A Way to Deal With Model-Plant Mismatch for a Reliable Diagnosis in Transient Operation

Abstract

Least-squares health parameter identification techniques, such as the Kalman filter, have been extensively used to solve diagnosis problems. Indeed, such methods give a good estimate provided that the discrepancies between the model prediction and the measurements are zero-mean, white, Gaussian random variables. In a turbine engine diagnosis, however, this assumption does not always hold due to the presence of biases in the model. This is especially true for a transient operation. As a result, the estimated parameters tend to diverge from their actual values, which strongly degrades the diagnosis. The purpose of this contribution is to present a Kalman filter diagnosis tool where the model biases are treated as an additional random measurement error. The new methodology is tested on simulated transient data representative of a current turbofan engine configuration. While relatively simple to implement, the newly developed diagnosis tool exhibits a much better accuracy than the original Kalman filter in the presence of model biases.