In this master's thesis, a new method of evolving neuro-fuzzy model identification of nonlinear systems with online recursive least-squares and instrumental variables identification of model parameters is presented. The identification method is comprised of the simultaneous estimation of the evolving neuro-fuzzy antecedent structure with an incremental clustering method and recursive optimization of the consecutive local linear submodels parameters from a sequential data stream. The clustering method also consists of mechanisms for online clusters adaptation, cluster formation, and merging.
Meanwhile, the model parameter optimization methods are modified to enable the identification of nonlinear output error models, which are excited with step input functions.
The purpose of this dissertation is to present a complete approach to evolving online identification of various nonlinear systems and the use of those models for system control.
The evolving neuro-fuzzy identification and its modifications are evaluated on sampled data, acquired from three representative nonlinear models; Hammerstein-Wiener, piecewise linear servomotor, and a theoretical heat exchanger model with nonlinear dynamics.
Those identified models are utilized to demonstrate a predictive control method based on the predictive functional control with particle swarm optimization, capable to incorporate the regulation error signal and all other restrictions directly in the criterion function.
|
