Neural networks approximation for sub-optimal gains of a nonlinear satellite attitude controller

This note presents a feasibility study of using a shallow neural network to compute attitude controller gains for satellite attitude maneuvers actuated by reaction wheels. The conventional solution proposed for the widely-used Lyapunov-based PD attitude controller gains ceases to be near time-optimal when the reaction wheels’ performance constraints are reached. Thus, for applications, whose success is sensitive to how fast the maneuvers are performed, it becomes advisable to use different gain values for different maneuver angles. We propose a method of calculating optimal and sub-optimal gains and train a neural network, which can be used on board of a nanospacecraft with limited computational resources to compute the controller gains for any given maneuver. Our numerical experiments corroborate that the gains learned by the neural network outperform those obtained from the conventional solution.