چكيده به لاتين
In this thesis, the Wiener model is considered to describe the nonlinear dynamic systems. This model is composed of a linear dynamic subsystem that is followed by a nonlinear static one. As inaccessibility of the connection point between the blocks, identification is started with an initial linear approximation of the main nonlinear system. Although the online identification is preferred in a large number of practical applications, but in the most articles, the modelling into the Wiener structure is performed in the offline mode. In this thesis, after an offline training period, system identification is fulfilled in the online mode using the new input-output data. In the first step, with a little time delay in the input line, MOESP-type subspace identifier and multilayer perceptron neural network are applied to describe the linear dynamic and nonlinear static subsystems, respectively. Stability of the Wiener model is demonstrated using a linear matrix inequality solution. Furthermore, convergence of the Wiener model outputs to the system outputs is guaranteed using the bounded learning rate in the Levenberg-Marquadt training method. In the next stage, identification algorithm is extended for nonlinear dynamic systems with input time delay. In this regard, a Laguerre filter network is used to model the linear dynamic block. In this phase, analysis on the upper-bound error and convergence of the model output are conducted using the Lyapunov theorem. On the other hand, the training procedure of the Wiener model is difficult in such cases that the system output is infected by the measurement noise. To solve this problem, in the final step, a committee neural network comprised of some least squares support vector machine is applied as the nonlinear static subsystem. With the suggested analysis, the synaptic weights of the output layer and the corresponding cost are computed in an optimal manner. Some practical nonlinear systems such as CSTR, pH, and 3-tank are used in order to show the capability of the proposed identification approaches. In addition, obtained results in comparison with the recently advanced methods proposed in literature show the desired performance of the suggested methods.
