汽车工程 ›› 2022, Vol. 44 ›› Issue (10): 1527-1536.doi: 10.19562/j.chinasae.qcgc.2022.10.007

所属专题: 底盘&动力学&整车性能专题2022年

• • 上一篇    下一篇

基于混合神经网络的汽车运动状态估计

高振海1,温文昊1,唐明弘1,张建2,陈国迎1()   

  1. 1.吉林大学,汽车仿真与控制国家重点实验室,长春  130022
    2.中国第一汽车集团有限公司智能网联开发院,长春  130000
  • 收稿日期:2022-04-17 修回日期:2022-05-13 出版日期:2022-10-25 发布日期:2022-10-21
  • 通讯作者: 陈国迎 E-mail:cgy-011@163.com
  • 基金资助:
    国家自然科学基金(51775236);国家重点研发计划项目(2017YFB0102600);吉林省自然科学基金(20210101064JC);吉林省科技发展计划项目(20200501009GX)

Estimation of Vehicle Motion State Based on Hybrid Neural Network

Zhenhai Gao1,Wenhao Wen1,Minghong Tang1,Jian Zhang2,Guoying Chen1()   

  1. 1.Jilin University,State Key Laboratory of Automobile Simulation and Control,Changchun  130022
    2.Intelligent Connected Vehicle Development Institute,China FAW Group Co. ,Ltd. ,Changchun  130000
  • Received:2022-04-17 Revised:2022-05-13 Online:2022-10-25 Published:2022-10-21
  • Contact: Guoying Chen E-mail:cgy-011@163.com

摘要:

针对现有车辆运动状态估计算法严重依赖动力学模型精度且在大的质心侧偏角工况下准确性难以保障的问题,本文提出了一种基于混合神经网络的车辆运动状态估计算法。通过分析车辆本身的动力学基本特性,设计了适合于车辆运动状态估计的HNN混合神经网络架构,实现了车辆运动状态的深度学习估计。基于多个标准工况组成的数据集与典型实车测试工况进行了网络训练与测试验证。结果表明,相比于传统算法,本算法基于神经网络实现了精准的无动力学模型的汽车运动状态估计,提高了估计精度,且对路面附着系数变化具有鲁棒性。

关键词: 车辆状态估计, 深度学习, 门控循环单元, 多层感知机, 混合神经网络

Abstract:

For the problem that the existing vehicle motion state estimation algorithm relies heavily on the accuracy of the dynamic model and the accuracy is difficult to guarantee under large slip angle, the paper proposes a vehicle motion state estimation algorithm based on the hybrid neural network (HNN). By analyzing the basic dynamic characteristics of the vehicle itself, an hybrid neural network architecture suitable for vehicle motion state estimation is designed, and the deep learning estimation of vehicle motion state is realized. Based on the dataset composed of multi standard operating conditions and typical real vehicle test conditions, network training and test verification are carried out. The results show that compared with the traditional algorithm, the proposed HNN algorithm realizes estimation of vehicle motion state without dynamic vehicle model, improves estimation accuracy, and is robust to road adhesion coefficient change.

Key words: vehicle state estimation, deep learning, gated recurrent unit, multilayer perceptron, hybrid neural network