TY - JOUR
T1 - A Novel Realistic Approach of Adaptive Beamforming based on Deep Neural Networks
AU - Mallioras, Ioannis
AU - Zaharis, Zaharias D.
AU - Lazaridis, Pavlos I.
AU - Pantelopoulos, Stelios
N1 - Publisher Copyright:
© 1963-2012 IEEE.
PY - 2022/10/1
Y1 - 2022/10/1
N2 - A new deep neural network (NN) approach applied to antenna array adaptive beamforming is presented in this article. A recurrent NN (RNN) based on the gated recurrent unit (GRU) architecture is used as a beamformer in order to produce proper complex weights for the feeding of the antenna array. The proposed RNN utilizes four hidden GRU layers and one extra layer for linear transformation. The produced weights are subsequently compared with respective weights derived by a null steering beamforming (NSB) technique in order to measure the accuracy of the RNN. The RNN training is performed by using a large dataset derived from an NSB technique applied to a realistic microstrip linear antenna array, in order to consider real-world effects, such as the nonisotropic radiation pattern of an array element and the mutual coupling between the array elements. The RNN performance is examined by using the root-mean-square error metric, whereas its beamforming performance is evaluated by estimating the mean value and the standard deviation of the divergences of the main lobe and nulls directions from their respective desired directions. A comparison between various NN structures and an overall study of the proposed RNN-based beamformer are also presented.
AB - A new deep neural network (NN) approach applied to antenna array adaptive beamforming is presented in this article. A recurrent NN (RNN) based on the gated recurrent unit (GRU) architecture is used as a beamformer in order to produce proper complex weights for the feeding of the antenna array. The proposed RNN utilizes four hidden GRU layers and one extra layer for linear transformation. The produced weights are subsequently compared with respective weights derived by a null steering beamforming (NSB) technique in order to measure the accuracy of the RNN. The RNN training is performed by using a large dataset derived from an NSB technique applied to a realistic microstrip linear antenna array, in order to consider real-world effects, such as the nonisotropic radiation pattern of an array element and the mutual coupling between the array elements. The RNN performance is examined by using the root-mean-square error metric, whereas its beamforming performance is evaluated by estimating the mean value and the standard deviation of the divergences of the main lobe and nulls directions from their respective desired directions. A comparison between various NN structures and an overall study of the proposed RNN-based beamformer are also presented.
KW - Adaptive beamforming (ABF)
KW - antenna beamforming
KW - deep learning
KW - neural networks (NNs)
KW - recurrent neural networks (RNNs)
KW - smart antennas
UR - http://www.scopus.com/inward/record.url?scp=85129623305&partnerID=8YFLogxK
U2 - 10.1109/TAP.2022.3168708
DO - 10.1109/TAP.2022.3168708
M3 - Article
AN - SCOPUS:85129623305
VL - 70
SP - 8833
EP - 8848
JO - IEEE Transactions on Antennas and Propagation
JF - IEEE Transactions on Antennas and Propagation
SN - 0018-926X
IS - 10
M1 - 9762632
ER -