Identificação de adulterantes no leite bovino por meio do uso de redes neurais artificiais e propagação de ondas mecânicas
Carregando...
Arquivos
Data
2021-08-11
Autores
Mazer, Rodrigo Hans
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
As adulterações alimentares causam prejuízos econômicos e de saúde, em âmbito mundial. O leite é um dos alimentos mais consumidos da humanidade por possuir grande poder nutricional e é alvo de inúmeras fraudes, difíceis de serem reconhecidas, especialmente quando são utilizados baixos percentuais de contaminações. Atualmente, as tecnologias capazes de inspeção de adulterantes no leite bovino são onerosas e lentas, questões motivacionais para a busca de novas alternativas para solucionar o problema. Nesse contexto, foi proposto um trabalho que desenvolveu um recipiente construído com diafragmas piezelétricos de baixo custo, propagando ondas mecânicas em amostras de leite contaminadas por bicarbonato de sódio, ureia e peróxido de hidrogênio, em proporções de 0,5% e 1%, cujos sinais elétricos foram adquiridos para processamentos computacionais. Por meio da utilização da técnica cromática, foi feita a decomposição dos sinais em três parâmetros que apresentaram características em que se foi possível diferenciar adulterações perante amostras de leite puro. Partindo-se desses dados, o presente trabalho utilizou oito ferramentas de Redes Neurais Artificiais (RNAs) de arquiteturas distintas para identificar cada amostra: Rede Neural Perceprom Multicamadas (RNPMC), Rede Neural de Retropropagaçãom em Cascata (RNRC), Rede Neural Elman de Retropropagação (RNER), Rede Neural de Reconhecimento de Padrões (RNRP), Rede Neural Probabilística (RNP), Rede Neural de Regressão Generalizada (RNRG), Rede Neural de Base Radial (RNBR) e Rede Neural de Base Radial Exata (RNBRE). Os dados, dispostos em 2 matrizes de 700 linhas por 4 colunas, foram divididos em 50% para o treinamento das redes e 50% o para teste, para todas as redes. Para efeitos comparativos individuais, também foram feitas execuções extras para as redes RNPMC, RNRC, RNER e RNRP, dividindo-se os dados em 75% para treinamento, 25% para validação e 25% para teste, pois as arquiteturas das mesmas permitem esta possibilidade. Os ensaios consistiram em 10 parametrizações diferentes para cada RNA, em que buscou-se os melhores desempenhos, evidenciados pelas quantidades e percentuais de acertos, erros quadráticos médios e regressões, para as devidas comparações individuais e entre as redes neurais. Os resultados das melhores parametrizações obtidas mostraram que as redes RNRP, RNER, RNPMC e RNRC foram capazes de identificar mais de 98% das amostras. As redes RNRG, RNP e RNBR obtiveram acertos de 86,71%, 86,57% e 85%, respectivamente, e a RNBRE, apenas 46,57%. Desta forma, a proposta desenvolvida demonstrou-se promissora e benéfica no apoio ao trabalho de identificação de adulterantes no leite, apresentando quatro opções com percentuais de acertos acima de 98%.
Food adulterations cause economic and health damage worldwide. Milk is one of the most consumed foods in humanity because it has great nutritional power and is the target of numerous frauds, difficult to be recognized, especially when low percentages of contamination are used. Currently, technologies capable of inspecting adulterants in bovine milk are costly and slow, motivational issues for the search for new alternatives to solve the problem. In this context, a study was proposed that developed a container built with low-cost piezoelectric diaphragms, propagating mechanical waves in milk samples contaminated by sodium bicarbonate, urea and hydrogen peroxide, in proportions of 0.5% and 1%, whose electrical signals were acquired for computational processing. Through the use of the chromatic technique, the signals were decomposed into three parameters that presented characteristics in which it was possible to differentiate adulterations toward samples of pure milk. Based on these data, the present work used eight tools of Artificial Neural Networks (ANNs) of different architectures to identify each sample: Perceprom Multilayer Neural Network (PMLNN), Cascade Backpropagation Neural Network (CBNN), Elman Backpropagation Neural Network (EBNN), Pattern Recognition Neural Network (PRNN), Probabilistic Neural Network (PNN), Generalized Regression Neural Network (GRNN), Radial Base Neural Network (RBNN) and Exact Radial Base Neural Network (ERBNN). The data, arranged in 2 matrices of 700 rows by 4 columns, were divided into 50% for training the networks and 50% for testing, for all networks. For individual comparative purposes, extra runs were also made for the PMLNN, CBNN, EBNN and PRNN networks, dividing the data into 75% for training, 25% for validation and 25% for testing, as their architectures allow this possibility . The experiments consisted of 10 different parameterizations for each ANN, in which the best performances were sought, evidenced by the amounts and percentages of correct answers, mean square errors and regressions, for the appropriate comparisons between individual and neural networks. The results of the best parameterizations obtained showed that the PRNN, EBNN, PMLNN and CBNN networks were able to identify more than 98% of the samples. The networks GRNN, PNN and RBNN obtained correct answers of 86.71%, 86.57% and 85%, respectively, and the ERBNN, only 46.57%. Thus, the developed proposal proved to be promising and beneficial in supporting the work of identifying adulterants in milk, presenting four options with percentages of correct answers above 98%.
Food adulterations cause economic and health damage worldwide. Milk is one of the most consumed foods in humanity because it has great nutritional power and is the target of numerous frauds, difficult to be recognized, especially when low percentages of contamination are used. Currently, technologies capable of inspecting adulterants in bovine milk are costly and slow, motivational issues for the search for new alternatives to solve the problem. In this context, a study was proposed that developed a container built with low-cost piezoelectric diaphragms, propagating mechanical waves in milk samples contaminated by sodium bicarbonate, urea and hydrogen peroxide, in proportions of 0.5% and 1%, whose electrical signals were acquired for computational processing. Through the use of the chromatic technique, the signals were decomposed into three parameters that presented characteristics in which it was possible to differentiate adulterations toward samples of pure milk. Based on these data, the present work used eight tools of Artificial Neural Networks (ANNs) of different architectures to identify each sample: Perceprom Multilayer Neural Network (PMLNN), Cascade Backpropagation Neural Network (CBNN), Elman Backpropagation Neural Network (EBNN), Pattern Recognition Neural Network (PRNN), Probabilistic Neural Network (PNN), Generalized Regression Neural Network (GRNN), Radial Base Neural Network (RBNN) and Exact Radial Base Neural Network (ERBNN). The data, arranged in 2 matrices of 700 rows by 4 columns, were divided into 50% for training the networks and 50% for testing, for all networks. For individual comparative purposes, extra runs were also made for the PMLNN, CBNN, EBNN and PRNN networks, dividing the data into 75% for training, 25% for validation and 25% for testing, as their architectures allow this possibility . The experiments consisted of 10 different parameterizations for each ANN, in which the best performances were sought, evidenced by the amounts and percentages of correct answers, mean square errors and regressions, for the appropriate comparisons between individual and neural networks. The results of the best parameterizations obtained showed that the PRNN, EBNN, PMLNN and CBNN networks were able to identify more than 98% of the samples. The networks GRNN, PNN and RBNN obtained correct answers of 86.71%, 86.57% and 85%, respectively, and the ERBNN, only 46.57%. Thus, the developed proposal proved to be promising and beneficial in supporting the work of identifying adulterants in milk, presenting four options with percentages of correct answers above 98%.
Descrição
Palavras-chave
Adulteração de leite bovino, Técnica cromática, Redes neurais artificiais, Reconhecimento de padrões, Adulteration of bovine milk, Chromatic technique, Artificial neural networks, Pattern recognition