Reinforcement-Learning-Based Fault-TolerantControl for a 2-DoF UAV System

摘要

This paper investigates fault-tolerant control for a 2-DoF Quanser Aero2 UAV system subject to propeller damage. An LQR controller with integral action (LQI) is first designed as the baseline con
troller under healthy conditions. Propeller faults are then introduced experimentally using single-cut, middle-cut, and symmetric-cut blade damage patterns. Hardware-in-the-loop results show that single-cut and middle-cut faults are largely tolerated by the LQI controller, whereas symmetric-cut faults cause significant thrust loss and actuator saturation. To improve performance under partial actuator degradation, a Soft Actor-Critic reinforcement-learning framework combined with an Extended Kalman Filter is proposed to estimate actuator effectiveness loss and compensate the control input. Experimental results show improved tracking performance and reduced overshoot for partially de graded symmetric-cut cases.