A Two-Stage Temporal Anomaly Detection Algorithm Based on Danger Theory

Abstract

Anomaly detection is to identify the abnormal states from time series, which is widely used for early diagnosis or warning in various areas. In recent years, there are many algorithms developed for anomaly detection. However, the time series from different areas have obvious specificities, which limits the generalization of the current anomaly detection approaches. In this study, we proposed a two-stage anomaly detection algorithm (TSAD) to identify discords in univariate time series. Based on Danger Theory, the TSAD approach determined an abnormal subsequence using a two-stage recognition strategy, recognition signal is referred as a danger signal to report the candidate abnormal subsequences, and the response signal is to determine if a candidate subsequence is an anomaly based on its statistical properties. To validate the effectiveness, we applied our developed approach on extensive time series. The simulation results indicate that TSAD is capable of identifying discords from time series. Besides, comparison with other typical algorithms also proved that it is outstanding to others.