Efficient Periodicity Analysis for Real-Time Anomaly Detection

The analysis of time series data is critical for the effective management of IT infrastructures, with real-time anomaly detection being of particular importance. A crucial component in any anomaly detection pipeline is the characterization of individual data streams so that the most appropriate algorithm can be deployed on a case-by-case basis. Periodicity is a characteristic that directly impacts the choice of algorithm; efficient identification of its presence, or absence, is essential. In the current paper, we develop a periodicity detector based on the Lomb-Scargle periodogram and analyze its performance when exposed to varying quantities of data. Our analysis uses the well-known F1-score to identify the lower limit on the data volume required for maximal detector performance. We then extend our periodicity detection method to include a statistical analysis based on the Wilson score to establish confidence in the detector result. We demonstrate that in the absence of ground-truth labels, our method can quantity the minimum amount of data and the minimum number of trials required to correctly, and definitively, identify periodicity within a data stream.