Estudo da utilização de algoritmos de deep-learning junto ao método dos elementos finitos para otimização de estruturas aeronáuticas reticuladas

Carregando...
Imagem de Miniatura

Data

2022-06-29

Autores

Pacheco, Matheus Felipe

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Estadual Paulista (Unesp)

Resumo

O presente trabalho se propõe a estudar o uso de técnicas de Deep Learning em conjunto com o método dos elementos finitos para resolver e otimizar estruturas simples, bidimensionais, treliçadas. Primeiramente, uma determinada geometria de treliça bidimensional com seções transversais variáveis é escolhida como caso de estudo; em seguida, o método dos elementos finitos é implementado usando a linguagem Python e usado para resolver o problema para uma população gerada aleatoriamente de conjuntos de áreas de seções transversais e um para um par de limites de deslocamento e tensão, que são gerados aleatoriamente dentro de um intervalo de valores. São utilizadas diferentes estratégias para gerar as populações, algoritmos de soma zero e diferentes distribuições estatísticas. Em seguida, uma rede é treinada para encontrar uma saída binária que indica se algum limite de tensão e deslocamento previamente definido foi excedido. Finalmente, essa rede treinada é usada como restrição para um algoritmo genético gerar famílias de áreas otimizadas, com objetivo de utilizar uma segunda rede que terá como entrada um valor para deslocamento máximo e tensão máxima e retornará as áreas otimizadas. Ao contrário da primeira rede neural, esta segunda rede neural possui uma saída não binária, sua resposta são os valores para as áreas que minimizam o peso da estrutura. A metodologia difere das usualmente utilizadas na literatura ao aplicar duas redes treinadas independentementes ao invés de uma. Para criar a arquitetura das redes, foram utilizadas as bibliotecas pré-implementadas em Python Keras, Tensor-Flow e Scikit-Learn. A metodologia é então utilizada para testar a influência de diferentes hiper parâmetros e diferentes estratégias de geração de população nas razões de convergência e significância dos resultados.
The present work proposes to study the use of deep learning techniques in conjunction with the finite element method to solve and optimize simple, two-dimensional, lattice structures. First, a given two-dimensional truss geometry with variable cross-sections is chosen as a case study; then the finite element method is implemented using the Python language and used to solve the problem for a randomly generated population of sets of cross-sectional areas and one for a pair of displacement and stress limits, which are randomly generated within a range of values. Different strategies are used to generate the populations, zero-sum algorithms and different statistical distributions. Next, a network is trained to find a binary output that indicates whether any previously defined voltage and displacement thresholds have been exceeded. Finally, this trained network is used as a constraint for a genetic algorithm to generate families of optimized areas, in order to use a second network that will have as input a value for maximum displacement and maximum stress and will return the optimized areas. Unlike the first neural network, this second neural network has a non-binary output, its response is the values for the areas that minimize the weight of the structure. The methodology differs from those usually used in the literature by applying two independently trained networks instead of one. To create the architecture of the networks, pre-implemented libraries in Python Keras, Tensor-Flow and Scikit-Learn were used. The methodology is then used to test the influence of different hyperparameters and different population generation strategies on the convergence and significance ratios of the results.

Descrição

Palavras-chave

Aprendizado do computador, Análise estrutural (Engenharia), Otimização matemática, Python (Linguagem de programação de computador), Análise de elementos finitos

Como citar