Aplicação de uma rede neural artificial NARX para obtenção do comportamento dinâmico de um atenuador de impacto de alumínio do tipo Honeycomb

• Ano de defesa: 2019
• Autor(a) principal: Souza, Felipe Maraschin Pereira de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Engenharia Mecânica; Programa de Pós-Graduação em Engenharia Mecânica; UFPB
• Palavras-chave em Português: Atenuador de impacto; Honeycomb; RNA; NARX; Impact Attenuator; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/19437

In the automobilist industry, one of the biggest concern is about the safety of the passengers. Due to it, in automobiles, there are various equipments used with the goal to, increasingly, make them safer. Among them, an important equipment is the impact attenuator, which has the purpose of absorbing the energy of the impact in case of collision. Many projects of impact attenuators applied on racing cars utilize the honeycomb as the raw material. This paper proposes a methodology that evaluates the acceleration, the absorbed energy and the deformation in function of the dimension of the attenuator using an Artificial Neural Network, a recurrent neural network (feedback) and a directly fed neural network with multiple layers (feedforward). For training and validation of the methodology proposed it was utilized the acceleration data, deformation and energy absorbed obtained from the numerical analysis made through the Finite Element Method (FEM) of the impact attenuator. The input parameters of the ANNs were based on the rules of the Formula SAE, regarding the speed, mass of impact and the minimal dimension of the attenuator. The results obtained were satisfactory, showing that both types of neural networks were able to learn the dynamic behavior of the impact attenuator, being the feedback network best performance for acceleration with the mean absolute percentage error of 3.9%, and the feedforward network best results for deformation and energy absorption MAPE of 6.3% and 2.29% respectively.