【作者】 陈建华；

【导师】 温巧燕；

【作者基本信息】 北京邮电大学 ， 信息安全， 2015， 硕士

【摘要】 银行设备的正常运行对银行业务正常运作非常重要,而银行运维系统的集中构建是对这些IT设备进行监控从而保障系统正常运行最重要的一环,目前银行对这些运维系统中产生的大量告警以及性能数据没有充分利用,告警的产生往往通过捕获而非预测。本文通过机器学习算法对这些数据进行挖掘进而告警预测,可以对银行故障进行主动告警排查,最大限度保证银行业务的正常运作。主要研究成果如下：通过对启发式搜索算法以及基于关联规则的特征选择(CFS)的分析研究,提出了基于双向搜索(BDS)与特征关联关系(FC)的特征选择算法BDS+FCA,该算法对BDS搜索得到的特征子集存在冗余以及CFS搜索效率不高的不足进行了改进；在分析了银行数据特点以及相关机器学习算法的基础上,选择适合银行故障数据特点的支持向量机(SVM)、朴素贝叶斯以及最大期望算法(EM)算法,将这些算法通过组合应用到系统中,得到了很好的故障预测效果；在理论研究的基础上通过对银行故障告警系统进行需求分析并完成详细的系统设计,同时将BDS+FCA特征选择算法以及SVM、朴素贝叶斯、EM算法应用到系统的核心模块中；最后利用J2EE、MongoDB、 Weka等技术实现了该系统,并对系统进行了功能测试以及模型有效性测试。对系统的测试结果表明,BDS+FCA算法可以对原始故障告警数据进行冗余特征的剔除,降低数据维度,提高机器学习分类算法的时间性能与预测准确率；而根据银行运维数据特征选择的机器学习算法作为告警系统的核心引擎,在银行告警数据比例很低的情况下取得较高的分类预测准确率,并且EM算法可以实现银行部分缺失属性值的填补。从而验证了BDS+FCA以及所选分类算法适合银行告警系统,对保障银行业务系统正常运行具有重要意义。更多还原

【Abstract】 The normal operation of bank equipment is very important to the bank business. And the bank operation system with centralized architecture is an important part of these IT equipment monitoring systems to guarantee the normal operation of the whole bank system. Currently, banks do not make full use of the performance data and alarm data produced by the operation systems. The alarms are usually captured not predicted. In this article, we can do data mining through machine learning algorithms and then predict alarms. In this way, the banking business can be ensured to the max extend. The main research results are as follows:Based on the analysis and research of heuristics feature selection algorithms and correlation based feature selection algorithms, this paper proposed a feature selection method called BDS+FCA combing bidirectional search (BDS) and feature correlation. The algorithm improves the shortcomings of BDS’s feature redundancy and CFS’s search inefficient. On the detailed study of the features of the bank data and certain kinds of machine learning algorithms, we choose Naive Bayes and Support Vector Machines and Expectation Maximization algorithms and apply them in the system and get good effects in return. On the basis of theoretical research, we complete the detailed system design with the demand analysis. In the system, we apply the BDS+FCA algorithm and combined classification algorithm to the core module. At last, we complete the implementation of the system through the J2EE, MongoDB and Weka technology. After that, we do the system test including functional test and model effectiveness test.The test results show that the BDS+FCA algorithm can improve the time performance and classification accuracy by exclude the redundant feature and the combined classification algorithm as the core engine can do good classification predication of the alarm system in the cases of low warning data proportion. Thus we verify the effectiveness and feasibility of the BDS+FCA and Combined Classification algorithms.更多还原