Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
NEURAL NETWORKS FOR AERODYNAMIC MODELING IN FLIGHT SIMULATIONS
Submission Author:
Victor Hugo Araújo Diniz , SP
Co-Authors:
Victor Hugo Araújo Diniz, Juliano A. B. Gripp, Flavio Luiz Cardoso-Ribeiro
Presenter: Victor Hugo Araújo Diniz
doi://10.26678/ABCM.COBEM2021.COB2021-0858
Abstract
The control laws of a fly-by-wire system are the algorithms present on the flight controls computer. They interpret the pilot commands and sensor measures to efficiently command the control surfaces, taking into consideration the operational limits and flight control envelope. The control of any system presumes the existence of a plant, which refers to the aircraft’s model, that holds utmost importance for the development of representative flight simulators and improvement of the control laws. The modeling of this plant is the responsibility of the flight mechanics team, which uses traditional approaches, based on the literature, wind tunnel, and industry experience. During the flight tests campaign, some maneuvers are selected for comparison of the modeling, generating data to be used in the identification of the system parameters. This process is developed with consolidated techniques, such as output error method, filter error method, equation error method, maximum likelihood, and, recently, neural networks. Neural networks can work as function approximators, creating a structure that is capable of predicting outputs based on input values, without needing an a priori knowledge of the system. Furthermore, they bypass the solution of equations of motion for the prediction of flight coefficients, becoming an attractive solution for the analysis of highly non-linear dynamics. In this work, a neural network is discussed for the estimation of the aircraft's longitudinal aerodynamic coefficients. The estimated parameters are used to evaluate the aerodynamic model performance in comparison to the real flight dynamics and aid in the proposition of a more suitable model. Due to the lack of available flight data, a computational model of an aircraft is described as an alternative for artificial flight data generation, and the data obtained from the model is conditioned to match the real signal conditioning obtained by flight data, based on the availability of aircraft sensors and the signal noise relations.
Keywords
Neural Network, Aerodynamic Model, aircraft parameter estimation
