Online Learning-Based Predictive Control of Crystallization Processes under Batch-to-Batch Parametric Drift

Abstract

This work considers a seeded fesoterodine fumarate (FF) cooling crystallization and presents the methodology and implementation of a real-time machine learning modeling-based predictive controller to handle batch-to-batch (B2B) parametric drift. Specifically, an autoencoder recurrent neural network-based model predictive controller (AERNN-MPC) is developed to optimize product yield, crystal size, and energy consumption while accounting for the physical constraints on cooling jacket temperature. Deviations in the kinetic parameters are considered in the closed-loop simulations to account for the B2B parametric drift, and two error-triggered online update mechanisms are proposed to address issues pertaining to the availability of real-time crystal property measurements and are incorporated into the AERNN-MPC to improve the model prediction accuracy. Closed-loop simulation results demonstrate that the proposed AERNN-MPC with online update, irrespective of the accessibility to real-time crystal property data, achieves a desired closed-loop performance in terms of maximizing product yield and minimizing energy consumption.

Online machine learning modeling and predictive control of nonlinear systems with scheduled mode transitions

“…where the above inequality is derived from the fact that the loss function L is bounded by M and the definition of regret in Equation (9). Using Equation ( 13), we have…”

“…Equation ( 11) is derived to connect regret with generalization error using the definition of regret in Equation (9). Specifically, the generalization error can be bounded by regret (the first term), the loss suffered by the optimal model h ?…”

“…3 In addition to the considerations of computational efficiency, online learning of machine learning models has demonstrated its benefits in improving model prediction and closed-loop performance in many real-time control systems. [4][5][6][7][8][9] However, an ongoing challenge for the practical implementation of online learning models in real-world chemical processes is their generalization performance on unseen data, for which a fundamental understanding needs to be developed. Generalization error bound is commonly used in statistical machine learning to quantitatively characterize the generalization performance of machine learning models.…”

“…Online learning algorithms have been widely used to develop machine learning models for large‐scale problems with a tremendous amount of data, since the training process is more computationally efficient than batch algorithms 3 . In addition to the considerations of computational efficiency, online learning of machine learning models has demonstrated its benefits in improving model prediction and closed‐loop performance in many real‐time control systems 4–9 . However, an ongoing challenge for the practical implementation of online learning models in real‐world chemical processes is their generalization performance on unseen data, for which a fundamental understanding needs to be developed.…”

Comparing Approaches for Explaining DNN-Based Facial Expression Classifications

“…In the study where the base-model was developed, the original images were detected and aligned using RetinaFace [46]. In order to give the model the most similar images as it was trained on as possible, we use the same alignment method for the KDEF dataset, using the implementation in [47]. Note that this is a different alignment method than used for the geometric features-based DNN.…”