posted on 2021-06-24, 12:38authored byDebanjan Ghosh, Prashant Mhaskar, John F. MacGregor
This paper presents a unique strategy
for integrating fundamental
process knowledge with measurement data to build a partial least squares
(PLS) model with improved estimation capability. To this end, variables
from two different sources are combined to create the predictor data
matrix for the PLS model. Measurement data from sensors is stored
and used as inputs to a modified first-principles model to generate
trajectory data of unmeasured variables. Then the traditional X data
matrix (built with measured data) is augmented with batch trajectory
data of the calculated variables. The PLS model built with this augmented
matrix is referred to as hybrid/augmented PLS, and this proposed methodology
is tested on a seeded batch crystallization process to illustrate
this straightforward but powerful approach to estimate the final crystal
size distribution. The efficacy of the proposed approach is demonstrated
using simulation studies by comparing the results with the standard
PLS and subspace based quality model.