Fault-tolerant Q-learning for discrete-time linear systems with actuator and sensor faults using input-output measured data

This study presents a novel output feedback Q-learning algorithm specifically designed for fault-tolerant control in real-time applications, circumventing the necessity for explicit system models or detailed actuator and sensor fault information. A significant benefit of this algorithm is its capability to simultaneously achieve optimality and stabilize systems with both actuator and sensor faults. Unlike traditional methods, it learns online using input-output data from the faulty system, bypassing the need for full-state measurements. We develop a unique expression of the Fault-Tolerant Q-function (FTQF) in the input-output format and derive a model-free optimal output feedback fault-tolerant control (FTC) policy. Furthermore, the algorithm's real-time implementation process is detailed, showing its adaptability in acquiring optimal output feedback FTC policies without prior knowledge of system dynamics or faults. The proposed method remains unaffected by excitation noise bias, even without a discount factor, and guarantees closed-loop stability and convergence to optimal solutions. Validation through numerical simulations on an F-16 autopilot aircraft underscores its effectiveness.