欢迎访问《汽车安全与节能学报》,

汽车安全与节能学报 ›› 2022, Vol. 13 ›› Issue (2): 309-316.DOI: 10.3969/j.issn.1674-8484.2022.02.011

• 智能驾驶与智慧交通 • 上一篇    下一篇

基于DRL的四轮独立驱动电动车辆的侧向车速估计

郑阳俊1(), 贺帅1, 帅志斌1,*(), 李建秋2, 盖江涛1, 李勇1, 张颖1, 李国辉1   

  1. 1.中国北方车辆研究所,北京100072,中国
    2.汽车安全与节能国家重点实验室(清华大学),北京100084,中国
  • 收稿日期:2021-09-09 修回日期:2022-01-19 出版日期:2022-06-30 发布日期:2022-07-01
  • 通讯作者: 帅志斌
  • 作者简介:*帅志斌,副研究员。E-mail: shuaizhibin@163.com
    郑阳俊(1997—),男(苗),湖南,硕士研究生。E-mail: Zheng_yangjun97@163.com
  • 基金资助:
    汽车安全与节能国家重点实验室开放基金课题(KF2018);国家自然科学基金项目(51975543)

Lateral velocity estimation for four-wheel-independent-drive electric vehicles based on deep reinforcement learning

ZHENG Yangjun1(), HE Shuai1, SHUAI Zhibin1,*(), LI Jianqiu2, GAI Jiangtao1, LI Yong1, ZHANG Ying1, LI Guohui1   

  1. 1. China North Vehicle Research Institute, Beijing 100072, China
    2. State Key Laboratory of Automotive Safety and Energy (Tsinghua University), Beijing 100084, China
  • Received:2021-09-09 Revised:2022-01-19 Online:2022-06-30 Published:2022-07-01
  • Contact: SHUAI Zhibin

摘要:

为精确估计车辆行驶状态,提出了一种四轮独立驱动电动车辆侧向车速估计方法。基于深度强化学习(DRL)范式,设计了侧向车速估计方法的架构;基于深度确定性策略梯度(DDPG)算法,设计了DRL智能体;采用循环神经网络,搭建了DDPG算法中的Actor网络和Critic网络。基于设计的奖励函数和训练场景,借助Matlab/Simulink软件,完成了算法的实现和训练;并通过在车辆双车道变换等实际行驶工况的仿真,进行了验证。结果表明:在经过了630次的学习训练之后,与扩展Kalman滤波方法相比,本文方法的估计精度提升40%。因而,本文方法能够在常用行驶工况中对车辆侧向车速进行估计。

关键词: 车辆动力学控制, 四轮独立驱动电动车辆, 侧向车速估计, 深度强化学习(DRL), 深度确定性策略梯度(DDPG)

Abstract:

A lateral-velocity estimation method was proposed for an electric vehicle with four-wheel independent-drive to estimate the vehicle motion states precisely. An architecture was designed for the lateral velocity estimation method based on the deep reinforcement learning (DRL) paradigm; A DRL agent was designed with deep deterministic policy gradient (DDPG) algorithm; The actor network and the critic network of the DDPG algorithm were constructed with the recurrent neural network (RNN). The algorithm was realized and trained in Matlab/Simulink with the designed award function and training scenarios; The algorithm effectiveness was verified by the simulation of practical driving maneuvers such as double-lane changing. The results show that after 630 episodes of training and learning, the proposed method improves the estimation accuracy by 40%, compared with that of the extended Kalman filter (EKF) method. Therefore, the proposed method can be used to estimate vehicle lateral velocity in general driving scenarios.

Key words: vehicle dynamics control, four-wheel-independent-drive electric vehicle, vehicle lateral velocity estimation, deep reinforcement learning (DRL), deep deterministic policy gradient (DDPG)

中图分类号: