Abstract: For improving the diagnosis speed and accuracy of a large-scale and complex system like satellite or spaceship, a hierarchical diagnosis model of satellite is proposed. A self organizing feature mapping neural network is adopted for the upper network, which is responsible for the preliminary fault localization and identification for the whole satellite. Generalized regression neural network(GRNN) is adopted for the lower network, which is responsible for accurately determining the localization and causes of faults for each subsystem of the satellite. The principal component analysis(PCA) is introduced to reduce the dimension of the original state variables. So, the number of the upper neural network neurons is reduced. The method is successfully applied to the fault diagnosis for the subsystems of a satellite. The accurate diagnosis result is obtained with improved efficiency.