Nonlinear Kalman Estimators for Low-Cost Bioprocess Monitoring with Unstructured Mechanistic Models

Abstract

Bioprocess monitoring is a critical component in the biopharmaceutical industry, essential for ensuring the consistent production of high-quality biopharmaceutical products. Despite significant advancements in bioprocessing technologies, monitoring techniques have not kept pace, leading to inefficiencies and increased production costs. Therefore, there is constant pressure for bioprocess monitoring strategies that are efficient and low-cost. The literature indicates that soft sensors based on nonlinear Kalman estimators (NKE) with unstructured mechanistic models (UMM) can enable fast and low-cost bioprocess monitoring. However, classic NKEs with UMM are limited to fast and low-cost bioprocess monitoring, and improvements are needed to handle different biomanufacturing conditions. Therefore, the main goal of this thesis is to enable the design and development of NKEs with UMM for fast and low-cost monitoring of highly nonlinear bioprocesses, such as viral vectors and monoclonal antibody productions in different biomanufacturing conditions. The thesis is structured around three main research questions. The first investigates how the joint estimation of states and unshared parameters by NKE with specific UMM can be achieved under biomanufacturing conditions for efficient monitoring. The proposed solution, called SANTO (Specific initiAl coNdiTiOn), enhances the performance of joint NKE by preventing the Kalman gain from being zero throughout the process. The second research question addresses the automatic tuning of all NKE components for new biomanufacturing conditions. The Batch Bayesian Auto-Tuning (BAT) approach is introduced, leveraging all available measured data to define a posterior distribution of NKE components, thereby enhancing the consistency and accuracy of the estimators. The third research question explores increasing the adaptability (generalization), simplifying the tuning process, and improving real-time parameter estimation of NKEs with generic UMM for highly nonlinear bioprocesses. The Hybrid Nonlinear Kalman Estimator (HNKE) is proposed as a novel hybrid Gaussian filter integrating a hybrid dynamic model (containing uncertainty quantification) and unscented transformation or cubature rule. This novel method allows for auto-initialization of state error covariance and real-time estimation of parameters and significantly reduces the data and tuning requirements. Empirical evaluations using bioprocess datasets related to different real-world biomanufacturing conditions validate the effectiveness of these approaches, demonstrating significant improvements in monitoring accuracy and efficiency. The findings contribute to advancing bioprocess monitoring towards Biomanufacturing 4.0, optimizing operations, and reducing production costs. This research provides a foundation for future developments in bioprocess monitoring, with potential applications in various biomanufacturing scenarios.