Modeling and Identification of Nonlinear DC Motor Drive Systems Using Recurrent Wavelet Networks

Abstract

The main objective of this research is to study the use of Recurrent Wavelet Networks (RWN) for the modelling and identification of nonlinear dynamic systems. Since the vast majority of physical processes and systems exhibit nonlinearities in their behavior, mathematical models may be difficult to obtain as processes may be affected by external operating conditions and a number of parameters may not be identified. Electromechanical systems are an example of nonlinear systems where parameters such as viscous and coulomb friction, and distributed inertias are often unknown. In such cases, a model is required that will capture the nonlinearities and the dynamics of the system. In this thesis, an online identification method is developed using structured Recurrent Wavelet Networks (RWN) in order to simultaneously identify linear and nonlinear mechanical parameters of an electromechanical system. Network learning is implemented using the gradient descent algorithm. Stability analysis is carried out based on the minimization of a Lyapunov function in order to obtain Adaptive Learning Rates (ALR) for training the network. Simulations are carried out to validate the performance of the proposed adaptive learning rate based modeling and identification technique. Search Terms: Wavelet Networks, Recurrent Wavelet Networks, DC Motor Parameter Identification, Friction Identification, Adaptive Learning Rates