Résumé: This paper presents a robust adaptive wideband beamforming approach based on a complex-valued deep convolutional neural network (CDCNN) for solving the problem of steering vector (SV) mismatches, named as CDCNN-CMR-SV algorithm. Firstly, via the complex convolution operation and complex batch normalization process, a complex convolution–normalization layer structure is constructed, which improves the feature extraction capability of complex-valued data and the speed of network training. On this basis, a CDCNN model is constructed to improve the ability to express the complex-valued domain broadband source model. Then, the interference plus noise covariance matrix is used as the input of the CDCNN, which is reconstructed by the focusing transformation method. The desired signal SV is used as the network label of CDCNN, which is corrected by solving the quadratic programming problem. Therefore, the mapping process is realized from neural network input data to label. Finally, a broadband beamforming weight vector is solved by the desired signal SV, which is predicted by the well-trained CDCNN. Simulation results show that the proposed algorithm has excellent beamforming performance, such as accurate beam pointing and strong interference suppression ability.

Langue: Anglais