Abstract

There are many schemes suitable for fault detection and isolation (FDI) such as observer-based methods, parity space and parameter estimation techniques. This thesis presents a neural network observer-based scheme for the actuator fault detection and isolation in the spacecraft attitude control. The features of neural network, such as its intrinsic nonlinearity property, its ability to learn, generalize and parallel processing make it suitable to model a non-linear dynamic system, such as the reaction wheel in our problem. We introduce three Elman recurrent networks and each of them is specific for modeling the dynamics of the wheel on each axis separately and independently. After each network has been trained approximately, it can give accurate estimation of the reaction torque generated by the wheel on each axis. Through some post-processing of the error signal between the actual and estimated output, we can get three residual curves for the FDI purpose. By comparing with a linear observer-based FDI scheme, the neural network observer-based scheme developed in this thesis does show advantages as demonstrated in the simulation results