Development of Clustering-Based Adaptive Soft Sensors for Industrial Distillation Columns

Abstract

The problem of adaptive soft sensor development with the use of clustering methods is considered in the examples of a reactive distillation process for the production of methyl tert-butyl ether and of crude distillation unit. We suggest to use clustering methods to assess whether updating the model parameters is expedient. An adaptive soft sensor operation algorithm using a “moving window” and clustering is proposed and tested on industrial data. The dependence of the soft sensor accuracy on the training sample window width is studied, and optimality criteria for the window width are considered. Our adaptive soft sensor with clustering is shown to be advantageous in accuracy and model parameter recalculation time over the traditional approach, where the model parameters are adapted at each step.