SENSOR ON-LINE PREDICTIVO DE PROPIEDADES EN EL PROCESO DE MEZCLADO DE GOMA MEDIANTE MODELOS DE REGRESIÓN CON SELECCIÓN DE VARIABLES

Abstract

This communication deals with the complex behavior of rubber mixture processes and the more accurate estimation of some properties of resulting rubber bands. The main issue is to develop an on-line soft sensor for estimating significant parameters related to rubber properties. The sensor would be able to avoid the continual discard of defective material, reducing its high costs associated. This can be achieved by detecting the unexpected process variations or even bad operating set points. The system is based on a “wrapper” scheme. First, a feature selection routine (backwards selection) is use to find the optimum feature subset from mixture process attributes, which will be utilized as inputs of linear regression model. Those attributes that better explain the dependent variables are determined in an iterative process and the most accurate solution will be finally selected. Our proposed sensor has several advantages, i.e. the use of a linear model provides wider and deeper knowledge of the industrial process and the backwards selection techniques allow us to obtain better parsimony models. Eventually, we demonstrate that the soft sensor is also able to establish the clear relations between the independent variables and rheometric parameters of rubber.