Recurrent Neural Network Based Prediction to Enhance Satellite Telemetry Compression

Because of the gradually increasing number of remote measured low and/or high frequency sampled parameters in space applications, aerospace mission operators have to make hard choices on which parameters at which sampling rates should be downlinked. On-board aerospace applications are characterized by limited storage and communication budgets, while lossless data compression schemes should be sufficient enough to enhance transmission efficiency and hence the whole aerospace mission. In this paper, a proposed two-stage lossless compression method for telemetry data is presented. The proposed method consists of a decorrelation stage and an entropy coding one. The Long-Short-Term Memory (LSTM) Recurrent Neural Network (RNN) is implemented as a predictor in the decorrelation stage of the proposed method, and an illustrative method of applying LSTM network for telemetry data samples prediction is presented and figured out. In experiments, different entropy coders: Rice codes, arithmetic method and Huffman algorithm are separately implemented at the second stage. The proposed method is tested by different real telemetry data sets of FUNcube satellite in frames of data words of 8-,10-,16-bits widths. Experimental results show that the proposed method improved compression efficiency based on a single stage of entropy coder: Rice codes, arithmetic code, and Huffman algorithm by a ratio up to: 98%, 21%, and 1.6%, respectively.