Revisit Dynamic ARIMA Based Anomaly Detection
Bonnie Zhu, S. Shankar Sastry · 2011
On the assumption that a model is correctly learned and built, the typical usage of ARIMA in anomaly detection compares data points with those predicated through the model to determine whether anomalies occur. Yet the time variability by the coefficients in those dynamic regression models is possibly indicative of whether anomalies are in the data set on which the ARIMA model builds. Thus we introduce a corresponding framework and a novel anomaly detection method that combines the Kalman filter for identifying the parameters of those dynamic models with a General Likelihood Ratio (GLR) test that is based on the former for detecting suspicious changes in the parameters and therefore the models. We illustrate the idea through experiments and show its promising potential in terms of accuracy and robustness.