(491f) Multi-Time-Scale Model Predictive Quality Control Via Multi-Parametric Programming

Featuring highly interactive cyber-physical systems, the modern manufacturing industry poses grand challenges and opportunities for advanced control techniques to achieve automated prescriptive decision making. One of the central research areas is the end-point quality control in non-continuous batch processes widely applied for value-added pharmaceuticals, chemicals, etc. [1] Theoretical approaches have developed using mechanistic model for batch quality control with mathematical descriptions over the entire batch trajectory [2-3]. Data-driven multivariate statistics-based inferential quality control provides another promising strategy which can circumvent process-specific model development [4-6]. Despite these advances, key research challenges remain on: (i) how to achieve good product quality prediction during initial operation given small amount of data? (ii) how to optimize operation over multiple time scales, not only taking optimal action to now but also forecasting to future trajectory over substantially large time span? and (iii) how to ensure the online computation efficacy?

To address these challenges, in this work, we present a two-level model predictive control and optimization strategy to integrate short-term and long-term decision making for batch quality control. It features a short-term controller which responds by seconds or minutes using model predictive control for disturbance rejection and setpoint tracking and a long-term control-aware optimizer which responds by hours with large time steps to oversee and predict the entire batch operation using moving horizon strategy. The optimizer generates optimal input trajectory setpoints for the controller, readjusted by an offline surrogate model [7] to bridge the time scale gap. The controller, optimizer, and surrogate model are all solved via multi-parametric programming [8], with a distinguished feature to generate offline explicit control/optimization laws as an affine function of process variables prior to online real-time implementation, thus substantially reducing online computational load. The controller and optimizer are also built on a unified mechanistic process model to underpin the consistency for multi-scale decision making. To ensure the accurate monitoring of real-time batch operation, incremental measured data are utilized to iteratively train the physical model parameters to address modeling uncertainty and to enhance process-specific features (e.g., reaction kinetics, mixing efficiency). The extension of this batch quality control strategy for safety-critical process systems will also be discussed by incorporating dynamic risk considerations [9]. A T2 batch reactor case study [10] will be utilized to showcase the potential and efficacy of the proposed approach to systematically account for interactions and trade-offs of multiple decision layers toward improving process efficiency and safety.

References

[1] Yabuki, Y., Nagasawa, T., & MacGregor, J. F. (2002). Industrial experiences with product quality control in semi-batch processes. Computers & chemical engineering, 26(2), 205-212.

[2] Mesbah, A., Landlust, J., Huesman, A. E. M., Kramer, H. J. M., Jansens, P. J., & Van den Hof, P. M. J. (2010). A model-based control framework for industrial batch crystallization processes. Chemical Engineering Research and Design, 88(9), 1223-1233.

[3] Choi, S. W., Morris, J., & Lee, I. B. (2008). Dynamic model-based batch process monitoring. Chemical Engineering Science, 63(3), 622-636.

[4] Rendall, R., Chiang, L. H., & Reis, M. S. (2019). Data-driven methods for batch data analysis–A critical overview and mapping on the complexity scale. Computers & Chemical Engineering, 124, 1-13.

[5] Jiang, Q., Yan, X., Yi, H., & Gao, F. (2019). Data-driven batch-end quality modeling and monitoring based on optimized sparse partial least squares. IEEE Transactions on Industrial Electronics, 67(5), 4098-4107.

[6] Kay, S., Kay, H., Mowbray, M., Lane, A., Mendoza, C., Martin, P., & Zhang, D. (2022). Integrating Autoencoder and Heteroscedastic Noise Neural Networks for the Batch Process Soft-Sensor Design. Industrial & Engineering Chemistry Research, 61(36), 13559-13569.

[7] Burnak, B., Diangelakis, N. A., Katz, J., & Pistikopoulos, E. N. (2019). Integrated process design, scheduling, and control using multiparametric programming. Computers & Chemical Engineering, 125, 164-184.

[8] Pistikopoulos, E. N., Diangelakis, N. A., & Oberdieck, R. (2020). Multi-parametric optimization and control. John Wiley & Sons.

[9] Ali, M., Cai, X., Khan, F., Pistikopoulos, E. N., & Tian, Y. Dynamic Risk-based Process Design and Operational Optimization via Multi-Parametric Programming. Under Review.

[10] Chemical Safety Board, T2 laboratories inc. reactive chemical explosion. https://www.csb.gov/t2-laboratories-inc-reactive-chemical-explosion/