Automatic feed phase identification in multivariate bioprocess profiles by sequential binary classification.

Area of Science:

• Biotechnology
• Machine Learning
• Process Engineering

Background:

• Accurate identification of feed phases in fed-batch fermentation is crucial for process optimization.
• Conventional multi-class machine learning (ML) approaches can be complex and susceptible to class imbalance.
• Existing methods often lack the parsimony and robustness needed for real-time industrial applications.

Purpose of the Study:

• To develop a new strategy for retrospective identification of feed phases using sensor data from Escherichia coli (E. coli) fed-batch fermentation.
• To improve model parsimony and accuracy by incorporating process knowledge into the classification system.
• To compare the performance of the proposed method against conventional static ML approaches.

Main Methods:

• A novel strategy employing a chain of binary classifiers with a one-way switch mechanism to enforce unidirectionality between feed phases.
• Utilizing regularized logistic regression (RLR) with Lasso for feature selection from high-dimensional sensor data.
• Training and comparing various soft computing classifiers (DT, k-NN, SVM, fuzzy classifier) with the proposed method.

Main Results:

• The proposed method significantly outperformed conventional static ML approaches in accuracy and robustness.
• Achieved near error-free feed phase classification, reducing misclassification rates by 39% to 98.2% in most test cases.
• Models using RLR-selected features showed superior performance and less sensitivity to lag parameter choice compared to expert-selected features.

Conclusions:

• The developed strategy offers a more parsimonious and robust approach to feed phase identification in fed-batch fermentation.
• Incorporating process knowledge through a unidirectional classifier chain enhances classification performance.
• The method demonstrates significant potential for improving the control and optimization of bioprocesses.