面向磁盘故障预测的机器学习方法比较  

Comparison of machine learning methods for disk failure prediction

作　　者：董勇[1,2] 蒋艳凰[1,2] 卢宇彤[1,2] 周恩强[1,2] 

机构地区：[1]国防科学技术大学计算机学院,湖南长沙410073 [2]高性能计算国家重点实验室,湖南长沙410073

出　　处：《计算机工程与科学》2015年第12期2200-2207,共8页Computer Engineering & Science

基　　金：国家863计划资助项目(2012AA01A301);国家自然科学基金资助项目(61272141;61303068;61120106005)

摘　　要：磁盘是保存数据的重要载体,提高磁盘的可靠性和数据可用性具有重要意义。现代磁盘普遍支持SMART协议,用来监控磁盘的内部工作状态。采用机器学习方法,分析磁盘的SMART信息,实现对磁盘故障的预测。所采用的机器学习方法包括反向神经网络、决策树、支持向量机以及简单贝叶斯,并采用实际磁盘SMART数据进行验证与分析。基于上述数据,对不同机器学习方法的有效性及其效果进行了对比。结果表明,决策树方法的预测率最好,支持向量机方法的误报率最低。As disk is one of the most important data storage device, it is significant to improve disks＇ reliability and data availability. Modern disks adopt the SMART protocol to monitor the internal operat ing status. We employ machine learning methods, including back-propagation neural networks, decision tree, supported vector machine and naive Bayes to analyze the SMART data of disks, which can predict disk failures. Real SMART data of disks are used in experiments to validate and analyze the effectiveness of those methods, and the effectiveness of different methods is compared. The results show that the decision tree method has best prediction rate while the supported vector machine method has the lowest false alarm rate.

关 键 词：磁盘 故障预测 机器学习 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]