大范围公路网交通态势估计驱动的非对称超大规模在线聚类技术研究

Nowadays, most of researches and applications on traffic situation assessment focus on the regional road network study where clustering technologies play an extremely important role in it. However, when we face the problem of traffic situation assessment for large-scaled road network, these existing clustering methods have some heavy shortcomings, such as low data throughput, very poor on-line clustering capability. To cope with these problems, this project aims to propose a very-large-scaled on-line clustering method, which is especially suitable for huge amounts of traffic situation flows. Based on the Fast Graph-based Relaxed Clustering (FGRC) algorithm proposed by the applicant, and by using the proposed incidence matrix approximation method, the applicant and his team plan to first develop the on-line FGRC algorithm, and then propose the consistent-set-based fast clustering aggregation algorithm. Based on these two novel algorithms, the final incidence matrix approximation and consistence-set fast clustering aggregation based very-large-scaled on-line graph-based relaxed clustering method will be also available. This novel method will be good at asymmetric information processing, very-large-scaled data throughput and on-line self-adaptive clustering. Furthermore, this proposed clustering method will provide a new way for effectively solving the scientific problems of situation monitoring, situation assessment, and emergency disposal for large-scaled road network. The project originates from the application researches of the applicant and his team, the outcomes of this project will be significant and meaningful to other related subjects as well. Therefore, the applicant believes that the study of this project has important academic and practical values in the fields of computational intelligence, pattern recognition and traffic management.

当前国内交通态势估计的研究与应用呈现出区域性之显著特点，聚类技术在其研究中占据着极其重要的作用。当面对大范围公路网交通态势估计研究时，现有的聚类方法会面临数据吞吐量弱、实时性和实用性差的困境。本课题拟针对此问题提出适合于海量交通态势流的超大规模在线聚类方法。课题组基于已有的快速图论松弛聚类算法，引入关联矩阵估计技术提出其在线图论松弛聚类版本，并在此基础上通过拟研究的基于一致集的快速聚类集成算法，提出本课题最终的基于关联矩阵估计和一致集快速聚类集成的超大规模在线图论松弛聚类方法。拟发展的此方法将具有非对称态势信息处理、超大规模数据吞吐量和在线自适应聚类等优点，并试图为解决大范围公路网交通态势的监测、估计和应急处理之科学问题提供新途径。本课题所要解决的问题来源于申请者及课题组的工程实践，但也具有普适意义，其成果对于计算智能、模式识别和交通管理等领域均具有重要的学术和应用意义。

在2013.01-2015.12执行国家自然科学基金（61202311）过程中，按计划开展了相关的研究，重点是为解决交通流状态监测与评估而进行适用的聚类、分类、大尺度数据处理技术和方法研究及应用。概括起来，本课题的成果和研究内容可归纳为如下的四个方面：1）基于知识迁移的学习技术研究。重点研究并提出了基于类原型与模糊隶属度联合杠杆作用的极大熵聚类技术、知识迁移模糊C均值聚类技术等，以及基于这些技术后继进行的路网节点交通流量状态监测实践研究。2)基于半监督的学习技术研究。重点研究并提出了归一化亲和度和惩罚项联合约束谱聚类算法、基于流形与成对约束联合正则化的半监督分类算法等，以及基于这些技术后继进行的路网节点交通流量状态监测实践研究。3)基于多视角的学习技术研究提出了熵加权多视角协同划分模糊聚类算法。4)大规模数据集学习技术研究。提出了基于最小包含球的异质空间大数据集快速相似度学习算法、基于共同决策方向矢量的多源迁移及其快速学习方法、大规模数据快速自适应同步聚类方法等一些适合大规模数据集学习的技术，为大范围公路网交通流状态监控提供了有效技术支撑。