Uso de sensores virtuais (soft sensors) para estimativa de impurezas em colunas de destilação de alta pureza.

Abstract

This work proposes a methodology for software implementation to make soft sensors. The goal is to estimate and predict the behavior of impurities in the bottom current of a highpurity distillation column, for 1,2-Dichloroethane or 1,2-DCE (C2H4Cl2) production. The data acquisition used in the construction of soft sensors was performed through a mathematical simulation of the process, with real industrial data taken from an industrial plant. A specific study involves the mathematical modeling, thermodynamics, evaluation of the steady state and the dynamic behavior of this process, simulated here in Aspen Plus and Aspen DynamicsTM software. Thus, the model provides the necessary data to infer the contents of the carbon impurities mentioned above, Tetrachloride (CCl4) and Chloroform (CHCl3), which are to be fixed approximately below 3000 and 400 ppm (parts per million) respectively. The methodology also covers selection algorithms of secondary variables, using multivariate statistical techniques: All Possible Regressions (TRP) and Principal Component Analysis (PCA). The data generated was checked in order to know whether to include or not a step for removal of outliers, so it was also included in the methodology one preprocessing data step. The ten best inference models were selected for each output concentration. With this information, these models do not use concentrations measurements in the feed streams but measurements of the temperature along the column. This is an important conclusion from the point of view of virtual sensors building, because in most of the literature reported these variables are crucial in getting good results. The training of soft sensors was done in a noisy environment, considering that simulated noise was inherent to measurements (Gaussian noise). In the final step, the soft sensors devices are constructed using an empirical modeling technique of artificial neural networks (ANN), which were generated ANN type Multilayer Perceptron (MLP). Several variations were also evaluated on the number of neurons and hidden layers of networks, employing as a stopping criterion the cross-validation technique. The developed soft sensors presented satisfactory errors from the engineering viewpoint, a good regression and a good mean square error. Finally, with these estimations it is expected to minimize and predict the transient behavior of the compounds in the referred process.