Abstract: A kind of satellite fault detection method was introduced to handle the problems of relying on rule database,insufficient multi parameter fusion and unbalanced distribution of data samples in satellite fault detection.The semi-supervised model was constructed based on time-series characteristics of satellite data and was proposed to achieve effective excavation of satellite data rules and data driven fault detection.Considering the temporal correlation between satellite data,this kind of fault detection method was proposed based on long short-term memory network.Also,a sliding window mechanism was involved for better predicting and detecting efficiency.Considering the correlation between multiple parameters as another dimension,temporal convolutional network(TCN) and auto-encoder network were used to excavate the correlation between historical data and different parameters at the same time.Experimental results show that the proposed model is superior to traditional fault detection models such as BP neural network in key indicators.

Key words: fault detection, temporal convolutional network, autoencoder, long short-term memory network, time-series modeling, semi-supervised learning