Fault Detection for Satellite Gyroscope Using LSTM Networks

Abstract

To handle the interference of attitude maneuver and measurement noise in gyroscope fault detection, a data-driven time series model based on long short-term memory (LSTM) with residual smoothing is proposed. First, a LSTM network is used to build a time series model, which achieves effective mining of attitude system data and tracking gyroscope output. And a sliding window mechanism is involved for better prediction. Then, the residuals between estimation data and real data are smoothed by exponentially weighted moving average (EWMA) to reduce the effect of measurement noise on fault detection. Finally, the fault is determined by comparing the smoothed residuals with the threshold. Simulation results show that the model proposed in this paper is effective in both fault scenarios of gyroscopes and has higher accuracy than traditional fault detection models such as BP and RBF neural networks.