【作者】 赵伟；

【导师】 都雪静；

【作者基本信息】 东北林业大学 ， 交通运输工程， 2025， 博士

【摘要】 智能车辆的大规模应用为提升道路交通安全和通行效率提供了重要保障。然而,在寒区冰雪环境下,由于路面摩擦系数降低、车辆操控稳定性下降及驾驶决策不确定性增加,智能车辆的运行安全面临严峻挑战,尤其是在换道过程中,风险更为突出。统计数据显示,冰雪环境下交通事故的发生率是干燥路面的5倍,且70%以上的交通事故涉及换道操作,如何有效控制换道风险已成为寒区道路交通安全防控工作的关键问题。换道作为驾驶过程中最常见且高风险的操作,其在冰雪路面环境下的特征规律尚不明晰,换道风险的识别、预测、评估及控制面临诸多挑战。因此,系统研究寒区冰雪路面智能车辆换道行为特征,构建精准的换道风险识别、预测与评估模型,并提出有效的换道风险控制策略,对于降低交通事故发生率、提升行车安全性与效率,以及保障冰雪道路交通系统的稳定运行具有重要理论价值和实际意义。本文以高速公路驾驶模拟数据为基础,分析冰雪路面智能车辆的换道行为特征,运用统计学和机器学习方法筛选关键的换道风险表征指标,并提出一种基于Stacking集成学习算法的冰雪路面智能车辆换道风险识别框架;采用基于注意力机制和残差网络连接的双向长短时记忆(Bi-directional Long Short-Term Memory,Bi LSTM)神经网络识别驾驶意图,并基于此运用轻量梯度提升机(Light Gradient Boosting Machine,Light GBM)算法在换道操作前实现智能车辆换道风险的主动预测;此外,利用环境-注意力神经网络(Environmental-Attention Network,EA-Net)模型预测智能车辆换道轨迹,在此基础上结合风险场理论动态评估换道过程中智能车辆潜在风险;最后,综合考虑冰雪路面智能车辆换道时机和最优轨迹,应用神经网络和控制方法对换道风险进行有效控制。本文的主要研究内容包括以下几个方面:首先,本文基于高逼真度驾驶模拟平台,采集了高速公路冰雪和干燥路面条件下的车辆运行数据、交互数据以及驾驶员眼动和生理数据,并制定了换道事件的提取规则,最终从冰雪路面和干燥路面分别提取了1790和2185个换道事件。随后,从换道的时间特性、空间特性以及驾驶行为特性三个方面分析了冰雪路面换道行为特征,并确定了典型的换道风险。采用C4.5决策树算法和相关性分析方法,筛选出了18个关键换道风险表征指标。基于Stacking集成学习算法构建了冰雪路面智能车辆换道风险识别模型,该模型的识别准确率达到了98.7%,显著优于其他传统模型。其次,为了实现换道操作前风险的主动预测,本文分析了冰雪路面驾驶意图阶段的驾驶行为与行车特性,并全面提取了驾驶意图特征指标,构建了基于注意力机制和残差网络连接的Bi LSTM驾驶意图识别模型,实现了驾驶意图的准确识别。通过对比不同换道意图时间窗下模型的识别效果,确定了最优时间窗长度。从时间和空间维度分别量化了换道风险,利用故障树分析法进行了综合风险量化,并通过K-means聚类算法将换道风险划分为低风险、中风险和高风险三类。在此基础上,基于Light GBM算法构建了换道风险主动预测模型,分析了不同换道模式下换道风险的预测结果。最后,通过消融实验验证了驾驶意图识别结果的引入能够显著提升换道风险预测的准确性。再次,为了实现换道过程中潜在风险的动态评估,本文从智能车辆运行特征和交互特征两个层面提取关键特征指标,构建了基于EA-Net的智能车辆换道轨迹预测模型,并结合混合密度网络估计车辆未来位置的概率分布。在此基础上,结合场强理论,构建了基于轨迹预测概率密度的风险场数学模型,通过对换道场景中相邻车辆的风险场重叠体积进行积分,量化换道过程中车辆间的冲突场强度,实现智能车辆换道潜在风险的动态评估。为了提高计算效率,采用了马尔可夫链蒙特卡洛算法优化了冲突场的计算过程。最后,通过实车实验数据验证了本文提出的换道风险评估模型的优越性。结果表明,该模型可以同时捕捉车辆在纵向和横向的多个冲突区域,能够动态评估车辆沿行驶轨迹运动过程中风险场的变化,相较于传统冲突指标(1/PET、1/MTTC和DRAC)评估方法,能够更精确地捕捉复杂多车交互下的风险动态变化。最后,为了确保智能车辆在冰雪路面复杂行车环境下安全换道,本文提出了一种综合考虑换道时机与最优轨迹的换道风险控制策略。采用BP神经网络实时预测换道执行时间,并结合换道综合安全距离确定换道时机,进而确定换道起始点。在此基础上,结合换道风险评估结果,采用多项式算法生成了初始换道轨迹,构建了多目标优化函数,并采用改进萤火虫算法对换道轨迹进行了优化。采用模型预测控制算法对最优换道轨迹进行了跟踪控制,确保智能车辆能够精准执行换道操作。最后,通过Carsim-Simulink联合仿真平台开展了换道风险控制实验。结果表明,车辆侧偏角和偏航率保持在较低水平,确保了换道过程中的偏航稳定性;同时最大跟踪误差小于0.3m,表明其具备较高的跟踪精度,能够有效控制冰雪路面智能车辆的换道风险。本文构建了冰雪路面智能车辆换道风险“识别-预测-评估-控制”的全过程研究体系,研究成果可为智能交通系统、自动驾驶等新兴技术在寒区冰雪环境下的应用提供理论支撑,为道路交通安全主动防控提供科学依据,并为冰雪环境下智能车辆的安全行驶奠定技术基础。更多还原

【Abstract】 The large-scale deployment of intelligent vehicles provides critical technological support for enhancing road traffic safety and operational efficiency.However,in cold regions with icy and snowy conditions,ensuring the safe operation of intelligent vehicles remains a significant challenge due to the reduced road friction coefficient,diminished vehicle handling stability,and increased uncertainty in driver decision-making.These risks are particularly pronounced during lane-changing maneuvers.Statistical data indicate that the incidence of traffic accidents in icy and snowy environments is five times higher than on dry roads,with over 70%of these accidents involving lane-changing operations.Effectively mitigating lane-changing risks has thus become a key challenge in traffic safety management for cold-region roads.Lane-changing,as one of the most frequent yet high-risk driving maneuvers,remains inadequately understood under icy and snowy conditions,posing substantial challenges in risk identification,prediction,assessment,and control.Therefore,systematically analyzing the lane-changing behavior of intelligent vehicles on icy and snowy roads,developing precise models for risk identification,prediction,and assessment,and formulating effective risk control strategies hold significant theoretical and practical value.These efforts contribute to reducing accident rates,enhancing driving safety and efficiency,and ensuring the stable operation of traffic systems in snowy and icy environments.Based on highway driving simulation data,this paper investigated the lane-changing behavior of intelligent vehicles in icy and snowy conditions.Statistical and machine learning methods were employed to identify key lane-changing risks characterization indicators,and a lane-changing risk identification framework was developed using a Stacking ensemble learning algorithm.To recognize driving intentions,a Bi-directional Long Short-Term Memory(Bi LSTM)neural network incorporating attention mechanisms and residual connections was implemented.Based on this,a Light Gradient Boosting Machine(Light GBM)algorithm was employed to proactively predict lane-changing risks before the maneuver was executed.Additionally,an Environmental-Attention Network(EA-Net)neural model was utilized to predict the lane-changing trajectory of intelligent vehicles.A risk field theory-based model was then formulated to dynamically assess potential risks during the lane-changing process.Finally,by comprehensively considering optimal lane-changing timing and trajectory,neural networks and control methodologies are integrated to effectively mitigate lane-changing risks.The main research contents include the following aspects:Firstly,a high-fidelity driving simulation platform was utilized to collect vehicle operation data,interaction data,along with driver eye movement and physiological data under highway conditions on both icy and snowy and dry roads.Extraction rules for lane-changing events were established,identifying 1,790 and 2,185 lane-changing events on icy and snowy and dry roads,respectively.Subsequently,the characteristics of lane-changing behavior on icy and snowy roads were analyzed from three perspectives:temporal,spatial,and driving behavior characteristics,leading to the identification of typical lane-changing risk.18 key lane-changing risk characterization indicators were selected using the C4.5 decision tree algorithm and correlation analysis.A lane-changing risk identification model for intelligent vehicles on icy and snowy roads was developed using the Stacking ensemble learning algorithm,achieving an identification accuracy of 98.7%,significantly outperforming other conventional models.Secondly,to enable proactive risk prediction before lane-changing maneuvers,driving behavior and vehicle operation characteristics during the lane-changing intention phase under icy and snowy conditions were analyzed.A comprehensive set of driving intention feature indicators was extracted,and a Bi LSTM-based driving intention recognition model incorporating attention mechanisms and residual network connections was constructed,ensuring accurate intention recognition.By comparing the model’s recognition performance across different lane-changing intention time windows,the optimal time window length was determined.Lane-changing risks were quantified from both temporal and spatial dimensions,and a comprehensive risk assessment was conducted using fault tree analysis.Additionally,the K-means clustering algorithm was employed to classify lane-changing risks into three levels:low,medium,and high.On this basis,a proactive lane-changing risk prediction model was developed using the Light GBM algorithm,and the predictive performance of lane-changing risks under different lane-changing patterns was analyzed.Finally,ablation experiments confirmed that incorporating driving intention recognition significantly improves the accuracy of lane-changing risk prediction.Thirdly,to achieve dynamic assessment of potential risks during lane-changing maneuvers,key feature indicators were extracted from two levels:vehicle operation characteristics and interaction characteristics.Based on these features indicators,an intelligent vehicle lane-change trajectory prediction model was constructed using the EA-Net neural network,while the probability distribution of the vehicle’s future position was estimated using the Mixture Density Network(MDN).On this basis,a risk field mathematical model based on trajectory prediction probability density was developed by incorporating field theory.The conflict field intensity between vehicles during lane-changing was quantified by integrating the overlapping volume of risk fields of adjacent vehicles in lane-changing scenarios,enabling the dynamic assessment of potential lane-changing risks for intelligent vehicles.To improve computational efficiency,the Markov Chain Monte Carlo(MCMC)algorithm was employed to optimize the calculation process of the conflict field.Finally,real vehicle experimental data were used to validate the superiority of the proposed lane-changing risk assessment model.The results indicate that the model can simultaneously capture multiple conflict zones in both longitudinal and lateral directions and dynamically assess changes in the risk field along the vehicle’s trajectory.Compared to traditional conflict indicators(1/PET,1/MTTC,and DRAC),the proposed approach can more accurately capture dynamic risk variations under complex multi-vehicle interactions.Finally,to ensure the safe lane-changing of intelligent vehicles under the complex driving environment on icy and snowy roads,an lane-changing risk control strategy was proposed,integrating both lane-changing timing and optimal trajectory.Lane-changing execution time was predicted in real-time using a BP neural network,while the comprehensive lane-changing safety distance was applied to determine the lane-changing timing and starting point.Based on the lane-changing risk assessment results,an initial lane-changing trajectory was generated using a polynomial algorithm,and a multi-objective optimization function was formulated.The improved firefly algorithm was then employed to optimize the lane-changing trajectory.To ensure precise execution of the lane-changing maneuver,a model predictive control algorithm was implemented for optimal trajectory tracking.Finally,a trajectory tracking experiment was conducted on the Car Sim-Simulink co-simulation platform.The results demonstrate that the vehicle’s sideslip angle and yaw rate remained at low levels,ensuring yaw stability during the lane-changing process.Additionally,the maximum tracking error was less than 0.3 m,indicating high tracking accuracy and effective control of intelligent vehicle lane-changing risk on icy and snowy roads.This paper has established a comprehensive research framework for the"identification-prediction-assessment-control"process of lane-changing risk for intelligent vehicles on icy and snowy roads.The findings provide theoretical support for the application of emerging technologies such as intelligent transportation systems and autonomous driving in cold-region environments,offer scientific evidence for proactive road traffic safety management,and lay a technical foundation for the safe operation of intelligent vehicles in icy and snowy conditions.更多还原