چكيده به لاتين
In real world, many physical systems are of noninteger order (also known as fractional order) or can be modelled by less parameters as compared to integer order modeling. On the other hand, neural networks are universal approximators that have been widely employed for the identification of multi-input single-output nonlinear systems. Besides, among block-oriented models, the Hammerstein model has attracted the most attention because of its simplicity and acceptable modeling. Moreover, in practice, measured data are usually contaminated by different types of noises and disturbances such as outliers. Outliers are defined as large deviations from the standard data range and may be referred to as large perturbations, impulsive noises, missing data, or non-Gaussian noises. Accordingly, the subject of this thesis is to achieve a simple, accurate and practical identification of nonlinear dynamic systems in presence of outliers, by the introduction of the Neuro-Fractional Order Hammerstein (NFOH) modelling. Herein, a two-stage frequency-time domain identification procedure has been proposed. In the first stage of identifiacation, recorded time domain data are transformed to the frequency domain and then fractional order and degree of the linear dynamic subsysytem of the Hammerstein model are identified on the basis of constrained optimization. Afterwards, the state space matrices/vectors of the NFOH are estimated in the time domain based on the Lyapunov stability theorem. Furthermore, when the recorded output is contaminated with the additive outliers, ununiform discrete Fourier transform is proposed in the frequency domain stage, and the piecewise/kernel-based Lyapunov function is introduced in the time domain identification. Stability and convergence analysis of the proposed modelling scheme are carried out and are verified throughout some simulational and experimental examples.
