A fault-tolerant model predictive control approach based on deep operator network

Abstract

This paper is concerned with the fault-tolerant control problem for unknown nonlinear systems in the model predictive control (MPC) framework. A modified deep operator network is designed to learn system dynamics from input-output data. However, the input data that contain fault information cannot be acquired directly in the presence of actuator faults. To overcome this difficulty, a mode simulation method is presented via adequate and uniform sampling of virtual fault information in a hyperspace. In this way, the system responses in different faulty modes are simulated to ensure excellent prediction accuracy of the modified network. Moreover, an improved fault estimation method is designed with historical input-output data of the modified network. Then, based on the fault estimates, the design problem of the fault-tolerant MPC controller is converted into a constrained optimization problem, which is further solved using an adaptive gradient descent method. Finally, two simulation experiments are provided to illustrate the validity of the proposed approach.