Inteligência artificial aplicada no controle de qualidade em linhas de produção

• Ano de defesa: 2013
• Autor(a) principal: Duarte, Júlia Bertelli
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; BR; Programa de Pós-graduação em Engenharia Mecânica; Engenharias; UFU
• Palavras-chave em Português: Controle de qualidade; Redes neurais auto-organizáveis; Algoritmos genéticos; Evolução diferencial; Sintomas vibroacústicos; Engenharia de produção - Controle de qualidade; Vibração; Redes neurais (Computação); Quality control; Self-organized map; Genetic algorithm; Differential evolution; Vibroacoustic symptoms; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: https://repositorio.ufu.br/handle/123456789/14927; https://doi.org/10.14393/ufu.di.2013.239

A great challenge in industry is to establish reliable criteria for quality control of products and services. However, many of these criteria have shortcomings due to the use of subjective parameters, such as hearing sensitivity of the experts in order to classify and define the final product quality. In this work a methodology based on vibroacoustics symptoms, self-organizing neural networks and optimization is presented to minimize the subjectivity effects of quality control in production lines and manufacturing. Initially, the sensitivity and efficiency of the Kohonen neural network (SOM) to segregate a set of signals using vibroacoustics symptoms correlated with nonconformities of the tested product was found. For the choice of the best symptoms, among a large set of possibilities, the SOM technique combined with heuristic optimization techniques such as Genetic Algorithm and Differential Evolution are used. The use of two optimization techniques to a case in which the groups of conformities and non-conformities were known led to good results in the segregation of groups with emphasis on the Differential Evolution, which required lower computational effort and resulted in lower values for the objective functions. To validate the proposed methodology a classical statistical analysis of the signals was performed. When comparing the results obtained by computational methods with classical analysis a good consistency between them was observed. So the computational methods to choose vibroacoustic parameters, symptoms, proved to be a good tool for quality control. At last, the developed methodology was used to segregate the signals of a small number of defective products of the signals of a large set of acceptable products. The aim of this was to study the possibility of using the methodology without prior knowledge of nonconforming products. In this case, there is a need for a network with a larger number of neurons to get good results. The results demonstrate that it is possible to construct a pass/no-pass symptoms vibroacoustic database to quality control purpose using only artificial intelligence, without the need of pre-existing signal sets of acceptable and not-acceptable products obtained through subjective analysis.