Wideband Adaptive Beamforming via Multiscale Channel Attention Convolutional Neural Network

Considering the problem that beam distortion caused by the coexistence of mainlobe and sidelobe interferences in the received data of sensor arrays, this article presents an effective wideband adaptive beamforming (WAB) method based on multiscale channel attention convolutional neural network (MACNN), named as WAB-MACNN algorithm. In the presented approach, a multiscale channel attention module is constructed to improve the prediction accuracy of beamforming weight vector. Specifically, via two branches with different scales, the attention weights of different feature channels can be better obtained to effectively strengthen significant features and weaken meaningless features for beamforming weight vector prediction. Then, with blocking matrix preprocessing (BMP) and interference-plus-noise covariance matrix (INCM) reconstruction, an efficient beamformer is used as the developed network training label to remove mainlobe interference and suppress sidelobe interferences. Finally, the well-trained model can rapidly and exactly output the predicted beamforming weight vector without complex matrix operations. Simulation results demonstrate that the presented algorithm can offer better beamforming performance with low time consumption under the coexistence of both mainlobe and sidelobe interferences.