Research On Real-Time Array Beamforming Method Based On Deep Learning

Beamforming technology controls the amplitude and phase of antenna array elements to emit or receive electromagnetic waves in specific directions or areas.It has the charac-teristics of high directionality,high gain,low sidelobe,wide scanning,and is widely used in radar systems,wireless communications,medical imaging,radio astronomy,and other fields.Traditional beamforming algorithms based on optimization algorithms calculate the element excitation close to the ideal direction pattern by several iterations.These al-gorithms have the characteristics of flexibility and strong adaptability.However,many iteration operations result in extremely high computational complexity,especially in the case of large arrays.In recent years,deep learning has gradually been applied in the elec-tromagnetic field due to its superior offline learning ability and nonlinear fitting ability.Beamforming methods based on deep learning not only ensure the performance of ra-diation patterns but also effectively optimize computational complexity,with enormous application value and research value.1.A linear array beamforming based on radial basis function neural networks is proposed.The dataset is provided by the MVDR algorithm.Radial basis neural network is constructed to analyze the mapping relationship between array guiding vector and array element excitation.It predicts the array excitation of a 16-element array antenna.Finally,MATLAB is used to verify the effectiveness of radial basis function neural network to predict the array excitation in different application scenarios.2.A Linear array beamforming method based on particle swarm optimization convo-lutional neural network is proposed.Aiming at the problems of weak generalization ability and long response time of existing deep learning algorithms,a convolutional neural net-work is constructed to solve the linear array fast beamforming problem.The extractor is used to extract the main features of the array steering vector matrix,and then the regres-sor is used to predicet the element excitation based on the extracted features.Meanwhile,in the training process,particle swarm optimization algorithm is introduced to optimize the hyperparameters of convolutional neural network extractor and optimize the general-ization performance of convolutional neural network.Experimental results demonstrate that the extractor-regressor structure can effectively improve the accuracy of element ex-citation prediction compared to the regressor structure.Furthermore,the computational complexity of the CNN is much lower than that of traditional beamforming methods.3.A planar array beamforming method based on self-paced learning residual neural network is proposed.Aiming at the problem of extremely high computational complexity in traditional planar array beamforming and the difficulty of achieving high-performance planar array beamforming using existing deep learning-based beamforming algorithms.Residual neural network is constructed to solve the problem of planar array fast beam-forming and the degradation of deep neural network performance.In the process of back propagation updating parameters,a self-paced learning strategy is adopted to prioritize learning easy samples.During iteration,more complex samples are gradually learned to construct a more robust and superior residual neural network.Experimental simulation results demonstrate that the residual neural network can significantly optimize the com-putational complexity of the beamforming algorithm while ensuring the performance of the planar array radiation pattern.