Desenvolvimento de um gerador de sinais de uma cuba de redução de alumínio como fonte de dados para aprendizagem de uma rede neural artificial.

Abstract

The electrolysis of aluminum is an electrochemical method that consists of producing metal through the mediation of an electric current and alumina dissolved in a cryolite bath at high temperature. Currently, the mathematical models that represent the dynamics of this method under the form of mass balance and energy balance are constituted of common connected differential equations, nonlinear and the changeable coefficients. This paper aims to develop a generator of signals in an electrolytic tank that will serve as source of data for the learning of artificial neural network, through a linear model of order ten, with eighteen parameters, in function of dac and dma, that describes the behaviour of the process of reduction of aluminum. The modeling and simulation of this model allowed the visualization and study of each variable and parameter involved. We determine the optimal parameters of the model and effect an analysis of sensitivity of the parameters, that indicates the dependence between these parameters and the impact of each parameter in the model. Through a typical relation between the resistance (R) and the alumina concentration (C) it was possible to develop an empirical model and an adjustment of the parameters was done. We verified that the model is adjusted to represent R and C. Developed software can be used to feed the artificial neural network that will return the parameters for each operation situation, allowing to determine new values of concentration of alumina, temperature of the bath and mass of the crust with no measurement needed.