سجاد رفيعي

Satellite missions, for a wide range of applications including communications, Earth observation and scientific research, face many challenges in the space environment, necessitating precise attitude maneuvering, control, and tracking for accomplishing their objectives. Notably, the attitude control system (ACS), equipped with sensors and actuators such as thrusters and reaction wheels, is more prone to faults compared to other subsystems. Consequently, the implementation of a fault-tolerant control system (FTCs) becomes paramount to ensure the reliability and safety of autonomous satellite operations. In the design of fault-tolerant control system, there are challenges such as requiring the system dynamic model and information about faults (location, magnitude, and type of fault). In this thesis, a fault-tolerant control approach based on reinforcement learning is designed for attitude control system of a rigid satellite that operates under challenging conditions with actuator faults and external disturbances. The thesis aims to to propose a model-free fault-tolerant controller using reinforcement learning, especially Q-learning algorithm. This fault-tolerant controller does not require the dynamic model of the satellite's attitude system or information about faults. It is worth noting that this controller operates independently of fault detection and identification, utilizing data extracted from the faulty attitude system trajectories. The stability of the proposed fault-tolerant controller is demonstrated using the Lyapunov technique, indicating its effectiveness in maintaining system stability and achieving desirable performance. Finally, the performance and efficiency of the proposed approach are eva‎luated through a comprehensive simulation.

يادگيري تقويتي , كنترل تحمل‌پذير عيب فعال , كنترل وضعيت , ماهواره صلب , عيب‌هاي عملگر , تشخيص و شناسايي عيب , يادگيري-كيو

Reinforcement learning , Active fault-tolerant control , Attitude control , Rigid satellite , Actuator faults , Fault detection and identification , Q-learning

sajad rafiee

hossein bolandi