INTELLIGENT IDENTIFICATION AND ADAPTIVE CONTROL OF NONLINEAR DYNAMICAL SYSTEMS

Abstract

Most of the physical systems in nature are highly nonlinear. Such systems are characterized by time-varying and complex behavior making it challenging to derive accurate models for understanding their behavior. Further developing a controller for such systems becomes more complex, rendering a conventional controller inadequate. Soft computing has become a viable alternative method for system dynamics modeling. Artificial Neural Networks (ANNs), among other methods, are notable for their exceptional capacity to approximate complex nonlinear functions and dynamically adjust to the changing behavior of the system. This thesis explores, the application of ANN-based methodologies for the identification and adaptive control of non linear dynamic systems. In this thesis, we have proposed several modifications to the existing ANN structures to improve the capability of ANN to handle dynamic systems. In particu lar, we have modified the structure of Recurrent Neural Networks (RNNs) and trained them using the back-propagation (BP) algorithm. The proposed design considered in this thesis is independent of the order of the system. Building on the novel identification structure, this work extends its application to design an adaptive controller for nonlinear dynamic systems. The controller is developed to operate online, enabling simultaneous identification and control. This ability makes the controller design robust enough to adapt in real time and capture the changing dynamics of the plant effectively. The performance of an ANN is influenced by the learning algorithm and structure. In this thesis, we have developed a constructive algorithm with an adaptive learning rate to dynamically optimize the structures of Feed Forward Neural Network (FFNN) and RNN. This approach enables efficient modeling and learning by facili tating dynamic growth of the network architecture. Furthermore, to optimize the parameters of the ANN and enhance the performance of the BP learning algorithm, we have developed an adaptive Particle Swarm Optimization (PSO)-based BP algorithm. The parameters of PSO such as inertia weight (w) and hyperparameters (c1 and c2) are also dynamically updated to improve the optimization capability of PSO. The proposed approaches are tested on various nonlinear benchmark systems such as liquid-level systems, Mackey glass series prediction prob iv lems, and various degrees of nonlinear plant equations. The stability and convergence of the update weight equations are derived in the sense of Lyapunov stability principles. The pro posed approaches are evaluated in terms of Mean Square Error (MSE), Mean Absolute Error (MAE), and Relative Mean Absolute Error (RMSE) against state-of-the-art neural structures in the literature. The robustness of the proposed approaches is validated by using parameter variations and disturbance. Detailed simulation analysis has been also carried out in the thesis to evaluate the performance of the proposed approaches. The results demonstrate the effec tiveness of these approaches in accurately learning the dynamics of nonlinear systems.