Identification of ground vehicle's aerodynamic derivatives using neural network

• Main Author: Ramli, Nabilah (Author)
• Format: Thesis
• Published: 2008-03.
• Subjects: TJ Mechanical engineering and machinery; Thesis
• Online Access: Get fulltext

Stability of a ground vehicle is dependent on its aerodynamic characteristics when encountered by sudden crosswind disturbances. Aerodynamic side force and yaw moment have been identified as the main influence on the sensitivity of a vehicle towards crosswind, which is largely shape related. A reliable identification technique is a prerequisite to estimate the aerodynamic side force and the yaw moment. One of the recent techniques in wind-tunnel testing is the use of a pure yawing motion test rig to simulate the transient behavior of a simple vehicle model in crosswind condition. Adapting the stiffness and damping approach, the lateral aerodynamic derivatives are evaluated from the identified systematics frequency and damping of a pure yawing motion. This research explores the alternative identification technique apart from the conventional method of using a spectral density plot to identify the systematics frequency and the logarithmic decrement of peak amplitude for estimating the system’s damping from a recorded impulse response data. The present study aims to design a multilayer feedforward neural network to carry out the estimation of natural frequency and damping ratio trained with the Bayesian Regularization training algorithm. The network properties studied are necessary to give insight on the optimum network architecture, the suitable input representation and the effect of noise. The possibility of using principal component analysis technique for reducing the network input dimension has also been explored. The results show that the neural network is able to approximate the natural frequency and the damping ratio of an impulse response data and also the ability of the network to handle noisy input data. The application of principal component analysis technique has been shown to reduce the network input dimension while maintaining good estimation results and shortening the network training period. This study demonstrates that the identification of the frequency and the damping of the system can be done using neural network and can be applied to any other similar systems.