• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (12): 74-86.doi: 10.3901/JME.2021.12.074

• 特邀专栏:汽车-道路相互作用动力学前沿问题 • 上一篇    下一篇

扫码分享

基于改进Keras模型的路面附着系数估计

林棻, 王少博, 赵又群, 蔡亦璋   

  1. 南京航空航天大学能源与动力学院车辆工程系 南京 210016
  • 收稿日期:2020-07-02 修回日期:2020-11-16 出版日期:2021-08-31 发布日期:2021-08-31
  • 通讯作者: 赵又群(通信作者),男,1968年出生,博士,教授,博士研究生导师。主要研究方向为车辆动力学与控制。E-mail:yqzhao@nuaa.edu.cn
  • 作者简介:林棻,男,1980年出生,博士,副教授,硕士研究生导师。主要研究方向为车辆动力学与控制。E-mail:flin@nuaa.edu.cn;王少博,男,1995年出生,硕士研究生。主要研究方向为车辆动力学与控制。E-mail:shaobo@nuaa.edu.cn;蔡亦璋,男,1996年出生,硕士研究生。主要研究方向为车辆动力学与控制。
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(NS2020016)

Road Friction Coefficient Estimation Based on Improved Keras Model

LIN Fen, WANG Shaobo, ZHAO Youqun, CAI Yizhang   

  1. Department of Vehicle Engineering, College of Energy and Power Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016
  • Received:2020-07-02 Revised:2020-11-16 Online:2021-08-31 Published:2021-08-31

摘要: 路面附着系数是车-路相互作用中最为关键的参数之一,它的精确获取也是车辆主动安全控制系统正常工作的前提和基础,为此提出一种基于改进Keras模型的路面附着系数估计方法。对车辆进行动力学分析,找出与路面附着系数相关的动力学参数作为神经网络模型的输入量;通过各种工况的仿真试验建立数据集;以Keras模型为基础,结合限幅递推平均滤波算法、K折验证、Dropout正则化与Sarsa强化学习,提出改进Keras模型的路面附着系数估计器。滤波算法用于除去神经网络模型输入量的噪声,K折验证用于扩大样本空间,Dropout正则化可以降低模型的过拟合现象,提高模型泛化能力,Sarsa强化学习可以改善路面附着系数预测量超过边界的问题。最后,通过仿真验证表明所设计的估计器在路面附着系数估计中的有效性与可靠性,提出的方法相比原Keras模型平均绝对误差降低了73%,均方根误差降低了58%。

关键词: 路面附着系数, 神经网络, Keras模型, 强化学习, 估计

Abstract: Road friction coefficient is one of the most important parameters in vehicle-road interaction, the accurate acquisition of road friction coefficient is the basis for proper functioning of the vehicle's active safety control system. A method for estimating road friction coefficient based on an improved Keras model is proposed. Conduct a vehicle dynamic analysis to find out the dynamic parameters related to road friction coefficient which as the input of the neural network model. A data set is established through simulation experiments by various driving conditions. Based on the Keras model, combined with a limiting recursive average filtering algorithm, K-fold verification, Dropout regularization, and Sarsa reinforcement learning, an improved Keras model for road friction coefficient estimator is proposed. The filtering algorithm is used to remove the noise of the neural network model's input. K-fold verification is used to expand the sample space. Dropout regularization can reduce the model's overfitting phenomenon and improve the generalization ability of the model. Sarsa reinforcement learning can deal with the problem that road friction coefficient beyond borders. The simulation verification shows the effectiveness and reliability of the designed estimator for road friction coefficient. Compared with the original keras model, the average absolute error and root mean square error of the proposed method are reduced by 73% and 58% respectively.

Key words: road friction coefficient, neural network, Keras model, reinforcement learning, estimation

中图分类号: