Multivariate statistical monitoring of an industrial SBR process. Soft-sensor for production and rubber quality

This work presents a full methodology to build and evaluate a soft sensor capable of monitoring the production of Styrene–Butadiene Rubber (SBR) in an industrial train of 7 continuously-stirred tank reactors. The aim is to develop a device for on-line estimation of production and quality variables using a multivariate statistical technique like partial least squared (PLS). Besides pursuing the soft sensor development, this paper attempts to provide a guide for similar developments by suggesting attention to several specific methodological points. In this regard, the following features are highlighted: i) since a wide range validation space is desired for this sensor and the actual plant cannot be arbitrarily disturbed, an existing complex fundamental model is used to explore different possible operating conditions; ii) the approach used to develop the soft sensor includes a distributed sampling of multivariate steady-state conditions to collect the calibration data set, and the use of a filter for excluding outliers; and iii) it is shown that the analysis of how the explained variability progress when latent variables are included in the model allows the detection of poor predictor variables providing the chance for improving the multivariable regression by eliminating interfering contributions. Few after-modeling techniques are also suggested to confirm the consistency of the calibrating data set and the model precision over the applicability domain.