基于车辆总线数据时序分析的驾驶行为辨识方法研究

Disorders of driver's behavior are the main causes to road traffic accidents. To solve those problems, researches on vehicle driver's behavior analysis, identification, and prediction are carried out with the driving experimental platform constituted by real vehicle. That is popular technical solution in current road traffic safety research areas. The applying project is to identify the driver's behavior on basis of data derived from vehicle data-bus. For that purpose, test data for analysis will be collected by real vehicle driving experimental platform. And the main theoretical method is time series analysis according to the features of real vehicle data. Specific content of the project includes characters analysis of data collected with the real vehicle driving experimental platform, identification modeling of driver's behavior, and valuation method for model parameters. Analyzing test data from the driving experimental platform for their system character, statistics, spectrum and dynamic is the premise to build the structure of an auto-regressive and moving-average (ARMA) model for driver's behavior identification. In order to meet the demand of behavior identification made by different drivers, recursive least squares method will be introduced to valuate and correct parameters of the ARMA model online. The applying project is to analyze and convert driver's behavior in view of system identification. Theoretical methods such as transportation engineering, vehicle engineering and mathematical statistics are used and interdisciplined through course of the research. In addition, the applying project may provide new idea and theoretical foundation for studying the new type of vehicle-embedded driving safety warning systems based on data derived from vehicle data-bus.

人的驾驶行为失常是致使道路交通事故发生的主要原因，采用实际车辆构建试验平台并开展机动车驾驶行为分析、辨识、预测是当前道路交通安全领域研究的热点。本研究以实车驾驶试验平台为依托，以时间序列分析方法为理论依据，通过对车辆总线时序数据特性分析、驾驶行为辨识建模、模型参数估值等内容开展理论研究，达到利用车辆自有数据即可辨识出驾驶行为的目的。分析实车驾驶试验数据的系统特性、统计特性、频谱特性和动态特性，是从系统辨识的角度确立驾驶行为自回归滑动平均模型结构的前提，利用递推最小二乘法的模型参数实时估值方法对模型参数进行自适应修正，可满足不同驾驶个体行为辨识的要求。本研究尝试从系统辨识角度对驾驶行为进行剖析和转化利用，综合运用了交通工程学、车辆工程学、数理统计等理论与方法，在学术上可促进信息科学与交通工程的学科交叉。在应用方面可为研发基于车辆总线数据的车载安全驾驶预警系统提供新思路和理论依据。

本项目尝试从系统辨识角度对驾驶行为进行剖析和转化利用，完成了以下研究内容：(1)驾驶行为实车试验数据采集系统研制。设计开发出以车载工控机为运算核心，以独立储能电池为供电系统、以低压触摸显示屏为人机交互外部设备的便携式车载实验平台。(2)面向驾驶行为辨识的实车驾驶试验设计、实施与数据采集。以城市道路交叉口附近驾驶人行为研究为重点，设计并实施驾驶人在多种驾驶意图导向下控制车辆运行的时间序列数据试验，为面向辨识的驾驶行为建模与预测提供数据支撑。(3)基于分形理论的驾驶行为数据特征分析。在区分不同驾驶行为状态方面，利用分形谱函数曲线散度指数，根据分解得出的特定层次特征频率或其他参数来进行区分。(4)基于贝叶斯网络的驾驶行为状态分析。采用节点及其概率网络来表征驾驶行为状态之间的映射关系，采用WinMine Toolkit软件快速构建驾驶行为贝叶斯网络；对生成的网络模型文件进行准确度评估与优化，运用贝叶斯推理算法分别计算单个观测节点的后验概率分布。(5)基于自回归滑动平均(ARMA)预测模型的交叉口车辆碰撞风险评估。利用车速预测值推算冲突方向车辆在交叉口内的行驶位移及其车间距离，评估车辆发生碰撞的风险。(6)直道超车过程大角度快速转向安全性分析。融合轮胎魔术公式和三自由度车辆动力学模型，建立小型轿车直道超车运动学模型。利用Matlab软件对提出的直道超车运动学模型进行仿真，结合系统仿真结果和试验数据变化曲线，研究45-65km/h车速范围内大角度快速转向过程的安全性。本项目的实施，丰富了面向安全预警的驾驶行为辨识研究方法与技术应用，对新型车载驾驶安全设备研发具有一定的启示作用。