Abstract: To solve the problem of actuator failure in satellite attitude control system, a fault diagnosis method based on recurrent neural network was proposed. The satellite attitude control system was modeled, fault analysis was carried out, continuous time fault data of star sensor and angular velocity gyro were collected. Six kinds of heterogeneous cyclic neural networks were designed to diagnose and classify the fault data, and the network effect was compared in terms of the network depth, feedback unit, activation function and training algorithm. The effect of deep loop neural network with GRU is better, the accuracy of fault diagnosis is 95.7%. The results show that, for time series satellite data, GRU and the recurrent neural network with a certain depth have better fault diagnosis effect.

Key words: satellite attitude, control system, recurrent neural network, fault diagnosis, GRU, deep learning