Abstract:

Hybrid precoding of millimeter wave Multi-Input Multi-Output (MIMO) is an important method to reduce hardware complexity and energy consumption.In order to reduce the complexity of optimization processing and improve spectral efficiency,a hybrid precoding fast optimization algorithm based on deep learning is proposed.The difference in signal-to-noise ratio between subchannels may lead to a poor bit error rate performance.The hybrid precoder is selected by geometric mean decomposition (GMD) of block diagonalization and training based on the deep neural network (DNN).The optimal selection of the precoder is regarded as the mapping relationship in the DNN to optimize the hybrid precoding process of the large-scale MIMO.The optimization problem of spectral efficiency is approximately reduced to the minimization of the Euclidean distance between all digital precoders and hybrid precoders,and the throughput is improved by using a limited number of RF links.Performance analysis and simulation results show that due to the improved gradient algorithm and single cycle iterative structure,the DNN based method can minimize the bit error rate (BER) of the millimeter wave MIMO and improve the spectral efficiency,while significantly reducing the required computational complexity.When the spectral efficiency is 50bps/Hz,the SNR can be saved by 3dB.If different schemes achieve the same bit error rate,the SNR can be saved by more than 5dB and have better robustness.

Key words: hybrid precoding, spectral efficiency, geometric mean decomposition, deep neural network, multi-input multi-output