On Line Failure Detection Using Kullback's Divergence

Abstract The purpose of this paper is the presentation of an on-line failure detection approach. This method has two parts: the first one consists in an identification of the AR coefficients using Instrumental Variable method. The model M 0 is identified in a growing size window and is used as a reference model. The second model M 1 is identified in a sliding window. The second part consists in a comparison between the models M 0 and M 1 with the aid of the Hinkley's cumulative sum test related to Kullback's divergence. This change detector gives us good results in all the studied cases: simulated scalar and real process signals