Among the causes that may lead to divergence of Kalman filter algorithm is a wrong estimation of the noise pertaining to either the state model or the measurement model. This paper deals with an on line identification of the values of these entities by a suitable adjusting of the filter gain. The result yields a new adaptive Kalman filter. This investigation is restricted to time invariant parameters. The proposal is mainly based on statistical properties of the innovation matrix of the filter and improves some previous results pointed out by Bellanger and Mehara. Particularly, the autocorrelation functions of the innovations is used. Furthermore, information about the quality of the autocorrelation parameter is incorporated through a weighted least squares methodology. While the estimation of the a priori noise statistics Q and R is obtained after assessing the optimal gain and innovation covariance of the filter.
展开▼
