Accurate Anomaly Interval Recognition and Fault Classification by Pattern Mining and Clustering

Abstract

To maintain the stability and reliability of a large-scale information system, monitoring its Key Performance Indicators (KPIs) time series and detecting their anomalies are very important. In practice, however, multivariate time series anomaly detection is challenging due to the large dimension of time series, diverse anomalous patterns, and their complex relationships. In addition, KPIs may exhibit different patterns when different types of faults occur, which aggravates the difficulty of anomaly detection. In this paper, we propose an accurate KPI anomaly detection and fault classification method, which can adapt to multiple metrics and diverse fault types. It can automatically extract common anomalous patterns from different KPI responses when faults occur and accurately determine the fault intervals. In this method, we do not need to deploy a lot of different anomaly detectors, and can conduct both anomaly detection and fault classification simultaneously. Experimental results on the real-world Exathlon benchmark dataset show that our algorithm can accurately recognize the anomaly intervals and classify the faults, with F1-score 0.94.