作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

ICV中联合驾驶风格的交通流数据有效性分析

  • 发布日期:2023-10-30

Validity Analysis of Traffic Flow Data for Combined Driving Style in ICV

  • Published:2023-10-30

摘要: 在智能网联汽车(ICV)中,提高驾驶数据的有效性是提升车辆安全性的基石。只有准确的、可靠的驾驶数据才能为车辆的安全性提供可靠的依据和支持。与传统的异常分析相比,ICV数据有效性分析面临着数据异常的多样性(传感器异常、驾驶行为、恶意篡改等)。如何将车辆自身数据特征、驾驶风格和交通流特征相结合,提供有效的数据异常检测方法,已成为智能网联汽车中新的问题。针对ICV系统,采用结合驾驶风格和交通流理论的方法,设计基于粒子群优化的TE-PSO-SVM数据有效性检算法,实现驾驶数据的有效检测。首先,定义驾驶风格识别系数Rad,设计驾驶风格量化模型;其次,建立交通流模型,由车辆状态数据融合驾驶风格和交通流理论通过LSTM网络对车辆速度预测;最后通过TE-PSO-SVM算法进行数据有效性检测。由于ICV数据的多样性,单一模型对多类型异常混合并存的场景中检测精度仍有局限问题,利用多个模型的优势构建模型池,并提出基于强化学习的模型选择算法(RLBMS)。通过对真实数据集highD的实验证明在不同噪声环境下TE-PSO-SVM算法模型的F1度量值相比于传统SVM模型平均提升约8.1个百分点;RLBMS算法模型在不同噪声环境下的F1度量值相比于模型池中检测率最高算法平均提高约1.7个百分点,进一步提高了数据有效性检测的准确率。

Abstract: In the intelligent Connected vehicle (ICV), improving the effectiveness of driving data is the cornerstone of improving vehicle safety. Only accurate and reliable driving data can provide a reliable basis and support for vehicle safety. Compared with traditional anomaly analysis, ICV data validity analysis faces a diversity of data anomalies (sensor anomalies, driving behavior, malicious tampering, etc.). How to combine the vehicle's own data characteristics, driving style, and traffic flow characteristics to provide an effective data anomaly detection method has become a new problem in intelligent networked vehicles. For the ICV system, a TE-PSO-SVM data validity detection algorithm based on particle swarm optimization is designed by combining driving style and traffic flow theory to realize effective detection of driving data. Firstly, the driving style recognition coefficient Rad is defined and the driving style quantitative model is designed. Secondly, a traffic flow model is established, which combines driving style and traffic flow theory with vehicle state data to predict vehicle speed through the LSTM network. Finally, the TE-PSO-SVM algorithm was used to detect the validity of the data. Due to the diversity of ICV data, the detection accuracy of a single model is limited in scenarios where multiple types of anomalies coexist. To make use of the advantages of multiple models, a model pool is constructed, and a model selection algorithm based on reinforcement learning (RLBMS) is proposed. Experiments on real data set highD show that the F1 metric value of the TE-PSO-SVM algorithm model is 8.1 percentage points higher than that of the traditional SVM model under different noise environments. Compared with the algorithm with the highest detection rate in the model pool, the F1 metric value of the RLBMS algorithm model in different noise environments is increased by about 1.7 percentage points on average, which further improves the accuracy of data validity detection.