Adapting to Nonlinear Transmitters With Hybrid Model Training for Neural Receivers

This paper proposes a novel hybrid model transfer learning approach designed for end-to-end OFDM neural receivers that effectively manage multiple channels and nonlinear transmitters. The hybrid model transfer learning method uses mixed Rayleigh channels and other obscured front-end models. This two-step process compensates for nonlinear front-end realizations and different channels, training a robust neural receiver. The neural receiver used is a deep complex convoluted network (DCCN), which replaces the conventional communication blocks with trainable layers that can correct the transmitter's nonlinear performance and other imperfections in the physical layer. This training approach improves the DCCN by 35% for bit error rate (BER), and training time can be reduced by 19% compared to other training approaches for the same tasks while adapting to different fading channels and being robust to noise in power amplifier models. Measurements on both a 28 GHz active phased array in package (AiP) and a GaN Hemt PA show that the trained DCCN can adapt to nonlinear behavior without sacrificing BER. This work demonstrates how training for multiple device operation states and channels helps develop a robust deep neural network capable of demodulating OFDM symbols subject to nonlinear distortions in multiple channel environments without retraining.