Adaptive Sliding Kalman Filter using Nonparametric Change Point Detection

This paper aims to develop an Adaptive Sliding Kalman Filter (ASKF) by fusing the concept of change detection in a data stream, adapting noise covariance matrices and the Sliding Kalman filter (SKF). Adaptive Kalman filtering (AKF) scheme modifies the noise covariance matrix (Q and R) value based on a window of past innovation sequence whereas SKF is a window based filtering technique which uses past information to obtain the present state estimate. However, the length of the window chosen for SKF and AKF is arbitrary and a scheme has been devised here to adapt this window length according to the data stream statistics. The change detection scheme chosen here does not make any assumption on the data distribution and is sequential in nature, such that a change is triggered whenever the underlying statistics of data crosses a pre-determined threshold. The key contribution of this work is toward the formulation of a mechanism by which the window length is made adaptive such that whenever a change is detected, the window length for SKF and AKF is curtailed and restarted in an oscillatory windowing fashion. The suggested filter is robust against temporary uncertainties and appropriate for reliable estimation of signals that may arise in many engineering areas. Real world experimental results demonstrate better estimation accuracy of the proposed method than that of others. 2016 Elsevier Ltd. All rights reserved.