Quadratically Optimal Augmented Identification and Filtration

Simultaneous evaluation of the whole set of the model parameters of different orders together with an ability to track unmodeled dynamics are desired features in the tasks of parameter estimation. A technique handling with the factors produced by an augmented covariance (ACM) or information (AIM) matrices is considered to be an appropriate tool for designing multiple model estimation. This is where the name augmented identification (AI) by using the least-squares method was taken. The method AI attains numerical stability of the calculation of the conventional least squares method while in the same time, fully extracts information contained in the observation. In order to track time varying parameters can be found that all the information pertinent to recursive identification and thus to data driven forgetting is concentrated in ACM as well as in AIM. In this thesis will be introduced how to selective forgetting to ACM should be applied in an effective way. It means forget only a portion of accumulated information which will be further modified by the newest data included in the regressor. In the estimation problems the knowledge of the inner states of the identified system is often required. Because the augmented identification belongs within the class so called prediction error method (PEM), some rational requirements can be deduced. As a result, state filter should constitute optimization procedure minimizing the predicted error of given state space model representation with respect to the vector of states. The proposed scheme will considerably extend the family of algorithms based on processing of ACM (AIM) about augmented filtering (AF). This all will establish a comprehensive concept of parametric estimation that compared with conventional approaches is characterized by versatility, low demands on a priori process information and by excellent numerical properties (robust against overparametrization, capable solving the multiple model problem)