Construção de um sensor virtual para a classificação de emissões de dióxido de enxofre em uma caldeira Kraft via algoritmo k-NN (k-Nearest Neighbours)

• Ano de defesa: 2021
• Autor(a) principal: Paolla Marlene Caetano da Cunha
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA QUÍMICA; Programa de Pós-Graduação em Engenharia Química; UFMG
• Palavras-chave em Português: Sensor virtual; Caldeira kraft; Dióxido de enxofre; k-Nearest Neighbors (KNN); Ensemble learning; Processo Kraft; Estudo de caso real; Engenharia química; Polpação alcalina por sulfato
• Link de acesso: http://hdl.handle.net/1843/38483

Industrial development is one of the main factors for environmental regulations in general. Certifications that standardize management and process parameters have been consolidated worldwide and are now seen as mandatory for sustainable industrial operation. The development of data-driven models has been strengthened by the massive generation of data by industrial processes in general all over the world, given the advances in the areas of instrumentation, informatics, and databases. This advance then contributes to the spread of data science applications in the industrial sector. One sub-area of data science concerns machine learning methods, which are usually simple to implement and are directly obtained from data sets. This work explored a supervised k-nearest neighbours (KNN) classification problem, focusing on gas emissions that pollute human health and the environment. Specifically, it proposes the construction of a soft sensor based on data for monitoring and, consequently, controlling sulphur dioxide (SO2) emissions in chemical recovery boilers in the kraft pulp industry. Regarding the methodology, the behaviour of the method with one input variable, with subsets of predictors, and with an ensemble learning, for six classes of SO2 was presented. The results presented are satisfactory, which is important to generate confidence in industrial implementations. The ensemble learning showed the best performance, with accuracy of 92% and geometric mean of 94.75% over the (independent) test set.