Abstract: In this paper, a novel representative feature auto-encoder (RFAE) model is proposed and applied to the unsupervised anomaly detection for the high-dimensional periodic time series telemetry data of a satellite power system. RFAE uses the improved stacked auto-encoder loss function and training algorithm, so that the model can learn the representative features of the same phase samples. Then, the samples are reconstructed according to the representative features and the reconstructed error is used to determine whether the samples are abnormal. In the experimental part, firstly, the synthetic data proves that the RFAE algorithm can effectively detect the anomalies of the high-dimensional periodic time series data. Then, the real telemetry data of a satellite power system from January to December 2014 is used to conduct experiment. The accuracy rate of the RFAE anomaly detection reaches 99%, and the detection effect is obviously better than those of other current anomaly detection algorithms.

Key words: Satellite power, Stacked auto encoders, Anomaly detection, Unsupervised learning