异常检测综述  

Survey of outlier detection technologies

作　　者：陈斌[1,2] 陈松灿[1] 潘志松[3] 李斌[2] 

机构地区：[1]南京航空航天大学信息科学与技术学院,江苏南京210016 [2]扬州大学信息工程学院,江苏扬州225009 [3]解放军理工大学指挥自动化学院,江苏南京210007

出　　处：《山东大学学报（工学版）》2009年第6期13-23,共11页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金项目(60903130;60603029);江苏省自然科学基金项目(BK2007074)

摘　　要：异常检测旨在检测出不符合期望行为的数据,因而适合应用于故障诊断、入侵和欺诈检测以及数据预处理等多个领域.针对目前众多的专用和通用异常检测方法,本文侧重对基于统计的主流异常检测方法进行了回顾,力图提供一个新的结构化的异常检测方法的认识框架,并依据其监督和无监督学习算法的原理进行了简单分类,特别对部分异常检测方法间的等价性进行了深入探讨.Outlier detection aims to detect those data that significantly deviate from the expected behavior, and thus is widely applied in many fields, such as, machine fault detection, intrusion detection, fraud detection and data preprocessing. Hence, thence exist many generic and special algorithms for outlier detection under the unsupervised and supervised learning framework. But up to now, there still has been no clear classification in this aspect. To provide a structural view, the review of the state-of-the-art statistics-based methods for outlier detection was focusedon, and a simple classification was given in this aspect. Moreover, the equivalence between some outlier detectors in depth is particularly discussed.

关 键 词：异常检测 统计 监督学习 无监督学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]