【作者】 张琨；

【导师】 崔胜民；

【作者基本信息】 哈尔滨工业大学 ， 车辆工程， 2013， 博士

【摘要】 智能汽车一直是现代汽车研究领域的热点和难点，伴随着控制理论的发展，越来越多新的控制理论和控制方法被应用于智能汽车的自主循迹控制，这使得如何根据不同的道路环境和行驶工况选择最适合的控制方法成为一门新的课题。本文在研究了智能汽车几种自主循迹横向控制方法和纵向控制方法的基础上，提出了智能汽车自适应径向基函数(Radial Basis Function, RBF)神经网络补偿横向控制方法和两类改进的纵向控制方法，并针对不同的行驶工况提出一种多控制方法变换策略，通过实验验证了控制策略的有效性。研究了目前国内外智能汽车自主循迹横向控制中常用的汽车转向几何学模型、汽车运动学模型和汽车动力学模型。针对各模型的特点，研究了非预瞄汽车转向几何学模型横向控制方法、基于预瞄的汽车转向几何学模型横向控制方法、汽车运动学模型光滑时变反馈横向控制方法、汽车动力学模型最优线性二次型调节(Linear Quadratic Regulator, LQR)横向控制方法、汽车动力学模型前馈最优LQR横向控制方法和基于预瞄的汽车动力学模型最优LQR横向控制方法，通过双移线仿真试验和圆形弯道仿真试验，从控制器的鲁棒性、对道路的特殊要求、转向超调量、稳态循迹误差以及各自的适宜场合总结了6种方法的优点与不足，为多控制方法的变换奠定了理论基础。针对名义汽车动力学模型在非线性区失效的问题，提出了智能汽车自主循迹自适应RBF神经网络补偿横向控制方法。分别研究基于模型不确定部分、基于模型整体以及基于模型分块的自适应RBF神经网络补偿方法。使用Lyapunov函数方法保证系统的稳定性，通过双移线仿真试验、圆形弯道仿真试验和连续弯道仿真试验比较3种补偿方法的优劣，最终选择基于模型不确定部分的自适应RBF神经网络补偿控制方法作为智能汽车在非线性区的自主循迹横向控制方法。在分析了智能汽车4种自主循迹纵向控制模型的基础上，提出2类改进的智能汽车自主循迹纵向控制方法，分别是数据拟合纵向控制方法和模糊神经网络(Fuzzy Neuron Network Control, FNNC)纵向控制方法。仿真试验结果表明，FNNC纵向控制方法在控制精度上要优于数据拟合纵向控制方法，但其输出的期望车速信号变化过于灵敏且存在一定程度的噪声污染。对此，使用db4小波强制阈值降噪方法对FNNC输出的期望车速信号进行处理，取得了良好的效果。提出了学习向量量化(Learning Vector Quantization, LVQ)神经网络行驶工况分类方法，多组试验结果表明，LVQ神经网络可以对复杂的行驶工况进行有效地分类。控制方法变换策略的理念是针对不同的LVQ行驶工况分类结果，选择与之最适合的控制方法，在保证循迹控制精度的前提下尽可能选择运算量小的控制方法。引入人—车闭环操纵性评价指标对所设计的智能汽车自主循迹控制器的控制效果进行评价，评价结果表明控制器的控制效果良好，但操纵负担略重。从驾驶模拟实验平台的工作原理、硬件组成和软件设计搭建了驾驶模拟实验平台。借助于CCD图像采集设备和车身状态数据采集设备，将控制器的仿真结果与实车实验数据进行对比，验证了控制器的有效性。更多还原

【Abstract】 Intelligent vehicle has always been the hotspot and difficulty point in the fieldof modern automobile research. With the development of control theory, more andmore new control theory and control method was applied to path tracking control ofthe intelligent vehicle, which made the question how to choose the most suitablecontrol method according to different road condition and driving condition become anew subject. An intelligent vehicle’s lateral control method based on RBF (RadialBasis Function) neuron network compensation and two kinds of improvedlongitudinal control method are proposed based on the research on several commonused methods in intelligent vehicle’s path tracking control. Additionally, a controlmethod of switching strategies based on different driving conditions is put forward,whose effectiveness is verified by several experiments.The geometic steering model, kinematic vehicle model and dynamic vehiclemodel that commanly used in intelligent vehicle’s path tracking lateral control arestudied.The non-preview geometric steering model lateral control method, thepreview geometric steering model lateral control method, the kinematic vehiclemodel smooth time-varying feedback lateral control method, the dynamic vehiclemodel based on optimal LQR (Linear Quadratic Regulator) method, the LQRmethod with a feed-forward term and the preview optimal LQR method up to6kinds of common control method in intelligent vehicle’s tracking control are studied.Through double lane change test and the circle course test, the respective advantagesand disadvantages of these6methods are studied, including: the controller’srobustness, special requirements of roads, the overshoot, the steady-state lateraltracking error, and their suitable situations, which has been the theoretical basis ofthe multiple control method switching theory.In order the to solve the lackness of the name dynamic vehicle model innonlinear area, a method based on RBF neural network compensation of theintelligent vehicle’s lateral path tracking control is raised.3kinds of RBF networkcompensation methods are studied, separetly based on model uncertain parts, modelpartitioned parts and model united parts. The stability of the system is analyzedbased on lyapunov function method, the advantages and disadvantages of these threekinds of compensation methods are compared through double lane change test,circle course test and a multi curvature road test. The RBF network compensationcontrol method based on model uncertain parts is finally chosen as the lateralcontrol method of an intelligent vehicle in nonlinear area.Two kinds of intelligent vehicle’s longitudinal tracking control methods which separately based on data fitting and FNNC (Fuzzy Neuron Network Control) areproposed on the basis of the analysis of those several common used intelligentvehicle’s speed control model. The data analysis shows that the longitudinal controlmethod based on FNNC is superior to longitudinal control method based on datafitting with respected to the control accuracy, but the output desired speed is toosensitive to the of change the input value and there is a certain degree of noisepollution. A method based on wavelet denoising is proposed in order to handle theoutput speed signal of FNNC, which has obtained good effect.A driving conditions classification method based on LVQ (Learning VectorQuantization) neuron network is raised, several groups of test data show that theLVQ neural network can effectively classify different driving conditions. Thecontrol method switch strategy’s theory is selecting different corresponding controlmethods according to different classification results of LVQ, trying to selectmethods with less calculation on the premise of guarantee the tracking controlaccuracy. The driver-vehicle closed-loop maneuverability evaluation index isintroducted in order to evaluate the designed longitudinal and lateral integratedcontroller of the intelligent vehicle’s path tracking. According to the evaluation testresults of the double lane change test and the serpentine test, the controller is withfarely good control effectiveness, but also with a heavy burden.The build process of the driving simulation platformis introduced from theworking theroy, hardware composition and software design aspects. With the aid ofCCD image acquisition equipment and the vehicle state acquisition equipment, theeffect of the controller is veritied through a real vehicle test, which shows that thetest results and simulation results are basically identical.更多还原