Batch process monitoring in pharmaceutical manufacturing relies on historical process data to effectively characterize multivariate process behavior and construct reliable monitoring models. However, in many industrial environments, particularly during new product development, process scale-up, and technology transfer, only a limited number of representative batches are available. Recent work has addressed this challenge using Bayesian Gaussian Process State-Space (GP-SS) models to generate synthetic batch trajectories from limited historical data. However, this strategy models each process variable independently and therefore does not explicitly account for the underlying cross-variable covariance structure, which may limit the representation of multivariate process behavior. This thesis investigates the integration of principal component analysis (PCA) into the GP-SS framework to enable synthetic batch generation in a latent representation. The approach is evaluated using a simulated benchmark process for penicillin production. The results show improvements in the realism and normal operating coverage of generated trajectories, together with modest improvements in cross-variable correlation preservation, although these improvements do not consistently translate into improved monitoring performance. These findings show that latent-space GP-SS can make limited manufacturing history more informative for process characterization and monitoring, yet remains fundamentally constrained by the information contained in the original data.

Synthetic data generation in latent spaces for multivariate batch process monitoring in small-data scenarios

CAPORALE, CHRISTINA MARIE
2025/2026

Abstract

Batch process monitoring in pharmaceutical manufacturing relies on historical process data to effectively characterize multivariate process behavior and construct reliable monitoring models. However, in many industrial environments, particularly during new product development, process scale-up, and technology transfer, only a limited number of representative batches are available. Recent work has addressed this challenge using Bayesian Gaussian Process State-Space (GP-SS) models to generate synthetic batch trajectories from limited historical data. However, this strategy models each process variable independently and therefore does not explicitly account for the underlying cross-variable covariance structure, which may limit the representation of multivariate process behavior. This thesis investigates the integration of principal component analysis (PCA) into the GP-SS framework to enable synthetic batch generation in a latent representation. The approach is evaluated using a simulated benchmark process for penicillin production. The results show improvements in the realism and normal operating coverage of generated trajectories, together with modest improvements in cross-variable correlation preservation, although these improvements do not consistently translate into improved monitoring performance. These findings show that latent-space GP-SS can make limited manufacturing history more informative for process characterization and monitoring, yet remains fundamentally constrained by the information contained in the original data.
2025
Synthetic data generation in latent spaces for multivariate batch process monitoring in small-data scenarios
Machine learning
Synthetic data
Latent variables
Batch monitoring
GP-SS models
File in questo prodotto:
File Dimensione Formato  
Caporale_Christina_CorrectedTitle.pdf

accesso aperto

Dimensione 6.38 MB
Formato Adobe PDF
6.38 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/115825