(278d) Quality Control of Penicillin Production Based on Multiple Data-Driven Models

Penicillin production plays an important role in pharmaceuticals and bio-chemicals industries for the production of low-volume but high-value-added secondary metabolites for antibiotics. In order to form the target product, it is common to grow the cells in a batch culture followed by a fed-batch operation to facilitate the synthesis of the antibiotic. The main advantage of batch operation in Penicillin production is consistently producing products of the desired quality.

Multiway analysis based inferential models are developed through batch-wise unfolding of process measurement trajectories in order to control the end-of-batch product quality [1]. The multiway analysis based inferential model, such as multiway partial least squares (MPLS), calls for the entire batch trajectory to predict the final quality in batch processes. Nevertheless, the process measurements beyond current sampling instance are not available, which are required for quality prediction. The unavailability of future trajectories is treated as a missing data problem in the framework of latent variable methods, which ignores the causal relationship between control action and process outputs [2]. Therefore, a data-driven dynamic model is recently developed to characterize the input-output dynamics for quality control in batch processes [3].

However, in the Penicillin process, the quality variables are also available intermittently, instead of only at the end, and existing approaches, including those that use a causal model do not utilize the improved information available through intermittent quality measurements. In addition, one global data-driven dynamic process model may not well characterize the essentially nonlinear dynamics in batch processes. Motivated by these considerations, this work addresses the problem of quality control for Penicillin production by developing a framework that utilizes the increased availability of quality measurements. To this end, the whole batch process is divided into several modeling phases based on the available intermediate and final product qualities. Then, each modeling phase is characterized by a localized multiple linear discrete time model, with a corresponding MPLS quality inferential model to predict the end-of-batch product quality. The developed multiple inferential and dynamic models are integrated with MPC framework so that causality and nonlinear relationships between the future inputs and outputs are accounted for in predicting the intermediate/final qualities and computing the control inputs.

The proposed multiple models based MPC results in significant improvement on final product quality in application to a highly nonlinear batch fermentation process, with comparison to trajectory tacking control and single model based MPC.

[1] Golshan, M., MacGregor, J. F., Bruwer, M. J., & Mhaskar, P. (2010). Latent Variable Model Predictive Control (LV-MPC) for trajectory tracking in batch processes. Journal of Process Control, 20(4), 538-550.

[2] Flores-Cerrillo, J., & MacGregor, J. F. (2004). Control of batch product quality by trajectory manipulation using latent variable models. Journal of Process Control, 14(5), 539-553.

[3] Aumi, S., Corbett, B., Clarke‐Pringle, T., & Mhaskar, P. (2013). Data‐driven model predictive quality control of batch processes. AIChE Journal, 59(8), 2852-2861.