Neural virtual sensor for determination of high-density polyethylene melt flow index and solids concentration in a loop slurry reactor

The high-density polyethylene (HDPE) production requires a precise control of process variables which are controlled by various instruments throughout the production process. Among all of them, two stand due to their importance: (i) the solids concentration (SC), which is measured by a nuclear mass measurer, and (ii) the melt flow index (MFI), that is done with a delay of 3 h from reactor. Accurate calculation and variable measurement are of great importance, and thus, the use of intelligent control tools is increasingly required. This study aimed to develop an artificial neural networks (ANNs)-based virtual sensor for the prediction of SC and MFI in the production of HDPE. Using real process data, principal component analysis was used to reduce the number of input variables and then several ANNs were trained. The results showed that the Levenberg-Marquardt algorithm was the most effective. The best result indicated that two ANNs, one for each variable, were necessary: The ANN for SC had an average error of 0.12%, and the ANN for the MFI had an average error of 3.8%. Finally, this study showed that a virtual neural sensor can be an accurate tool for predicting variables in real industrial processes.