FAULT DETECTION BY ADAPTIVE NONLINEAR FILTERING.

Abstract

In a lintar iyitern perturbed by Gaussian noise, the state can be estimated from the observations by using Kalman filter. However, if a fault develops in the system at any random time, the Kalman filter will not be able to track the fault and large errors will develop in the state estimate. Consequently, the innovations process will no longer be white. If the random time of occurrence is coaiidered as a state then the system of state equations become nonlinear. In this paper, the Fujisaki, Killianpur arid Kunita nonlinear filtering results have been applied to obtain a representation for the stage estimite given the observations. The non-white nature of the innovations process has been modelled as an autoregressive process and an adaptive scheme has been proposed to improve the filter performance.