Study On Super-resolution DOA Estimation Methods Based On Deep Learning For VHF Radar

In the modern battlefield environment,with the development and use of stealth aircraft and anti-radiation missiles,the Very-high Frequency(VHF)radar has highlighted its unique advantages and functions.However,the VHF radar has a long wavelength,limited array aperture,wide beamwidth,and poor angular resolution.Therefore,for long-distance lowelevation targets,the direct signal and multipath signal scattered back by the target are received by the array from the main lobe of the beam,and it is difficult to distinguish the direct signal and multipath signal from the dimensions of the time domain,frequency domain,and spatial domain.Multipath signal distribution is related to terrain parameters,and has time-varying and irregular characteristics.These problems greatly reduce the detection performance of VHF radar under low elevation angle conditions.Therefore,how to improve the detection performance of VHF radar under low elevation angle conditions has always been a research hotspot and difficulty in the radar industry.This dissertation takes the low elevation angle measurement of the VHF radar as the research background,and on the basis of previous research work,studies the super-resolution direction of arrival(DOA)estimation methods based on deep learning for the VHF radar.The main research content is summarized as follows.1.The technology of altitude measurement of low elevation target is studied,and process and analyze the real data of the VHF radar in different environments.First,two multipath signal models are introduced,that is classic multipath signal model and perturbation multipath signal model.The classical multipath signal model models the direct signal and the multipath signal as a spatially adjacent coherent source within a beamwidth,but only considers a specular reflection signal.However,the multipath signal received by the actual array radar includes not only the specular reflection signal,but also the complex ground diffuse reflection signal.The perturbation multipath signal model accurately describes the effect of the complex multipath signal on each element of the array.The amplitude and phase interference effects of the array received signal are effectively modelled.Then,five super-resolution algorithms,suiting for VHF radar low-elevation altitude measurement,are introduced in detail.The advantages and disadvantages of the algorithms are compared.Finally,combining the real data of the VHF radar in two different environments,the amplitude and phase features of the array received data and the spatial spectrum features of multiple super-resolution algorithms under low elevation conditions are analyzed,and verify the performance of multiple super-resolution algorithms with multiple flight data.2.A novel low-elevation DOA estimation method based on spatial feature reversal is studied.Under low elevation conditions,physical models are usually difficult to accurately model multipath signals,so the unsupervised learning model is used to solve the problem of DOA estimation.Firstly,analyze the relationship between the spatial distribution features of the classic multipath signal model and the DOA.Then,the learning principle and function of the autoencoder are introduced in detail,and a parameter estimation method based on spatial feature reversal is proposed.The method uses an autoencoder to deeply mine the potential key spatial features of the data received by the array,and builds a mapping library between the potential features and the DOA.By projecting the mined potential features to the feature library,the DOA is estimated by reversal operation.Computer simulation experiments and real data verify that the proposed method has high estimation accuracy and good generalization performance under the background of coherent sources,which is suitable for estimation problems under low elevation conditions.3.A novel DOA estimation method for low-elevation targets based on supervised phase enhancement is studied.The existence of multipath signals will reduce the significance of the amplitude and phase features of the direct signal.The supervised learning model is combined with the DOA estimation problem to improve the significance of the phase features of the direct signal,and improve the estimation accuracy of the algorithm.Firstly,according to the array receiving data model,qualitatively analyze the relationship between the performance of the physical-driven super-resolution DOA estimation algorithms and the phase features,and combine the computer simulation experiment to quantitatively analyze the amplitude/phase error sensitivity of the super-resolution algorithm performance.The results fully demonstrate that the enhancing phase feature is the key to improve the performance of physical-driven algorithms.Then,combining the deep neural network and the deep convolutional neural network,a super-resolution DOA estimation algorithm with supervised phase enhancement is proposed,which uses the neural network to enhance the phase features of the direct signal,and reconstructs the data with the enhanced phase features and the original amplitude features.The DOA is estimated by classic super-resolution algorithms.Computer simulation experiments analyze the estimation performance and generalization of the proposed algorithm under the conditions of different signal-to-noise ratio parameters and angle separation parameters.Finally,part of the real data is used to verify the reliability of the proposed algorithm.4.A novel low-elevation DOA estimation method based on self-paced phase enhancement is proposed.In order to improve the generalization of the phase enhancement model,the self-paced learning algorithm is combined with the phase enhancement model.Firstly,the principle of self-paced learning algorithm is introduced in detail,and combine the self-paced learning(SPL)algorithm with the supervised phase enhancement model to improve the learning performance and generalization of the neural network by optimizing the neural network learning process.Then,combining the self-paced learning algorithm with three neural network learning models,a novel DOA estimation algorithm based on self-paced phase enhancement is proposed.Finally,the self-paced learning process is explained via real data.The measurement performance of the three self-step learning networks before and after the phase enhancement is compared with a large number of measurement results and statistical results,which verify that the neural network combined with the self-paced learning algorithm has a better fitting performance and generalization.5.A novel super-resolution DOA estimation method based on deep sparse angle separation learning is proposed.For the problem of DOA estimation of coherent sources within a beamwidth,the angle separation parameter is introduced to analyze the relationship between the angle separation features of space adjacent coherent sources and the DOAs of coherent signals.Then,the classification learning model is used to mine the angle separation from the real and imaginary features of the array received data or over-completed spatial spectrum features.A novel DOA estimation method combining the angle separation features for spatially adjacent coherent sources is proposed.Finally,a large number of computer simulation experiments verify that the estimation accuracy and generalization of the proposed method are better than the existing multiple physical-driven and data-driven algorithms under matched or mismatched parameter conditions.