The application of neural networks to spacecraft control

This thesis investigates how two neural network-based control techniques can be applied to a specific spacecraft control problem. The neural networks used are simple backpropagation networks, consisting of one or more tansigmoidal neurons (neurons with tanh transfer functions) in a hidden layer, and a linear neuron in the output layer. The neural network control techniques investigated here are Direct Model Inversion and Indirect Model Inversion. The spacecraft control problem is that of reducing the vibrations of a spacecraft payload. The source of the vibrations is a mass imbalance in one of the reaction wheels of the spacecraft. Four components are represented in the spacecraft model. These are rigid body inertia, solar array flexure, fuel slosh and payload vibration. A simple sinusoidal signal is used to model the disturbance torque produced by the reaction wheel mass imbalance. The complete model is broadly based on the Solar Heliospheric Observatory (SOHO) that is due for launch in 1995. Each of the neural network control techniques used is shown to be successful in reducing the effects of the disturbance torques on the spacecraft payload. However, in each case, a simple positional feedback gain term provides more effective and reliable control.