This thesis treats different aspects on robust controller design and model identification techniques. The controller design technique proposes frequency-domain specifications for achieving a fixed-structure controller with user specified optimality criteria. The optimization based design method is iterative and it is based on direct shaping of the frequency responses without a need to explicitly design any weighting filters in contrast to classic loop-shaping methods. Computational efficiency has been taken into account by utilizing linear matrix equations for characterizing the frequency responses in the time-domain. The proposed controller design method can be used for designing any type of linear controllers, e.g. PID-type controllers, for identified linear systems.
Support vector regression (SVR) has several inherent excellent features that can with advantage be utilized in robust system identification. One of these is the usage of Vapnik’s İ-insensitive loss function that gives robustness and insensitivity to overtraining. Other features are the automatic computing of the parameters used in SVR and the convex optimization that guarantees to always find the global optimum.
SVR has in this thesis been tailored by modifying the kernel function to better fit several common model identification problems. These are identification of state-dependent parameter models or quasi-ARX models, smoothness priors models of linear systems and nonlinear Wiener models. All these proposed identification methods have been applied to examples of different systems. The results have been either as good or even better compared to other corresponding methods.
|Publication status||Published - 2019|
|MoE publication type||G5 Doctoral dissertation (article)|