Enhancing Cyber-Security in Industrial Control: An Encrypted MPC Framework with Neural Network-Based Generalization

Cybersecurity presents a significant challenge in networked industrial control systems. To address this issue, we propose a fully homomorphic encryption scheme designed to secure control signals, thereby ensuring the network security of Model Predictive Control (MPC) systems. In response to the encryption requirements, the control output is approximated as a linear combination of basis functions which is realized by employing a neural network. This approach facilitates the computation of constrained approximate optimal solutions within the ciphertext space, therefore mitigating the computational overhead associated with online optimization in MPC and maintaining MPC's optimization performance. The basis functions are trained using data from multiple similar objectives, e.g. same type devices, or similar mechanisms, each under variant predictive control parameters. In practice, these basis functions can be reused, necessitating only the design of a set of weighting coefficients tailored to each device's performance requirements. The proposed method is demonstrated through its application to chemical process control. Simulation results validate the effectiveness and practicality of the approach.