In aerospace applications, Low Earth Orbit (LEO) satellites often lose some parts of raw data, which will disturb pattern recognition on satellite time series data and decline accuracy. A novel MR-GRU model is proposed, which can achieve a high accuracy on satellite time series data while some data are randomly missing. The MR-GRU model directly trains a recurrent neural network on the incomplete time series data, instead of the traditional way that tries to complement the missing data. The common Gated Recurrent Unit (GRU) model is improved to MR-GRU model. Two terms are expanded, i.e. masking term and attenuation term. The masking term is applied to the input at each time, and the attenuation term is applied to the input and output of each hidden unit. Consequently, the inherent time characteristics of time series data are ensured by the MR-GRU model, while the accuracy of pattern recognition is increased. According to the experimental results on satellite time series data, it is shown that the MR-GRU model is superior to the traditional models.