WS-CACNN algorithm for robust adaptive beamforming

Abstract

The increased signal bandwidth and steering vector mismatch error lead to degradation of beamforming performance. Besides, the focusing transformation and convex optimization affect the real-time performance. To address aforementioned problems, this paper presents a channel attention atrous convolutional neural network (CACNN)-based wideband signal (WS) beamforming method to accurately and quickly predict the weight vector, named WS-CACNN. Firstly, the proposed CACNN introduces atrous convolutions to decrease the quantity of network parameters and extract the spatial feature information of the sample covariance matrix, which are achieved without adding more parameters by enlarging the receptive field. Secondly, a channel attention mechanism is introduced to extract important amplitude and phase information, thereby improving the prediction accuracy of the beamforming weight vector. Then, using the robust adaptive wideband beamforming weight vector as the training label, maximizing output signal-to-noise ratio is able to be achieved through training the proposed CACNN network. Finally, the sample covariance matrix is input of a well-trained CACNN model which outputs the near-optimal weight vector. The simulation results indicate that the WS-CACNN algorithm not only outputs near-optimal wideband beamforming weight vector but also achieves excellent real-time performance.