Research On Heterogeneous Clutter Suppression For Airborne Radar

The echo of the airborne radar contains a large amount of ground clutter when performing air-detection and ground-detection tasks.In order to improve the moving target detection performance of airborne radar in the background of strong clutter,especially for the detection of low-velocity moving targets in the main lobe clutter region or weak targets in the side lobe clutter region,it is required to suppress the clutter to prevent the target from being annihilated.Space-time adaptive processing（STAP）combines two-dimensional sampling information from both space and time domain to adaptively form oblique notches at clutter locations to suppress clutter.To keep the average performance loss no more than 3 d B compared to the optimum filter,the statistical STAP method requires at least twice the system degrees of freedom of the independent and identically distributed（IID）training samples to estimate the clutter statistical properties.However,in practice,both the non-ideal radar configuration and heterogeneous clutter environment lead to the heterogeneity of clutter,which makes it difficult to meet the training sample requirement,and thus the clutter suppression performance of STAP method is degraded severely.In this dissertation,based on the clutter intrinsic sparsity and the sparse recovery theory,we research on the methods of heterogeneous clutter suppression with limited samples.The main research contents are as follows:Ⅰ.The sparse recovery-based STAP（SR-STAP）algorithms based on the approximate penalty term of l₀-norm are studied.The sparse recovery function of clutter with l₀-norm is NP-hard and cannot be solved.The most common solution is to replace l₀-norm with l₁-norm to describe the clutter sparsity.However,l₁-norm is not only related to the number of non-zero terms in the coefficient vector,but also related to the amplitude of the non-zero terms.When dealing with the problems of clutter suppression in the airborne radar,the main lobe clutter power is much stronger than the side lobe clutter.The SR-STAP algorithms with l₁-norm will focus on the power spectrum estimation of the main lobe clutter and ignore the power spectrum estimation of the side lobe clutter,resulting in the deviation between the estimated clutter covariance matrix and the ideal clutter covariance matrix.Thus,the performance of clutter suppression and moving target detection are degraded.To solve the above problems,two SR-STAP algorithms with l₀-norm approximation penalty terms are proposed in Chapter 2 of the dissertation:the SR-STAP algorithm with log summation penalty function and the SR-STAP algorithm with iterative re-weighting l₂-norm.With limited training samples,the clutter power spectrums estimated are close to the ideal clutter power spectrum in terms of both clutter power and clutter position,and the performance of the proposed algorithms are effective on heterogeneous clutter suppression.Ⅱ.An iterative reweighted algorithm based on the sparse Bayesian framework is studied.SR-STAP algorithms based onl_p-norm（0≤p≤2）requires the fine tuning of the regularization parameterλ.The regularization parameterλis an important parameter in the sparse recovery algorithms,which is used to controlling the tradeoff between the sparsity of the clutter signal and the data fitting error.Either too high or too low regularization parameter will degrade the performance of SR-STAP algorithms.In addition,most SR-STAP algorithms are hardly guaranteed to have global convergence property.To solve the above problems,Chapter 3 of the dissertation combines the Bayesian framework with iterative reweighted algorithm,and then an algorithm that does not need to set the regularization parameterλand converges globally is proposed:M-SBL-IRl_2,1 algorithm.The algorithm constructs the upper-bound auxiliary function of the marginal likelihood function with the conjugate function in the Bayesian framework,and then uses the majorization-minimization（MM）algorithm to iterative update the hyper-parameters in the sparse Bayesian learning algorithm.TheM-SBL-IRl_2,1 algorithm achieves the effective heterogeneous clutter suppression without the fine tuning of the regularization parameterλ,while has the property of fast global convergence.Ⅲ.A fast marginal likelihood maximization algorithm for the off-grid problem is studied.Most of SR-STAP algorithms do not consider the off-grid problem that may be faced when constructing the space-time dictionary matrix.When using the mismatched STAP dictionary for sparse recovery of clutter,the accuracy of coefficients is reduced due to the deviation between the clutter scattering points and the pre-defined grid points,and the clutter covariance matrix cannot be accurately estimated with the sparse representation and the mismatched dictionary matrix,which leads to the degradation of the clutter suppression and target detection performance.The degree of grid mismatch directly determines the performance of SR-STAP algorithms.The larger the degree of grid mismatch,the more the performance loss.In Chapter 4,the multiple fast marginal likelihood maximization（M-FMLM）algorithm:iteratively estimate the clutter support subspace in the Bayesian framework,and reconstruct clutter with the clutter support subspace.The M-FMLM algorithm does not need to consume much computational complexities and memory space,and it can not only deal with the heterogeneous clutter suppression in the absence of off-grid problem,but also in the presence of off-grid problem.When in the presence of off-grid problem,the performance loss caused by the off-grid problem can be mitigated by reducing the grid-interval.Ⅳ.A robust fast marginal likelihood maximization algorithm is investigated.Most of the SR-STAP algorithms require joint estimation of the sparse coefficient matrix A and the noise powerσ².However,these algorithms are not jointly convergent with the sparse coefficient matrix A and the noise powerσ².When the number of atoms of the dictionary matrix is far larger than the system degrees of freedom,the update formula for the noise power is not accurate enough.Inaccurateσ² will affect the estimation of the clutter subspace in the M-FMLM algorithm,resulting in performance loss of the M-FMLM algorithm.To solve the above problems,the multiple robust and fast marginal likelihood maximization（M-RFMLM）algorithm is proposed in Chapter 5.The Gamma distribution is introduced in the hierarchical Bayesian framework to integrate the noise powerσ² out.Taking the place of update the hyper-parameter and the noise powerσ²,we update the ratio between the hyper-parameter and the noise powerσ² in the M-RFMLM algorithm.Intuitively,the algorithm no longer estimates the clutter power and noise power corresponding to each grid points,but the clutter-to-noise ratio（CNR）corresponding to each grid points.The M-RFMLM algorithm is a complement to the M-FMLM algorithm,which inherits the advantages of the M-FMLM algorithm and is more robust than the M-FMLM algorithm.