APPLICATION OF RECURRENT NETWORKS FOR SPACECRAFT ATTITUDE CONTROL

Abstract

There is a need to perform a control allocation to map the torque commands from the body frame to the actuator frames in a spacecraft attitude control problem. When the number is more than three, this mapping, done by a pseudoinverse, is not unique and depends on the cost function used to perform the allocation. The infinity norm allocation is a specific control allocation method that emphasizes minimizing the maximum torque actualized by the spacecraft’s momentum exchange devices. It also allows access to the full torque envelope, giving approximately an extra 20 percent of torque capacity for the system to utilize over the traditional calculation method. The infinity norm allocation is not generally used as it requires an iterative optimization to be performed. An alternative is to use a recurrent neural network to solve the allocation problem. Implementation of a recurrent neural network can solve the pseudoinverse of the torque transformation matrix of a spacecraft’s momentum exchange devices. The recurrent neural network is shown to improve performance over the conventional allocation scheme for both reaction wheel and control moment gyro actuated spacecraft.