Abstract
Real-time monitoring of bioprocesses is hindered by sparse, heterogeneous measurements of key biological states, such as biomass, substrate, and product concentrations. Extended Kalman Filter (EKF)-based soft sensors offer a physics-grounded solution but are sensitive to limited observability, sensor bias, and process-model mismatch-conditions common in industrial fermentations. This work proposes a Levelized Hybrid Estimation Architecture (LHEA) that systematically enhances physics-based state estimation through increasing robustness and adaptivity while preserving model transparency and regulatory interpretability. The approach is evaluated using the KTB1 benchmark simulation model for continuous lovastatin production under an industrially realistic, cost-constrained sensor configuration combining dissolved oxygen, biomass proxy, volume measurements, and sparse HPLC product assays. Three estimator levels are investigated: (L1) a baseline EKF, (L2) a bias-augmented EKF for sensor-drift robustness, and (L3) a hybrid EKF with physics-constrained residual learning driven by sparse assays. Results show that L1 provides stable state reconstruction under nominal conditions but is sensitive to bias and model mismatch. L2 effectively isolates measurement bias, while L3 adapts to evolving process kinetics and achieves the lowest estimation errors under mismatch. These findings demonstrate that a structured, levelized integration of machine learning can significantly enhance soft-sensor reliability without sacrificing interpretability, providing a practical pathway toward robust digital twins.