车联网中数据融合的研究

发布时间：2018-06-20 14:49

本文选题：据融合 + 机器学习　；参考：《北京交通大学》2017年硕士论文

【摘要】：随着社会的发展,道路中机动车的数量也在迅速增加。与其密切相关的智能交通系统也在快速发展。但传统智能交通系统存在固有的缺点:道路交通相关信息只能在特定位置被检测,仅能获取类似交通流量等较为简要数据;路上固定基础设施需要耗费大力维护。已经存在的道路识别和预测方法存在以下不足:1)所利用的数据不够充足;2)输入特征基本是手工提取;3)所用的学习方法次优。本文着重研究车联网中大数据融合在智能交通服务中的具体应用…-道路故障检测系统。车联网作为物联网在智能交通系统中的具体应用,能够使道路中的众多车辆实现互通,不仅能够使车辆节点之间进行通信,车辆与路边单元之间也能进行通信共同实现车联网大数据环境下的新型智能交通信息体系。本文通过结合车联网中关键技术车载自组织网络(VANET,Vehicular Ad Hoc Network)来实现大数据技术在交通服务系统中的应用。接下来根据智能交通信息服务对实时性和准确性的首要需求,将系统工作过程分为数据收集和数据分析两个阶段进行研究分析。首先,深入研究了车联网中VANET在大数据环境下所面临的通信负载问题的相关解决方案,通过对车辆节点进行分簇,形成各个群组,设计相关通信协议以群组的传播方法收集数据。实现以低通信负载来接收道路中的车辆节点信息,避免了与道路中的每一车辆进行点对点通信,这一部分对应上文所说的数据收集部分。其次,在数据分析阶段利用机器学习方法实现相应交通服务功能。本文经过一系列对比分析后选择了支持向量机(SVM,Support Vector Machine)和深度神经网络(DNN,Deep Neural Network)两种方法,并对这两种算法进行相应改进使其适用于本文所应用的车联网的数据融合的实际问题中。在道路故障检测系统中,作为一个分类问题,分别应用SVM和DNN对道路故障进行识别检测;作为一个回归问题,用DNN对故障位置进行判断预测。最后,本文利用微型的交通软件VISSIM模拟产生实验所需的交通数据。并对数据进行处理以及归一化等操作,利用本文提出的方法对道路故障进行识别和位置预测,分析输入特征以及训练数据对系统判别精度的影响。并且对比了不同的DNN结构对结果的影响,分析其中的过拟合现象。实验结果表明,该模型切实可行,在故障识别问题中基本能够达到95%以上的检测率,位置估测也能够满足实际的应用。这样的实验结果可以作为智能交通服务应用功能实现的参考,并且其中位置估计对于道路交通管理和维护可以有良好的辅助作用。
[Abstract]:With the development of society, the number of motor vehicles in the road is also increasing rapidly. The intelligent transportation system, which is closely related to it, is also developing rapidly. However, the traditional intelligent transportation system has its inherent disadvantages: road traffic related information can only be detected in a specific location, and can only obtain relatively simple data such as traffic flow, and the road fixed infrastructure needs to be maintained with great effort. The existing road recognition and prediction methods have the following shortcomings: 1) the data used is not enough) the input feature is basically extracted by hand) the learning method used in this paper is suboptimal. This paper focuses on the specific application of large data fusion in intelligent transportation service. -Road fault detection system. As the concrete application of the Internet of things in the intelligent transportation system, the vehicle networking can make many vehicles in the road interoperate, not only make the communication between the vehicle nodes, The communication between the vehicle and the roadside unit can also realize the new intelligent transportation information system under the big data environment. In this paper, the application of big data technology in traffic service system is realized by combining with the key technology of vehicle network, vehicle Ad Hoc network. Then, according to the first requirement of real-time and accuracy of intelligent transportation information service, the working process of the system is divided into two stages: data collection and data analysis. First of all, the paper deeply studies the solutions to the communication load problem of VANET in the big data environment, and forms each group by clustering the vehicle nodes. Design related communication protocols to collect data by group propagation. A low communication load is implemented to receive the vehicle node information in the road, thus avoiding point-to-point communication with each vehicle in the road. This part corresponds to the data collection section mentioned above. Secondly, in the data analysis stage, the machine learning method is used to realize the corresponding traffic service function. After a series of comparative analysis, two methods, support Vector Machine (SVM) and Deep Neural Network (DNN), are selected in this paper, and these two algorithms are improved to be applicable to the practical problems of data fusion in the network of cars in this paper. In the road fault detection system, as a classification problem, SVM and DNN are used to identify and detect road faults, and as a regression problem, DNN is used to judge and predict the fault location. Finally, the traffic data needed in the experiment are generated by using the micro traffic software VisSIM. The method proposed in this paper is used to identify and predict the road faults, and the influence of input features and training data on the system discriminating accuracy is analyzed. The effects of different DNN structures on the results are compared, and the phenomenon of overfitting is analyzed. The experimental results show that the model is feasible, and the detection rate of more than 95% can be basically achieved in the problem of fault identification, and the location estimation can also meet the practical application. The experimental results can be used as a reference for the application of intelligent transportation services, and the location estimation can be helpful to traffic management and maintenance.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN929.5;U495

【参考文献】