Model-based Algorithms for Nonlinear System Identification

In this thesis three algorithms for the estimation of parameters which occur nonlinearly in dynamic systems are presented. The first algorithm pertains to systems in discrete-time regression form. It is shown that the task of finding an update law for the parameter estimates can be solved numerically by the formulation of a quadratic programming problem. The algorithm does not depend on analytical knowledge of the regressor function. In particular, a neural system model can be used to approximate the required regression form for systems which cannot easily be transformed into this form analytically. The second algorithm makes use of model-based parameterizations. It is shown that if some of the system parameters occur linearly, or enter the model multiplicatively, an update law for these parameters can be found analytically. The third algorithm makes use of convex properties of the regression function and applies to a class of continuous-time systems. It is demonstrated how the algorithms can be modified to make them robust in the presence of a bounded disturbance. The performance of the algorithms and the nature of the parameter convergence are illustrated in simulations of a magnetic bearing system and a low velocity friction model. Thesis Supervisor: Anuradha M. Annaswamy Title: Assistant Professor