طراحي ‌كنترل‌كننده مرزي تطبيقي مبتني بر يادگيري تقويتي عميق براي ماهواره‌هاي انعطاف‌پذير

In this dissertation, the issue of simultaneous attitude an‎d vibration control of flexible satellites is addressed using a boundary control approach combined with deep reinforcement learning. The flexible satellite under consideration is equipped with honeycomb-structured solar panels. First, the governing dynamic equations are derived using Hamiltonʹs principle an‎d the calculus of variations. Subsequently, an innovative Lyapunov function is designed, an‎d the boundary controller is developed by directly applying the governing partial differential equations of the system. The asymptotic stability of the system is then proven using the extended LaSalle’s invariance principle. Following this, the controllerʹs performance is eva‎luated by discretizing the governing equations of the system using the assumed modes method an‎d performing numerical simulations. Simulation results in various scenarios, including active an‎d passive control, all indicate the achievement of simultaneous attitude an‎d vibration control for the flexible satellite. Furthermore, to overcome modeling limitations an‎d practical challenges such as external disturbances an‎d uncertainties, deep reinforcement learning is utilized. The automatic tuning of the boundary controller’s gains is carried out using an actor-critic neural network combined with temporal-difference learning techniques. The asymptotic stability of nonautonomous system with actuator saturation is then verified using the Barbalatʹs lemma. Numerical simulation results, including attitude angle tracking in the presence of disturbances an‎d uncertainties, demonstrate that the proposed method effectively enhances the robustness an‎d performance of the controller against disturbances an‎d modeling errors. Additionally, the training of the controller results in a reduction of settling time an‎d control effort. Finally, by designing an‎d constructing a fixture an‎d panel, the rigid satellite simulator at the university is upgraded to a flexible satellite simulator. The dynamic equations governing the simulator are first derived, an‎d then the implementation of the boundary controller in two different scenarios is carried out, one with fixed controller gains an‎d the other with an adaptive boundary controller based on deep reinforcement learning with online training. Lastly, the implementation of actuator failure is examined using the adaptive boundary control law. The results of the implementation are compared with numerical simulation results, confirming the accuracy of the extracted dynamic equations an‎d validating the theoretical aspects. Furthermore, obtaining accurate implementation results emphasizes the practical applicability of the innovations an‎d objectives of the dissertation. This research presents a novel approach for the attitude an‎d vibration control of flexible satellites by combining boundary control an‎d deep reinforcement learning, representing a significant step forward in enhancing the stability an‎d performance of space systems.

5