Inteligência artificial aplicada a modelagem de processos da indústria de alimentos
Arquivos
Data
2021-08-05
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Viçosa
Resumo
A torra e o forneamento são responsáveis pelas transformações observadas em alguns alimentos, tais como café e pães. Dentre essas modificações, a cor é aquela de maior destaque pois consiste num indicador da evolução do processo e tem sido utilizada no desenvolvimento de ferramentas de controle. Dessa forma, a modelagem do escurecimento não enzimático decorrente dos processos de torra e forneamento possibilita o desenvolvimento de sistemas de visão computacional para acompanhamento e classificação desses processos. A modelagem fenomenológica, baseada na abordagem da cinética de reações, permitiu uma maior compreensão dessas mudanças de cor. Porém, os modelos resultantes apresentam grandes limitações de ordem prática para sua aplicação nos processos industriais, em especial no contexto da indústria de quarta geração (Indústria 4.0), a qual preconiza o uso de sistemas e equipamentos inteligentes. Nesse sentido, a adoção de técnicas de inteligência artificial (AI), em especial as redes neurais convolucionais (CNN), parece ser o caminho a ser seguido. Assim, neste trabalho foram introduzidas técnicas de modelagem por AI do escurecimento não enzimático decorrente dos processos de torra e forneamento de café e pães, respectivamente. A adoção de CNN com reduzido número de camadas convolucionais resultou numa redução no consumo de memória de mais de 90%. Além disso, o sistema híbrido formado por uma CNN e uma máquina de vetores de suporte (SVM) resultou na redução de 93% no tempo de convergência. Foram identificados mais 20% dos núcleos convolucionais seletivos a cores, o que evidencia a capacidade das CNN de extrair características de cores e utilizá-las para classificação das amostras. Ao classificar amostras de pães e café, as CNN apresentaram exatidão superior a 98,0% e 95%, respectivamente, superando arquiteturas tradicionais. A CNN também foi capaz de estimar o tempo necessário para o final do processo de torra de café com Root Mean Square Error (RMSE) de 0,4 min. As CNN se mostraram uma poderosa ferramenta não invasiva e não destrutiva para modelagem do escurecimento não enzimático decorrente dos processos de torra e forneamento. Palavras-chave: Torra. Forneamento. Redes Neurais Convolucionais. Escurecimento Não Enzimático. Aprendizado Profundo.
Roasting and baking are responsible for the transformations observed in some foods, such as coffee and bread. Among these changes, color is the most prominent because it is an indicator of the evolution of the process and has been used in the development of control tools. Thus, the modeling of non-enzymatic browning resulting from roasting and baking processes enables the development of computer vision systems to track and classify these processes. Phenomenological modeling, based on the reaction kinetics approach, has allowed a better understanding of these color changes. However, the resulting models have major practical limitations for their application in industrial processes, especially in the context of the fourth generation industry (Industry 4.0), which advocates the use of intelligent systems and equipment. In this sense, the adoption of artificial intelligence (AI) techniques, especially convolutional neural networks (CNN), seems to be the way to be followed. Thus, in this work we introduced AI modeling techniques of non-enzymatic browning resulting from coffee and bread roasting and baking processes, respectively. The adoption of CNN with reduced number of convolutional layers resulted in a reduction in memory consumption of more than 90%. In addition, the hybrid system consisting of a CNN and a support vector machine (SVM) resulted in a 93% reduction in convergence time. More than 20% of the color-selective convolutional kernels were identified, which highlights the ability of CNNs to extract color features and use them for sample classification. When classifying bread and coffee samples, the CNNs showed accuracy of over 98.0% and 95%, respectively, outperforming traditional architectures. CNN was also able to estimate the time required for the end of the coffee roasting process with Root Mean Square Error (RMSE) of 0.4 min. CNNs proved to be a powerful non-invasive and non-destructive tool for modeling non- enzymatic browning arising from the roasting and baking processes.Keywords: Roasting. Baking. Convolutional Neural Networks. Non-Enzymatic Browning. Deep Learning.
Roasting and baking are responsible for the transformations observed in some foods, such as coffee and bread. Among these changes, color is the most prominent because it is an indicator of the evolution of the process and has been used in the development of control tools. Thus, the modeling of non-enzymatic browning resulting from roasting and baking processes enables the development of computer vision systems to track and classify these processes. Phenomenological modeling, based on the reaction kinetics approach, has allowed a better understanding of these color changes. However, the resulting models have major practical limitations for their application in industrial processes, especially in the context of the fourth generation industry (Industry 4.0), which advocates the use of intelligent systems and equipment. In this sense, the adoption of artificial intelligence (AI) techniques, especially convolutional neural networks (CNN), seems to be the way to be followed. Thus, in this work we introduced AI modeling techniques of non-enzymatic browning resulting from coffee and bread roasting and baking processes, respectively. The adoption of CNN with reduced number of convolutional layers resulted in a reduction in memory consumption of more than 90%. In addition, the hybrid system consisting of a CNN and a support vector machine (SVM) resulted in a 93% reduction in convergence time. More than 20% of the color-selective convolutional kernels were identified, which highlights the ability of CNNs to extract color features and use them for sample classification. When classifying bread and coffee samples, the CNNs showed accuracy of over 98.0% and 95%, respectively, outperforming traditional architectures. CNN was also able to estimate the time required for the end of the coffee roasting process with Root Mean Square Error (RMSE) of 0.4 min. CNNs proved to be a powerful non-invasive and non-destructive tool for modeling non- enzymatic browning arising from the roasting and baking processes.Keywords: Roasting. Baking. Convolutional Neural Networks. Non-Enzymatic Browning. Deep Learning.
Descrição
Tese de Doutorado defendida na Universidade Federal de Viçosa.
Palavras-chave
Café - Processamento, Escurecimento, Reação Maillard, Alimentos - Indústria, Imagens digitais, Inteligência artificial, Aprendizado do computador
Citação
COTRIM, Weskley da Silva. Inteligência artificial aplicada a modelagem de processos da indústria de alimentos. 2021. 120 f. Tese (Doutorado em Ciência e Tecnologia de Alimentos) - Universidade Federal de Viçosa, Viçosa-MG. 2021.