Desenvolvimento de sensor virtual para predição do teor de cal livre no clínquer em uma fábrica de cimento

• Ano de defesa: 2019
• Autor(a) principal: Roberta de Souza Lima Magalhães
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Clínquer; Cimento; Qualidade; Predição; Sensor virtual; Predição (Lógica); Engenharia quimica; Cal
• Link de acesso: http://hdl.handle.net/1843/RAOA-BCTLSR

Due to the significant percentage of clinker in cement, the quality of clinker produced has a direct impact on cement quality and should be monitored continuously. However, clinker analysis in cement plants is not done in real time, with sampling and analysis usually exceeding two hours. The free lime content is considered a key indicator to evaluate clinker quality. Thus, the prediction of free lime value through virtual sensors based on online process data available in the cement industry presents itself as an interesting and low cost alternative to estimate clinker quality. The objective of this work is to propose a virtual sensor based on an empirical model to predict the free lime content in the clinker, from the cement plant operational data. The operational data modeling, using only online process data, was developed with a virtual sensor approach using multiple linear regression techniques and artificial neural networks - MLP and RBF type. The empirical model based on multiple linear regression presented better performance in relation to models based on artificial neural networks. It was verified in the results analysis of the determination coefficient and mean square error, as well as the residuals analysis, for the models obtained by multiple linear regression and by artificial neural networks. Considering the data base used in this study, composed of industrial data, and the results available in the literature, it can be concluded that the model obtained by multiple linear regression, considering 93 regression variables, was the best model obtained, explaining 73.09 % and 71.92% of the variation in free lime in the clinker (adjust and validation steps, respectively).