欢迎来到天天文库
浏览记录
ID:37379875
大小:447.42 KB
页数:11页
时间:2019-05-22
《基于半监督学习的行为建模与异常检测》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、维普资讯http://www.cqvip.comISSNl000—9825.C0DENRUXUEWE-mail:jos@iscas.ac.cnJournalofSoftware,Vo1.18,No.3,March2007,PP.527—537http://www.jos.org.cnDOI:10.1360~os180527Tel,lax:+86一l0-62562563@2007byJournalofSoftware.Allrightsreserved.基于半监督学习的行为建模与异常检测李和平,胡占义+I吴毅红,吴福朝(中国科学院自动化研究所
2、模式识别国家重点实验室,北京100080)BehaviorModelingandAbnormalityDetectionBasedonSemi-SupervisedLearningMethodLIHe.Ping,HUZhan—Yi,WUYi—Hong,WUFu—Chao(NationalLaboratoryofPa~ernRecognition,InstituteofAutomation,TheChineseAcademyofSciences,Bering100080,China)+Correspondingauthor:Plan:+86—
3、10-62616540,E-mail:huzy@nlpr.ia.ac.caLiHP,HuZY,WuYH,WuFC.Behaviormodelingandabnormalitydetectionbasedonsemi-supervisedlearningmethod.JournalofSoftware,2007,18(3):527-537.http://www.jos.org.cn/lO00—9825/18/527.htmAbstract:Asimpleandeficientmethodbasedonsemi—supervisedlearni
4、ngtechniqueisproposedforbehaviormodelingandabnormalitydetection.Themethodiscomposedofthefollowingsteps:(1)Dynamictimewarping(DTVObasedspectralclusteringmethodisusedtoobtainasmallsetofsamplestoinitializethehiddenMarkovmodels(HMMs)ofnorma1.behaviors;(2)TheHMMs’parametersaref
5、urthertrainedbythemethodofiterativelearningfromalargedataset;(3)Maximumaposteriori(MAP)adaptationtechniqueisusedtoestimatetheHMMs’parametersofabnormalbehaviorsfromthoseofnormalbehaviors;(4)ThetopologicalstructureofHMMisfinallyconstructedtodetectabnormalbehaviors.Themaincha
6、racteristicoftheproposedmethodisthatitCanautomaticallyselectthenumberofnormalbehaviorpatternsandsamplesfromthetrainingdatasettobuildnormalbehaviormodelsandCanefectivelyavoidtherunningriskofover-fittingwhentheHMMsofabnormalbehaviorsarelearnedfromsparsedata.Experimentalresul
7、tsdemonstratetheeffectivenessoftheproposedmethodincomparisonwithotherrelatedworksintheliterature.Keywords:behaviormodeling;abnormalitydetection;semi—supervisedlearning;HMM(hiddenMarkovmodels);computervision摘要:提出了一种基于半监督学习的行为建模与异常检测方法.该算法包括以下几个主要步骤:(1)通过基于动态时间归整(DTW)的谱聚类方法获
8、取适量的正常行为样本,对正常行为的隐马尔可夫模型(HMM)进行初始化;(2)通过迭代学习的方法在大样本下进一步训练这些隐马尔可夫模型参数;(3)以监督的方式,利用最大后验(MAP
此文档下载收益归作者所有