Deep learning based fuzzy-MPC controller for satellite combined energy and attitude control system

• Published in: Advances in space research Vol. 74; no. 7; pp. 3234 - 3255
• Main Authors: Aslam, Sohaib, Chak, Yew-Chung, Jaffery, Mujtaba Hussain, Varatharajoo, Renuganth, Razoumny, Yury
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2024
• Subjects: Attitude Control; CEACS; Constraints; Deep Learning; FMPC Control; Neural Network; Takagi-Sugeno fuzzy Model; Neural Network; Constraints; FMPC Control; Attitude Control; Takagi-Sugeno fuzzy Model; Deep Learning; CEACS
• ISSN: 0273-1177
• DOI: 10.1016/j.asr.2024.07.034

•A linear model-based controller is designed for the CEACS attitude regulation.•Takagi-Sugeno fuzzy model with parallel distribution approach has been used.•The fuzzy-MPC controller achieves the desired CEACS attitude pointing accuracy.•A replica of the fuzzy MPC controller is designed by using deep layer neural network (D-FMPC).•The D-FMPC controller performs equally well as FMPC controller but significantly reduces computational burden. Combined Energy and Attitude Control System (CEACS) reduces the size and mass budgets of typical satellites and consequently, increases their payload capacity. CEACS uses flywheels for a dual purpose, i.e., as both energy storage and attitude control device. This maiden work attempts to introduce a novel Deep-Learning capability of the fuzzy-Model Predictive Control (FMPC) controller for CEACS. The design approach for the fuzzy-MPC controller uses the Takagi-Sugeno (T-S) fuzzy model of satellite attitudes and computes the control torque through a parallel distribution compensation (PDC) approach. However, the MPC controller offers a high computational burden, and it becomes a significant problem for smaller satellites having limited computational power. Therefore, in this research work, a novel Deep-Learning-based fuzzy-MPC controller (D-FMPC) is designed for the CEACS attitude regulation subject to higher initial angles, actuator constraints, parametric uncertainties, and external disturbance torques. Here, the deep-layer neural network is trained offline with the MPC controller data to replicate the FMPC controller, thus ensuring its controllability. Numerical results validate that the D-FMPC controller successfully mimics the FMPC controller and produces the desired pointing accuracy effectively with smooth transient response and without violating the attitude control actuator constraints. The results also validate that the D-FMPC controller offers significantly reduced computational burden than the FMPC controller. Therefore, the novel Deep-Learning solution provides a feasible platform for applying more complicated and sophisticated attitude control techniques for the CEACS attitude regulation in small satellites as an example.