EXTENDED KALMAN FILTER (EKF) AND ARTIFICIAL NEURAL NETWORKS (ANN) FOR ESTIMATE OF CONCENTRATION IN A NON ISOTHERMIC CSTR OF PRODUCTION OF PROPYLENE GLYCOL
Marcelo da Silva Pedro1; Walter Yanko de Aragão Brandão1; Paulo Romero de Araujo Mariz1; Emanuella Francisca de Lacerda Vieira1; Jéssica Oliveira de Brito Lira1; Arioston Araújo de Morais Júnior1; Leopoldo Oswaldo Alcázar Rojas1; Jonas Ladson Marinho da Silva Santos1
1 UFPB - Universidade Federal da Paraíba
doi:10.20906/CPS/CILAMCE2017-0243
Resumo
The most relevant piece of equipment in the chemical industry is the chemical reactor. Those equipments usually demand high amounts of energy during its operation, and considerable costs to install and maintain. That makes necessary a decent knowledge about the operation at any time and the remote monitoring of different variables inside the reactor. Several physical sensors are present in a chemical plant, providing measurements with good precision and low sampling times, with the flow, temperature, level, and others standing out. However, when the measurement of chemical and biochemical variables is desired (e.g. biomass, average particle diameter, and product concentration) several difficulties are present, with the need for offline sampling usually being necessary. To overcome these and other difficulties the digital sensors were developed, which are mathematical models implemented in different software solutions making use of secondary physical measurements. In that sense, this study proposes the creation of virtual sensors for the concentration inference in the production process of Propylene Glycol (C3H8O2). The developed methodology was created in the following order: the mathematical modelling was initially proposed; the process was simulated in stationary and transient regimes using the software solution MATLAB - Simulink®; the soft sensors were constructed employing two different kinds of modeling, the first being a semi-empirical approach known as the extended Kalman Filter (EKF), and the second one following an empiric approach known as Artificial Neural Networks (ANN). The results obtained with the soft sensor through the artificial neural networks (CAANN)-with levenberg marquardt training algorithm-presented better performance compared to CAEFK, given the fact that ANN is a powerful non-linear identification algorithm and has a high capacity of generalization.The results presented were quantified using the error criteria MSE and RSME, where it was observed smaller values of estimation error for the
Palavras-chave: chemical reactors; soft sensor; extendend filter Kalman; articial neural network