Attitude control of a nanosatellite system using reinforcement learning and neural networks

This paper describes a robust and efficient attitude determination and control system on-board a nanosatellite that makes use of the concepts of neural networks and reinforcement learning to develop an attitude control algorithm which can provide the required torque for stabilization of the satellite body along all three axes. The control system under consideration takes data from six sun sensors (one on each of the panels of the satellite body), a magnetometer and a gyroscope, placed inside the satellite, as input. It also requires the input from an on-board GPS module, which is run once per orbit (ideally) due to constraints of electric power in a nanosatellite system. The system consists of multiple stages, the first of which is running the orbit propagator. This will make use of the latest position and velocity vectors obtained from the GPS module for estimating the current position of the satellite, since the GPS module cannot be run on each iteration of the algorithm. The proposed system makes use of a neural network to perform the task of the orbit propagator, by forming a non-linear function for position estimation. Next, using the position vector of the satellite, the ideal orientation of the satellite is estimated in terms of the ideal magnetic field vector (using the IGRF model) and the ideal sun vector (in the orbit frame). These are then fed into the controller along with the measured magnetic field vector and the sun vector (in the body frame) to get the required torque. The controller mainly consists of two neural networks to give the torques which will help in the stabilization of the satellite. The neural networks in the controller are trained using reinforcement learning and temporal difference learning, using a modification of the actor - critic algorithm in reinforcement learning. The controller will be trained before the launch of the satellite using Software in the Loop (SIL) simulations of the desired orbit of the satellite to tune the parameters of the neural networks. Further, once the satellite is in orbit, the controller will be tuned after fixed intervals of time to adjust to any changes in the environment in the orbit.