Rotor aerodynamic model analysis by functional-link neural networks

A neural network is a biologically inspired mathematical model that has the ability to learn through training. It has many advantages over current computational mathematical models including the ability to learn and generalize from their environment. This thesis has applied Back Propagation Neural Network (BPNN) to perform nonlinear function mapping problems, including airfoil coefficient problems of NASA Ames Research center. The results have shown that BPNN mapping is faster than the traditional Computational Fluid Dynamic (CFD) method. Moreover,BPNN mapping is more accurate for a given set of nonlinear functions in a rather short time frame.To improve the convergence (accuracy, or speed, or both) of a BackPropagation neural network, different transfer functions as well as an innovative method have been studied. This method is called functional-link neural network in which some appropriate functions are chosen as additional inputs of the input layer.The results of the selected specific engineering problems have shown the following conclusions:1. Depending on the properties of the nonlinear function, the selection of appropriate transfer function can greatly improve the convergence of the back propagation neural network.2. Based on the concept of Professor Pao's functional-link neural networks, adding some appropriate functions as part of the input layer can greatly improve the convergence of the BPNN. Meanwhile, also served as partIVof the input layer, Gaussian functions can improve the convergence for certain types of nonlinear functions. However, this method should be applied with caution because of the existence of an over-fitting problem, which could lead to unreasonable interpolation.