个性化驾驶员模型及其在驾驶行为评估中的应用

发布时间：2018-01-11 07:29

本文关键词：个性化驾驶员模型及其在驾驶行为评估中的应用　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：车辆测试是车载控制系统研发的重要环节。建立合适的驾驶员模型取代真实的驾驶员进行车辆测试,可以大大缩短车载控制系统的研发周期。由于驾驶风格会直接影响测试结果,因此,建立能够体现驾驶风格的驾驶员模型成为研究的主要内容。本课题提出通过模仿真实驾驶行为建立个性化驾驶员模型的思路和方法。所建立的个性化驾驶员模型能在保留驾驶风格的前提下实现较好的速度跟随性能。考虑到实际汽车测试数据的高度离散性和突变性,本课题采用学习控制的直接逆模型法,再结合神经网络实现该模型。并用此模型代替驾驶员或常规的PID模型进行联邦测试循环(Federal Test Procedre,FTP)标准工况的实验室测试。以VTD以及FTP标准为背景的测试结果和企业应用验证所提建模方法的有效性。随后在此基础上对驾驶员驾驶风格量化分析和异常驾驶行为检测进行了初步的探索。论文的主要内容包括:(1)基于驾驶行为模仿的个性化驾驶员建模。为了在车辆测试中融入不同的驾驶风格而使测试结果更接近于真实,提出一种基于汽车测试数据和神经网络的个性化驾驶员建模方法。考虑到实际数据的高度离散性和局部突变性,以局部性神经网络的典型代表小脑模型关节控制器(Cerebellar model articulation controller,CMAC)网络为基础,采用学习控制的直接逆模型法来建立驾驶员模型,并用此模型代替驾驶员或常规的PID模型进行联邦测试循环(Federal Test Procedure,FTP)标准工况的实验室测试。另一方面,为方便起见,建立并采用基于神经网络和汽车测试数据的汽车模型来代替真实汽车进行模拟测试。以VTD以及FTP标准为背景的测试结果验证了所提方法的有效性。采用个性化驾驶员模型能够在体现驾驶风格情况下完成标准工况测试,使实验室测试更接近于真实。(2)基于相空间重构的驾驶风格定量评估。首先,对驾驶行为进行标准化处理以消除环境因素的影响;接着,将标准化后的驾驶行为进行相空间重构分析,提出一种基于关联维数的驾驶风格指数,用于定量评估驾驶的激进程度;最后,将驾驶风格指数应用于驾驶风格的识别,仿真结果验证了所提方法的有效性。(3)基于非监督特征学习的异常驾驶行为检测。首先,基于驾驶行为标准化系统和已有研究,对不同的异常驾驶行为进行分析以及仿真;然后,基于稀疏自编码器神经网络建立异常驾驶行为检测系统。为了获取更具鲁棒性的特征,在稀疏自编码器的基础上融入去噪编码方法。同时,在整个网络中加入"Dropout"训练技术,以减少因过拟合而带来的预测误差。最后,采用对比试验验证所提方法的有效性。
[Abstract]:Vehicle testing is an important part of vehicle control system research and development. Establish appropriate driver model instead of real drivers to carry out vehicle testing. Can greatly shorten the vehicle control system research and development cycle. Because driving style will directly affect the test results, so. It is the main content of the research to establish the driver model which can embody the driving style. This paper puts forward the idea and method of establishing the individualized driver model by imitating the real driving behavior. It can achieve better speed following performance on the premise of keeping driving style. Considering the high discreteness and mutation of the actual vehicle test data. This paper adopts the direct inverse model method of learning control. Combining with neural network, the model is implemented, and the model is used to replace the driver or the conventional PID model to carry on the federal test cycle and the Federal Test Procedre. FTP). Laboratory testing under standard operating conditions. Test results based on VTD and FTP standards and enterprise applications to verify the validity of the proposed modeling methods. Then, quantitative analysis and anomaly of driver driving style are made on this basis. The main contents of this paper are as follows:. (. 1) individualized driver modeling based on the imitation of driving behavior. In order to incorporate different driving styles into the vehicle test, the test results are more realistic. This paper presents a personalized driver modeling method based on vehicle test data and neural network, which takes into account the high discreteness and local mutation of the actual data. Cerebellar model articulation controller Cerebellar model articulation controller is represented by local neural network. Based on CMAC network, the direct inverse model of learning control is used to establish the driver model. This model is used to replace the driver or the conventional PID model for the Federal Test Procedure. FTP standard operating conditions for laboratory testing. On the other hand, for convenience. The vehicle model based on neural network and vehicle test data is established and used to replace the real vehicle for simulation test. The test results based on VTD and FTP standards verify the effectiveness of the proposed method. The individualized driver model can complete the standard working condition test under the condition of embodying the driving style. Make the laboratory test more close to the real. (2) quantitative evaluation of driving style based on phase space reconstruction. Firstly, the driving behavior is standardized to eliminate the influence of environmental factors; Then, the phase space reconstruction analysis of the standardized driving behavior is carried out, and a driving style index based on correlation dimension is proposed, which is used to quantitatively evaluate the degree of radicalization of driving. Finally, the driving style index is applied to the recognition of driving style. The simulation results show that the proposed method is effective and based on unsupervised feature learning to detect abnormal driving behavior. Based on the standardized driving behavior system and existing research, different abnormal driving behaviors are analyzed and simulated. Then, based on the sparse self-encoder neural network, the abnormal driving behavior detection system is established. In order to obtain more robust features, the sparse self-encoder is integrated into the denoising coding method. At the same time. The "Dropout" training technique is added to the whole network to reduce the prediction error caused by over-fitting. Finally, the effectiveness of the proposed method is verified by a comparative experiment.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U491.25

【相似文献】