Event-triggered Model Predictive Controller with Neural Network for Autonomous Rendezvous and Proximity on Elliptical Orbits

Autonomous rendezvous and proximity operation (ARPO) is the basis for various on-orbit services. However, aiming for high-value targets and space debris located on elliptical orbits, ARPO becomes challenging for conventional guidance and control methods, due to the complexity and strong nonlinearity of dynamics. This article proposes an encoder-decoder network with an interconnecting branch between layers to enhance the ability of the neural network to operate ARPO on elliptical orbits. The neural network is trained through behavioral cloning, with expert trajectories collected from the optimization results of model predictive control. The performance of the proposed network architecture is effectively improved compared with the existing network architectures while the computational overhead is greatly reduced compared with model predictive control. We further propose an event-triggered neural network controller. It uses the neural network to calculate control inputs under normal circumstances to save computing resources and switches to model predictive control to ensure safety and improve control accuracy when events are triggered. To prevent the control input provided by the neural network from exceeding predefined boundaries, constraint violation corrections are added to ensure the safety of the transfer process. Adaptive performance enhancement is implemented to optimize the steady-state relative distance. This mechanism adaptively determines the control thresholds for model predictive controllers and neural network controllers. The proposed approach selects the appropriate controller according to the designed event-triggered conditions, thereby achieving a balance between efficiency and accuracy. Simulation with environment perturbations and sensor measurement noise demonstrates the effectiveness and robustness of the proposed controller.