جستجوی مقالات مرتبط با کلیدواژه « safe reinforcement learning » در نشریات گروه « فنی و مهندسی »

In this article, a method for designing a fault-tolerant optimal attitude tracking control (FTOATC) for a quadrotor UAV subject to component and actuator faults is presented. The proposed fault-tolerant method is based on safe reinforcement learning (SRL) and is capable of ensuring input and state constraints without the need for prior knowledge of the quadrotor dynamics. To this end, the proposed optimal method is presented with a dual neural network (NN) structure consisting of identifier-critic neural networks. In the identifier NN update law, in addition to considering the variable forgetting factor dependent on measurement noise, the experience response method is used, which increases convergence speed and robustness to measurement noise and reduces estimation error. In this method, solving the constrained FTOATC problem is equivalent to solving an unconstrained optimal stabilization problem for an augmented system, where control input constraints and states are guaranteed by selecting suitable cost functions on the input signal and appropriate control barrier functions (CBF)on the states, respectively. Furthermore, fault detection is performed without the need for any model or filter bank, simply by comparing the residual value of the Hamilton-Jacobi-Bellman (HJB) equation with a predetermined threshold. The Uniformly Ultimately Boundedness (UUB) of identifier and critic NN weight errors and, as a result, the convergence of the control input to the neighborhood of the optimal solution are all proved by Lyapunov theory and the performance of the method is validated through simulation results.