Research On Mm Wave Hybrid Beamforming Technology Based On Deep Learning

Millimeter waves combined with massive multiple-input multiple-output(MIMO)antennas use beamforming technology to focus signals in a specific direction,which can significantly improve the quality of communication.Full-digital beamforming can achieve optimal spectral efficiency,but requires an RF link per antenna,which is expensive and power-hungry.Millimeter-wave massive MIMO arrays generally use a hybrid beamforming structure including both digital and analogy components.On the one hand,the hardware cost can be reduced.On the other hand,by designing an efficient algorithm to jointly solve the digital and analog beamforming matrices,the spectral efficiency can be approached to an full-digital array.The traditional hybrid beamforming method based on optimization algorithm can achieve spectral efficiency close to full-digital beamforming through multiple iterations,but the computational complexity is significantly increased.The method based on deep learning,due to the offline training and online reasoning,can take into account high spectral efficiency and low computational complexity,which has attracted the attention of researchers.In the existing hybrid beamforming methods based on deep learning,the training data are all generated by traditional optimization algorithms,so the spectral efficiency cannot exceed traditional optimization algorithms.In addition,the robustness of deep learning methods under imperfect channel state information and how to further optimize neural network parameters need to be further studied.Therefore,the main work done in this thesis is as follows:1.An efficient millimeter-wave hybrid beamforming method based on deep residual network is studied.Firstly,a conditionally constrained millimeter-wave hybrid beamforming neural network structure is designed,and the predicted hybrid beamforming matrix is constrained according to the constant modulus constraint of the RF phase shifter and the transmit power constraint of the baseband digital beamformer.Then,the negative spectral efficiency is introduced as a part of the loss function to simultaneously realize the minimization of the distance loss between the hybrid beamforming matrix and the alldigital beamforming matrix and the maximization of the spectral efficiency.This method solves the problem of suboptimal spectral efficiency of existing deep learning methods through conditional constraints and joint loss function design.2.Aiming at the problem that the spectral efficiency of hybrid beamforming decreases significantly when the channel state information is imperfect,a millimeter-wave hybrid beamforming algorithm based on feature extraction network is designed.In this thesis,the main features of the imperfect channel matrix are extracted through a feature extraction network by exploiting the sparse property of the millimeter-wave channel matrix.Then,according to the extracted channel matrix characteristics,the follow-up millimeterwave hybrid beamforming is carried out.The experiment verifies that the spectral efficiency of hybrid beamforming remains robust under imperfect channel conditions by extracting the channel matrix.3.The computational complexity of beamforming matrix based on real neural network is relatively high,and the millimeter wave hybrid beamforming based on complex number neural network is studied.Different from the real neural network which processes the real and imaginary parts of complex numbers separately,a complex-to-complex neural network is designed in this thesis to directly obtain the complex parameters of the hybrid beamforming matrix.Complex neural network solves the problem of parameter redundancy in real neural network through parameter sharing.The simulation results confirm that the complex neural network can improve the spectral efficiency of mm Wave hybrid beamforming.More importantly,thanks to the characteristics of complex neural networks,complex neural networks have smaller parameters and lower computational complexity.