Research On Direction-of-Arrival Estimation With Coprime Array

Direction-of-arrival(DOA)estimation is an essential topic in array signal processing and has attracted tremendous interest in many fields,e.g.,wireless communication,radar,sonar and acoustics.Compared with the traditional uniform array which satisfies the spatial Nyquist sampling theorem,the recently proposed coprime array is preferred to provide an increased number of degrees of freedom(DOFs)and a larger array aperture without increasing the number of physical sensors.Based on the literature review of the latest research progress on DOA estimation with coprime array,we focus on the requirement of orientation accuracy,system robustness and real-time performance in practical applications,and propose highaccuracy model-driven and data-driven DOA estimation algorithms with coprime array.The main content of this thesis is summarized as follows:1.A robust DOA estimation algorithm based on variational inference with coprime array is proposed.Considering gain-phase errors and outliers,the augmented second-order virtual array signal derived from the coprime array is formulated as a mixture of perturbed data and outliers.By leveraging the sparsity of sources in the angular domain,the overcomplete secondorder steering matrix is formulated and the problem of DOA estimation with coprime array is formed as a problem of sparse signal representation.Then,the variational inference algorithm is designed based on the characteristic of sources and outliers.The proposed algorithm can realize accurate DOA estimation with increased number of DOFs even in the existence of gainphase errors and outliers.Besides,it does not require the priori knowledge of the number of sources and has good performance with small number of snapshots.Moreover,a prior-grid scheme with coprime array is proposed,where the prior grids for DOA estimation are derived based on the mixture of von Mises distribution rather than the uniform manner in the existing variational Bayesian inference framework.The proposed scheme can reduce the computational complexity since fewer grid points are utilized in each iteration.2.A gridless Bayesian inference algorithm for DOA estimation with coprime array is proposed.Considering the limitation of the basis mismatch on DOA estimation performance,we assume that the power of the sources follows Bernoulli-Laplace distribution,and regard the DOAs as the variable whose Probability Distribution Function(PDF)is represented by mixtures of von Mises PDFs,yielding the continuous parameterization of the direction information.The estimated number and power of sources are updated according to the designed greedy strategy.Besides,based on the prior information of DOAs,we obtain the initialization of DOAs which speeds up the convergence of the algorithm and improves the accuracy of estimation.Especially,the proposed algorithm can realize accurate DOA estimation with low Signal-to-Noise Ratio(SNR).3.A deep learning based DOA estimation method is proposed with coprime array.Considering DOA estimation performance of model-based methods will deteriorate drastically in the unfavorable environment,we extract the feature from real and imaginary parts of the nonredundant augmented second-order virtual signals,and design a Deep Neural Network(DNN)to learn the non-linear relationship between the DOAs and extracted features from data set at different SNRs.The method can realize accurate and high-resolution DOA estimation in the unfavorable environment characterized by low SNR,limited snapshots and array model mismatches.Besides,considering the high cost of acquiring plenty of samples and the difficulty in acquiring accurate labels in the continuous angular domain,we design a deep Bayesian Neural Network(BNN)based DOA estimation method,which can realize robust DOA estimation with small samples by introducing uncertainty.