Robust Identification of Linear Systems.
BALLISTIC RESEARCH LABS ABERDEEN PROVING GROUND MD
Pagination or Media Count:
The first two sections of this report survey the current techniques of identification and time series analysis for linear, time invariant, single-input, single-output systems with noisy measurements. Typically, this additive noise has been modeled either as being completely specified e.g., a Gaussian with known parameters or as being required to satisfy general conditions such as whiteness and non-correlation with the input. Departures from the assumed noise model sometimes cause severe deterioration in the efficiency of available identification algorithms. Since it seems reasonable to have some, if not complete, knowledge of the operating environment, it is assumed in this report that the measurement noise w sub k has a distribution FW 1-eKw eCw, where K, is a completely specified distribution and C. belongs to some broad class of distribution. In the third section, a robust scheme for estimating the system cross correlations is proposed in order to desensitize the performance of the identification algorithm to the distribution of w sub k. Extensive computer simulations show that the proposal provides a robust identification technique which has good uniform behavior over a variety of distribution for w sub k. Author
- Statistics and Probability