Beamforming Algorithms Based On Deep Learning

As a key technology in array signal processing technology,adaptive beamforming technology is widely used.However,with the explosive growth of communication data in the 5G era,it is difficult for traditional adaptive beamforming algorithms to achieve a satisfactory speed.On the other hand,with the improvement of GPU computing power,newly proposed algorithms based on machine learning are developing rapidly in many research fields.In this thesis,beamforming algorithms based on deep learning are proposed.Through the establishment of a novel antenna array model based on the coordinate axis and the realization of the LMS algorithm,a data set for artificial neural network training is obtained,which contains 20000 pairs of target direction and the corresponding weight vector.On the basis of the training data,3 kind of variants are chosen for the network training,including the BP neural network based beamforming,DNN based neural beamforming and a novel design of CNN based beamforming.Simulation results show that the beamforming algorithms on the basis of BP neural network and CNN can shorten the calculation time to 1/10 to 1/5 times comparing with tradition LMS algorithm while ensuring extremely high accuracy,which means that beamforming based algorithms have great feasibility and superiority.