Data synchronization techniques and their impact on the prediction performance of automated recalibrated soft sensors in bioprocesses

Innovative soft sensor concepts can recalibrate automatically when the prediction performance decreases due to variations in raw materials, biological variability, and changes in process strategies. For automatic recalibration, data sets are selected from a data pool based on distance-based similarity criteria and then used for calibration. Nevertheless, the most appropriate data sets often are not reliably selected due to variances in the location of landmarks and process length of the bioprocesses. This can be overcome by synchronization methods that align the historical data sets with the current process and increase the accuracy of automatic selection and recalibration. This study investigated two different synchronization methods (dynamic time warping and curve registration) as preprocessing for the automatic selection of data sets using a distance-based similarity criterion for soft sensor recalibration. The prediction performance of the two soft sensors without synchronization was compared to the variants with synchronization and evaluated by comparing the normalized root mean squared errors. Curve registration improved the prediction performance on average by 24% (Pichia pastoris) and 9% (Bacillus subtilis). Using dynamic time warping, no substantial improvement in prediction performance could be achieved. A major factor behind this was the loss of information due to singularities caused by the changing process characteristics. The evaluation was performed on two target variables of real bioprocesses: biomass concentration prediction in P. pastoris and product concentration prediction in B. subtilis.