Submitted by Automa??o e Estat?stica (sst@bczm.ufrn.br) on 2016-06-14T20:19:26Z No. of bitstreams: 1 IgorGuedesReboucas_DISSERT.pdf: 4109459 bytes, checksum: 9d05bc4fe3e1208e2195bfe4405550de (MD5) / Approved for entry into archive by Arlan Eloi Leite Silva (eloihistoriador@yahoo.com.br) on 2016-06-15T23:42:40Z (GMT) No. of bitstreams: 1 IgorGuedesReboucas_DISSERT.pdf: 4109459 bytes, checksum: 9d05bc4fe3e1208e2195bfe4405550de (MD5) / Made available in DSpace on 2016-06-15T23:42:40Z (GMT). No. of bitstreams: 1 IgorGuedesReboucas_DISSERT.pdf: 4109459 bytes, checksum: 9d05bc4fe3e1208e2195bfe4405550de (MD5) Previous issue date: 2015-12-17 / Este trabalho consiste na elabora??o de uma Rede Neural Artificial (RNA) com o fim de modelar o comportamento de comp?sitos quando submetidos a carregamento de fadiga. A proposta ? desenvolver e apresentar um modelo misto, que associa uma equa??o anal?tica (Equa??o de Adam) ? estrutura da RNA. Tendo em vista que os comp?sitos geralmente apresentam comportamentos semelhantes quando sujeitos a carregamentos flutuantes, essa equa??o visa estabelecer um padr?o pr?-definido de compara??o para um material gen?rico, a fim de que a RNA ajuste o comportamento de outro comp?sito a esse padr?o. Dessa forma, a RNA n?o precisaria aprender por completo o comportamento de determinado material, pois a Equa??o de Adam faria boa parte do trabalho. Este modelo foi utilizado em duas arquiteturas de rede diferentes, modular e perceptron, com o objetivo de analisar a sua efici?ncia em estruturas distintas. Al?m das diferentes arquiteturas, foram analisadas as respostas geradas a partir de dois conjuntos de dados diferentes ? com tr?s e duas curvas S-N. Esse modelo tamb?m foi comparado com os resultados da literatura especializada, que utilizam uma estrutura convencional de RNA. Os resultados consistem em analisar e comparar algumas caracter?sticas, como a capacidade de generaliza??o, a robustez e os Diagramas de Goodman, desenvolvidas pelas redes. / This work consists basically in the elaboration of an Artificial Neural Network (ANN) in order to model the composites materials? behavior when submitted to fatigue loadings. The proposal is to develop and present a mixed model, which associate an analytical equation (Adam Equation) to the structure of the ANN. Given that the composites often shows a similar behavior when subject to float loadings, this equation aims to establish a pre-defined comparison pattern for a generic material, so that the ANN fit the behavior of another composite material to that pattern. In this way, the ANN did not need to fully learn the behavior of a determined material, because the Adam Equation would do the big part of the job. This model was used in two different network architectures, modular and perceptron, with the aim of analyze it efficiency in distinct structures. Beyond the different architectures, it was analyzed the answers generated from two sets of different data ? with three and two SN curves. This model was also compared to the specialized literature results, which use a conventional structure of ANN. The results consist in analyze and compare some characteristics like generalization capacity, robustness and the Goodman Diagrams, developed by the networks.