Classifier algorithms for tuning multi-model soft sensors : application to the estimation of quality variables in a continuous industrial process

Abstract

In this work, a multi-model soft sensor (SS) is proposed to estimate non-measurable variables in continuous processes. The proposed approach involves a first stage of clustering, using Gaussian mixture models, to identify the clusters that represent the multiple working conditions of the process. Then, for each identified cluster, multivariate linear regression sub-models are calibrated. Finally, the required non-measurable variable is estimated through a linear combination of the estimations from each sub-model. The weight coefficients for each sub-model are calculated using a classification algorithm. The performance of four different classification algorithms is evaluated in terms of the capability of their resulting multi-model soft sensor to estimate the mass conversion in a numerical simulation of a continuous emulsion polymerization for industrial production of Styrene-Butadiene Rubber. The results showed that the classifier model plays an important role in the multi-model soft sensor performance. Furthermore, a multi-model soft sensor that assigns the weights through Gaussian mixture models performs better than cases where a multi-layer perceptron, a linear discriminant analysis, or a K-nearest neighbors are used.