A Two-Stage Deep Learning Receiver for High Throughput Power Efficient LEO Satellite System With Varied Operation Status

Low Earth orbit satellites are expected to be one of the biggest suppliers of wireless communication within the coming years. For this to happen 5G and 6G networks, are crucial to be implemented in satellite communication. This comes with the problem of power-efficient transmissions. This paper exploits recent advances in complex-valued deep learning to cope with this challenge. The proposed approach is based on the autoencoder structure, where a legacy orthogonal frequency division modulation (OFDM) transmitter is used as an encoder and a deep complex convoluted network (DCCN) is used as decoder/receiver. Different from other state-of-the-art receiver architectures based on one-stage trained neural networks, our proposed DCCN adopts a two-stage training scheme, where the first stage is trained using AWGN channel and a fixed non-linear front end. The second stage uses a transfer learning to adapt to the flat fading channels and the front end model can be changed to compensate for different front ends, significantly reducing training time.This allows for power-efficient transmission at different operation statuses (e.g. radiated power levels and steering angles) without compromising the bit error rate in both average white Gaussian noise (AWGN) and flat fading channels. A K-band (28 GHz) active phased array in package (AiP) transmitting a 5G NR OFDM signal with a bandwidth of 100 MHz was used as the main front end test vehicle for validating the proposed DCCN. Satisfactory bit error rates were achieved while the AiP was driven into saturation with high power efficiency at different power levels and steering angles. This work demonstrates, for the first time, the promising capability of deep neural networks in processing varied operation staged non-linear OFDM waveforms in the form of an auto-decoder receiver.