Machine Learning Based Waveform Reconstruction Demodulation Method for Space Borne AIS Signal

Due to high speed of low earth orbit (LEO) satellites and wide footprint that satellite antennas may cover, space-borne AIS (automatic identification system) receivers encounter challenges, such as large Doppler offset and multiple signal collision, which lead to low success rates of signal demodulation. In this paper, we propose a Machine Learning based waveform reconstruction demodulation method with good performance. Firstly, this paper adopts a novel Least-Squares method which can reduce the estimation error of Doppler shift to less than 3Hz. Secondly, it introduces a Machine Learning based waveform classification & reconstruction method tailored for GMSK modulation characteristics, so that to recover the strongest signal from the received overlapped AIS signals. Compared with conventional AIS demodulation algorithms, the proposed method displays superior anti-noise and anti-interference performance. Simulation results show that when signal-to-interference (SIR) ratio is greater than 9dB, the success rate of decoding the strongest signal in an overlapped AIS signal can reach almost 100%. Even in a very poor environment where both SNR and SIR are 5dB respectively, the demodulation success rate of the proposed method can still reach around 46%, which is much better than any other demodulation methods we have observed so far.