一种基于日志聚类的多类型故障预测方法  

A Multi-type Failure Prediction Method Based on Log Clustering

作　　者：王卫华 应时[1] 贾向阳[1] 王冰明 程国力 WANG Weihua;YING Shi;JIA Xiangyang;WANG Bingming;CHENG Guoli(State Key Laboratory of Software Engineering,Computer School,Wuhan University,Wuhan 430072,Chin)

机构地区：[1]武汉大学计算机学院软件工程国家重点实验室,武汉430072

出　　处：《计算机工程》2018年第7期67-73,共7页Computer Engineering

基　　金：国家自然科学基金"基于Saa S软件运行日志分析的软件性能问题的在线识别和诊断方法"(61672392);国家重点研发计划项目"国产化高等级病原微生物模式实验室建设及管理体系研究"(2016YFC1202204)

摘　　要：现有故障预测方法的日志事件多数是无规律交错存在的,且不同类型的故障事件所涉及的事件数量与时间范围存在一定差异。为在故障预测时能够提供故障相关信息,提出一种基于频繁日志事件序列对多种不同类型的故障进行预测的方法。以日志事件序列之间的最长公共子序列作为其相似性度量,使用聚合层次聚类算法挖掘与故障事件相关的频繁事件序列。在频繁事件序列的基础上生成故障事件预测规则,并给出一种对故障事件预测规则进行过滤的方法,将过滤后的规则应用到测试集上进行故障预测。实验结果表明,该方法不仅能够进行有效的故障预测,而且可以平衡故障预测的准确率和召回率。Most of the existing fault prediction methods have log event staggered, and there are certain difference in the number of events and the time range involved in different types of faule events. In order to provide fault-related information during fault prediction,a method for predicting multiple different types of faults based on frequent log event sequences is proposed. Taking the Longest Common Subsequence（ LCS） between log event sequences as its similarity measure, an aggregational hierarchical algorithm is used to mine frequent event sequences related to failure events. A fault event prediction rule is generated on the basis of a sequence of frequent events,and a method of filtering a fault event prediction rule is provided, and the filtered rule is applied to a test set for fault prediction. Experimental results show that this method can not only effectively predict failures,but also balance the accuracy and recall rate of failure prediction.

关 键 词：故障事件预测 频繁事件序列 规则过滤 聚合层次聚类算法 超级计算机 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]