Navegando por Assunto "Rede neural de regressão geral"

Agora exibindo 1 - 2 de 2

Acesso aberto (Open Access)
Síntese de superfícies seletivas de frequência multicamadas via otimização bioinspirada
(Universidade Federal do Pará, 2019-08-23) LIMA, Wirlan Gomes; ALCÂNTARA NETO, Miércio Cardoso de; http://lattes.cnpq.br/0549389076806391; BARROS, Fabrício José Brito; http://lattes.cnpq.br/9758585938727609
The analysis of electromagnetic devices via computer software usually demands high computational cost and high processing time. In certain situations, to meet certain design objectives, finding the optimal structural parameters can take days or even weeks when done by trial and error when seeking accurate answers in highly complex structures. In this scenario, bioinspired computation (BIC) tools are strong allies in saving time, computational cost and, consequently, money. To enhance the power and efficiency of these tools, hybrid methods have been developed in which neural networks work in conjunction with optimization algorithms to obtain even more satisfactory and accurate results. In this context, this work presents the use of two multiobjective bioinspired hybrid optimization models for the design and synthesis of multilayer frequency selective surfaces (FSS). Initially, an electromagnetic investigation of the unit cell of the patch-like structures that will compose the multilayer FSS is made, which are a triangular loop and a solid diamond printed on fiberglass substrate (FR-4). The computer simulations were performed with the aid of CSTR○ Micro Wave Studio software, whose finite integrals (FIT) numerical technique is used. Three filters with distinctive characteristics that cover the C, X and Ku bands are designed. The synthesis process consists of tuning the objectives of the structures inserted in the cost function of the optimization algorithms. The modeling of the structures is performed by a general regression neural network (GRNN) and the optimization process is performed by the algorithms. The computational simulations for calculating the electromagnetic (EM) data of the multilayer FSS were performed using the CSTR○ software. The optimized values returned by the hybrid models were also simulated using Ansoft 𝐷𝑒𝑠𝑖𝑔𝑛𝑒𝑟𝑇𝑀 HFSS software to evaluate the previously obtained results. Good agreement between the simulated results was observed, showing a reduction in the processing time of the structures, besides showing that the GRNN-AG Multi model stood out in relation to the GRNN-MOCS, presenting errors in relation to the design objectives for the simulations. in CSTR○ of 0.44%, 0.254% and 0.387% for filters 1, 2 and 3, respectively, which is the most efficient hybrid model for multi-layer FSS optimization.
Acesso aberto (Open Access)
Síntese de superfícies seletivas de frequência para micro-ondas utilizando otimização multiobjetivo bioinspirada
(Universidade Federal do Pará, 2015-08-19) ALCÂNTARA NETO, Miércio Cardoso de; D'ASSUNCÃO, Adaildo Gomes; http://lattes.cnpq.br/4159638862269940; CAVALCANTE, Gervásio Protásio dos Santos; http://lattes.cnpq.br/2265948982068382
The evolution of computing has made possible substantial advances in research related to engineering and industrial projects. In these areas, the use of computational tools has been intensified for simulation and obtaining certain parameters of the project. However, the growing demand for precision and the gradual increase of the complexity of the structures and systems, results in a simulation process increasingly time-consuming, because the evaluation of a single criterion can consume several hours, as well as several days or even weeks. Therefore, a method that maximizes the time of simulation and optimization, can thus save time and money. In this context, bioinspired computing (BIC), presents accurate and efficient, where many traditional computational methods fail and consists of new mechanism to address such difficulties. Thus, in this work, a study about some of the algorithms used today for BIC design and optimization of general problems in engineering and industry. From now on, one sees develop a metaheuristic optimization code to provide lower-cost computational multiobjective and, consequently, less time for data processing. Initially, an electromagnetic research of frequency selective surfaces with triangular patch elements is done by computer simulations. The numerical analysis is carried out using a full-wave technique based on finite integrals implemented on commercial software performed for simulations in electromagnetism. The synthesis process consists of tuning the resonant frequency of the structures and the bandwidth according to the objectives in the cost function optimization algorithms. The modeling of structures is performed by an artificial neural network and optimization process is performed by meta-heuristics algorithms. The results obtained by these codes are compared to simulated ones by commercial software and measured. A good agreement between simulated and measured results was obtained, as well as a substantial reduction in the processing time of the structures. Finally, conclusions and proposals for further works are presented.