Beam Sweeping and Beam Selection Using Neural Network Based Approach For Millimeter Wave Channels

Next generation wireless communication uses millimeter wave (mmWave) band which offers an exceptional opening for increasing the data rate and throughput. Millimeter wave bands suffer from poor propagation losses. Because of these losses it is difficult to obtain channel state information (CSI) accurately. Accurate alignment of transmitter and receiver can be done by the ‘beam sweeping’ and then ‘beam selection’ procedure mentioned in the 5G NR standard. Because of dynamic nature of the wireless channel the transceiver needs to be continuously updated for these changes and hence there is a need for a network which learns itself from the history or training samples as input. Several machine learning based channel estimation schemes are proposed in the literature. The advantage of the accurate channel estimation comes at the cost of complex system structure and reduced energy efficiency. Different types of beam management methods like analog, digital and combination of it Hybrid methods are proposed in standards. To start the research work, authors have studied analog beamforming design with two neural network viz., generator and discriminator forming Generative Adversarial Network (GAN). The neural network is trained with the samples containing GPS location of the receiver and optimal beam pair index. The performance of the model is evaluated for top-K accuracy of selecting beam from K beam pairs, with highest Reference Signal Received Power (RSRP), spectral and energy efficiency plots. By simulating the environment in MATLAB authors found that proposed model achieves nearly 90% accuracy with the less than half beam pair index only compared to other exhaustive search algorithms.