Single-hidden-layer fuzzy recurrent wavelet neural network: Applications to function approximation and system identification

When a fuzzy wavelet neural network (FWNN) has a large number of neurons in its consequent part, it may not be able to effectively follow fast variations in the process. This paper aims to develop a single-hidden-layer fuzzy recurrent wavelet neural network (SLFRWNN) for the function approximation and identification of dynamic systems. In the proposed framework, the consequent part of each fuzzy rule is developed by a single neuron with the capability of storing the past data of the network. To guarantee the convergence and to speed up the process of the on-line training algorithm, the optimal learning rate (OLR) is introduced based on Lyapunov stability theory. The modifications allow the SLFRWNN to be much faster than FWNN and hence it is more appropriate in real-time applications. The efficiency of the modified model is investigated using three different types of wavelet families in approximating a benchmark piecewise function. Furthermore, two nonlinear dynamic plants are considered to demonstrate the feasibility of the SLFRWNN in the system identification. The results indicate that SLFRWNN achieves higher accuracy and needs a lower number of parameters than the other models.