適応的カルマンフィルターによる異常値検出について

抄録

In this paper a new and simple method for on-line detection of the time of occurrence of abnormality is proposed. The adaptive Kalman filter (AKF) coupled with a generalized likelihood ratio test is derived incorporating this method for the case of both the unknown abnormal and observed variables being vectors. Each state variable is clearly defined as a function of the unknown abnormal variable. Physical relations among the state variables and among the other variables of the AKF are also formulated. To show its effectiveness, adaptability and ability to detect abnormalities, the derived AKF is applied to a synthetically generated time sequence having an abnormality occurring at an unknown point in the sequence. The proposed method results in a more practical on-line algorithm for the AKF and much less computer loads.