Effect of data availability on in-line adaptation of fermentation mechanistic models under abnormal conditions

초록

Industrial fermentation processes are vulnerable to abnormal events such as contamination and process disturbances, which can rapidly degrade the consistency between mechanistic models and actual plant characteristics. This study presents an in-line, data-adaptive model adaptation framework that integrates Bayesian parameter estimation with a mechanistic fed-batch fermentation model to maintain model reliability under abnormal conditions. Abnormal batch behavior is represented using empirical decay functions embedded within Monodtype kinetics, enabling explicit modeling of performance deterioration. Upon abnormality detection, Bayesian updating is applied using only the truncated dataset available at the early stage of an abnormal batch, providing robust and uncertainty-aware parameter estimates under data-scarce conditions. As additional process data become available, the framework transitions to a genetic algorithm-based population optimization approach, which achieves higher accuracy when sufficient data support reliable parameter identification. The proposed framework was evaluated using 30 industrial-scale contaminated fermentation batches. Under early-stage, sparse-data conditions, the Bayesian approach reduced post-detection prediction errors by 26% and 44% for substrate and product concentrations, respectively, compared with GA-based fitting. When larger datasets were available, GA-based optimization provided superior predictive accuracy, improving prediction accuracy by 25% and 19% for substrate and product concentrations, respectively, compared with Bayesian estimation. Unlike conventional single-method approaches, the proposed framework introduces a data-adaptive switching strategy between Bayesian inference and evolutionary optimization, enabling robust and accurate model adaptation across varying data availability conditions.

키워드