(344b) Machine Learning Methods in Process Control | AIChE

(344b) Machine Learning Methods in Process Control

Authors 

Zhang, Z. - Presenter, University of California, Los Angeles
Wu, Z., University of California Los Angeles
Rincon, D., University of California, Los Angeles
Christofides, P., University of California, Los Angeles
Process Systems Engineering (PSE) has been commissioned to merge mathematics and chemical engineering from about 50 years ago onwards [1]. Since the beginning of this effort, different mathematical approaches were used for tackling chemical engineering problems, and many of these approaches were materialized in real applications in the computational era. Artificial Intelligence (AI) is one particular approach that has been persistent along the decades with many ups and downs [2]. Among many problems that are found in AI during its initial phase are time-consuming and computationally expensive methodologies, and the impossibility of implementing neural networks with many layers [2]. Since the beginning of this century, many breakthroughs have been made in the area of machine learning and deep learning [3], which led to numerous important applications in human visual pattern recognition and natural language processing. Along with the rapid development of machine learning algorithms and computing resources/platforms, many free and open-source software libraries for machine learning applications have been created, which contribute to the broader use of machine learning techniques in classical engineering fields in addition to computer science and engineering.

In this work, feedforward neural networks (FNN) and recurrent neural networks (RNN) are first employed to model steady-state input-output nonlinear relationship and dynamical nonlinear systems, respectively. Moreover, advanced recurrent neural network architectures such as Long short-term memory (LSTM) [4] and Gated recurrent units (GRUs) that can overcome the issue of vanishing gradient are also applied in dynamical nonlinear system modeling. Additionally, since most of the systems in process control engineering are controlled in a dynamic and online manner, which requires the real-time calculation and implementation of control actions, parallel computing of ensemble regression models on a high-performance computing cluster is introduced to assign calculation tasks of multiple machine learning models to distributed processors to improve computational efficiency and prediction accuracy. To further improve the performance of machine learning models, online adaptation and training are employed using real-time data sets collected from multiple sensors to reduce modeling error and account for model uncertainties.

Due to the strong approximation capabilities and user-friendly, open-source machine learning libraries developed nowadays, machine learning models have great potential for process control and operations, for example, process fault detection, cybersecurity, control and real-time optimization. Specifically, we demonstrate that machine learning modeling method is able to approximate a broad class of nonlinear systems with a sufficiently small modeling error, and therefore, the closed-loop system under the controller that incorporates machine learning models is demonstrated to achieve desired stability and performance. The effectiveness of the proposed implementation of machine learning model in process control and the enhancement of computational efficiency via parallel computing are finally demonstrated through chemical process examples.

[1] Ramkrishna, D., and Neal R. A. Mathematics in chemical engineering: A 50 year introspection. AIChE Journal, 2004, 50: 7-23.

[2] Venkatasubramanian, V. The promise of artificial intelligence in chemical engineering: Is it here, finally?. AIChE Journal, 2019, 65: 466-478.

[3] Schmidhuber, J. Deep learning in neural networks: An overview. Neural networks, 2015, 61: 85-117.

[4] Wang, Y. "A new concept using LSTM Neural Networks for dynamic system identification." In Proceedings of American Control Conference (ACC), pp. 5324-5329. IEEE, 2017.