Novo método de síntese de FSS multibanda baseado em aprendizado de máquina para sistemas de comunicação sem fio

• Ano de defesa: 2021
• Autor(a) principal: Fontoura, Leidiane Carolina Martins de Moura
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: FSS bioinspirada; FSS multibanda; Filtro de frequências; Síntese de FSS; Aprendizado de máquina
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/44631

The present work is a study on the application of supervised machine learning with the decision tree algorithm in the synthesis of frequency selective surfaces, or simply FSS. For this, the sunflower (Helianthus annuus) was used as a base element, being an original and simplified geometry, with frequency response characteristics similar to those of fractal structures. The thesis work is thus divided in two parts: the proposed element characterization and synthesis of the multiband FSS. Initially, the evolution of geometry and design equations are presented. The intermediate and the proposed structures are numerically characterized using the commercial software Ansoft Designer, manufactured, and experimentally characterized, with good agreement between the simulated and measured results. In the second step, the sunflower geometry is partially modified to define parameterization variables. The Ansoft Designer numerically characterizes the value of each variable of the new geometry, and it generates the frequency responses without repetition. The decision tree algorithm performs the dataset classification and evaluation, and the random forest algorithm validates and confirms the results. This process and the synthesis of the FSS using the decision tree algorithm occur in less than 10 seconds, with accuracy greater than 90%, meeting the desirable criteria, under two different scenarios. The decision tree algorithm learns simple decision rules inferred from training data, with simple and easy to implement calculations, written in the Python language. The accuracy is a parameter used to measure the quality of prediction in training and validation of the algorithm. Based on these scenarios, two FSS are manufactured and experimentally characterized, obtaining results with good agreement. The designed and fabricated FSS structures have closely spaced operation bands. Thus, it is observed that the agility and precision of this classification algorithm make the synthesis of the structures particularly interesting. Intuitive implementation, simplicity in training and validation, and an efficient data analysis model are highlighted.