470662 Improved Soft Sensors for Mixed Culture System Monitoring
To address this challenge, spectra-based soft sensor was proposed  to estimate individual biomass concentration in a mixed culture. Although this approach is quick and sufficiently accurate, models are built from data with large number of variable compared to the number of samples, which makes models less accurate and less robust. This work is intended to address this issue.
In this work, models were developed to predict individual biomass concentration for co-culture mixtures of Scheffersomyces Stipitis & Methylomicrobium buryatense, and Escherichia coli & Saccharomyces cerevisiae using partial least squares (PLS). For outlier detection, principal component analysis (PCA) was applied. The main goal of this work was to investigate the effect of variable selection and experimental design on the performance of the PLS soft sensor. Variable selection methods studied in this work are variable importance in projection (VIP), synergy-iPLS, genetic algorithm, and recursive-VIP or R-VIP. Performances of soft sensors with different variable selection methods and different model training scenarios are compared and discussed.
 Stone K & Shah D., “A novel soft sensor approach for estimating individual biomass in mixed cultures”. Annual AIChE Meeting, Salt lake city, UT(2015)