Dynamic self-correcting key performance indicator anomaly detection algorithm

Abstract

ABSTRACT The operation and maintenance of the background system is always an important link to ensure the system’s high availability. With the increasing number of background systems, their operation, and maintenance have to develop from the initial huge-crowd strategy to the direction of intelligence. The key to intelligent operation and maintenance is the abnormal detection of key performance indicators (KPI), such as CPU utilisation. However, the existing KPI anomaly detection algorithms not only cannot select the dynamic threshold under the non-parametric methods but also have no false-positive correction mechanism to correct the false alarms. In order to overcome the above shortcomings, this work proposes a dynamic self-correcting Key Performance Indicator (KPI) anomaly detection algorithm, hereafter referred to as DSCAD. To the best of our knowledge, in the field of KPI anomaly detection, the DSCAD algorithm is the first dynamic threshold algorithm that does not rely on the assumption of normal distribution. Compared with the existing KPI anomaly detection methods, the F-score of the DSCAD algorithm increased by 3% and had the best performance.