Online Identification of Recurring Changepoints

This work aims at accurately detecting recurring changes as quickly as possible after their occurrence. We consider a general framework in which the changepoints and the duration of the data in each state after a change, are unknown. Assuming that the statistical model of the data is known, we adopt the windowed-cumulative sum (W-CUSUM) test to solve this problem. In this work, we also propose the False-Alarm Density (FADE) and False-Alarm Correction Time (FACT) metrics to characterize the false-alarm performance of the test. We present simulation results for a Gaussian mean-shift problem to demonstrate the performance of the W-CUSUM test.