Machine Learning in Process Control, Safety and Operations

Machine learning is creating new paradigms and opportunities in the design of advanced process control systems for chemical processes. Traditionally, model predictive control (MPC), a constrained optimization-based control problem formulation that is the gold standard employed in advanced control of chemical processes, is formulated with linear data-based empirical models and is used to compute control actions to maintain optimal process operation while accounting for process and control actuator constraints. However, chemical processes are inherently nonlinear and often require nonlinear models in order to be controlled efficiently. Nonlinear first-principles process modeling provides a direct way for accounting for nonlinear process behavior in the control system design but it may be cumbersome and difficult to implement in complex industrial processes which are not well-understood.  Machine learning tools like recurrent neural networks and ensemble learning provide an efficient way to build nonlinear dynamic models from data that can be used in the model predictive control system, thereby improving control system performance, process operational safety and process operation. In addition to revealing nonlinear dynamic process relationships from data, machine learning tools can address classification problems such that the ones arising in diagnosing process faults and cyber-attacks, thus providing a broad array of topics where machine learning can make an impact.

In this talk, we will primarily present our research work on the use of machine-learning tools in developing nonlinear model predictive control methods that ensure optimal control system performance and process operational safety, as well as establishing cybersecurity for nonlinear processes. Specifically, we will present: a) a machine-learning-based predictive control  framework that integrates recurrent neural networks within MPC and utilizes ensemble learning and parallel computing for enhanced prediction accuracy and computational efficiency, b) machine-learning-based MPC structures for nonlinear processes, which address simultaneously closed-loop stability and performance, and ensure process operational safety in the sense of guaranteed avoidance of unsafe operating conditions, and c) a two-tier detector-controller architecture that uses a machine learning-based classification detector to ensure process robustness with respect to a broad set of cyber-attacks. Throughout the talk, we will present applications of our methods to chemical processes of industrial interest to demonstrate their applicability and performance in meeting next-generation industrial goals related to improving process economics, safety and cyber-security. We will conclude the presentation by discussing the use of machine learning for data reduction and real-time operational decision making in deposition processes, additive manufacturing and an experimental electrochemical reactor.