An online soft sensor method for biochemical reaction process based on JS-ISSA-XGBoost

Abstract

Background A method combining offline techniques and the just-in-time learning strategy (JITL) is proposed, because the biochemical reaction process often encounters changing features and parameters over time. Methods Firstly, multiple sub-databases in the fermentation process are constructed offline by an improved fuzzy C-means algorithm and the sample data are adaptively pruned by a similarity query threshold. Secondly, an improved eXtreme Gradient Boosting (XGBoost) method is used on the online modeling stage to build soft sensor models, and the multi-similarity-driven just-in-time learning strategy is used to increase the diversity of the model. Finally, to improve the generalization of the whole algorithm, the output of the base learner is fused by an improved Stacking integration model and then the predictive output is performed. Results Applying the constructed soft sensor model to the problem of predicting cell concentration and product concentration in Pichia pastoris fermentation process. The experimental results show that the root mean square error of the cell concentration is 0.0260, the coefficient of determination is 0.9945, the root mean square error of the product concentration is 2.6688, and the coefficient of determination is 0.9970. It shows that the proposed method has the advantages of timely prediction and high prediction accuracy, which validates the effectiveness and practicality of the method. Conclusion The JS-ISSA-XGBoost is an extensive and excellent soft measurement model that meets the practical needs for real-time monitoring of parameters and prediction of control in biochemical reactions.