Desenvolvimento de sensor virtual utilizando redes neurais artificiais na destilação de bebidas fermentadas

• Autor(a) principal: Lima, Tiago Hora Alves de
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFS
• Texto Completo: http://ri.ufs.br/jspui/handle/riufs/17051
• topic: Bebidas destiladas; Sensores virtuais; Redes neurais artificiais; Distilled beverages; Soft sensors; Neural networks; ENGENHARIAS::ENGENHARIA QUIMICA

The need to monitor industrial processes reliably to ensure products quality contrasts with the high cost of available equipments and the difficulty of obtaining information in real time. In relation to Brazilian distilleries producing spirits, this is an even greater challenge, since most are small, often family-owned enterprises. In this context, virtual sensors are a viable alternative for their ability to provide information about quality variables such as composition from easily measurable primary variables such as temperature and pressure using mathematical models programmed into hardware devices. With this objective, this work develops a soft sensor to infer the composition of ethanol in a batch distillation process in the production of distilled beverages of tropical fruits, overcoming the difficulty in the execution of the distillation cuts, made by a craft way. The proposed sensor is based on an artificial neural network of the feedforward multilayer perceptron type, which applies the Levenberg-Marquardt algorithm in the optimization of its parameters. For its construction were used data from binary distillations produced in the laboratory, composed of ethanol and water and with initial concentration close to the initial concentration of fermented fruit must. The built sensor presented excellent results, with mean absolute error (EAM) of 0,0140 to 0,0311 in mass fraction in the experiments performed. In order to prove its efficiency in the situation of interest, the sensor was tested in the fermented must distillations of mango, watermelon and jabuticaba, also produced in the laboratory, obtaining an excellent performance, with EAM of 0.0160 to 0.0324. In this way, the proposed virtual sensor proved capable of inferring the ethanol composition over time in a reliable way, being a viable alternative for the efficient monitoring.