Adaptive Ensemble-Based Hyperparameter-Free Just-In-Time Learning for Robust Cell Culture Process Monitoring
- Publisher:
- Elsevier
- Publication Type:
- Journal Article
- Citation:
- Procedia Computer Science, 2025, 264, pp. 157-166
- Issue Date:
- 2025-01-01
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The production process in the biopharmaceutical industry relies on various Process Analytical Technologies (PAT) to ensure safety and consistency. Raman spectroscopy is a commonly used PAT, enabling real-time monitoring and control of cell culture processes by correlating Raman signals with key variables through data-driven models. To enhance the performance of these Raman models, just-in-time learning (JITL) can be employed to learn from a relevant subset of historical data, allowing JITL to adapt quickly to data generated from dynamic environments, such as bioreactors. This ensures the models remain accurate and effective in real-time applications. A critical and highly sensitive parameter in JITL is the number of data points used to train the model, commonly denoted as k. However, most studies determine k based on past experiences or trial and error and apply the same k value for all experiments, which is not ideal for data with constant changes, such as bioprocess data, leading to suboptimal outcomes. This study addresses this gap by developing a two-layer ensemble-based JITL method that combines predictions from models with different k values and automatically adjusts the weights of individual predictors to adapt to changes in the data, eliminating the need for manual adjustments from the user. Thirty-eight experimental bioreactor runs using different cell lines and feed media were conducted to evaluate the algorithm's performance. The findings demonstrate that the ensemble approach consistently achieved superior performance, eliminating the need to manually select and tune the highly sensitive hyperparameter k, creating a hyperparameter-free Just-In-Time Learning approach for users. This approach is particularly appealing in scenarios where hyperparameter tuning is not feasible due to limited available data and frequently changing environments requiring constant tuning of hyperparameters.
Please use this identifier to cite or link to this item: