Additive Change Detection in Nonlinear Systems With Unknown Change Parameters

We study the change detection problem in partially observed, nonlinear systems [which satisfy the hidden Markov model (HMM) property]. The change parameters are assumed unknown, and the changes can be slow or sudden. A partially observed system needs to be tracked first before changes can be detected. Sudden changes result in significant loss of track. These can be detected easily using the increase in tracking error (TE) or observation likelihood (OL) or using a CUSUM-type method applied to either of these. However, slow changes (which result in small loss of track) often get missed. We propose here a statistic that uses the tracked component of the change to detect it and, hence, detects slow changes faster than TE or OL. We show, both analytically and through simulations, that this statistic complements OL and TE for change detection