【作者】 覃明贵；

【导师】 朱扬勇；

【作者基本信息】 复旦大学 ， 计算机软件与理论， 2010， 博士

【摘要】 智能交通系统是有效地集成信息技术、数据通讯技术、电子传感技术、电子控制技术以及计算机数据处理技术的地面运输管理体系,是当前研究与应用的热点。其中,大规模交通数据管理、整合和挖掘是一项关键技术。数据挖掘是从大量数据中寻找其规律的技术,是目前最强有力的计算机数据分析技术之一。交通数据挖掘技术的研究是智能交通技术和数据挖掘技术领域最活跃的研究方向之一。交通数据挖掘的主要目的是寻找交通数据中的规律,为智能交通系统的设计提供技术支持,有利于缓解交通拥挤、优化交通路网运行,促进交通健康稳定发展。交通流量、交通拥堵状况和交通流分布预测和分析是目前智能交通数据挖掘研究中的三个重要问题,对于智能交通系统的交通信号管理与控制、交通流诱导、动态交通分配等方面有着重要的意义,在智能交通系统设计和实现中起着重要作用。当前智能交通数据挖掘研究的重点在于如何设计有效的挖掘算法,主要有两个方面的难题：一方面,由于交通流数据的特殊性,使得现有的数据挖掘算法无法直接在大规模交通流数据中高效实现；另一方面,由于没有根据领域知识设计专门的挖掘算法,造成挖掘结果无法满足应用需求。本文针对当前智能交通数据挖掘技术研究领域中存在的问题,在交通流量预测、交通拥堵事件挖掘和交通流分布模式挖掘等几个方面开展研究,提出了相应的挖掘算法,并将这些方法应用于智能交通数据挖掘系统中。本文取得的主要研究成果如下：1)针对短时十字路口交通流量预测问题设计实现了基于组合模型的挖掘算法及时、准确地识别和预测道路交通的状态是智能交通系统实现动态交通管理的重要前提。交通流量是交通流的重要特性之一,智能交通系统的控制和诱导需要对道路网络交通流量进行准确、快速的预测。本文针对路口短时交通流量预测问题,提出了基于交通流量序列分割和神经网络组合模型的交通流量预测算法CITFF (Combined Intersection Traffic flow Forcast), CITFF算法首先采用聚类方法对交通流量在流量大小和时间上进行序列分割,然后再采用神经网络对各个交通流模式进行描述和预测。实验证明基于组合模型的预测方法具有较高的预测精度。2)构建了道路交通流模式库并设计了相应的交通流拥堵事件挖掘算法如何应对城市现代化带来的交通拥堵问题,是交通管理者迫切需要解决的问题。道路交通的拥堵事件检测是智能交通领域研究的关键技术。本文通过对道路交通流数据的分析,构建了道路交通流模式库,并给出一个结合同向斜率树(Same-Directed Slope Tree, SDS-Tree)逐层分类表示交通流数据的方法。基于构建的交通流模式库,提出了一种高效的道路交通流的拥堵事件挖掘算法Detection-CS (Detection of Continual Stream of Traffic Flow),同时对算法效率和空间复杂度进行了详细分析。Detection-CS算法首先对当前实时交通数据进行特征提取,通过对交通流模式库进行匹配,获取前k个有效反馈,并根据反馈的交通路况信息进行分析,结合路况分层模式信息,给出当前路况的实时检测信息,实现对交通路况检测。为提高挖掘算法的效率,根据交通流模式库的路况分层信息,建立了多层索引结构,减少算法的搜索空间,从而实现算法优化。结合实际需要,算法进一步给出随着时间推移如何更新交通流模式库的方法,通过逐步替换使用频度最少的信息和更新新出现的路况信息,保证交通流模式库的有效性。在真实数据集上的实验表明,与现有算法相比,Detection-CS算法对于当前解决交通路况的实时检测具有很好的效率和较高的准确度。3)提出了一个道路交通流分布模式挖掘算法道路网络上运行的交通流具有不同的空间分布模式,根据交通流运行的空间分布特性,对道路交通网络进行实时、动态的交通区域划分是当前智能交通系统的研究热点之一。本文对分布在道路网络空间中的环形感应线圈检测器检测的交通流数据进行空间聚类分析,设计了一个高效的交通流空间聚类算法SPANBRE (Efficient Clustering Algorithm for Spatial Data with Neighborhood Relations),自底向上生成道路交通流的空间簇,使具有相似性质且具有空间关联性的交通流数据对象聚成一簇,用以发现道路交通流的空间分布模式。SPANBRE算法无需执行复杂的空间连接和空间合并操作,实验证明具有良好的时间效率。4)设计实现了一个基于数据挖掘技术的综合智能交通系统道路交通数据挖掘技术的研究对于智能交通系统的交通信号管理与控制、交通流诱导、动态交通分配等方面有着重要的意义。本文将上述挖掘方法应用于智能交通系统中,设计并实现了一个基于数据挖掘技术的综合智能交通系统。该系统已经实际获得应用,为道路交通管理提供了有效的工具。更多还原

【Abstract】 Intelligent Transportation System (ITS) is an integrated transportation and management system with high accuracy and efficiency, which integrates the advanced information technology, communication technology, electronic sensor technology, electronic control technology and computer data processing technology. In this research area, management, integration and mining of massive traffic data are key technologies. Data mining, one of the most powerful data analysis techniques, is an important technique when exploring the common rule from large volume data. The goal of the data mining in ITS is to mine the potential rule behind the traffic data and provide helpful guidance for the design of the ITS. So systems based on data mining can be used to alleviate traffic jam, optimize the traffic road network, and accelerate the traffic development healthily and steadily. The prediction and analysis of the traffic volume, traffic jam and traffic flow distribution are the three important questions of the traffic data mining research, which are significant to the traffic signal management and control, traffic flow inducement and dynamic traffic flow distribution and also play an important role in ITS design and implementation.How to design an efficient mining algorithm is now the key issue of the intelligent traffic data mining research, which involves two following aspects. The first difficulty is to apply the existing data mining algorithm directly to the mass of the traffic flow data for its specific characteristic. The second aspect is that the mining results cannot satisfy the application requirements for the lacking of domain knowledge. Aiming at these problems in the filed of ITS data mining, this thesis proposes the corresponding efficient mining algorithms and applies these algorithms to implement ITS to improve the performance of existing ITS by researching on the intersection traffic flow, traffic flow jam mining and traffic flow distribution pattern mining,. The achievements of this thesis are summarized as follows:1) Design the corresponding algorithm based on combined models through the analysis of the problem at the intersection traffic flow short-term predictionIt is an important premise of dynamic traffic management in ITS to identify and predict the status of the traffic flow instantly and accurately. Traffic volume is one of the main features of the traffic flow; therefore, it is necessary to predict the traffic volume in road network. Aiming at the problem, we propose a traffic volume prediction CITFF(Combined Intersection Traffic Flow Forecast) algorithm, which based on traffic volume sequence partition and neural network model, and divides the traffic volume into different patterns along the volume and time dimension by clustering, and then describes and predicts the traffic flow status according to these different patterns. The experiment results on real data sets demonstrate that our algorithm based on the combination model is much accurate.2) Construct a road traffic flow pattern database and design a traffic flow jam mining algorithmTraffic jam detection is the key technology in ITS research. By the analysis of the traffic flow data, we construct a traffic flow pattern database and propose a traffic flow data description based on the Same-Directed Slope Tree (SDS-Tree). Correspondingly, an efficient traffic flow jam mining algorithm named Detection-CS, which based on the traffic flow pattern database is proposed. Detection-CS(Detection of Continual Stream of Traffic Flow) algorithm extracts the feature of the real-time traffic data and obtains the first k effective feedback through matching the feature with the traffic flow pattern database. Detection-CS then points out the current traffic status According to the feedback. To improve the efficiency, we build a multilayer index structure based on the traffic flow layered information, which can minimize the searching space. This thesis also presents a method to update the traffic flow pattern database to ensure the pattern database efficiently by replacing the seldom-used traffic flow data with the new ones gradually. The experiments on the real data sets show that Detection-CS exhibits higher efficiency and superior accuracy compared with some famous algorithms.3) Analyze the time and space characteristics of the traffic flow data and propose a traffic flow distribution pattern mining algorithmThe traffic flow in the road network has different time and space distribution pattern, so now it is one of the hot topics in ITS to partition real-time and dynastically the traffic area in the road network. We design an Efficient Clustering Algorithm for Spatial Data with Neighborhood Relations (SPANBRE) through clustering the traffic flow data from the loop inductor coils distributing over the road network. SPANBER builds the traffic flow space cluster from bottom up to make the traffic flow data with similar characteristics and space relationship into one cluster. SPANBER can find out the space distribution pattern of the traffic flow. It needs no complex space connecting and combining operation, and the experiment shows SPANBER has high efficiency.4) Design an integrative Intelligent Transportation System based on the data mining techniqueThe study of the traffic flow data mining technology is meaningful to the traffic management and control, traffic flow induction, dynamic traffic allocation. In this thesis, we apply above mining algorithms to implement a comprehensive Intelligent Transportation System based on the data mining technique. It has been put into successful practice in a few projects in several cities from medium to large scale, and also provides an efficient tool for the traffic management.更多还原