaoguan,Guangdong,511200
  • Received:2025-09-01 Online:2026-10-25 Published:2026-11-10
  • Contact: YE Jie

摘要:

高精度动力学模型是基于模型的强化学习框架实现智能汽车路径跟踪控制器参数自动整定的关键基础。针对物理模型因简化假设导致精度不足、数据驱动模型泛化性差的问题,提出一种物理内嵌高斯过程(PEGP)模型,将物理模型的预测状态嵌入高斯过程的特征空间,引导高斯过程学习物理模型的建模误差,实现物理信息与数据驱动的融合。实车试验结果表明,PEGP模型的预测精度和泛化性能优于物理模型与纯数据驱动模型等基准模型。在此基础上,提出一种PEGP-RL控制器参数自动整定方法,该方法仅需少量实车交互即可实现控制器参数的有效整定,中低速工况下的实车验证结果表明,其跟踪性能与人工多次整定的控制器相当。

关键词: 车辆动力学, 智能汽车, 物理内嵌高斯过程模型, 路径跟踪控制, 控制器参数整定, 基于模型的强化学习

Abstract:

A high-fidelity vehicle dynamic model is a critical foundation for model-based reinforcement learning frameworks for automatic parameter tuning of path tracking controllers in intelligent vehicles. A PEGP model was proposed to address the limited accuracy of simplified physics-based models and the poor generalization of purely data-driven approaches. In this method, the predictive states of a physical model were embedded into the feature space of the Gaussian process. The unmodeled dynamics of the physical model were thereby learned by the Gaussian process, and physical principles were integrated with data-driven learning. Real-vehicle tests demonstrated that the proposed PEGP model achieved superior prediction accuracy and generalization performance compared with the physics-based and purely data-driven baseline models. Based on this model, a PEGP-RL automatic parameter tuning method was developed, by which the controller parameters were effectively tuned with only a small amount of real-vehicle interaction data. Real-vehicle experiments at low-to-medium speeds show that the proposed approach achieves tracking performance comparable to that of a controller tuned through multiple rounds of manual adjustment.

Key words: vehicle dynamics, intelligent vehicle, physics-embedded Gaussian process model(PEGP), path tracking control, controller parameter tuning, model-based reinforcement learning

中图分类号: