【作者】 温娟；

【导师】 贺国光；

【作者基本信息】 天津大学 ， 系统工程， 2005， 硕士

【摘要】 随着高速公路的发展和大城市交通问题的日益严重,建立交通应急系统的要求也越来越迫切,而有效的交通事件检测是应急系统成功运行的第一步,也是关键之所在,所以对交通事件检测算法的研究就成为热点问题。本文试图把粗集理论和神经网络理论相结合,建立一种全新的粗集预处理数据的神经网络交通事件自动检测算法,主要研究内容包括以下几个方面:总结现有的交通事件自动检测算法,对其性能进行比较分析,并详细地列举出评价算法优劣的评价指标和评价准则,提出本文欲建立算法的基本原理。对粗集理论进行数据挖掘的若干问题进行了比较深入地研究,如属性特征的选择、连续属性离散化、和约简算法寻优等。对神经网络检测交通事件的几种算法也做了介绍和比较,详细研究了网络结构的选择和训练问题,为建立本文的算法提供了基础。提出了建立粗集预处理数据的神经网络交通事件自动检测算法,该算法充分发挥了粗集理论和神经网络理论各自的优势。神经网络具有并行处理、网络全局、信息分布存储等特点,可通过训练、学习产生一个非线性映射,自适应地对数据进行聚类,同时具有较好的抑制噪声干扰的能力和较强的鲁棒性。其缺点是当输入信息空间的维数较大时,网络不仅结构复杂,而且训练时间也很长。而粗集可对数据进行属性约简和值约简,消除样本中的噪声和冗余对象。这两者的结合不仅可以减小网络的规模,同时通过消除对象冗余可减少网络的训练和学习负担,还可以通过消除噪声提高神经网络预测的准确性。本文算法中的神经网络采用的是多层前向反馈神经网络,并引入了动量项,采用自适应的学习率,可以获得更好的性能。为检验算法的有效性,建立了事件仿真模型,获得所需数据,并用获得的交通流数据来训练和检测本文算法,对已有神经网络算法和传统检测算法与本文算法做了对比试验和分析,分析结果表明,本文算法可获得较高的检测率和较低的误报率,对评价指标的协调性也比其它算法更好。更多还原

【Abstract】 With the development of highway and traffic problem being serious day by day inbig cities, the demand of establishing Emergence Management Systems (EMS)becomes urgent. Effective detection for traffic incident is the first step and the keycomponent of operating the EMS successfully;therefore, the study on detectionalgorithm for traffic incident has become a hot issue. The study tries to combine therough set theory and neural network to establish a new automatic detection algorithmfor traffic incident using neural network based on rough set filtration. Main researchwork is listed as follows:The existing automatic detection algorithms for traffic incident are summarized,the performances of it are compared and analyzed, the index for evaluatingalgorithm’s performance is listed,and then the basic principle of the algorithm in thethesis is expounded.A number of problems in data mining using rough set had been studied deeply inthe thesis,for example, feature selection, consecutive feature discretization, seekingthe optimum deduction algorithm. Several algorithms using neural network to detecttraffic incident also had been introduced and compared. The structure’s selection andtraining of neural network were studied in particular, which were the basic ofestablishing the algorithm in the thesis.A new automatic detection algorithm for traffic incident using neural networkbased on rough set filtration was proposed, which has the advantages of neuralnetwork and rough set theory. Neural network has several characteristics, such asdisposing datum in parallel, aiming for overall function, storing informationdistributed and so on, which can produce a non-linear map by training and learning,cluster data adaptively, with the abilities of restraining the noise’s disturbance andgood robustness. Its shortcomings are that when the space dimension of the inputinformation is larger, not only complex the network’s structure is, but also the networkneed more time to train. Rough set can deduct feature and value of the data,eliminatethe noise and redundant targets in sample. The merge not only reduced the scale of thenetwork, lessened burden of training and learning by eliminating redundant targets,but also improved the accuracy of detection by eliminating noise. The network in thethesis has multi-layer feed forward architecture;the motion vector and the adaptivelearning rate are also adopted for better performance.In order to verify the effect of the algorithm proposed in the thesis, simulationmodel of traffic incident was established to obtain the traffic data which was used totrain and test the algorithm. Comparing and analyzing had been done among thealgorithm in the thesis, the existing neural network and traditional detectionalgorithms. The results show that the algorithm in the thesis has higher detection rateand lower false detection rate. Its coordination ability to evaluation index is alsobetter than other algorithms.更多还原