Deep-learning-based beamforming for rejecting interferences

Antenna arrays have been widely used for space and space-time processing to estimate the desired signals in the presence of narrowband and wideband interferences. Estimating the optimal weight vector is a challenging problem in space and space-time processing due to its high computational complexity. The problem escalates when the number of antennas and taps of filters for time-domain processing increases. In this study, a convolutional neural network is employed to estimate nearly-optimal weight vectors even when the number of available snapshots of the received signal is as low as 400 and a signal-to-noise ratio as low as -5 dB. Unlike the conventional processors, the authors proposed method requires no prior knowledge about the direction of arrival of the desired signal. In addition, due to its parallel architecture, the computational complexity of the neural processors is reasonable using a single graphical processing unit on a PC to run the algorithms. Through simulations of the uniform linear array in the presence of narrowband and wideband interferences, they demonstrate that the output signal-to-interference plus noise ratio (SINR) is very close (by <0.5 dB) to the max SINR obtained by the optimal weight vector of antenna arrays.