Cross-Plant Monitoring Model Transfer for Fed-Batch mAb Production under Limited Target Data

Establishing reliable multivariate statistical process monitoring (MSPM) after transfer of a biopharmaceutical process is difficult when only a few normal target-site batches and no labeled faulty batches are available. This work contributes to addressing two key challenges in biopharmaceutical manufacturing: (i) developing a reduced first-principles simulator for two industrial 2000 L fed-batch monoclonal-antibody processes and (ii) transferring MSPM models between source and target plants that, while similar in scale, differ in operating conditions, batch recipes, configurations, and sensor systems. The simulator couples cell-population kinetics, reaction-informed effective metabolite rates, gas–liquid mass transfer, acid–base chemistry, prescribed feed events, and pH and dissolved-oxygen control. Its effective coefficients were calibrated against industrial trajectories and frozen before held-out, batch-conditioned validation; it is used as a fit-for-purpose engineering scenario generator rather than as a uniquely identified mechanistic digital twin. We next develop recursive principal component analysis (RPCA) transfer variants for the few-target-batch setting and evaluate four strategies as completed target batches become available. Site-wise target scaling can amplify near-zero cross-batch variance in plateau- and control-dominated regions, whereas source-anchored variance shrinkage stabilizes this failure mode. Because the trajectories are nonstationary, autocorrelated, and non-Gaussian, we replace global, time-invariant parametric F and χ2 limits with empirical, within-batch-time-resolved Hotelling’s T2 and squared prediction error (SPE) limits. Process monitoring approaches are compared using measured normal histories and a common set of simulated normal and engineering-defined fault cases; paired evaluations show that source-anchored shrinkage and pooled scaling provide the leading scarce-target-data trade-off, chiefly through improved false-alarm control while preserving fault sensitivity.