(719c) Process and Controller Performance Monitoring Using Machine Learning Methods | AIChE

(719c) Process and Controller Performance Monitoring Using Machine Learning Methods

Authors 

Abdullah, F. - Presenter, University of California, Los Angeles
Wu, Z., University of California Los Angeles
Christofides, P., University of California, Los Angeles
Process monitoring systems have been playing an important role in maintaining safe, efficient and sustainable operation of industrial chemical processes. While traditional process monitoring methods rely heavily on human intervention and classical statistical techniques, recent works have demonstrated the possibility of using machine learning techniques for process monitoring and decision making to minimize human participation in routine tasks of the manufacturing process [1, 2, 3]. Compared to the statistical process monitoring approaches such as cumulative sum and exponential weighted moving average [4], which may not be effective for nonlinear processes due to the “curse of dimensionality”, machine learning, a method of data analysis that has been successfully applied to address classification and nonlinear regression problems in classical engineering fields, provides a powerful toolkit for process and controller performance monitoring when a large number of data samples covering all possible process outcomes are available and the process behavior is nonlinear. At this stage, how to incorporate machine learning methods in performance monitoring of advanced process control systems, e.g., model predictive controllers (MPC), for (large-scale) nonlinear systems has not been investigated.

In this work, we will focus on exploring the use of deep-learning neural networks (NN) and data dimensionality reduction techniques like nonlinear principal component analysis methods (e.g., [5]) to exploit high-dimensional data collected using in-situ and ex-situ sensing to study and model the relationship between the process control inputs/operational parameters and the process outputs. As a result, we will determine what operational parameters (or linear/nonlinear functions of the original operational parameters) included in the data sets have the strongest effect on the process outputs (and in turn on the product properties), and therefore decide what are the critical data sets that can be analyzed on the edge or to upload on the cloud for further analysis. We will then use the data sets at the cloud level collected from different processes and our machine learning tools to monitor entire plant behavior and plant-wide control performance to reduce energy consumption, reduce product waste and optimize product quality. Additionally, a deep feed-forward NN will be developed for MPC performance monitoring where we train the NN for many different controller performance scenarios/examples and then use it in real-time with process real-time measurements to monitor controller performance.

[1] Amini, M., & Chang, S. (2018). A review of machine learning approaches for high dimensional process monitoring. In Proceedings of the 2018 Industrial and Systems Engineering Research Conference, Orlando, FL.

[2] Qin, S. J. (2012). Survey on data-driven industrial process monitoring and diagnosis. Annual reviews in control, 36(2), 220-234.

[3] Aldrich, C., & Auret, L. (2013). Unsupervised process monitoring and fault diagnosis with machine learning methods. London: Springer.

[4] Ellis, M., & Christofides, P. D. (2014). Performance monitoring of economic model predictive control systems. Industrial & Engineering Chemistry Research, 53(40), 15406-15413.

[5] Dong, D., & McAvoy, T. J. (1996). Nonlinear principal component analysis—based on principal curves and neural networks. Computers & Chemical Engineering, 20(1), 65-78.