Abstract

Essential ingredients for fault-tolerant control are the ability to represent system behaviour following the occurrence of a fault, and the ability to exploit this representation for deciding control actions. In this talk, we discuss a fault-tolerant control approach based on a model predictive control (MPC) framework with a probabilistic model, namely a Gaussian process (GP), with the aim of more flexible identification and effective control under fault conditions (possibly unexpected ones). This approach is characterised by a probabilistic representation of the system, the online learning to update the representation from input-output data, and the optimisation over this up-to-date model in a predictive controller for an enhanced control action.

Such a control framework is implemented and its fault-tolerant capability is investigated. Through computer simulations, the GP model in the MPC is demonstrated to be able to capture various fault conditions that change the system’s input-to-state behaviour and the system’s dynamic behaviour. Given the updated GP model, the MPC is also shown to be effective in the fault scenarios investigated. Additionally, control actions under cost functions with different probabilistic cost over states demonstrate the cautious control action from an expected-quadratic cost, as well as the effective exploratory actions from a risk-seeking cost. The effectiveness of the said framework is further evaluated on a high-fidelity flight control benchmark, the GARTEUR RECOVER benchmark. Both longitudinal control and full-state control demonstrate the controller’s capability to tolerate actuator faults like ‘stuck-in-position’ and ‘runaway’.