Kalman Filter Changepoint Detection and Trend Characterization
Scott Kuzdeba, Brandon Hombs, Jeremy D.W. Greenlee, Frank H. Guenther · 2019
In this paper we describe a data-driven change detection algorithm based on the Kalman filter. The algorithm models trends in the underlying data through the use of a Kalman filter. A learning rate is applied to the Kalman filter gain to allow for trends to be locked in. This forces the estimates to rely more on the prediction and less on the observation as time goes on. Statistical thresholds are set to detect a change, i.e. changepoint, and a new Kalman filter is started to track the new trend in the data. We show results using neural electrocorticography time-series data.