Sensores virtuais utilizando redes neurais artificiais para estimação da consistência e do freeness de celulose

Abstract

To remain competitive, the process industries need to produce with high quality levels and at the same time optimize energy and raw material consumption. To achieve these objectives, real-time monitoring and control of as many parameters as possible is necessary, including process variables and product quality indicators. Although instruments for real-time measurement of most variables already exist in pulp refining, some are still obtained through laboratory tests. This is due to technical limitations and acquisition costs of physical sensors, such as measuring pulp consistency in the highconsistency refining process and measuring drainage capacity (freeness). Considering that these two variables depend on others, for which there are already existing measurement systems and data collections, the development of soft sensors for their estimation is a good alternative to overcome the limitations of current techniques. In this work, three types of artificial neural networks, functioning as soft sensors, are used in a comparative manner to estimate consistency and freeness: the Multilayer Perceptron (MLP), the neural network with Radial Basis Function (RBF) and the Extreme Learning Machines (ELM). For performance comparison purposes, the variables were also estimated using multiple linear regression by the ordinary least squares (OLS) and partial least squares (PLS) methods. Most of the performances obtained were promising, highlighting the ELM network that presented the best estimation results for consistency and freeness with a mean absolute percentage error of 2.44% and 3.44% respectively.