基于小波分析的云计算在线业务异常负载检测方法
Wavelet-based approach for anomaly detection of online services in cloud computing systems

随着越来越多的在线业务被迁移到基于云的平台上，如何检测云平台上在线业务的异常运行状态成为了一个重要的问题。现有方法通过分析在线业务的实时负载数据来判断业务是否存在异常，在应对由程序异常或突发用户访问引起的异常负载时存在准确率低、误报率高的问题。该文提出并实现了一种面向云计算在线业务的异常负载检测方法。该方法利用小波分析技术，将原始负载数据分解成频率不同的多条曲线，并利用统计分析技术，通过检测各个频率上的异常增长或降低来判断负载是否存在异常。实验结果表明：同现有方法相比，该方法更准确，同时可以大大降低误报率。
Abstract：As an increasing number of online services have migrated into the cloud, anomaly detection has now become an important problem. Existing efforts detect anomalies by mining real-time workloads; however, the accuracy of such approaches cannot be assured in case of user spikes and application errors. This paper presents a wavelet-based online anomaly detection approach that uses discrete wavelet transforms to decompose real-time workload traces into multiple curves with different frequencies and then applies statistical analysis to the decomposed traces to detect the workload anomalies. Tests show that this approach is more accurate with a lower false-alarm rate than existing approaches.