A Machine Learning-Based Approach to Cybersecurity and Safety of Model Predictive Control Systems
Automated real-time operations of industrial process control systems depend heavily onaccurate information and reliable communication of decision variables, state variables, and measurement data. Augmentation in sensor information and network-based communication increases the complexity of the control problem in terms of modeling structure and accuracy, process safety, as well as vulnerability to cybersecurity threats. Production plants collect a large amount of operational and instrumentation data to be used for monitoring, control, and troubleshooting. The potential application of these data goes beyond preventative maintenance and fault detection, especially with increased digital connectivity and increased computing power. With the rise of big data, researchers are equipped to explore more robust systems that will improve production and computation efficiency, operational process safety as well as cybersecurity in many industrial applications. As various machine-learning algorithms have demonstrated success in a wide range of engineering applications, the development of rigorous and systematic integration of nonlinear process control and machine-learning methods has become the focus of this research.
Large-scale industrial processes face many control and operating challenges such as highdimensionality, information structure constraints, complex interacting process dynamics, and uncertainties in the system. To this end, the need of accounting for multivariable interactions and input/state constraints has motivated the development of model predictive control (MPC), and subsequently highlights the need for a process model that describes the process dynamics accurately. More specifically, distributed and decentralized control structures demonstrate better computational efficiency compared to the centralized control framework. On the other hand, traditional approaches to process safety through process design considerations and hazard analysis are independent from the design of control systems. MPC provides a framework to account for process operational safety constraints, and is able to provide simultaneous closed-loop stability and safety. Furthermore, cybersecurity has become another leading cause of process safety incidents, and is becoming increasingly important in chemical process industries as cyber-attacks have grown in sophistication and frequency. As intelligent cyber-attacks have access to control system information, researchers are motivated to develop cyber-attack detection and resilient operation strategies to address cyber-security issues beyond fault diagnosis.
This dissertation presents the use of machine learning techniques in MPC, and providesvarious methods of designing MPC systems for improved cyber-security and operational safety for nonlinear chemical processes. Integrated detect-control architectures for Lyapunov-based MPC, economic MPC, and distributed and decentralized MPC systems are presented to address several types of intelligent cyber-attacks with machine-learning-based detection algorithms and resilient control and mitigation strategies. Data-driven nonlinear dynamic models are developed for large-scale processes consisting of multiple subsystems, and are used as the predictive model in distributed and decentralized MPC systems. Distributed MPC systems designed with control Lyapunov-barrier functions (CLBF) to guarantee closed-loop stability and safety properties are presented, and machine-learning methods to characterize the barrier function used in a CLBF-MPC are developed with statistical stability and safety analyses. Nonlinear chemical process examples are numerically simulated to demonstrate the effectiveness and performance of the proposed control methods throughout the dissertation.