Identification of ground vehicle aerodynamic derivatives Application of neural network with principal component analysis

Abstract

Principal component analysis is introduced to a multilayer neural network application to identify the ground vehicle aerodynamic derivatives. The main aim is to reduce the size of neural network input nodes while maintaining the neural network's accuracy. The design process is carried out using simulated data based on an equivalent system's dynamics and the designed neural network then is applied to wind tunnel measured data. Measured impulse response data of a simple ground vehicle body recorded from an oscillating test rig are used to estimate the aerodynamic loads acting on it. The aerodynamic loads are modeled in the form of stiffness and damping acting on the system. The neural network was trained using Bayesian Regularization training algorithm. The results using the optimized neural network input node are benchmarked against the aerodynamic derivatives identified using conventional method and full size neural network. Both neural network methods are shown to be able to estimate the aerodynamic derivatives as good as the conventional method with the advantage of a more direct method compared to conventional technique. Further more, the optimized neural network input node has a much smaller network size compared to the full size neural network.