(375e) Cyber-Attack Detection and Resilient Control Using Physics-Informed Machine Learning

Digital transformation across industries has significantly expanded the potential of cyber-attacks, bringing new challenges for secure and stable operation of industrial manufacturing systems [1]. Traditional model-based detection methods are based on accurate first-principles models. However, due to the nonlinearity and interactions of process variables in real-world chemical processes, it is difficult to derive first-principles models for complex chemical processes. To that end, machine learning (ML)-based detection methods have garnered considerable interest in recent years for the identification of various types of cyber-attacks [2]. Data-driven detection strategies require a large volume of high-quality data with different types of cyber-attacks reported in incidents and under different operating conditions. However, it is generally challenging to obtain such a representative dataset for industrial chemical processes. To improve generalization performance and reduce the data requirement of data-driven detectors, physics-informed neural networks (PINNs), a synergistic approach that integrates data with physical knowledge, have attracted increasing attention in the machine learning community [3,4]. Despite the success of PINN in the process modeling field, at this stage, the incorporation of physics-informed machine learning techniques into cyber-attack detection has not been investigated. In addition to the detection of cyber-attacks, the design of resilient control schemes plays a crucial role in reliably operating chemical processes and mitigating the impact of cyber-attacks [5]. However, how to use a priori knowledge about cyber-attacks to efficiently construct resilient control systems remains an open question.

To mitigate the impact of sensor cyber-attacks in chemical processes, this work presents a framework that develops physics-informed machine learning (PIML)-based detectors and resilient controllers for improving closed-loop performance of nonlinear system under cyber-attacks. Specifically, a physics-informed recurrent neural network (PIRNN) -based detection system is first developed by incorporating the knowledge of cyber-attacks in terms of their attacking patterns into the loss function of PIRNN. Furthermore, a priori knowledge of measurement noise is embedded in the training of PIRNN to reduce false positives in the presence of sensor noise. The novel PIRNN detector not only improves the accuracy and reliability of detection but also relaxes the requirement of the amount of training data. Subsequently, we develop a knowledge-guided extended Kalman filter to provide an estimated state for post-attack resilient control. Finally, a nonlinear chemical process example under min-max and surge attacks is used to demonstrate the superiority of the proposed physics-informed detection and resilient control scheme against conventional, purely data-driven machine learning methods.

References:

1. Parker, S., Wu, Z., & Christofides, P. D. (2023). Cybersecurity in process control, operations, and supply chain. Computers & Chemical Engineering, 171, 108169.

2. Wu, Z., Albalawi, F., Zhang, J., Zhang, Z., Durand, H., & Christofides, P. D. (2018). Detecting and handling cyber-attacks in model predictive control of chemical processes. Mathematics, 6(10), 173.

3. Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3(6), 422-440.

4. Zheng, Y., Hu, C., Wang, X., & Wu, Z. (2023). Physics-informed recurrent neural network modeling for predictive control of nonlinear processes. Journal of Process Control, 128, 103005.

5. Wu, Z., Chen, S., Rincon, D., & Christofides, P. D. (2020). Post cyber-attack state reconstruction for nonlinear processes using machine learning. Chemical Engineering Research and Design, 159, 248-261.