Optimized actuator-fault-tolerant control using reinforcement learning for attitude dynamic system of quadrotor unmanned aerial vehicle.

This study proposes a fault-tolerant optimized backstepping (OB) control for the attitude system having actuator failure of quadrotor unmanned aerial vehicle (QUAV). The design principle of OB technology is to perform reinforcement learning (RL) in every backstepping subsystem, so that the entire backstepping control can attain optimization performance. However, owing to the complication and intricateness of OB control algorithm, it will make the QUAV attitude control tenderly happen system failures when executing the long-time running and high-intensity works. For devising the OB attitude control to have the fault-tolerate capability, the RL is integrated with an adaptive estimation of the proportion term of unknown efficiency factor and bias signal of fault actuator model. So as to achieve the integration smooth, the RL training laws of critic and actor networks are deduced in accordance with gradient descent method of a simple positive function that is equivalent to Hamilton-Jacobi-Bellman (HJB) equation, so that it can significantly reduce the complexity of RL algorithm. Finally, effectiveness of this attitude control scheme is demonstrated through Lyapunov stability proof and simulation example.