Anomaly detection for drug product manufacturing considering data limitations and shifts: A case study on industrial freeze-dryers

Abstract

Ensuring the quality of biopharmaceutical products requires robust manufacturing processes and reliable monitoring systems. In industrial applications with real data, traditional data-driven anomaly detection methods often face challenges due to data scarcity and data shifts. To address these challenges, we propose the application of Statistic Alignment (SA) as a domain adaptation technique within the broader framework of transfer learning. A methodology is presented incorporating SA as an effective precursor for One-Class Support Vector Machine (OCSVM) based anomaly detection. Using industrial data from two parallel freeze-dryers, we investigate two cases: (1) transferring a model within the same machine to handle data shifts caused by maintenance, and (2) transferring a model between machines to assess cross-system transferability. We evaluate three SA methods—Mean Alignment, Standard Alignment, and Correlation Alignment—while also exploring data requirements for effective alignment. Moreover, we propose a heuristic-based hyperparameter tuning method for OCSVM to further enhance model performance. Our results demonstrate that SA allows model transfer between domains with F1 scores around 0.9, offering a promising solution for enhancing model robustness in dynamic biopharmaceutical production environments.