Research On Millimeter Wave Array Direction Finding Based On Deep Learning

Direction of Arrival(DOA)estimation is a key problem in the field of array signal processing.Millimeter wave is widely adopted in the field of communication and radar due to their primary advantages such as wide bandwidth,narrow beams,and high spatial resolution.However,due to the high frequency and short wavelength of millimeter wave,traditional direction findings,such as interferometer direction finding,spatial spectrum estimation direction finding and amplitude comprised direction finding are difficult to apply to direction finding of wide space domain.In addition,in the actual application process,due to the influence of various channel amplitude and phase errors,antenna position errors,antenna pattern inconsistencies and other system errors,the direction finding accuracy of traditional direction finding methods will decrease.Focusing on the problem of high-precision millimeter wave direction finding in a wide space domain,this paper uses deep learning's powerful learning ability in nonlinear fitting problems to establish an end-to-end direction finding network based on deep learning.First,collect correlation matrices through simulation and prototype experiments,and transform the matrices to obtain the original sequence output of the network.The network part is constructed based on a fully connected network.A dense and compact network structure containing a large number of hidden neurons further enhances the representational power of the network.The network output passes through a specific non-negative activation function to obtain the final angle prediction value.During network training,the data ratio of the training and test sets is fixed at 3:1,and the test data is drawn at equal intervals,the Root Mean Square prop(RMSProp)optimization method is used as the optimizer to minimize the loss function,and the gradient backward propagation is used to optimize the entire network.First of all,a preliminary feasibility verification is carried out using simulation data,and the results show that the network can effectively fit the angle attributes in the correlation matrix.Then,the feasibility of the prototype experimental data is further verified,and explored the influence of different input correlation matrices on the fitting angle of the neural network and the influence of the power of the correlation matrix on the network fitting.Finally,through a comprehensive comparison between the more representative Multiple Signal Classification(MUSIC)algorithm in traditional methods and deep learning methods,in order to facilitate the explanation of the gap between the two methods,a more detailed comparison and analysis of the correlation matrix of a specific group is also carried out.The results show that,no matter under the condition of low signal-to-noise ratio or under different types of correlation matrix input,the methods based on deep learning have achieved a more stable and highprecision fitting of the angle of the correlation matrix.