Research On Array Signal Processing Method Based On Compressed Sensing

Array signal processing has a wide range of applications in the radar field.Array radar can realize target localization,target tracking and real-time imaging of static scenes under long-distance,all-day and all-weather conditions.It is the main equipment for battlefield information acquisition.Therefore,it is extremely important to propose novel array signal processing methods under complex conditions to improve target estimation accuracy and signal processing speed.Compressed sensing algorithms can utilize the sparsity of signals to achieve distortion-free reconstruction of signals in the case of downsampling.By utilizing the concept of sparsity flexibly,the theory of compressed sensing provides a new perspective for array signal processing.The compressed sensing algorithms can utilize the received signals with fewer or even single snapshots to realize the super-resolution estimation of DOA and the superresolution imaging of the targets.In addition,the compressed sensing algorithms can be used to solve the system error parameters through reasonable modeling.Based on the problems existing in traditional array signal processing methods,this paper conducts research on problems such as poor algorithm estimation accuracy,long computing time and model mismatch.Utilize algorithms such as tensor decomposition,Bayesian inference and atomic norm minimization to achieve high-performance estimation of array parameters.The main research content of this paper can be divided into the following four parts:Part Ⅰ: Traditional compressed sensing algorithms typically have the following issues: Gridbased algorithms leads to poor DOA estimation accuracy under high signal-to-noise ratio conditions;DOA estimation is often based on matrix operations without exploiting the highdimensional structure of the data.Multiple snapshot signals often result in high computational complexity.In addition,radar systems may experience element faulty,leading to data loss and degraded estimation performance.The first part studies the gridless estimation of DOD and DOA for bistatic MIMO radars.The complex-valued domain tensor model and the real-valued domain tensor model of the system under faulty conditions were established,which takes advantage of the high-dimensional structure of the signal.Moreover,it is proved through derivation that the model meets the conditions of PARAFAC decomposition.And proved that under the condition of the faulty array remains symmetrical,the tensor can be transformed into the real-valued domain through forward-backward averaging and unitary transformation,which further improves the algorithm speed.Atomic norm minimization enables data completion and gridless angle estimation,improving the accuracy of DOD and DOA estimation.In the paper,the estimation performance of different algorithms is statistically analyzed in terms of the RMSE between the estimation results and the setup parameters,which verifies the effectiveness of the proposed algorithms.In addition,the estimation time of different algorithms is statistically analyzed,which also verifies the conclusion of the paper.Part Ⅱ: In the first part of the study,it was found that the real-valued domain processing of tensor has an advantage over the complex-valued domain processing in terms of correlation target discrimination and computational speed.However,the tensor can realize the realvalued domain transformation only when the faulty array remains symmetric.To address this contradiction,asymmetric data loss is performed on the tensor to construct a symmetric tensor,which satisfies the conditions for tensor real-valued domain conversion.The algorithm proposed in this part has the advantages of real-valued domain signal processing,i.e.,strong discriminative ability of the relevant target and fast operation speed.In the paper,the RMSE between the estimation results and the setup parameters are counted to verify the performance of the proposed algorithm.The probability of successful detection of the correlation target is compared between the proposed algorithm and the traditional algorithm,which highlights the performance of the proposed algorithm.Part Ⅲ: Traditional model-driven algorithms based on either matrix or tensor modeling cannot avoid the large computational pressure.The trained network does not need to be iterated and optimized during the computation process,which has a large advantage in terms of computational speed.In this part,an novel unsupervised learning DOA estimation method is designed for sparse arrays.Deep learning algorithms need to utilize labeled datasets for supervised training of networks,which cannot get rid of the dependence on labeled datasets.And DOA datasets with discretized spatial spectra as labels usually lead to the convergence of DOA estimation accuracy under high signal-to-noise ratio conditions.Aiming at the above contradictions,this part proposes a DOA estimation algorithm based on Resnet and unsupervised learning,which realizes the fast gridless estimation of the DOA.Compared to the atomic norm minimization algorithm,the estimation speed of the proposed algorithm is improved by about 40 times.The RMSEs of different algorithms are counted through the Monte Carlo experiments to verify the performance of the proposed algorithm in terms of estimation accuracy.Part Ⅳ: The systems discussed above in this paper are all digital multi-channel array systems,which are are still overstretched in the face of the demand for miniaturization and lightweighting.In this part,a single-channel real-aperture radar 3D imaging system is designed based on the digital coding metasurfaces.Under the background of sparse scene imaging,a compressed sensing model is established based on the modulation characteristics of digital coding metasurfaces.And according to the characteristics of the coupling between metasurface elements,the model is improved.Therefore,three-dimensional imaging is converted into a compressed sensing problem.The scene vector and coupling parameters are solved by sparse Bayesian inference,which improves the quality of the three-dimensional imaging.The final simulation verifies the performance of the proposed method in 3D imaging.