广东工业大学毕业论文

计算机学院毕业论文II本科毕业设计（论文）BP神经网络的异常点检测应用可行性研究学院计算机学院专业软件工程年级班别2006级（4）班学号3106007039学生姓名蔡东赟指导教师王丽娟2010年5月IIIIV摘要异常点数据是指数据集中与众不同数据。这部分数据的量小，但是对于我们的日常生产生活的影响极大。因此，异常点检测被广泛应用于网络入侵检测，金融保险，天气预报以及新药研制等领域。相对于大量的正常数据挖掘而言，异常点检测被称作小模式数据挖掘。BP算法是一种常用的数据挖掘算法。但是BP算法进行实际数据的异常点数据挖掘过程中存在：实际数据的维数较高，存在冗余特征的干扰，以及在高维特征下，数据量不充分的问题。因此，本文分析BP神经网络处理各种数据的情况，并得到以下结果。（1）BP神经网络能够较好的分离特征单一的仿真数据；但是（2）特征相似性较大的数据集，难以分离判断；（3）正常数据不充分或者不具有代表性，因此正常数据类学习不充分，从而导致异常无法判断。针对以上问题，本文提出了以下的改进措施：（1）BP算法前进行特征约简（映射）从中选取有益于异常检测的特征（2）多神经网络融合，不同神经网络识别不同的特征，相互取长补短，融合后得到最终的结果。关键字：异常，BP，异常点检测，神经网络注：本设计（论文）题目来源于教师的国家级（或部级、省级、厅级、市级、校级、企业）科研项目，项目编号为：。VVIAbstractOutlierdataisthedatasetdifferentdata.Thispartofthesmallamountofdata,butforourdailyproductionandlifeofgreat.Therefore,theanomalydetectioniswidelyusedinnetworkintrusiondetection,finance,insurance,weather,andnewdrugdevelopmentandotherfields.Relativetothelargenumberofnormaldatamining,theanomalydetectionmodeliscalleddataminingsmall.BPalgorithmisacommonlyuseddataminingalgorithm.ButtheBPalgorithmtorealdataoutliersexistinthedataminingprocess:thehigherthedimensionoftheactualdata,thereareredundantfeaturesoftheinterference,andhigh-dimensionalfeature,theissueofinadequatedata.Therefore,thispaperanalyzesavarietyofBPneuralnetworkprocessingofdata,andtogetthefollowingresults.(1)BPneuralnetworkcanbetterseparationcharacteristicsofasinglesimulationdata;but(2)thecharacteristicsofsimilarlargedatasets,separationisdifficulttojudge;(3)normaldataisnotsufficientornotrepresentative,sothenormaldataclasslearningisnotsufficient,leadingtoabnormalcannotjudge.Tosolvetheaboveproblem,thispaperproposesthefollowingimprovements:(1)BPalgorithmbeforefeaturereduction(map)benefitfromanomalydetectionfeaturesselected(2)integrationofmultipleneuralnetworks,differentneuralnetworktorecognizethedifferentcharacteristicsofeacheachother,thefinalfusionresult.KeyWords：Outliers-Data，BP，Algorithms，NeuralNetworksVIIVIII目录1引言................................................................................................................................................................11.1背景.....................................................................................................................................................11.2传统已有异常点算法介绍............................................................................................................11.2.1基于统计学的异常点检测算法.......................................................................................11.2.2基于距离的异常点检测算法...........................................................................................21.2.3基于密度的算法..................................................................................................................31.2.4基于偏差的异常点检测.....................................................................................................51.2.5基于聚类的异常点检测算法...........................................................................................62基于属性特征在异常点检测中的研究...............................................................................................73BP神经网络介绍.......................................................................................................................................93.1模型简介...........................................................................................................................................93.2计算各层节点输出.........................................................................................................................93.3修正权值........................................................................................................................................104异常检测中BP神经网络的设计.......................................................................................................134.1可微阈值单元................................................................................................................................134.2单个BP网络结构设计...............................................................................................................134.3BP神经网络学习过程的基本步骤..........................................................................................145实验研究....................................................................................................................................................175.1研究使用的数据库介绍..............................................................................................................175.2训练方案一实验：把bp神经网络相似性代替距离算法相似度量.............................175.3训练方案二实验：用单个神经网络对训练数据库整体特性进行学习......................185.4训练方案三实验：多神经网络各种形式训练及其决策..................................................195.4.1实验设计思路.....................................................................................................................195.4.2实验方案及步骤................................................................................................................205.4.3实验分析..............................................................................................................................225.4.4实验失败原因分析......................................