基于深度学习的云计算系统异常检测方法  

Anomaly Detection for Cloud Computing Systems Based on Deep Learning

作　　者：任明[1] 宋云奎[2] REN Ming;SONG Yun-kui(China UnionPay Co.,Ltd.,Shanghai 201201,China;Institute of Software,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国银联股份有限公司,上海201201 [2]中国科学院软件研究所,北京100190

出　　处：《计算机技术与发展》2019年第5期54-57,共4页Computer Technology and Development

基　　金：国家自然科学基金(61602454)

摘　　要：在云计算服务中,异常检测是防止意外系统停机,并确保终端用户服务可靠性的关键技术。虽然操作控制台日志记录了云计算系统的运行时状态信息,但现有的云计算系统管理技术主要在出现问题后分析原因,而不能提前检测异常。因此,文中提出了一种基于深度学习的云计算系统异常检测方法。首先,提取日志模式,使用聚类将相似格式和内容的日志聚集为模式组;然后,将每个模式作为单词,将离散模式集作为文档,从而降低日志维度,得到低维度特征空间;最后,使用递归神经网络,即长短时记忆,处理训练过程中标记数据的"稀缺性",捕获跨序列的依赖性,获得系统状态的鲁棒异常信号。使用Web系统日志进行实验的结果表明,与现有方法比较在检测复杂异常时具有更高的检测准确性。In cloud computing services,anomaly detection is a key technology to avoid system outages and guarantee the service reliability of end users.Although operating consoles record the runtime status of cloud computing systems,existing management technologies locating the root causes of occurred faults cannot detect anomalies in advance.Therefore,we propose a deep learning-based anomaly detection approach for cloud computing systems.First,we cluster logs to groups according to the format and content of logs,and then extract execution patterns from the clusters.Second,we regard a pattern as a word and a pattern set as a set to lower the dimensionality of features in logs.Third,we use a recurrent neural network to address the sparsity of labelled data instances and the dependencies between series,and then define robust anomaly signatures and detect anomalies.We use the logs of a web system to validate the proposed approach,and the experimental results demonstrate that this approach has higher precision than existing ones in detecting complex anomalies.

关 键 词：异常检测 日志分析 文本挖掘 递归神经网络 云计算 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]