Classificação de castas em vinhos do Douro usando redes neuronais artificiais
Ficheiros
Data
2015-05-11
Autores
Título da revista
ISSN da revista
Título do Volume
Editora
Resumo
No geral, o interesse da química é o estudo das propriedades, constituição e transformações de substâncias e materiais, sendo os alimentos um foco de grande interesse, estando o vinho dentro desse grande grupo.
O objetivo geral deste trabalho consiste em propor métodos de classificação de vinhos monocasta da região Demarcada do Douro com base em parâmetros químicos. Um primeiro objetivo prende-se com a discriminação da casta predominante num conjunto de amostras constituído por dois tipos de vinho monocasta (Touriga Nacional e Aragonês). Os dados das amostras resultaram de um anterior estudo que envolveu a análise por High-Performance Liquid Chromatography de vinhos das duas castas referidas e que resultou na recolha dos valores de antocianinas em cinquenta e três vinhos diferentes (vinte e um de Aragonês e trinta e dois de Touriga Nacional). O segundo objetivo é discriminar os vinhos por idade, primeiramente no geral (levando em conta as cinquenta e três amostras) e de seguida mais particular, fazendo a classificação por idade apenas nas amostras de cada casta (para as vinte e uma de Aragonês e para as trinta e duas de Touriga Nacional, em separado). Os classificadores usados baseiam-se em redes neuronais artificiais que se confirmaram como uma ferramenta muito útil e fiável na discriminação da casta predominante e da idade dos vinhos.
Os resultados gerais foram bastante aceitáveis, a rondar os 95% de acerto quanto à casta predominante. Independentemente da casta, e na tentativa de discriminação dos vinhos de todas as idades obtiveram-se resultados a rondar os 70%. Na separação dos vinhos entre os mais novos e os mais antigos (um, dois, três e mais que três anos, atingiram-se valores médios de 90%, visto que os vinhos com mais idade começam a ter características, à base de antocianinas, um pouco mais variáveis comparadas com os vinhos mais recentes.
A pouca abundância de amostras foi o que mais penalizou o estudo, visto que uma rede neuronal artificial funciona melhor quanto mais amostras existirem principalmente para realizar o treino. Quanto maior for o número de amostras usadas na fase de treino melhor será o processo de aprendizagem e melhores resultados se obtêm na classificação de novas amostras.
The interest of chemistry is the study of the properties, constitution and transformation of substances and materials, been the food is an item of big interest and wine is in this large group. The general objective of this work is to propose methods of classification of single variety wines from Douro region, based in chemical parameters. The first objective relates to caste prevailing discrimination in a group of samples constituted by two types of single variety wines (Touriga Nacional and Aragonês). Data samples resulted from a previous study which involved High-Performance Liquid Chromatography analyze of two varieties wines that resulted on a collection of values of anthocyanins in fifty-three different wines (twenty-one of Aragonês and thirty-two of Touriga Nacional). The second objective is to discriminate wines by age, first in general (taking into account 53 samples) and then more particularly, making the classification by age only in the samples of each variety (for 21 of Aragonês and 32 of Touriga Nacional separately). The classifiers used are based on artificial neural networks which were confirmed as a useful and reliable in predicting the predominant variety and age of the wine. The overall results were quite acceptable, around 95% hit as the predominant variety. Independently of caste, and in an attempt to discriminate wines of all ages yielded results of around 70%. In the separation among younger and older (one, two, three and more than three years), there were reached mean values of 90%, whereas older wines start having characteristics, based of anthocyanins, slightly more variable compared with older wines. The low abundance of samples was what most penalized the study, whereas an artificial neural network works better as more samples exist, principally for realize the train. As larger is the number of samples used in train phase, will be better the process of learning and will be better results in classification of new samples.
The interest of chemistry is the study of the properties, constitution and transformation of substances and materials, been the food is an item of big interest and wine is in this large group. The general objective of this work is to propose methods of classification of single variety wines from Douro region, based in chemical parameters. The first objective relates to caste prevailing discrimination in a group of samples constituted by two types of single variety wines (Touriga Nacional and Aragonês). Data samples resulted from a previous study which involved High-Performance Liquid Chromatography analyze of two varieties wines that resulted on a collection of values of anthocyanins in fifty-three different wines (twenty-one of Aragonês and thirty-two of Touriga Nacional). The second objective is to discriminate wines by age, first in general (taking into account 53 samples) and then more particularly, making the classification by age only in the samples of each variety (for 21 of Aragonês and 32 of Touriga Nacional separately). The classifiers used are based on artificial neural networks which were confirmed as a useful and reliable in predicting the predominant variety and age of the wine. The overall results were quite acceptable, around 95% hit as the predominant variety. Independently of caste, and in an attempt to discriminate wines of all ages yielded results of around 70%. In the separation among younger and older (one, two, three and more than three years), there were reached mean values of 90%, whereas older wines start having characteristics, based of anthocyanins, slightly more variable compared with older wines. The low abundance of samples was what most penalized the study, whereas an artificial neural network works better as more samples exist, principally for realize the train. As larger is the number of samples used in train phase, will be better the process of learning and will be better results in classification of new samples.
Descrição
Dissertação de Mestrado em Engenharia Informática
Palavras-chave
Redes neuronais (Artificiais) , Vinho , Castas