Researches On Underwater Acoustic DOA Estimation And Matched Field Processing Algorithms Based On Sparse Bayesian Learning

Direction of arrival(DOA)estimation and matched field processing(MFP)of underwater acoustic signals are important research fields in array signal processing,and are of great significance for underwater acoustic active/passive detection,positioning and recognition.This dissertation mainly conducts researchs on underwater acoustic DOA estimation and MFP location based on sparse Bayesian learning(SBL),in order to improve the computational efficiency and the accuracy of DOA estimation in nonideal conditions for SBL methods,and improve the robustness to noise and environmental mismatch of the MFP algorithm with low computational workload.1)To enhance the computational efficiency with satisfied estimation accuracy for SBLbased DOA methods,an adaptive grid refinement sparse Bayesian learning(AGRSBL)DOA estimation algorithmis proposed.The guiding idea of this method is to gradually make the rough grid points refine around the DOA region adaptively during the iteration process.The Bessel K sparse prior distribution under the SBL framework is established,and a novel fixedpoint iteration rule of signal power is derived using the EM algorithm.Then,the AGR process maximizes the modified logarithm of joint probability density function to determine the area of adaptive grid refinement,and some conditions and criteria for grid refinement are combined to insert new grid points adaptively,so that the coarse grid points are gradually refined around the area of potential DOA.After the AGR process,a post-processing refinement DOA search is performed to further reduce the off-grid DOA error.In addition,the proposed method is extended to DOA estimation of broadband signals by combining multi-frequency array data.The proposed method improves the utilization efficiency of grid points and reduces the presence of pseudo-spectrum compared with traditional methods.The simulation results show that this method has higher computational efficiency and estimation accuracy than the classical off-grid SBL algorithms in the case of low SNR and limited snapshots.2)To enhance the DOA estimation performance of sparse Bayesian learning under low SNR and small number of snapshots,a sparse Bayesian learning DOA estimation algorithm based on the generalized double Pareto(GDP)prior distribution is proposed.The generalized double Pareto distribution has the characteristics of sharp peak at the origin and heavy-tailed distribution,and its sparse-constraint property is stronger than the prior distribution used in the traditional SBL algorithm,which has the advantage of improving the performance of sparse reconstruction estimation.However,the GDP distribution does not belong to the conjugate distribution of Gaussian distribution,so it cannot be directly used in SBL for complex signals.Therefore,we establish a Gaussian Scale Mixture(GSM)model of the prior distribution of GDP for the complex signal in SBL so that the prior distribution of GDP can obtain the solvable SBL process for the complex signal,and use the EM algorithm to derive the hyperparametric update rules in the Type II Bayesian estimation.In addition,in order to reduce the off-grid deviation,a post-processing high-precision DOA estimation process under the GDP distribution is also derived.A large number of simulations have proved the excellent performance of the proposed algorithm under low SNR,small number of snapshots,and small angle interval cases.3)To improve the robustness to noise and environmental parameter mismatch with a low computational workload for SBL-based MFP methods,a noise-free fast sparse Bayesian learning(NFFSBL)MFP algorithm is proposed to locate the underwater sound source.The algorithm first introduces noise precision parameters with Gamma prior into the signal prior.Then treats the noise precision as a redundant variable to integrate it so that the posterior probability distribution of the signal is a Student-t distribution with heavy-tailed characteristics,which improves the robustness of external disturbances of environmental parameter mismatch,and also eliminates the impact of inaccurate noise estimation.Then,the sequence procedure is adopted to maximize the logarithmic joint probability density function,which selects an atom to add,delete,and update according to the change of the objective function in each iteration.Thus the algorithm gradually selects the required atoms and estimates the corresponding signal power.The sequence procedure does not involve the inversion and multiplication of high-dimensional matrices,so its computational efficiency is very high.A large number of simulations show that the proposed NFFSBL algorithm has advantages of the robustness against noise and environmental parameter mismatch than classical MFP methods and is more efficient than the robust SBL-based MFP algorithm.4)The SWell Ex-96 experimental data and the underwater data collected in the Guishan Island sea of Zhuhai are used to test the proposed algorithms,the results compared with other algorithms further verify the effectiveness of the proposed algorithms.