Modelo de redes neurais artificiais para previsão de frequência natural em lajes de concreto armado

• Ano de defesa: 2022
• Autor(a) principal: Marangoni, Bruno
• Orientador(a): Kripka, Moacir
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade de Passo Fundo
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Civil e Ambiental
• Departamento: Faculdade de Engenharia e Arquitetura – FEAR
• País: Brasil
• Palavras-chave em Português: Redes neurais (Computação); Dinâmica estrutural; Lajes de concreto; Concreto armado
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA CIVIL
• Link de acesso: http://tede.upf.br:8080/jspui/handle/tede/2494

In structural designs, it is necessary to analyze the dynamic external actions and determine the effect and relevance of the vibrations caused, so that, for example, excessive cracking and strong vibration can be controlled, based on the proper sizing of the structural components or by the limit determined by the natural frequency. The methodologies to obtaining natural frequency demand specific technical knowledge, are of high financial cost due to the equipment and software and require considerable time for execution. With the objective of developing an alternative tool to measure this property in concrete structures in a simplified way, reducing the execution time and the costs with the tests, it is proposed to study and elaborate a model to predict the natural frequencies of the slabs using Artificial Neural Networks (ANN). For this, network architectures were elaborated according to the database composed by in loco readings of ribbed and precast slabs. The supervised training, validation and testing of the networks were carried out with the (11-6-1) and (8-5-1) architectures, respectively. Among the hyperparameters analyzed, early stopping was responsible for stopping the training at the point where the network presented the best performance, consequently the lowest error value through the metrics used. The networks could be tested with 30% of the input data, comparing them with their predictions. It is possible to state that the network made it possible to predict frequencies satisfactorily for the low-frequency slabs. For slabs with higher frequencies, the network was not able to accurately predict, since a need to manipulate missing data in the database was detected.