Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Identification of aerodynamic coefficients of a flexible fixed-wing aircraft using deep learning
Submission Author:
Rodrigo Costa do Nascimento , SP , Brazil
Co-Authors:
Rodrigo Costa do Nascimento, Éder Alves de Moura, Thiago Rosado de Paula, Vítor Fernandes, Luiz Carlos Góes, ROBERTO GIL ANNES DA SILVA
Presenter: Vítor Fernandes
doi://10.26678/ABCM.COBEM2023.COB2023-2314
Abstract
System identification is a process of building mathematical models of dynamic systems from observed data and has become increasingly popular due to their ability to learn complex relationships from input-output data. This task is crucial for many applications, including control systems, robotics, and financial markets and, to this work, in aerospace. The identification of aircraft parameters consists of estimating the values of physical parameters that govern the behavior of an aircraft. This information is essential for designing and testing aircraft control systems and for improving the accuracy of flight simulators. An increasingly adopted approach in system identification is the use neural networks, due to their ability to learn complex relationships from input-output data. However, there are still many challenges that need to be addressed to improve the accuracy and reliability of neural network-based system identification methods. Other important question in adopting neural network in this area is its black-box description of the modeled system. Classical methods of system identification are based on paradigms from statistics, where they strongly rely on prediction error methods altogether with an ordinary differential equation that describes the system behavior, and the identification process consist of defining the constants for a given reference model. On the other side, neural networks typically are a set of multiple interconnected layers, where each node calculates a weighted sum of its inputs, followed by a non-linear activation function, which do not preserve any relation with real world system. In addition to it, when aircraft is flexible, more degrees of freedom are involved and more complex is the identification process. Thus, some alternatives have been proposed with the objective of taking advantage of the generalization power of models obtained with neural networks, while preserving the ability to identify the physical parameters of the analyzed system. This work proposes the analyses of the capacity of a deep learning model, a neural network with multiple hidden layers, to identify the parameters of a flexible fixed-wing aircraft, with data obtained in-flight. Thus, the objective is to propose an architecture for the identification of aerodynamic coefficients and compare these results with classical methods.
Keywords
Parameter identification, System Identification, Deep learning, Artificial neural networks (ANN)
