基于改进K-Medoids的组合聚类算法及异常值检测研究  

Research on combinatorial clustering algorithm and anomaly detection based on improved K-Medoids

作　　者：贺玉海[1,2,3] 周庆琨 程埮晟 王勤鹏 HE Yuhai;ZHOU Qingkun;CHENG Tansheng;WANG Qinpeng(School of Naval Architecture, Ocean and Energy Power Engineering, Wuhan University of Technology, Wuhan 430063, China;Key Laboratory of Ship Power Engineering Technology Transportation Industry, Wuhan University of Technology, Wuhan 430063, China;Electronic Control Sub Laboratory of National Engineering Laboratory of Naval Architecture and Ocean Engineering Power Systems, Wuhan University of Technology, Wuhan 430063, China)

机构地区：[1]武汉理工大学船海与能源动力工程学院,湖北武汉430063 [2]武汉理工大学船舶动力工程技术交通行业重点实验室,湖北武汉430063 [3]武汉理工大学船舶与海洋工程动力系统国家工程实验室电控分实验室,湖北武汉430063

出　　处：《大连理工大学学报》2022年第4期403-410,共8页Journal of Dalian University of Technology

基　　金：国家自然科学基金资助项目(51009112).

摘　　要：采用聚类算法和异常值检测算法进行车辆轨迹信息的提取与挖掘,在交通控制与管理、道路拥堵时空分析与治理、用户出行线路规划与推荐,以及自动驾驶决策规划等应用中具有重要意义.针对现有聚类算法和异常值检测算法参数难以控制、算法存在随机性的不足,提出基于K-Medoids与DBSCAN组合的聚类算法.通过对模拟十字交叉路口数据集的训练,得到一个交叉路口最佳聚类模型,并用真实轨迹数据集验证、优化该模型.然后,将交叉路口区域内一段时间内的轨迹聚类数据流进行可视化再现,取得了异常轨迹少、聚类速度快的聚类效果,同时比较选择出算法各个参数的最优值.最后,通过参数传递使DBSCAN算法能够更精确地识别出异常轨迹,为交通治理与自动驾驶决策提供指导.The extraction and mining of vehicle trajectory information using clustering algorithm and anomaly detection algorithm are of great significance in applications such as traffic control and management,spatial and temporal analysis and management of road congestion,user travel route planning and recommendation,and autonomous driving decision planning.A clustering algorithm based on a combination of K-Medoids and DBSCAN is proposed to address the shortcomings of existing clustering algorithms and anomaly detection algorithms,which are difficult to control the parameters and have randomness.Through training on simulated four-exit intersection datasets,an optimal clustering model for intersections is obtained,and the model is validated and optimized with real trajectory datasets.Then,the trajectory clustering data flow in the intersection area over some time is reproduced visually,and the clustering effect of fewer abnormal trajectories and faster clustering is achieved,while the optimal values of each parameter of the algorithm are selected by comparison.Finally,the parameter transfer enables the DBSCAN algorithm to identify the abnormal trajectories more accurately and provide guidance for traffic management and autonomous driving decisions.

关 键 词：车辆轨迹 聚类分析 异常值检测 相似性度量 DBSCAN算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]