Deep Learning Driven Non-Orthogonal Precoding for Millimeter Wave Communications

Hybrid beamforming (HB) based millimeter wave (mm-Wave) transmission is a promising technology for providing low latency and high data rate transmission. However, due to the high path loss, the performance of mmWave-based HB is significantly impacted by channel degradation, especially when the number of the available transmission path is less than the transmission streams. Driven by deep learning (DL) based beamforming, we propose a non-orthogonal precoding based HB scheme for the deterioration of channel transmission. Instead of directly applying the neural network as the beamforming function, we present the insight and theoretical analysis of the beamforming strategy in the DL based HB. By studying the beamforming behavior, we designed and formulated deep learning driven hybrid beamforming method. Compared with the existing DL-based hybrid beamforming, the neural network in the proposed method acts as a general neural codebook, which avoids network training even channel changed. In the overloaded channel condition, the proposed deep learning driven non-orthogonal precoding can solve the problem of channel degradation, which improves the robustness for hybrid beamforming. Simulation results show that the proposed method can achieve significant performance improvement over existing ones in overloaded transmission conditions. We also designed and implemented our scheme over the FPGA platform, which shows a much lower hardware cost than other existing schemes.