Utilização de algoritmos de aprendizagem de máquina para predição de absorção de água nas carcaças de frango: aplicação de redes neurais artificiais

• Ano de defesa: 2021
• Autor(a) principal: Pinto, Mariana Camargo Leal
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Londrina; Brasil; Programa de Pós-Graduação em Tecnologia de Alimentos; UTFPR
• Palavras-chave em Português: Frango de corte; Carne - Resfriamento; Alimentos - Umidade; Redes neurais (Computação); Broilers (Chickens); Meat - Cooling; Food - Moisture; Neural networks (Computer science); CNPQ::CIENCIAS AGRARIAS::CIENCIA E TECNOLOGIA DE ALIMENTOS::TECNOLOGIA DE ALIMENTOS; Tecnologia de Alimentos
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/24950

The absorption of water in chicken carcasses has been researched by several authors regarding specific interferences during the stages of the chiller cooling process. Meeting the limits foreseen in the legislation for the absorption of water in the carcasses of birds and seeking the closest values foreseen by the legislation, imply economic benefits for the industries, increasing the profitability and quality of the product. In the proposed study, the machine learning algorithm, one of the models of artificial neural networks, was used to predict the values and classify the variables with the greatest influence of water absorption in chicken carcasses of an industrial plant, helping plant managers to make decisions. As a result, it was possible to identify the ANN with the best performance for water absorption, and the criteria used to evaluate the results obtained were the correlation coefficient (R) between the predicted and experimental values and the mean quadratic error (RMSE). For the dripping test, the correlation coefficients, both for Cross Validation and for Percentage Split, presented close values (R 0.3902 and R 0.4207), which are considered low when compared to the results obtained for water absorption. Regarding the prediction of water absorption values in carcasses, the application of Multilayer Perceptron had a good evolution as adjustments were made to the algorithm using the Weka software, presenting correlation coefficients of 0.95, using 1 hidden layer with four (4 ) neurons and three thousand (3,000) training periods. The neural network algorithm identified the variables - average mass, pre-chiller retention time, pre-chiller temperature and scalding temperature; which influence the water absorption process and the literature corroborates the variables raised in the study.