Off-lattice estimation model and fast algorithm for phased array spatial domain super-resolution

By combining off-grid estimation models and fast algorithms with structured grid localization and adaptive step size to accelerate gradient iteration, the contradiction between high estimation accuracy and low computational complexity in large-scale phased array radar is resolved, achieving high real-time performance and high applicability in two-dimensional angular super-resolution.

CN118980995APending Publication Date: 2024-11-19SHANGHAI SPACEFLIGHT ELECTRONICS & COMM EQUIP RES INST
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411040909.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing spatial super-resolution methods struggle to balance high estimation accuracy and low computational complexity in large-scale phased array radars, especially in two-dimensional estimation scenarios where computational complexity increases significantly, making it difficult to meet real-time requirements.

Method used

By employing an off-grid estimation model and a fast algorithm, the guide vector space is approximated through a compact linear representation. Combined with structured grid localization and adaptive step size to accelerate gradient iterative estimation, two-dimensional angular super-resolution of the target is achieved.

Benefits of technology

While ensuring estimation accuracy, it significantly reduces computational complexity, is applicable to arrays with arbitrary geometric arrangements, improves real-time performance and applicability, and is suitable for regular arrays and non-ideal antenna arrays.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118980995A_ABST
    Figure CN118980995A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of radar target direction finding and positioning, in particular to an off-lattice estimation model and a fast algorithm for phased array spatial domain super-resolution. According to the invention, the off-grid estimation model based on approximate linear representation and the adaptive step acceleration gradient iteration algorithm based on grid constraint are adopted; the method has the following advantages and positive effects that the problem that the estimation precision is highly coupled with the grid density in a conventional airspace super-resolution method is avoided through the off-grid estimation, the provided off-grid estimation algorithm based on iteration only relates to simple matrix multiplication and addition operation, the iteration speed is high, and the method is suitable for large-scale popularization and application. According to the method, the calculation cost is effectively saved while the estimation precision is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of radar target direction finding and positioning, and in particular to a grid-off estimation model and a fast algorithm for phased array spatial domain super-resolution. Background Art

[0002] With the rapid development of swarm intelligence technology and the application of drone swarms in actual combat, the target situation faced by air defense systems has shown significant characteristics of scale and density. The premise of countering dense swarm attacks is that radar can stably track and finely distinguish swarm targets to support accurate interception and intention recognition. However, the angular resolution of radar is limited by the array aperture. Under the constraints of limited hardware cost, platform loading space and weight, and high concealment requirements, the array size cannot be significantly increased. It is necessary to use airspace super-resolution algorithms to improve the resolution of swarm targets and the angle measurement accuracy.

[0003] Among the existing spatial super-resolution methods, grid-based algorithms represented by multiple signal classification (MUSIC), sparse modeling and sparse representation are the most widely used. The estimated values ​​of such conventional methods can only appear on the pre-defined angle grid. To improve the estimation accuracy, the grid must be fine enough and the computational complexity is significantly increased. However, in the two-dimensional estimation scenario, the grid size will grow in a quadratic manner, resulting in a significant increase in computational complexity. For large-scale phased arrays, its real-time performance is often difficult to meet application requirements. Existing solutions include three categories:

[0004] 1) Based on regular arrangements such as rectangular arrays (or first using virtual array transformation or manifold separation to map the original array into a regular virtual array), dimensionality reduction is achieved through two-dimensional angle decoupling and independent estimation. After dimensionality reduction, the grid can be further refined to improve estimation accuracy.

[0005] 2) Utilizing the regularity of array arrangement, low-complexity parameter estimation is achieved through analytical calculation methods such as root finding, which does not require grid division, the estimated value is continuous, and the accuracy and computational complexity are not mutually restricted;

[0006] 3) Based on Taylor expansion, the steering vector is approximated and an off-grid model for joint estimation of grid position and grid deviation is constructed. Its complexity has increased, but its accuracy is no longer limited by the degree of grid refinement. However, these methods are either only applicable to regular arrays or ideal arrays with known analytical expressions of steering vectors, or the estimation accuracy is still limited by transformation errors or approximation errors, and their practicality needs to be improved.

[0007] In summary, it is difficult for existing spatial super-resolution methods to achieve a satisfactory compromise between applicability, estimation performance, and real-time processing (especially for large-scale phased array two-dimensional spatial super-resolution problems), which has imposed many restrictions on their application. Summary of the invention

[0008] The purpose of the present invention is to solve the contradiction between high estimation accuracy and low computational complexity when large-scale phased array radar performs spatial super-resolution on dense targets. A grid-free estimation model and a fast algorithm for phased array spatial super-resolution are provided. The method is oriented to beam domain dimension reduction data, approximates the guidance vector space based on a compact linear representation, and adopts structured grid positioning and adaptive step-size accelerated gradient iterative estimation to achieve two-dimensional angle super-resolution and continuous and accurate estimation of the target. The model is suitable for arrays of arbitrary geometric arrangement (including sparse phased arrays with random elements) or actual antenna arrays with non-idealities such as array errors; the structured grid positioning step supports MUSIC type algorithms or block sparse algorithms, and the iterative precision estimation step is established based on a non-parametric model, so the proposed method is still effective under the conditions of highly correlated sources and a small / single number of snapshots. Compared with conventional grid algorithms, the proposed method can obtain excellent estimation performance without fine grids, and its real-time performance is greatly improved.

[0009] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:

[0010] A grid-out estimation model and fast algorithm for phased array spatial super-resolution, the specific steps of which are as follows:

[0011] S1: Perform beam domain transformation and pre-whitening processing on the multi-channel data acquired by the digital phased array radar according to the angular region of interest of the current wave position;

[0012] S2: discretize the two-dimensional directional sine space according to the required angular resolution, and construct a basis matrix for off-grid estimation based on an approximate linear representation model;

[0013] S3: A grid estimation model is established based on the basis matrix, and a structured MUSIC algorithm or a block sparse algorithm is used to locate the directional sinusoidal grid of each signal source, and the steering vector representation coefficient is estimated at the same time;

[0014] S4: Calculate the precise steering vector estimate of each signal source using the representation coefficient, and construct a local optimization problem for the two-dimensional sine of each signal source;

[0015] S5: Adopt the adaptive step-size accelerated gradient iteration algorithm based on grid constraints to solve the two-dimensional directional sine estimation value of each signal source respectively. After the iteration is terminated, the arc sine is solved to further obtain the accurate two-dimensional angle estimation value of each target.

[0016] Furthermore, step S1 includes the following contents:

[0017] The M×N dimensional multi-channel sampling raw data X obtained by the digital phased array radar is multiplied by the M′×M beam domain transformation matrix T on the left.H We get X′=T H X, where M is the number of array elements and N is the number of sampling snapshots; T = [a(α1, β1)a(α2, β2)…a(α M′ ,β M′ )], M′ is the number of beams (its value should generally be greater than the number of targets but much smaller than the number of array elements), (α m′ ,β m′ ) is a two-dimensional spatial angle that can be flexibly selected within the spatial domain of interest at the current wave position; a(α m′ ,β m′ ) is the array steering vector corresponding to the m′th (m′=1,...,M′) beam pointing angle, and its expression is a(α,β)=exp(j·2π(l x sinα+l y sinβ) / λ wl ), where the two-dimensional spatial angle (α, β) is defined as the angle between two orthogonal normal planes to the plane where the array is located, and its value range is [-π / 2,π / 2]; l x =[x1…x M ] T and l y =[y1…y M ] T is the array coordinate vector, (x m ,y m ) is the two-dimensional coordinate of the mth (m=1,...,M) array element, λ wl In fact, the process of beam domain transformation not only realizes spatial filtering in the direction of interest, but also completes data dimension reduction, which is more conducive to the real-time processing of large-scale phased arrays.

[0018] Furthermore, since the beam domain transformation process will cause the noise component in the data to no longer be Gaussian white noise, it is necessary to pre-whiten X′, that is, multiply the matrix (T H T) -1 :X″=(T H T) -1 X′.

[0019] Further, step S2 includes the following contents:

[0020] Define the two-dimensional directional sine f = sinα, g = sinβ, and divide J two-dimensional grids in the directional sine space of interest, with the minimum interval Δf and Δg determined according to the required angular resolution. Then for any grid point (f j ,g j ), its neighborhood The steering vector corresponding to any two-dimensional sine in the matrix can be linearly represented by the eigenvector of the matrix in the following formula:

[0021]

[0022] Among them, b(f,g)=a(α,β)=exp(j·2π(l x f+l y g) / λ wl ); It is related to the grid spacing Δf and Δg. b α (f) = exp(j·2π·l x f / λ wl ), b β (g) = exp(j·2π·l y g / λ wl ), represents the Hadamard product of matrices. Specifically, C α and C β The expressions of the p-th row and q-th column (p, q = 1, ..., M) elements of α,pq =2Δfsinc(2ΔfΔx pq / λ wl ) and C β,pq =2Δgsinc(2ΔgΔy pq / λ wl ), and then the p-th row and q-th column element of matrix C is:

[0023] C pq =4Δf·Δgsinc(2ΔfΔx pq / λ wl )·sinc(2ΔgΔy pq / λ wl )

[0024] Where Δx pq =x p -x q , Δy pq =y p -y q , sinc(x)=sin(πx) / (πx).

[0025] This results in the standard orthogonal basis matrix for off-grid estimation:

[0026]

[0027] in, v m Arrange the mth descending eigenvalue λ of the matrix C m The corresponding eigenvector of; the number of columns K of the matrix is ​​determined by the following formula:

[0028]

[0029] Among them, ε is a sufficiently small value to ensure the completeness of the basis matrix, thereby reducing the approximation error of the linear representation. Generally, ε is taken as 10 -3 ~10 -5 . Obviously, when K = M, is a set of complete orthogonal bases for the steering vector b(f, g) corresponding to any two-dimensional azimuth sine within the neighborhood , that is, there exists a set of representation coefficient vectors r m (f, g) (m = 1,..., M) such that When K < M, then is an approximately complete orthogonal basis for the steering vector b(f, g) within the neighborhood , that is Among them,

[0030] In order to match the data preprocessing process in step 1, the basis matrix is further transformed as follows:

[0031]

[0032] In addition, for a real array with unknown non-idealities (element position error, channel amplitude-phase error, mutual coupling error, etc.), the integral form of H(f j , g j ) in the above model can be replaced by its discrete version: (where is a set of real array steering vectors actually measured within the neighborhood ) to ensure the effectiveness of the model.

[0033] Furthermore, step 3 includes the following content:

[0034] Based on the basis matrix obtained in step 2, the data X″ preprocessed in step 1 can be modeled as:

[0035]

[0036] Among them, j d is the grid point index closest to the true two-dimensional direction sine value of the d-th target (d = 1,..., D, D is the total number of targets) ; is the signal component in this model, where the representation coefficient of the echo steering vector of the d-th target at the j d -th grid point is introduced N″ is the noise component in the data.

[0037] According to the above model, by determining the grid index j dThe rough azimuth cosine estimate of the target can be obtained by estimating the coefficient vector (hereinafter referred to as r d ) can further obtain the precise guidance vector of the target, which can then be used to estimate the two-dimensional direction cosines; however, the array manifold part in the proposed model changes from the single-column guidance vector of the traditional model to a multi-column basis matrix It shows obvious structural characteristics, and the estimation algorithm needs to be adaptively adjusted. To this end, the present invention proposes two grid positioning and representation coefficient vector estimation methods suitable for different scenarios:

[0038] 1) When the echoes are weakly correlated and the number of snapshots is sufficient, a modified MUSIC algorithm is used, and its two-dimensional spatial spectrum is given by P(f j ,g j )=1 / λ min (G(f j ,g j )) is given, where λ min (·) represents the operation of taking the minimum eigenvalue of a matrix; For the M′×(M′-D)-dimensional noise subspace estimation, the covariance matrix X″X″ of the sampled data can be obtained by H / N performs eigendecomposition and extracts the eigenvectors corresponding to the smallest M′-D eigenvalues. By searching for Ρ(f j ,g j ) spectrum peak, we can find the index j of the target direction sine in the grid d ; r d The normalized estimate of By The eigenvector corresponding to the minimum eigenvalue of is given. To further save computational cost, the conventional MUSIC algorithm can also be used to determine j d , and then use the above method to further estimate its corresponding

[0039] 2) In the case of strong echo correlation or coherence, few snapshots or a single snapshot, a block sparse method is used, and its structured redundant dictionary is composed of [E1″…E′ J ′] is given, and the corresponding signal matrix is a block sparse matrix with D non-zero blocks. Therefore, S can be estimated by block sparse recovery algorithms such as combined norm minimization. j ″(The estimated value is recorded as ), and determine the index j according to the position of the non-zero block d ; Non-zero block That is, for r d ·s d The estimated value (sd is a 1×N-dimensional signal vector), which is equivalent to the coefficient vector r d is a set of linear combinations of .

[0040] Furthermore, step 4 includes the following:

[0041] Based on the grid index j obtained in step 3 d and coefficient vector estimates, and further calculate the steering vector estimates (for the MUSIC algorithm described in step 3) or the estimated cluster after linear scaling (Block sparse class method described in step 3).

[0042] The following optimization problem is further constructed to obtain the optimal estimate of the sine of the d-th target two-dimensional direction:

[0043]

[0044] in,

[0045] Furthermore, step 5 includes the following:

[0046] Given that the objective function h(f,g) of the optimization problem in step 4 is in the neighborhood For concave functions that are Lipschitz continuous, the optimal estimate can be obtained by accelerated gradient iteration as follows:

[0047]

[0048] Wherein, the superscript i indicates the i-th iteration; is the estimated value of the d-th target in the i-th iteration, and is a pair of intermediate quantities; η i Determines the proportion of momentum component in the iteration process; μ f and μ g is the step size, which mainly affects the convergence speed; the gradient of h(f,g) with respect to f and about g They are given by the following formulas:

[0049]

[0050] Wherein, Im{·} represents the imaginary part of the complex number. In order to speed up the iteration speed, the present invention further proposes an adaptive step size adjustment mechanism based on grid constraints, namely:

[0051]

[0052] Among them, 0<ρ<1, ρ is generally 1 / 3~1 / 2. gThe update rules of are similar and will not be described here.

[0053] The above iteration is initialized by To achieve this, the iteration is terminated when the normalized gradient is sufficiently small, that is:

[0054]

[0055] Among them, max(·) represents the maximum operation, and γ is usually 10 -3 After the iteration is terminated, the two-dimensional angle deviation estimate of the d-th target is given by Given.

[0056] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0057] The proposed method avoids the problem of high coupling between estimation accuracy and grid density in conventional spatial super-resolution methods through off-grid estimation. The proposed off-grid estimation algorithm based on iteration only involves simple matrix multiplication and addition operations and has a fast iteration speed, so that the method of the present invention effectively saves computing costs while ensuring estimation accuracy.

[0058] The grid positioning and coefficient vector estimation steps of the proposed method provide effective algorithms for weakly correlated signals, large snapshot conditions and strongly correlated signals, small snapshot conditions, respectively, so that the method of the present invention can ensure high reliability and practicality in different application scenarios.

[0059] The two-dimensional angle off-grid estimation model constructed by the proposed method can effectively eliminate the influence of approximation error on estimation accuracy by reasonably selecting parameters in the process of basis matrix construction, and its scope of application is extended from the ideal array limited by the traditional Taylor expansion model to the real array with unknown non-idealities.

[0060] As a fast implementation scheme for large-scale phased array spatial super-resolution, the proposed method has no special requirements on the geometric configuration of the antenna array and is widely applicable to various types of regularly arranged phased arrays or sparse phased arrays with randomly drawn elements. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0062] Figure 1 The present invention is a flowchart of the method.

[0063] Figure 2This is the iterative convergence process of the estimated value of the method proposed in the present invention.

[0064] Figure 3 This paper analyzes the operation speed and estimation accuracy of the method proposed in this invention under different grid densities.

[0065] Figure 4 Comparison of the basis matrix dimensions of the model proposed in this invention under different parameter values.

[0066] Figure 5 The estimation accuracy of the proposed method is compared with that of the traditional method under different signal-to-noise ratios.

[0067] Figure 6 The time cost comparison between the method proposed in the present invention and the traditional method under different signal-to-noise ratios is shown in FIG. DETAILED DESCRIPTION

[0068] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] refer to Figure 1 The present embodiment discloses a grid estimation model and a fast algorithm for phased array spatial domain super-resolution, including the following steps:

[0070] Step 1: According to the angular region of interest of the current wave position, the multi-channel data acquired by the digital phased array radar is transformed in the beam domain and pre-whitened;

[0071] Step 2: discretize the two-dimensional directional sine space according to the required angular resolution, and construct a basis matrix for off-grid estimation based on an approximate linear representation model;

[0072] Step 3: Establish a grid estimation model based on the basis matrix, and use a structured MUSIC algorithm or a block sparse algorithm to locate the directional sinusoidal grid of each signal source, and estimate its steering vector representation coefficient;

[0073] Step 4: Calculate the precise steering vector estimate of each signal source using the representation coefficient, and construct a local optimization problem for the two-dimensional sine of each signal source;

[0074] Step 5: Adopt the adaptive step-size accelerated gradient iteration algorithm based on grid constraints to solve the two-dimensional directional sine estimation value of each signal source respectively. After the iteration is terminated, the inverse sine is solved to further obtain the accurate two-dimensional angle estimation value of each target.

[0075] The above steps are described in detail as follows:

[0076] Furthermore, step 1 includes the following:

[0077] The M×N dimensional multi-channel sampling raw data X obtained by the digital phased array radar is multiplied by the M′×M beam domain transformation matrix T on the left. H We get X′=T H X, where M is the number of array elements and N is the number of sampling snapshots; T = [a(α1, β1)a(α2, β2)…a(α M′ ,β M′ )], M′ is the number of beams (its value should generally be greater than the number of targets but much smaller than the number of array elements), (α m′ ,β m′ ) is a two-dimensional spatial angle that can be flexibly selected within the spatial domain of interest at the current wave position; a(α m′ ,β m′ ) is the array steering vector corresponding to the m′th (m′=1,...,M′) beam pointing angle, and its expression is a(α,β)=exp(j·2π(l x sinα+l y sinβ) / λ wl ), where the two-dimensional spatial angle (α, β) is defined as the angle between two orthogonal normal planes to the plane where the array is located, and its value range is [-π / 2,π / 2]; l x =[x1…x M ] T and l y =[y1…y M ] T is the array coordinate vector, (x m ,y m ) is the two-dimensional coordinate of the mth (m=1,...,M) array element, λ wl is the working wavelength. Furthermore, since the beam domain transformation process will cause the noise component in the data to no longer be Gaussian white noise, X′ needs to be pre-whitened, that is, multiplied by the matrix (T H T) -1 :X″=(T H T) -1 X′.

[0078] Taking a 64×64 rectangular phased array as an example, when the antenna unit spacing is half a wavelength, its two-dimensional beam width is about 1.6°. If the current wave position only cares about the targets within the ±2° airspace, and the number of targets within this range does not exceed 20 batches, then the corresponding (α m′ ,β m′) can be selected from 0°, ±0.8°, ±1.6° (a total of 25 combinations), that is, the number of beams M′ is 25. At this time, the number of data rows after beam domain transformation and pre-whitening processing is 25, which is significantly lower than the number of rows of 4096 of the original sampling data X, which can effectively reduce the computational complexity of subsequent processing; taking a 4×4 formation of drones with a spacing of 30 meters at a distance of 3 kilometers from the radar as an example, its angle interval is about 0.57°, and the airspace coverage does not exceed ±0.9°. After beam domain transformation, the airspace coverage and multi-channel processing freedom meet the detection requirements of this wave target.

[0079] Furthermore, step 2 includes the following:

[0080] Define the two-dimensional directional sine f = sinα, g = sinβ, divide J two-dimensional grids in the directional sine space of interest, and its minimum interval Δf, Δg is determined according to the required angular resolution. Taking the above example, the target angle interval does not exceed 1 / 3 of the beam width (the corresponding directional sine interval is about 0.01), and a certain margin is reserved, and Δf = Δg = 0.003, then the total number of two-dimensional grids J in the angular region of interest is about 200.

[0081] Construct a matrix C whose p-th row and q-th column elements are:

[0082] C pq =4Δf·Δg·sinc(2ΔfΔx pq / λ wl )·sinc(2ΔgΔy pq / λ wl )

[0083] Where Δx pq =x p -x q , Δy pq =y p -y q , sinc(x)=sin(πx) / (πx). Perform eigendecomposition on the matrix C and obtain the eigenvector v corresponding to the mth descending eigenvalue m . Then for the jth grid point (f j ,g j ), the standard orthogonal basis matrix for off-grid estimation is constructed based on the following formula:

[0084]

[0085] in, b(f,g)=a(α,β)=exp(j·2π(l x f+l y g) / λ wl); the number of columns K of the matrix is ​​determined by the following formula:

[0086]

[0087] Among them, ε needs to be sufficiently small to ensure the completeness of the basis matrix, thereby reducing the approximation error of the linear representation. Here, it is taken as 10 -4 .

[0088] In order to match the data preprocessing process, the basis matrix is ​​further transformed as follows:

[0089]

[0090] In addition, for real arrays with unknown non-idealities (element position error, channel amplitude and phase error, mutual coupling error, etc.), Directly through Perform eigendecomposition and obtain the eigenvector corresponding to the mth descending eigenvalue (where For the neighborhood A set of real array steering vectors actually measured in the experiment is used to ensure the validity of the model.

[0091] Furthermore, step 3 includes the following:

[0092] The premise of realizing off-grid estimation is to first determine the sine of the true two-dimensional direction of each target The closest grid index j d (d=1,...,D, D is the total number of targets), and estimate the target's true guidance vector In the basis matrix The linear representation coefficient vector under (hereinafter referred to as r d ),Right now To this end, the present invention provides a grid positioning and representation coefficient vector estimation method for two different application scenarios:

[0093] 1) When the echoes are weakly correlated and the number of snapshots is sufficient, a modified MUSIC algorithm is used, and its two-dimensional spatial spectrum is given by P(f j ,g j )=1 / λ min (G(f j ,g j )) is given, where λ min (·) represents the operation of taking the minimum eigenvalue of a matrix; For the M′×(M′-D)-dimensional noise subspace estimation, the covariance matrix X″X″ of the sampled data can be obtained by H / N performs eigendecomposition and extracts the eigenvectors corresponding to the smallest M′-D eigenvalues. By searching for Ρ(f j ,g j ) spectrum peak, we can find the index j of the target direction sine in the grid d ; r d The normalized estimate of By To further save computational cost, in this embodiment, the conventional MUSIC algorithm is first used to determine j d , and then use the above method to further estimate its corresponding

[0094] 2) In the case of strong echo correlation or coherence, few snapshots or a single snapshot, a block sparse method is used, and its structured redundant dictionary is composed of [E1″…E′ J ′] is given, and the corresponding signal matrix is a block sparse matrix with D non-zero blocks. Therefore, S can be estimated by block sparse recovery algorithms such as combined norm minimization. j ″(The estimated value is recorded as ), and determine the index j according to the position of the non-zero block d ; Non-zero block That is, for r d ·s d The estimated value (s d is a 1×N-dimensional signal vector), which is equivalent to the coefficient vector r d is a set of linear combinations of .

[0095] Furthermore, step 4 includes the following:

[0096] Based on the grid index j obtained in step 3 d and coefficient vector estimates, and further calculate the steering vector estimates (for the MUSIC algorithm described in step 3) or the estimated cluster after linear scaling (Block sparse class method described in step 3).

[0097] Considering that there are generally no multiple targets in the same grid under the defined grid density, the optimal estimate of the two-dimensional sine of the d-th target is obtained by solving the following optimization problem:

[0098]

[0099] in, As long as the grid spacing is not too coarse, the objective function h(f,g) can be guaranteed to be in the neighborhood near the target grid. Concave functions that are internally Lipschitz continuous.

[0100] Furthermore, step 5 includes the following:

[0101] The optimal solution to the optimization problem in step 4 can be obtained by accelerated gradient iteration. First, initialize Among them, the superscript indicates the number of iterations; is the estimated value of the d-th target in the i-th iteration, and is a pair of intermediate quantities; η i Determines the proportion of momentum component in the iteration process; μ f and μ g is the step size, which mainly affects the convergence speed.

[0102] The estimated value during the iteration is updated according to the following formula:

[0103]

[0104] in, and are the gradients of h(f,g) with respect to f and g, respectively, given by:

[0105]

[0106] Where Im{·} represents the imaginary part of the complex number. In order to speed up the iteration, the step size is adaptively updated during the iteration process. Specifically, the following adaptive step size adjustment mechanism based on grid constraints is adopted:

[0107]

[0108] Here, ρ is 1 / 3 to 1 / 2. g The update rules of are similar and will not be described here.

[0109] The iteration is terminated when the normalized gradient is sufficiently small, that is:

[0110]

[0111] Among them, max(·) means taking the maximum operation, and γ is 10 -3 After the iteration is terminated, the two-dimensional angle deviation estimate of the d-th target is given by Given.

[0112] In summary, the grid-free estimation model and fast algorithm for phased array spatial super-resolution proposed in the present invention are suitable for accurate resolution and angle estimation of radar dense targets under high real-time requirements. This method is applicable to arrays of arbitrary geometric arrangement (including sparse phased arrays with randomly drawn elements) or actual antenna arrays with non-idealities such as array errors; the estimation process consists of core steps such as structured grid positioning and adaptive step-size accelerated gradient iteration. The former supports MUSIC-type algorithms or block sparse algorithms, and the latter is based on a non-parametric model. Therefore, the proposed method is still effective under the conditions of highly correlated sources and a small / single number of snapshots. Compared with conventional grid-based algorithms, the method of the present invention can obtain excellent estimation performance without a fine grid, and its real-time performance is greatly improved.

[0113] The effects of the present invention are verified by the following examples and simulation experiments.

[0114] Figure 2 The azimuth estimation process of the adaptive step-size accelerated gradient iteration method proposed in the present invention is shown (ρ = 0.2, γ = 10 -3 ), where each set of curves represents the estimated value and true value of a target. It can be seen that due to the limited density of the angle grid, there is a certain error between the initial estimated value obtained in the grid positioning step and the true value; however, using the proposed algorithm, it only takes an average of more than 10 iterations for the estimated value to converge to a higher accuracy, thus fully ensuring the real-time performance of the off-grid estimation.

[0115] Figure 3 The estimation accuracy and processing speed of the proposed method under different grid densities are further demonstrated (the horizontal axis τ in the figure represents the ratio of the grid spacing to the beam width in the directional sine domain), where the two-dimensional angles of the target are randomly generated and uniformly distributed in 2000 trials, the signal-to-noise ratio (SNR) is 10 dB, and the number of snapshots is 256; ε = 10 is taken when constructing the basis matrix -5 , the iterative algorithm takes ρ = 0.4, γ = 10 -3 Obviously, under the same grid density, the computational complexity of the proposed method is comparable to that of the conventional MUSIC algorithm, but its accuracy is not sensitive to the grid density. In other words, to achieve the same estimation accuracy, the proposed method does not require fine grid division, and the time cost consumed is much lower than that of conventional MUSIC.

[0116] Figure 4 The dimension K of the basis matrix involved in the off-grid estimation model proposed in the present invention is analyzed under different ε values; it can be seen that as the grid density decreases, the basis matrix needs to be appropriately expanded to ensure the required approximation accuracy and avoid affecting the estimation accuracy.

[0117] Figure 5The estimation accuracy of the method of the present invention under different ε values ​​(τ = 1 / 8) and the MUSIC algorithm under different τ values ​​are further compared, where the true direction of the target is also randomly given in each test. Obviously, the estimation accuracy of the MUSIC algorithm based on the classical model is strongly related to the grid density. Even if the grid division has been refined to 1 / 100 beam width, its accuracy is still significantly worse than that of the method of the present invention with a grid spacing of 1 / 8 beam width under high signal-to-noise ratio; at the same time, when ε is small enough, the method of the present invention can approach the optimal estimation, that is, at this time, the influence of the approximation error in the off-grid estimation model proposed by the present invention can be ignored.

[0118] Depend on Figure 6 The time cost comparison in shows that, under the same grid density, the time cost of the method of the present invention is only slightly higher than that of the MUSIC algorithm (e.g. Figure 2 As shown in Figure 2, the adaptive accelerated gradient iteration step only requires more than 10 iterations on average); however, at the same accuracy level, the speed of the method of the present invention is more than one order of magnitude faster than the MUSIC algorithm (by Figure 5 It can be seen that the grid density required by the MUSIC algorithm is far more than 1 / 8 beam width).

[0119] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A grid estimation model and fast algorithm for phased array spatial super-resolution, characterized in that: The method comprises the following steps: S1: Perform beam domain transformation and pre-whitening processing on the multi-channel data acquired by the digital phased array radar according to the angular region of interest of the current wave position; S2: discretize the two-dimensional directional sine space according to the required angular resolution, and construct a basis matrix for off-grid estimation based on an approximate linear representation model; S3: A grid estimation model is established based on the basis matrix, and a structured MUSIC algorithm or a block sparse algorithm is used to locate the directional sinusoidal grid of each signal source, and the steering vector representation coefficient is estimated at the same time; S4: Calculate the precise steering vector estimate of each signal source using the representation coefficient, and construct a local optimization of the two-dimensional sine of each signal source; S5: Adopt the adaptive step-size accelerated gradient iteration algorithm based on grid constraints to solve the two-dimensional directional sine estimation value of each signal source respectively. After the iteration is terminated, the arc sine is solved to further obtain the accurate two-dimensional angle estimation value of each target.

2. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 1, characterized in that: The step S1 is specifically as follows: The M×N dimensional multi-channel sampling raw data X obtained by the digital phased array radar is multiplied by the M′×M beam domain transformation matrix T on the left. H We get X′=T H X, Where M is the number of array elements, N is the number of sampling snapshots; T = [a(α1, β1)a(α2, β2)…a(α M′ ,β M′ )], M′ is the number of beams, (α m′ ,β m′ ) is a two-dimensional spatial angle that can be flexibly selected within the spatial domain of interest at the current wave position; a(α m′ ,β m′ ) is the array steering vector corresponding to the m′th (m′=1,...,M′) beam pointing angle, and its expression is a(α,β)=exp(j·2π(l x sinα+l y sinβ) / λ wl ) gives, The two-dimensional spatial angle (α, β) is defined as the angle between two orthogonal normal planes to the plane where the array is located, and its value range is [-π / 2, π / 2]; l x =[x1…x M ] T and l y =[y1…y M ] T is the array coordinate vector, (x m ,y m ) is the two-dimensional coordinate of the mth (m=1,...,M) array element, λ wl is the working wavelength; Since the beam domain transformation process will cause the noise component in the data to no longer be Gaussian white noise, X′ is pre-whitened, that is, the matrix (T H T) -1 :X″=(T H T) -1 X′.

3. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 2, characterized in that: The step S2 is specifically as follows: Define the two-dimensional directional sine f = sinα, g = sinβ, divide J two-dimensional grids in the directional sine space of interest, and its minimum interval Δf, Δg is determined by the required angular resolution; then for any grid point (f j ,g j ), its neighborhood The steering vector corresponding to any two-dimensional sine in must be linearly represented by the eigenvector of the following matrix: Among them, b(f,g)=a(α,β)=exp(j·2π(l x f+l y g) / λ wl ); It is related to the grid spacing Δf and Δg. b α (f) = exp(j·2π·l x f / λ wl ), b β (g) = exp(j·2π·l y g / λ wl ), represents the Hadamard product of matrices; Specifically, C α and C β The expressions of the p-th row and q-th column (p, q = 1, ..., M) elements of α,pq =2Δfsinc(2ΔfΔx pq / λ wl ) and C β,pq =2Δgsinc(2ΔgΔy pq / λ wl ), and then the p-th row and q-th column element of matrix C is: C pq =4Δf·Δgsinc(2ΔfΔx pq / l wl )·sinc(2ΔgΔy pq / l wl ) where, Δx pq = x p - x q , Δy pq = y p - y q , sinc(x) = sin(πx) / (πx); This results in the standard orthogonal basis matrix for off-grid estimation: in, v m Arrange the mth descending eigenvalue λ of the matrix C m The corresponding eigenvector of; the number of columns K of the matrix is ​​determined by the following formula: Among them, ε is a sufficiently small value to ensure the completeness of the basis matrix and thus reduce the approximation error of the linear representation. Generally, ε is 10 -3 ~10 -5 ; When K = M, is a neighborhood a complete orthogonal basis for any two-dimensional azimuth sine corresponding steering vector b(f,g) within, that is, there exists a set of representation coefficient vectors r m (f,g) (m = 1,..., M) such that When K < M, then is an approximate complete orthogonal basis for the steering vector b(f,g) within the neighborhood that is where In order to match the data preprocessing process in step S1, the basis matrix is ​​further transformed as follows: In addition, for real arrays with unknown non-idealities, the integral form of H(f j ,g j ) can be replaced by its discrete version: in For the neighborhood A set of real array steering vectors actually measured in the simulation is used to ensure the validity of the model.

4. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 3, characterized in that: The step S3 includes the following contents: Based on the basis matrix obtained in step 2, the preprocessed data X″ in step 1 can be modeled as: Among them, j d is the true two-dimensional sine value of the direction closest to the dth target (d = 1, ..., D, D is the total number of targets) The grid point index of ; is the signal component under this model, in which the d-th target echo steering vector is introduced at the j-th d The representation coefficients at the grid points N″ is the noise component in the data; According to the above model, by determining the grid index j d The rough azimuth cosine estimate of the target can be obtained by estimating the coefficient vector (hereinafter referred to as r d ) can further obtain the precise guidance vector of the target, which can then be used to estimate the two-dimensional direction cosines.

5. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 4, characterized in that: Step S3 is as follows: When the echoes are weakly correlated and the number of snapshots is sufficient, a modified MUSIC algorithm is used, and its two-dimensional spatial spectrum is given by Ρ(f j ,g j )=1 / λ min (G(f j ,g j )) is given, where λ min (·) represents the operation of taking the minimum eigenvalue of a matrix; For the M′×(M′-D)-dimensional noise subspace estimation, the covariance matrix X″X″ of the sampled data can be obtained by H / N performs eigendecomposition and extracts the eigenvectors corresponding to the smallest M′-D eigenvalues; by searching for Ρ(f j ,g j ) spectrum peak, we can find the index j of the target direction sine in the grid d ; r d The normalized estimate of By The eigenvector corresponding to the minimum eigenvalue of is given; in order to further save the computational cost, the conventional MUSIC algorithm can also be used to determine j d , and then use the above method to further estimate its corresponding 6. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 4, characterized in that: Step S3 is as follows: In the case of strong echo correlation or coherence, few snapshots or a single snapshot, a block sparse method is used, whose structured redundant dictionary is composed of [E1″…E′ J ′] is given, and the corresponding signal matrix is a block sparse matrix with D non-zero blocks; therefore, S can be estimated by block sparse recovery algorithms such as combinatorial norm minimization. j ″(The estimated value is recorded as And determine the index j according to the position of the non-zero block d ; Non-zero block That is, for r d ·s d The estimated value (s d is a 1×N-dimensional signal vector), which is equivalent to the coefficient vector r d is a set of linear combinations of .

7. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 5 or 6, characterized in that: The step S4 includes the following contents: Based on the grid index j obtained in step 3 d and coefficient vector estimates, and further calculate the steering vector estimates (for the MUSIC algorithm described in step S3) or the estimated cluster after linear scaling (the block sparse method described in step S3); The following optimization problem is further constructed to obtain the optimal estimate of the sine of the d-th target two-dimensional direction: in, 8. The off-grid estimation model and fast algorithm for phased array spatial super-resolution according to claim 5 or 6, characterized in that: The step S5 includes the following contents: Given that the objective function h(f,g) of the optimization problem in step S4 is in the neighborhood For concave functions that are Lipschitz continuous, the optimal estimate can be obtained by accelerated gradient iteration as follows: Wherein, the superscript i indicates the i-th iteration; is the estimated value of the d-th target in the i-th iteration, and is a pair of intermediate quantities; η i Determines the proportion of momentum component in the iteration process; μ f and μ g is the step size, which mainly affects the convergence speed; the gradient of h(f,g) with respect to f and about g They are given by the following formulas: Where Im{·} represents the imaginary part of the complex number. To speed up the iteration, an adaptive step adjustment mechanism based on grid constraints is adopted, namely: Among them, 0<ρ<1, ρ is generally 1 / 3~1 / 2, μ g The update rules of are similar to those of , which will not be described here; The above iteration is initialized by To achieve this, the iteration is terminated when the normalized gradient is sufficiently small, that is: Among them, max(·) represents the maximum operation, and γ is usually 10 -3 After the iteration is terminated, the two-dimensional angle deviation estimate of the d-th target is given by Given.

Citation Information

Cited By

  • Power distribution network multi-branch line fault positioning method and system based on time domain difference

    CN120009669A