面向纵向自动驾驶的仿人驱动控制网络模型  

Humanoid-driven Control Network Model for Longitudinal Autonomous Driving

作　　者：高镇海[1,2] 于桐 孙天骏 唐明弘 高菲 赵睿[1,2] GAO Zhenhai;YU Tong;SUN Tianjun;TANG Minghong;GAO Fei;ZHAO Rui(State Key Laboratory of Automobile Simulation and Control,Jilin University,Changchun 130022;College of Automotive Engineering,Jilin University,Changchun 130022)

机构地区：[1]吉林大学汽车仿真与控制国家重点实验室,长春130022 [2]吉林大学汽车工程学院,长春130022

出　　处：《机械工程学报》2023年第18期251-262,共12页Journal of Mechanical Engineering

基　　金：国家自然科学基金(52202495,52202494,51775236);吉林大学研究生创新基金(451230411061);吉林大学长沙汽车创新研究院自由探索(CAIRIZT20220106);中央高校基本科研业务费专项资金(2022-JCXK-24)资助项目。

摘　　要：自动驾驶车辆已经成为当前汽车行业的研究热点。运动控制算法在很大程度上决定自动驾驶汽车的安全性和驾乘人员的接受度。自动驾驶控制算法的精度日益提高,然而其与驾驶员控制风格的一致性仍然较低,这会降低驾乘人员的体验感与接受度。此外,现有方法仍存在大量需要人工标定的参数,导致算法部署时的工作量较大。针对以上问题,提出车辆纵向驱动控制仿人神经网络(Vehicle longitudinal drive control human-like neural network,LCN),并以此为基础构建了仿人机理的控制模型,其控制风格与驾驶员的一致性较高且能够实现参数的自学习标定。LCN的设计基于对人类驾驶行为的分析,其将驾驶员的控制机理与数据驱动方法相结合,并通过独特的网络架构设计将驾驶员机理中的容差控制特性与时延响应特性融入LCN。试验结果表明,所提出模型的控制风格更接近人类,且能够实现对车辆动力学特性的自估计与参数的自标定。Autonomous vehicles have become a current research hotspot in the automotive industry.Motion control algorithms remarkably determine the safety and passenger acceptance of autonomous vehicles.The accuracy of autonomous driving control algorithms is increasing,yet their consistency with the driver's control style is still low,which can reduce the driver's experience and acceptance.In addition,existing methods still have a large number of parameters that need to be manually calibrated,which leads to a high workload in algorithm deployment.A vehicle longitudinal drive control human-like neural network(LCN)is proposed,and a control model with humanoid mechanism is constructed based on it,which has a highly homogeneous control style and is capable of self-learning calibration of parameters.The design of the LCN is based on the analysis of human driving behavior.It combines the driver's control mechanism with a data-driven approach,and incorporates tolerance control characteristics and time-delay response characteristics in the driver's mechanism in the network architecture.The experimental results show that the control style of this model is closer to that of humans,and it is able to achieve self-estimation of vehicle dynamics and self-calibration of parameters.

关 键 词：自动驾驶 仿人机理 自学习标定 运动控制 容差控制 

分 类 号：U46[机械工程—车辆工程] TP399[交通运输工程—载运工具运用工程]