| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 3.27 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O presente trabalho visa conciliar a temática das redes neuronais artificiais à previsão de
trajetórias de aeronaves. Especificamente, centra-se à volta da aplicação de métodos de
regressão através de algoritmos de Máquina de Suporte Vetorial com base nos Mínimos
Quadrados (ou LS-SVM Least-Square Support Vector Machines) na previsão de trajetórias de
quatro dimensões de uma aeronave comercial.
Inicialmente, é feito um enquadramento da temática de aprendizagem automática e redes
neuronais através de uma contextualização teórica, progredindo para um desenvolvimento de
algumas funções aplicadas no tema das redes neuronais artificiais. Dentro destas funções,
destacam-se o Método Kernel e as Funções de Bases Radiais (ou RBF - Radial Basis Function),
aplicadas também na obtenção de resultados práticos. Na segunda parte deste capítulo, são
desenvolvidos os temas de LS-SVM, projeto SESAR (responsável pela implementação de
trajetórias 4D no céu europeu), e um dos métodos de aquisição de dados de trajetória em
aeronaves comerciais, o Automatic Dependent Surveillance-Broadcast (ADS-B). Toda esta
explicação teórica, dá lugar à apresentação dos conteúdos formais e matemáticos por detrás
dos métodos implementados, tais como: Condições de Otimalidade, Geração de Erro Gaussiano,
Funções Kernel aplicadas, bem como uma explicação de como a Localização Aeroespacial
decorre numa aeronave comercial.
Por fim, os dados foram simulados e validados num contexto de Previsão de Trajetória,
permitindo assim obter, no final, uma função que preveja a trajetória da aeronave em todo o
seu percurso. Para isso, foram utilizadas três trajetórias distintas, criadas através da
implementação do Algoritmo de Simulação Numérica RK-Butcher, bem como um ambiente de
navegação de quatro dimensões, em que a aeronave define a sua trajetória inicialmente,
através de um conjunto de waypoints definidos anteriormente. No final, é possível observar
não só o sucesso da implementação desta metodologia, bem como fazer um estudo comparativo
entre os diferentes métodos utilizados.
Após uma reflexão sobre os resultados obtidos, e analisando o conteúdo de todo o trabalho
através de vários prismas, é possível fazer uma apreciação bastante positiva do projeto
desenvolvido. Não só os resultados foram de acordo com o pretendido, tendo sido obtidas todas
as funções de previsão com sucesso, mas também foi possível obter uma visão geral acerca da
vasta aplicação que estes métodos detêm, não só na área da Engenharia Aeronáutica, mas como
em qualquer setor de investigação científica.
The present work aims to reconcile the neural networks subject, with the aircraft trajectory prediction theme. Specifically, it revolves around the application of Least-Squares Support Vector Machines methods in the prediction of 4-dimensional trajectories of commercial aircrafts. Initially, this work explains the guidelines regarding the machine learning and neural networks subjects through a theoretical explanation, gradually progressing to a more detailed elucidation of some of the functions applied in artificial neural networks. Within these functions, we highlight the Kernel Method and Radial Basis Functions, which were essential in order to obtain the practical results. In the last part of this chapter, the LS-SVM topic is approached theoretically, as well as the SESAR Project, which is responsible for the implementation of 4D trajectories in the European sky and the main trajectory data acquisition method in commercial aircraft, the Automatic Dependent Surveillance-Broadcast (ADS-B). All this theoretical explanation gives rise to the enunciation of the formal and mathematical content of this project, such as: Optimality Conditions, generation of Gaussian Error, Kernel Functions applied, as well as an explanation of how the Aerospace Localization methods work in a commercial aircraft. Lastly, all the data was simulated and validated in a context of Trajectory Prediction, allowing the algorithm to build a function that could predict the trajectory of the aircraft in its complete course. In order to achieve that, three different trajectories were used, created through the implementation of the Numeric Simulation Algorithm of RK-Butcher, as well as an environment of 4D trajectory navigation, in which the aircraft defines its trajectory through the definition of previously mentioned waypoints. In the end, it is possible to observe not only the success of the implementation of this method, but to make a comparative study between the different parameters used as well. After some reflection on the results obtained, and analyzing the content of this work through different states of mind, it is possible to make a very positive appreciation of the project that was developed. Not only the results matched the purposes defined in the beginning, but it also made possible to obtain a wider overview about the vast applications that these methods possess, not only in the field of Aeronautical Engineering, but in all the subjects of scientific investigation.
The present work aims to reconcile the neural networks subject, with the aircraft trajectory prediction theme. Specifically, it revolves around the application of Least-Squares Support Vector Machines methods in the prediction of 4-dimensional trajectories of commercial aircrafts. Initially, this work explains the guidelines regarding the machine learning and neural networks subjects through a theoretical explanation, gradually progressing to a more detailed elucidation of some of the functions applied in artificial neural networks. Within these functions, we highlight the Kernel Method and Radial Basis Functions, which were essential in order to obtain the practical results. In the last part of this chapter, the LS-SVM topic is approached theoretically, as well as the SESAR Project, which is responsible for the implementation of 4D trajectories in the European sky and the main trajectory data acquisition method in commercial aircraft, the Automatic Dependent Surveillance-Broadcast (ADS-B). All this theoretical explanation gives rise to the enunciation of the formal and mathematical content of this project, such as: Optimality Conditions, generation of Gaussian Error, Kernel Functions applied, as well as an explanation of how the Aerospace Localization methods work in a commercial aircraft. Lastly, all the data was simulated and validated in a context of Trajectory Prediction, allowing the algorithm to build a function that could predict the trajectory of the aircraft in its complete course. In order to achieve that, three different trajectories were used, created through the implementation of the Numeric Simulation Algorithm of RK-Butcher, as well as an environment of 4D trajectory navigation, in which the aircraft defines its trajectory through the definition of previously mentioned waypoints. In the end, it is possible to observe not only the success of the implementation of this method, but to make a comparative study between the different parameters used as well. After some reflection on the results obtained, and analyzing the content of this work through different states of mind, it is possible to make a very positive appreciation of the project that was developed. Not only the results matched the purposes defined in the beginning, but it also made possible to obtain a wider overview about the vast applications that these methods possess, not only in the field of Aeronautical Engineering, but in all the subjects of scientific investigation.
Description
Keywords
Algoritmos de Máquina de Suporte Vetorial Aprendizagem Automática Aprendizagem Semi-Supervisionada Funções de Bases Radiais Inteligência Artificial Navegação Por Waypoints Previsão de Trajetórias Redes Neuronais Trajetórias 4d