Method and device for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm
By using the Fourier dictionary matrix and G-S decomposition fast SBL algorithm in ISAR imaging, the problem of high computational complexity under segmented observation conditions is solved, and efficient and accurate high-resolution imaging is achieved.
Patent Information
- Application Number
- CN202210944216.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-05
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-08-05
AI Technical Summary
The prior art during ISAR imaging under segmented observation conditions has high computational complexity, resulting in a long imaging time, and the approximation method of the fast SBL algorithm reduces imaging accuracy.
The sparse Bayesian learning algorithm based on Fourier dictionary matrix is adopted, combined with G-S decomposition and fast Fourier transform, and the inverse matrix in the iteration process of sparse signals is quickly solved, reducing the computational complexity and improving the computational efficiency.
High-resolution imaging results accuracy is achieved, the computational complexity is reduced by several orders of magnitude, and the imaging efficiency is significantly improved.
Smart Images

Figure CN115453528B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of radar technology, and in particular relates to a method and a device for realizing segmented observation ISAR high-resolution imaging based on a fast SBL algorithm. Background Art
[0002] Since Inverse Synthetic Aperture Radar (ISAR) can image non-cooperative targets and provide physical information such as the shape, structure, material, motion state and mass distribution of the target, this unique advantage makes it increasingly valued by military powers around the world in the field of national defense. High-resolution images can intuitively provide richer target information and are the key technology of ISAR imaging. There are two ways to improve the range resolution. The first is to achieve the resolution of the target scattering points in the distance by transmitting a linear frequency modulation signal with a large time bandwidth and performing matched filtering and pulse compression at the receiving end, but this method is costly. Therefore, the second method is generally chosen, that is, to improve the range resolution by performing broadband synthesis of sparse sub-band data through interpolation or extrapolation, but the core of this method is to use staged missing data for accurate spectrum or parameter estimation. The azimuth resolution depends on the coherent accumulation angle. The so-called coherent accumulation refers to the high correlation between pulses in a certain direction, so that they can achieve energy accumulation through certain transformations and achieve a resolvable effect. In fact, due to the mobility of non-cooperative targets, it is also difficult to obtain large coherent aperture data. Although the small viewing angle simplifies the signal form and improves the imaging efficiency, it is not easy to meet the requirements of high resolution. In the face of limited short-aperture data, mining effective prior information and improving resolution are one of the research hotspots of ISAR imaging in recent years. In the ISAR system, radar tasks are generally divided into target tracking, scanning and broadband imaging modes. By transmitting narrow-band pulses to perform beam scanning within a wide viewing angle range, the target is captured and tracked, and then a broadband signal is emitted to perform coherent accumulation of a certain number of pulses to achieve high-resolution imaging of the target. In actual multi-target imaging, the single-antenna radar system needs to continuously alternate and switch between different functions to ensure the timeliness of radar imaging information perception. The design of multi-function phased array radar is to achieve scheduling between multiple radar tasks. However, it is generally very difficult to continuously observe the observation scene or target for a long time. The azimuth aperture observation of the target often appears discontinuous, resulting in the staged loss of observation data, that is, sparse aperture. In the broadband networked radar system, due to the limitations of the existing system, for a single target, each traditional device can usually only obtain limited time observation, corresponding to the sparse aperture observation of the target. In addition, in actual radar work, due to external environmental interference or its own influence, the radar's measured observation data in certain specific times is invalid. For observation data that is periodically missing in the range dimension or azimuth dimension, or called segmented observation data, the imaging results of the traditional range Doppler (RD) algorithm have strong ghosting, grating lobes and side lobes, resulting in a decrease in image resolution. Therefore, the study of segmented observation ISAR high-resolution imaging has important practical significance.
[0003] In view of the sparse characteristics of radar target echoes, the radar imaging model is converted into a sparse representation model, and the sparse reconstruction method is used to optimize and solve the radar target parameters, thus developing radar imaging technology based on sparse representation theory. Since the development of sparse representation theory, many sparse reconstruction algorithms have been developed. Among the many algorithms, the sparse Bayesian learning (SBL) algorithm has stronger robustness and higher estimation accuracy, so it has aroused the research interest of researchers in both theory and application. The SBL algorithm uses the prior information of the signal and image to solve the posterior probability of the sparse signal parameters through the Bayesian reasoning method. In the Bayesian calculation solution process, the signal parameters can be automatically estimated without manual parameter intervention. Moreover, since the SBL algorithm takes into account the noise statistical information of the signal, it can often obtain results that are better than other sparse reconstruction algorithms under low signal-to-noise ratio conditions.
[0004] The SBL algorithm achieves signal reconstruction through iteration. However, each iteration of the SBL algorithm contains a matrix inversion operation, and the dimension of the matrix is the same as the length of the observation data. If the traditional direct inversion method is used to solve it, its computational complexity is proportional to the cube of the length of the observation data. Therefore, when the number of observation data samples is large, the calculation time is often very long. In response to this problem, many scholars have proposed some fast SBL algorithms, but these fast algorithms all use some approximations, which will reduce the accuracy of the results. If used for segmented observation ISAR imaging, the accuracy of the results will be worse. Summary of the invention
[0005] In order to solve the above problems existing in the prior art, the present invention provides a method and device for realizing segmented observation ISAR high-resolution imaging based on a fast SBL algorithm. The technical problem to be solved by the present invention is achieved through the following technical solutions:
[0006] The embodiment of the present invention provides a method for implementing segmented observation ISAR high-resolution imaging based on a fast SBL algorithm, comprising the steps of:
[0007] S1. Modeling the segmented observation data according to the valid sampling samples and the missing sampling samples in the segmented observation data to obtain a segmented observation data model;
[0008] S2, based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal precision value and the noise precision value, using fast Fourier transform to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm;
[0009] S3, based on the inverse matrix to be found, using the Levinson-Durbin iterative algorithm to solve the GS decomposition factor of the inverse matrix;
[0010] S4, combining the GS decomposition factors of the inverse matrix, using the fast Fourier algorithm to solve the mean and diagonal elements of the covariance of the posterior distribution during the iteration of the sparse Bayesian learning algorithm on the sparse signal;
[0011] S5. When it is determined that the mean does not meet the convergence condition, a new signal accuracy value and a new noise accuracy value are calculated based on the diagonal elements of the covariance and the mean, and the process returns to step S2; when it is determined that the mean meets the convergence condition, the mean is output.
[0012] In one embodiment of the present invention, step S1 comprises:
[0013] The segmented observation data is divided into several segments according to the positions of the missing samples, and the segmented observation data is expressed as:
[0014]
[0015] The effective data is expressed as:
[0016]
[0017] in, N s(i) is the i-th segment data y s(i) The length, N g(i) is the valid data y g(i) The length, N m(i) For missing data y m(i) The length of , q is the number of segments of segmented observation data.
[0018] In one embodiment of the present invention, step S2 comprises:
[0019] S21. Under the segmented observation data model, the Fourier basis of the (k+1) column in the Fourier dictionary matrix is expressed as:
[0020]
[0021] Among them, ω k =2πk / K, k=0,…,K-1, K is the product of the super-resolution multiple and the observed data; For The corresponding Fourier basis is, is the valid data of the i-th segment, Represents a length of N g(i) A complete Fourier basis for N o(i) is the length bias based on the segmented observation data model;
[0022] S22. Based on the Fourier dictionary matrix, in combination with the signal precision value and the noise precision value, the inverse matrix to be calculated is calculated by fast Fourier transform:
[0023]
[0024] in, For a dimension of N g ×N g The identity matrix, β is the noise accuracy value, is the Fourier dictionary matrix, Λ is 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, is a q×q Hermitian-block-Toeplitz matrix;
[0025]
[0026] Among them, R i,j for One dimension is N g(i) ×N g(j) The sub-matrix of
[0027]
[0028] Among them, r m For R i,j Elements in
[0029] S23, using two permutation matrices and five custom matrices, the inverse matrix to be found is represented by the inverse matrix to be found in the target form:
[0030]
[0031] in, and is the permutation matrix, A custom matrix.
[0032] In one embodiment of the present invention, step S3 includes:
[0033] S31, based on the inverse matrix to be determined in the target form, calculate the inverse matrix using GS decomposition:
[0034]
[0035] in, and is the permutation matrix, is a custom matrix;
[0036] S32, calculating the displacement representation of the inverse matrix:
[0037]
[0038] in, Based on Custom variables for Based on Custom variables for and is the GS decomposition factor;
[0039] S33. Based on the displacement representation, calculate the GS decomposition of the inverse matrix:
[0040]
[0041] in, is a block Toeplitz matrix, and is the GS decomposition factor, is the block displacement matrix, M = max[N g(1) ,N g(2) ,…,N g(q) ],N g(i) is the valid data y g(i) Length;
[0042] S34, using the Levinson-Durbin iterative algorithm to calculate the GS decomposition factor of the GS decomposition formula.
[0043] In one embodiment of the present invention, step S4 includes:
[0044] S41. Combine the GS decomposition factors of the inverse matrix and use fast Fourier transform to solve the diagonal elements of the covariance. The expression of the vector composed of the covariance and its diagonal elements is:
[0045]
[0046] ε=diag(Σ)
[0047] Among them, ε=diag(Σ) means that the vector ε is composed of the elements on the diagonal of the matrix Σ in order, and Λ is composed of 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, β is the noise accuracy value, is the Fourier dictionary matrix, is the inverse matrix;
[0048] S42, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the mean, the expression of the mean is:
[0049]
[0050] Among them, β is the noise accuracy value, is the Fourier dictionary matrix, is the inverse matrix, Λ is 1 / γ k The diagonal matrix constructed in sequence, The total valid data.
[0051] In one embodiment of the present invention, step S41 includes the steps of:
[0052] S411, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform to solve the diagonal matrix:
[0053]
[0054] in, It represents a K-point fast Fourier transform of the vector in brackets. and is a custom vector based on the piecewise observation data model, is the Fourier dictionary matrix, is the inverse matrix;
[0055] S412. Calculate the covariance using the diagonal matrix, the signal precision value, and the noise precision value:
[0056]
[0057] Among them, ε k is the (k+1)th value of the matrix ε, which is a vector of the elements on the diagonal of the covariance matrix in order, δ k is the (k+1)th value of the diagonal matrix δ, β is the noise accuracy value, γ k is the signal accuracy value.
[0058] In one embodiment of the present invention, step S42 includes the steps of:
[0059] S421, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the first vector:
[0060]
[0061] in, is the inverse matrix, is the total valid data;
[0062] S422. Calculate a second vector using the first vector and the Fourier dictionary matrix:
[0063]
[0064] in, It represents a K-point inverse fast Fourier transform of the vector in brackets. is the Fourier dictionary matrix, is the first vector;
[0065] S423. Calculate the mean value by combining the second vector and the noise precision value:
[0066]
[0067] Among them, β is the noise accuracy value, Λ is calculated by 1 / γ k The diagonal matrix constructed in sequence, is the second vector.
[0068] In one embodiment of the present invention, the convergence condition is:
[0069]
[0070] Among them, δ is the set convergence threshold.
[0071] Another embodiment of the present invention provides an ISAR high-resolution imaging device for segmented observation based on a fast SBL algorithm, comprising:
[0072] A segmented observation data model building module is used to model the segmented observation data according to valid sampling samples and missing sampling samples in the segmented observation data to obtain a segmented observation data model;
[0073] The inverse matrix calculation module is used to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm using fast Fourier transform based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal precision value and the noise precision value;
[0074] A GS decomposition factor solving module is used to solve the GS decomposition factor of the inverse matrix by using the Levinson-Durbin iterative algorithm based on the inverse matrix to be solved;
[0075] A module for solving the diagonal elements and mean of the covariance, for solving the mean of the posterior distribution and the diagonal elements of the covariance in the process of iterating the sparse signal by the sparse Bayesian learning algorithm by using the fast Fourier algorithm in combination with the GS decomposition factor of the inverse matrix;
[0076] A judgment module is used to calculate a new signal accuracy value and a new noise accuracy value based on the diagonal elements of the covariance and the mean when it is judged that the mean does not meet the convergence condition, and return the inverse matrix calculation module to be calculated; when it is judged that the mean meets the convergence condition, output the mean.
[0077] Compared with the prior art, the present invention has the following beneficial effects:
[0078] The imaging method of the present invention adopts a sparse Bayesian learning algorithm based on a Fourier dictionary matrix, realizes high-resolution imaging of periodically missing data, and the imaging result has high accuracy; the inverse matrix in the iteration of the sparse Bayesian learning algorithm is represented by GS decomposition, avoiding the high computational complexity caused by directly calculating the inverse matrix; when the inverse matrix is involved, it can be quickly solved by fast Fourier transform, the amount of calculation is reduced by several orders of magnitude, and the computational efficiency is high; thus, the imaging method reduces the computational complexity on the basis of ensuring the accuracy of the imaging result, and greatly improves the imaging efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0079] Figure 1 A schematic diagram of a process for implementing a segmented observation ISAR high-resolution imaging method based on a fast SBL algorithm provided by an embodiment of the present invention;
[0080] Figure 2 A schematic diagram of a segmented observation data model provided for an embodiment of the present invention.
[0081] Figure 3 A schematic diagram of reconstruction results of different algorithms provided in an embodiment of the present invention;
[0082] Figure 4 A performance curve diagram of various algorithms when the length of observation data is different provided by an embodiment of the present invention;
[0083] Figure 5 A performance curve diagram of various algorithms with different observed data missing rates provided by an embodiment of the present invention;
[0084] Figure 6 A performance curve diagram of various algorithms when the number of segments into which the observation data is divided is different provided by an embodiment of the present invention;
[0085] Figure 7 A HRRP of complete Yak-42 data and an imaging result diagram using the range-Doppler algorithm and the SBL algorithm provided in an embodiment of the present invention;
[0086] Figure 8 An HRRP of segmented Yak-42 data and an imaging result diagram using a range-Doppler algorithm and a FD-GSBL algorithm are provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0087] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.
[0088] Embodiment 1
[0089] This embodiment first introduces a sparse signal reconstruction model and an SBL algorithm.
[0090] 11) Sparse signal reconstruction refers to the process of solving signal x based on measured data y under the premise that signal x has a certain sparsity; its system model can be described by an underdetermined linear system with noise, such as:
[0091] y=Dx+e
[0092] in, is the observational data; is an overcomplete dictionary matrix and K>>N; x refers to the sparse signal to be reconstructed, that is, most elements in the vector x are zero; Refers to observation noise.
[0093] 12) Among the many sparse signal reconstruction algorithms, the sparse Bayesian learning (SBL) algorithm has attracted the interest of researchers for its high accuracy. This is because SBL is a very important Bayesian statistical optimization algorithm, which is developed on the basis of Bayesian theory and realizes signal reconstruction from a statistical perspective. That is, under the SBL framework, the signal to be restored satisfies a certain prior distribution, and then the posterior distribution information of the signal is obtained through Bayesian analysis. The introduction of the SBL framework is as follows:
[0094] 12a) Prior distribution of the signal.
[0095] In order to effectively improve the sparsity of the signal, a hierarchical Bayesian prior model is usually used to describe the signal in SBL. The first layer of the hierarchical model is the modeling of the signal x and the noise e. Assume that x obeys a zero-mean covariance complex Gaussian distribution Λ, and e obeys a zero-mean covariance complex Gaussian distribution β -1 I, then the probability density function (PDF) is:
[0096]
[0097] Among them, x k represents the (k+1)th element of vector x and each element in vector x is independent of each other, γ k represents the signal accuracy (inverse variance), Λ is a function of 1 / γ k The diagonal matrix formed in sequence, e n represents the (n+1)th element of vector e, and β represents the noise accuracy.
[0098] The second layer of the hierarchical model is γ k The modeling of and β, they are all gamma distributions, and their PDFs are:
[0099]
[0100] Gamma(·) indicates that the variable in the brackets follows a gamma distribution, and a and b are γ k The shape and scale parameters of , c and d are the shape and scale parameters of β, respectively, which are called hyperparameters. To obtain a wide range of hyper-priors, hyperparameters are usually set to small positive constants, Γ(a) represents the gamma function.
[0101] 12b) The posterior distribution of the signal.
[0102] Based on the prior distribution and measurement data y introduced in 12a), the posterior distribution of x can be obtained by using the Bayesian formula and the expectation maximization (EM) algorithm. Its posterior distribution can be analytically expressed as a complex Gaussian distribution:
[0103]
[0104] Among them, the covariance of the complex Gaussian distribution Σ=(βD H D+Λ -1 ) -1 , mean μ=βΣD H y.
[0105] According to the Woodbury matrix identity, Σ and μ can be expressed as:
[0106] Σ=(βD H D+Λ -1 ) -1 =Λ-βΛD H Q -1 DΛ
[0107] μ=βΛD H Q -1 y
[0108] Where, Q = I + βDΛD H .
[0109] 12c) SBL algorithm iteration process.
[0110] In SBL, signal reconstruction is achieved through iteration. The optimal mean of the posterior distribution introduced in 12b) is the reconstructed signal. The following are the iterative steps of the SBL algorithm, which is called direct inversion SBL (DI-SBL):
[0111]
[0112] Q=I+βDΛDH
[0113] Σ (j) =Λ-β (j) ΛD H Q -1 DΛ
[0114] ε (j) =diag(Σ (j) )
[0115] μ (j) =β (j) ΛD H Q -1 y
[0116] Among them, ε=diag(Σ) means that ε is a vector consisting of the elements on the diagonal of the matrix Σ, ε k represents the (k+1)th element of ε; Represents the γ obtained after the jth iteration k ; μ and ∑ represent the mean and covariance of the posterior probability of signal x respectively; where the unknown parameter γ of μ k and β are called hyperparameters and can be solved by the maximum expectation algorithm; ||·||2 represents the l2 norm.
[0117] From the above DI-SBL iterative process, it can be seen that the key step of the SBL single iteration process is to calculate ε and μ, but the calculation process requires solving Q -1 The computational complexity of the traditional direct inversion method is proportional to the cube of the matrix dimension. The dimension of the matrix Q is the same as the dimension of the observation vector y. If there are many observation data, the calculation time is often very long, which is difficult to achieve in actual engineering.
[0118] In segmented observation ISAR imaging, although SBL can achieve high-resolution imaging with high accuracy, it has the above-mentioned problems, resulting in high time cost. Therefore, this embodiment proposes a fast SBL method based on Fourier dictionary to achieve segmented observation radar high-resolution imaging, which is referred to as FD-GSBL. Because in the SBL algorithm based on Fourier dictionary, the inverse matrix Q to be calculated is a Hermitian-Block-Toeplitz matrix, Q can be solved by using Gohberg-Semencul (GS) decomposition -1 , avoiding the large computational complexity caused by direct inversion. In addition, based on the GS decomposition factor, ε and μ can be solved by FFT / IFFT, which greatly shortens the calculation time.
[0119] See also Figure 1 , Figure 1A schematic diagram of a process for implementing a segmented observation ISAR high-resolution imaging method based on a fast SBL algorithm provided by an embodiment of the present invention. The imaging method specifically comprises the following steps:
[0120] S1. Modeling the segmented observation data according to valid sampling samples and missing sampling samples in the segmented observation data to obtain a segmented observation data model.
[0121] See also Figure 2 , Figure 2 A schematic diagram of a segmented observation data model provided by an embodiment of the present invention. For segmented observation data, its signal reconstruction schematic diagram is as follows Figure 2 As shown in the figure, the gray box represents the valid sampling samples, and the white box represents the missing sampling samples. According to the position of the missing sampling samples, all the segmented observation data can be divided into q segments, and the i-th segment data y s(i) The length is N s(i) , effective data y g(i) The length is N g(i) , missing data y m(i) The length is N m(i) . Then all segmented observation data and valid data can be expressed as:
[0122]
[0123] in, N s(i) is the i-th segment data y s(i) The length, N g(i) is the valid data y g(i) The length, N m(i) For missing data y m(i) The length of , q is the number of segments of segmented observation data.
[0124] Obviously, and N s(i) =N g(i) +N m(i) , and the missing rate of this data is
[0125] S2, based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal precision value and the noise precision value, the fast Fourier transform is used to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm. Specifically including the steps:
[0126] S21. Representing a Fourier dictionary matrix under the segmented observation data model.
[0127] First, define a length offset N o(i) :
[0128]
[0129] Because this embodiment uses a dictionary composed of Fourier bases, when there is missing data, the dictionary matrix is not a complete Fourier dictionary. Represents the Fourier dictionary matrix, the Fourier dictionary matrix The Fourier basis of the (k+1)th column in is expressed as:
[0130]
[0131] Among them, ω k =2πk / K, k=0,…,K-1, K is the product of the super-resolution multiple and the total observation data length. For The corresponding Fourier basis.
[0132] It is expressed as:
[0133]
[0134] in, is the valid data of the i-th segment, Represents a length of N g(i) A complete Fourier basis, N o(i) is the length bias of the model based on segmented observation data.
[0135] It is expressed as:
[0136]
[0137] S22. Based on the Fourier dictionary matrix, in combination with the signal precision value and the noise precision value, a fast Fourier transform is used to calculate the inverse matrix to be calculated.
[0138] Based on the Fourier dictionary matrix in step S21 In (12b), Q can be expressed as:
[0139]
[0140] in, is the inverse matrix to be found, For a dimension of N g ×N g The identity matrix, β is the noise accuracy value, is the Fourier dictionary matrix, Λ is 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, is a q×q Hermitian-block-Toeplitz matrix.
[0141] Hermitian-Block-Toeplitz Matrix It is expressed as:
[0142]
[0143] In the above formula, R i,j for One dimension is N g(i) ×N g(j) The submatrix of is expressed as:
[0144]
[0145] Among them, r m For R i,j The elements in r m The calculation formula is:
[0146]
[0147] From the above formula, we can see that r m By k The K-point fast Fourier transform (FFT) is quickly calculated, and its computational complexity is O(Klog2K).
[0148] based on The inverse matrix to be found can be calculated. Obviously, the inverse matrix to be found is also a Hermitian-block-Toeplitz matrix and has the same structure as The structure is the same.
[0149] S23. Using two permutation matrices and five user-defined matrices, the inverse matrix to be determined is represented by an inverse matrix to be determined in a target form.
[0150] First, we will use two permutation matrices to transform the inverse matrix It can be represented in two block matrix forms.
[0151] Specifically, given The submatrix Q in i,j is a dimension of N g(i) ×N g(j) The Toeplitz matrix can be written in two block matrix forms:
[0152]
[0153] in,
[0154] Representative Q i,j One of the dimensions is (Ng(i) -1)×(N g(j) -1), and is also a Toeplitz matrix.
[0155] Based on the submatrix Q i,j The following seven matrices are defined. The elements of these matrices come from each Q i,j For ease of introduction, superscript<i,j> Represents the matrix Q i,j Elements in, for example, and Represents Q i,j q0 and q1 in
[0156]
[0157]
[0158] Similarly, define and Since the inverse matrix is a Hermitian matrix, so we have and
[0159] Using the above seven custom matrices and two permutation matrices, the inverse matrix to be found can be Written in the following target form:
[0160]
[0161] in, and is the permutation matrix, A custom matrix.
[0162] It should be noted that among the seven custom matrices mentioned above, and Therefore, when writing the inverse matrix to be found into the target form, only five custom matrices are used.
[0163] S3. Based on the inverse matrix to be determined, the GS decomposition factor of the inverse matrix is solved by using the Levinson-Durbin iterative algorithm.
[0164] S31. Based on the inverse matrix to be determined in the target form, calculate the inverse matrix using GS decomposition.
[0165] Specifically, the inverse matrix Q to be solved in step S23 is Ng Using the block matrix inversion formula, we can get the inverse matrix:
[0166]
[0167] in, and is the permutation matrix, A custom matrix.
[0168] S32. Calculate the displacement representation of the inverse matrix.
[0169] because By q 2 The dimension of the sub-matrices is (N g -q)×(N g -q) matrix, then It can also be divided into q 2 In the form of sub-matrices, represent One of the dimensions is (N g(i) -1)×(N g(j) -1). By using the permutation matrix and Two matrices can be constructed:
[0170] and
[0171] One of their sub-matrices is:
[0172] and
[0173] Next, define a dimension N g ×N g The block displacement matrix like:
[0174]
[0175] Among them, S N is a unit lower triangular displacement matrix of dimension N×N, that is
[0176] Block displacement matrix Available
[0177]
[0178] Then, the inverse matrix The displacement representation The formula is as follows:
[0179]
[0180] make
[0181]
[0182]
[0183] Therefore, the displacement represents for:
[0184]
[0185] in, Based on Custom variables for Based on Custom variables for and is the GS decomposition factor. From the displacement representation The expression of the inverse matrix The displacement rank is 2q.
[0186] S33. Based on the displacement representation, calculate the GS decomposition formula of the inverse matrix.
[0187] Specifically, the dimension N in step S32 is used. g ×N g The block displacement matrix and displacement representation Inverse Matrix It is expressed as:
[0188]
[0189] Where M = max[N g(1) ,N g(2) ,…,N g(q) ].
[0190] Define a Toeplitz matrix and a block Toeplitz matrix like:
[0191]
[0192] Among them, t N and Respectively represent a length of N and N g Vector.
[0193] Then, the inverse matrix The GS decomposition expression is:
[0194]
[0195] in, is a block Toeplitz matrix, and is the GS decomposition factor, is the block displacement matrix, M = max[N g(1) ,N g(2) ,…,N g(q) ],N g(i) is the valid data y g(i) Length.
[0196] The above formula is called the inverse matrix GS decomposition, and It is called the decomposition factor of GS decomposition.
[0197] S34, using the Levinson-Durbin iterative algorithm to calculate the GS decomposition factor of the GS decomposition formula.
[0198] Influenced by the Levinson-Durbin (LD) algorithm, this embodiment proposes an iterative method to calculate the inverse matrix GS decomposition factors. The details are as follows:
[0199] make
[0200]
[0201] enter: and
[0202] Initial Iteration:
[0203] Iteration process:
[0204]
[0205] in,
[0206] Output:
[0207] Then, using W in step S31 q and V q The calculation formula for solving W q and V q .
[0208] S4. In combination with the GS decomposition factors of the inverse matrix, the fast Fourier algorithm is used to solve the diagonal elements and mean of the covariance of the posterior distribution during the iteration of the sparse Bayesian learning algorithm on the sparse signal.
[0209] In this embodiment, the fast Fourier algorithm includes fast Fourier transform or inverse fast Fourier transform.
[0210] Specifically, based on the Fourier dictionary, the vector ε and mean μ composed of the covariance Σ and its diagonal elements are expressed as:
[0211]
[0212] ε=diag(Σ)
[0213]
[0214] Among them, ε=diag(Σ) means that the vector ε is composed of the elements on the diagonal of the matrix Σ in order, and Λ is composed of 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, β is the noise accuracy value, is the Fourier dictionary matrix, is the inverse matrix; The total valid data.
[0215] S41. In combination with the GS decomposition factors of the inverse matrix, the diagonal elements of the covariance are solved by using fast Fourier transform.
[0216] Because in the covariance formula Σ, Λ is a diagonal matrix, the calculation of the diagonal elements of the covariance Σ can be divided into two steps:
[0217]
[0218] Among them, ε k and δ k denote the (k+1)th value of the diagonal matrix ε and the target matrix δ, respectively. is the Fourier dictionary matrix, is the inverse matrix. And because the signal accuracy value γ k The sum of the noise accuracy values β can be calculated by (12b), which only requires a fast calculation of the target matrix δ.
[0219] S411. Combine the GS decomposition factors of the inverse matrix and use fast Fourier transform to solve the diagonal matrix.
[0220] Specifically, substitute the Fourier dictionary matrix in step S21 and the inverse matrix in step S31 into the diagonal matrix In, we get:
[0221]
[0222] in, represent The i-th subvector of represent The i-th subvector of .
[0223] The (k+1)th value of is:
[0224]
[0225] in, Representation Matrix The sum of all elements on the mth diagonal of . It can be quickly solved by the following method:
[0226] If N g(i) ≥N g(j) ,make Its calculation expression is:
[0227]
[0228] Among them, if but Note that the superscript i here refers to the quantity The elements in .
[0229] Obviously c ij The right side of the expression is the product of a Toeplitz matrix and a vector. Their product can be efficiently realized by FFT / IFFT.
[0230] If N g(i) ≥N g(j) +2, so The calculation formula is:
[0231]
[0232] If N g(i) <N g(j) , Can be calculated by the same method.
[0233] From the above The calculation formula shows that Can be quickly calculated by FFT:
[0234] make
[0235]
[0236] Among them, 0 means that the vector A zero vector of dimension N, 0 ij Represents a dimension (N o(j) -N o(i) -N g(i) +1), 0 ji Represents a dimension (N o(i) -N o(j) -N g(j) ). Then, by Perform K-point Fast Fourier Transform FFT calculation It can be obtained by the same method. Therefore, the diagonal matrix δ can be quickly calculated by fast Fourier transform FFT, that is:
[0237]
[0238] in, It represents a K-point FFT (Fast Fourier Transform) of the vector in brackets. and is a custom vector based on the piecewise observation data model;
[0239] S412. Calculate the diagonal elements of the covariance using the diagonal matrix, the signal precision value, and the noise precision value.
[0240] Specifically, after the diagonal matrix δ is calculated, the diagonal matrix δ, the noise accuracy value β and 1 / γ are used to calculate the diagonal matrix δ. k The dot product of ε is used to calculate the diagonal elements of the covariance Σ. That is:
[0241] ε=diag(Σ)
[0242]
[0243] Among them, ε k is the (k+1)th value of the matrix ε, which is a vector of the elements on the diagonal of the covariance matrix in order, δ k is the (k+1)th value of the diagonal matrix δ, β is the noise accuracy value, γ k is the signal accuracy value.
[0244] S42, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the mean. Specifically comprising the steps of:
[0245] S421, combining the GS decomposition factor and the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the first vector:
[0246]
[0247] in, is the inverse matrix, is valid data.
[0248] Specifically, substitute the inverse matrix in step S31 into The expression of The right side of is the product of some Toeplitz matrix and a vector, so It can be quickly calculated by fast Fourier transform FFT / inverse fast Fourier transform IFFT.
[0249] S422. Calculate a second vector using the first vector and the Fourier dictionary matrix:
[0250]
[0251] Specifically, Divided into q segments, the length of the i-th segment is N g(i) .make in, but The calculation formula is:
[0252]
[0253] in, It means to do K-point IFFT on the vector in brackets. is the Fourier dictionary matrix, is the first vector.
[0254] S423. Calculate the mean value by combining the second vector and the noise precision value.
[0255] Specifically, through β and 1 / γ k The dot product of is:
[0256]
[0257] Among them, β is the noise accuracy value, Λ is calculated by 1 / γ k The diagonal matrix constructed in sequence, is the second vector.
[0258] This embodiment utilizes the SBL algorithm based on the Fourier dictionary, where the inverse matrix to be calculated is a Block-Toeplitz matrix, which can be represented by GS decomposition. At the same time, the operation involving the inverse matrix can be quickly calculated by FFT / IFFT, which reduces the computational complexity while ensuring the accuracy of the imaging result, reduces the amount of operation by an order of magnitude, and achieves better real-time performance.
[0259] S5. When it is determined that the mean does not meet the convergence condition, a new signal accuracy value and a new noise accuracy value are calculated based on the diagonal elements of the covariance and the mean, and the process returns to step S2; when it is determined that the mean meets the convergence condition, the mean is output to complete high-resolution imaging.
[0260] Specifically, set the convergence threshold δ and determine whether the μ value obtained in each iteration meets the convergence condition:
[0261]
[0262] If the convergence condition is not met, the iteration continues; if the convergence condition is met, the optimal mean value can be obtained, that is, the reconstructed signal, to achieve high-resolution imaging.
[0263] This embodiment achieves high-resolution imaging of staged missing data by applying the Fourier dictionary-based SBL algorithm to segmented observation ISAR imaging, and the imaging result has high accuracy. The fast algorithm proposed in this example uses GS decomposition to represent the inverse matrix in the SBL iteration, avoiding the high computational complexity caused by directly calculating the inverse matrix. At the same time, the formula involving the inverse matrix can be quickly solved by FFT, the amount of calculation is reduced by several orders of magnitude, and the computational efficiency is high. Therefore, the computational complexity is reduced on the basis of ensuring the accuracy of the imaging result, and the imaging efficiency is greatly improved.
[0264] Embodiment 2
[0265] On the basis of the first embodiment, this embodiment further illustrates the method for implementing segmented observation ISAR high-resolution imaging based on the fast SBL algorithm of the first embodiment through simulation experiments.
[0266] SBL algorithm parameter settings: initial value Hyperparameters a = b = c = d = 10 -6 ; Convergence threshold δ = 10 -3 ; Frequency sampling factor K / N s =4. In order to clearly see the performance of the imaging method of the FD-GSBL algorithm proposed in the present invention, some existing typical sparse signal reconstruction methods are added in this embodiment for comparison, including the fast iterative adaptive iterative algorithm (FIAA), orthogonal matching pursuit (OMP), S-ESBL and DI-SBL algorithms. Here, the S-ESBL algorithm is a proposed approximate fast SBL algorithm, and the DI-SBL refers to a direct calculation SBL algorithm.
[0267] Simulation experiment: The simulated observation data comes from a simulated signal with 25 random frequency points and a signal-to-noise ratio of 10dB.
[0268] Measured data experiment: The measured observation data comes from the Yak-42 aircraft. The radar used to collect ISAR data operates in the C-band, the frequency band is 400 MHz, and the pulse repetition frequency is 300 Hz. The range window has 256 sampling points, and the imaging time contains 256 pulses.
[0269] In order to show the reconstruction performance of the algorithm, the normalized root mean square error (nRMSE) of signal reconstruction is defined as:
[0270]
[0271] in, represents the reconstructed signal value, and x represents the real signal value.
[0272] 1.2 Experimental content and results
[0273] The first step is to use MATLAB R2020b to reconstruct the simulated observation data. The simulated data is 512 in length, divided into 4 segments, and the missing rate is 50%. Figure 3 , Figure 3 A schematic diagram of reconstruction results of different algorithms provided in an embodiment of the present invention, wherein: Figure 3 (a) is the reconstruction result of FIAA algorithm; Figure 3 (b) is the reconstruction result of the OMP algorithm; Figure 3 (c) is the reconstruction result diagram of the S-ESBL algorithm; Figure 3 (d) is the reconstruction result of DI-SBL algorithm; Figure 3 (e) is the reconstruction result diagram of the FD-GSBL algorithm.
[0274] Table 1 shows the signal reconstruction time and normalized root mean square error of the above algorithms.
[0275] Table 1 Comparison of signal reconstruction time and normalized root mean square error of various algorithms
[0276] FIAA OMP S-ESBL DI-SBL FD-GSBL Reconstruction time / s 0.7910 0.0548 5.8342 28.2271 1.5619 nRMSE 0.3230 0.6772 0.4500 0.2498 0.2498
[0277] The second step is to do a Monte Carlo experiment to compare the performance graphs of the algorithms under different parameters. Figure 4 , 5 , as shown in Figure 6. Figure 4 A performance curve diagram of various algorithms when the length of observation data is different provided by an embodiment of the present invention, wherein Figure 4 (a) is the reconstructed calculation time, and the time value on the curve is the logarithm; Figure 4 (b) is the normalized root mean square error of the reconstruction. Figure 4 (c) Graph is the variance of the normalized root mean square error. Figure 5 The performance curves of various algorithms provided by the embodiment of the present invention when the observed data missing rate is different are as follows: Figure 5 (a) Figure, Figure 5 (b) Figure and Figure 5 (c) The graph represents the computation time, normalized RMS error, and variance, respectively. Figure 6 The present invention provides a performance curve diagram of various algorithms when the number of segments of observation data is different, wherein Figure 6 (a) Figure, Figure 6 (b) Figure and Figure 6(c) The graph represents the computation time, normalized RMS error, and variance, respectively.
[0278] The third step is to use the software MATLAB R2020b to image the measured data. In order to compare with the imaging effect of the segmented observation data, this embodiment provides an imaging result diagram of the complete observation data. Figure 7 , Figure 7 An imaging result diagram of a complete "Yak-42" data provided by an embodiment of the present invention, wherein: Figure 7 (a) is the high-resolution range image (HRRP), Figure 7 (b) is the imaging result of the traditional range-Doppler algorithm. Figure 7 (c) is the imaging result of the DI-SBL algorithm. For the segmented observation data, assuming that the Yak-42 data is missing in the azimuth dimension, its MR is 50%. Please refer to Figure 8 , Figure 8 An HRRP of segmented "Yak-42" data and an imaging result diagram using a range-Doppler algorithm and a FD-GSBL algorithm provided by an embodiment of the present invention, wherein: Figure 8 (a) is the HRRP diagram of the segmented Yak-42 data. Figure 8 (b) is the imaging result using the range-Doppler algorithm. Figure 8 (c) is the imaging result using the FD-GSBL algorithm. The range dimension oversampling factor for complete data and missing data imaging is 4. The dynamic display range of all imaging images is 40dB.
[0279] Table 2 gives the average running time of the DI-SBL and FD-GSBL implementations in the above segmented observation measured data imaging.
[0280] Table 2 Comparison of average running time of DI-SBL and FD-GSBL in segmented observation of measured data imaging
[0281] DI-SBL FD-GSBL Time / s 8.9184 1.9420
[0282] 1.3 Results Analysis
[0283] from Figure 3It can be seen that when the two frequency values of the signal differ by one minimum frequency resolution unit, the signal reconstruction results of the FIAA, DI-SBL and FD-GSBL algorithms are very good, indicating that they have higher resolution. However, the signal reconstruction results of the S-ESBL and OMP algorithms are poor. The signal reconstruction results of the DI-SBL and FD-GSBL algorithms are the same. In addition, the nRMSE of the various algorithms listed in Table 1 also proves the above conclusion, that is, the nRMSE of the FIAA, DI-SBL and FD-GSBL algorithms are relatively small, and the nRMSE of the DI-SBL and FD-GSBL algorithms are the same. By comparing the calculation time in Table 1, it can be obtained that the calculation time of the OMP algorithm is the shortest, and FD-GSBL is more than 18 times faster than DI-SBL.
[0284] Figure 4 , 5 , 6 show the impact of some variables on the performance of the algorithm. Figure 4 (a) and Figure 4 (b) It can be seen that as the total length of the observed data increases, the algorithm's calculation time becomes longer and nRMSE gradually decreases; for different MRs, the larger the MR of the data, the less valid data. The single iteration time of the S-ESBL, DI-SBL and FD-GSBL algorithms decreases, but when the convergence threshold is reached, the total number of iterations increases. Figure 5 (a) and Figure 5 As shown in (b), as MR increases, the calculation time of the algorithm decreases within a certain range, while nRMSE increases. The calculation time of the FD-GSBL algorithm is several times shorter than that of the DI-SBL algorithm. In addition, when MR is greater than 40%, the nRMSE of the S-ESBL algorithm becomes larger and increases rapidly with the increase of MR. This is because S-ESBL adopts a certain approximation. The more missing samples, the worse the reconstruction effect; when MR is greater than 70%, the nRMSE of FIAA becomes larger; the nRMSE of DI-SBL and FD-GSBL also increase, but the increase is very small. Even when MR is 80%, the error value is acceptable. From Figure 6 (a) and Figure 6 (b) It can be seen that the q value only affects the computational complexity of FIAA and FD-GSBL, but has no effect on the reconstruction error value. And as the q value increases, the computational time of FIAA and FD-GSBL becomes longer. This is because the shift rank of the inverse matrix of the covariance matrix in the FIAA algorithm and the inverse matrix in the FD-GSBL algorithm are both 2q. The larger the value of q, the higher the computational time. In addition, in order to compare the stability of the algorithm, the variance diagram of the nRMSE of the algorithm is also Figure 4 (c) Figure 5 (c) and Figure 6It is given in (c). Obviously, the variance values of these algorithms are all less than 0.03, indicating that various algorithms have good stability.
[0285] In order to verify the effectiveness of the fast algorithm proposed in this embodiment, the measurement data of the "Yak-42" aircraft were imaged using the traditional range-Doppler algorithm and the FD-GSBL algorithm. The high-resolution range image and imaging results of the complete measurement data are shown in Figure 2. Figure 7 As shown, the high-resolution range image and imaging results of the segmented measurement data are shown in Figure 8 As shown. Figure 7 It can be seen that the imaging results of the range-Doppler algorithm have a higher sidelobe level, and the imaging results of the DI-SBL algorithm are better. Figure 8 It can be seen that compared with the imaging results of the complete data, the imaging results of the range-Doppler algorithm for the segmented data have a higher sidelobe level, while the imaging results of the FD-GSBL algorithm proposed in this paper are better, indicating that the FD-GSBL algorithm has a higher imaging resolution.
[0286] Table 2 shows the computation time of the DI-SBL and FD-GSBL algorithms in the segmented measured experiment. It is obvious that the computation time of the FD-GSBL algorithm is very short compared with that of the DI-SBL. Because the length of the measured data used in the experiment is only 128, the MR is 50%, and the amount of effective data is small, the acceleration effect of the FD-GSBL algorithm is not obvious. In summary, it can be seen that even in the case of large MR, the FD-GSBL algorithm can obtain good imaging results and the computation time is short.
[0287] Embodiment 3
[0288] Based on Example 1, this embodiment also provides a segmented observation ISAR high-resolution imaging device based on a fast SBL algorithm. The imaging device includes: a segmented observation data model establishment module, an inverse matrix calculation module, a GS decomposition factor solution module, a covariance diagonal element and mean solution module, and a judgment module.
[0289] The segmented observation data model establishment module is used to model the segmented observation data according to the valid sampling samples and missing sampling samples in the segmented observation data to obtain the segmented observation data model. The inverse matrix calculation module is used to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal accuracy value and the noise accuracy value, using the fast Fourier transform. The GS decomposition factor solving module is used to solve the GS decomposition factor of the inverse matrix based on the inverse matrix to be calculated using the Levinson-Durbin iterative algorithm. The diagonal elements and mean solution module of the covariance is used to combine the GS decomposition factor of the inverse matrix, and use the fast Fourier algorithm to solve the diagonal elements and mean of the covariance of the posterior distribution in the process of iterating the sparse signal by the sparse Bayesian learning algorithm. The judgment module is used to calculate a new signal accuracy value and a new noise accuracy value based on the diagonal elements of the covariance and the mean when it is judged that the mean does not meet the convergence condition, and return the inverse matrix calculation module to be calculated; when it is judged that the mean meets the convergence condition, output the mean.
[0290] The ISAR high-resolution imaging device for segmented observation based on the fast SBL algorithm provided in the embodiment of the present invention can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be repeated here.
[0291] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.
Claims
1. A method for realizing segmented observation ISAR high-resolution imaging based on a fast SBL algorithm, characterized in that: Includes steps: S1. Modeling the segmented observation data according to the valid sampling samples and the missing sampling samples in the segmented observation data to obtain a segmented observation data model; S2, based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal precision value and the noise precision value, using fast Fourier transform to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm; S3, based on the inverse matrix to be found, using the Levinson-Durbin iterative algorithm to solve the GS decomposition factor of the inverse matrix; S4, combining the GS decomposition factors of the inverse matrix, using the fast Fourier algorithm to solve the mean and diagonal elements of the covariance of the posterior distribution during the iteration of the sparse Bayesian learning algorithm on the sparse signal; S5. When it is determined that the mean does not meet the convergence condition, a new signal accuracy value and a new noise accuracy value are calculated based on the diagonal elements of the covariance and the mean, and the process returns to step S2; when it is determined that the mean meets the convergence condition, the mean is output.
2. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 1, characterized in that: Step S1 includes: The segmented observation data is divided into several segments according to the positions of the missing samples, and the segmented observation data is expressed as: Valid data is represented as: in, N s(i) is the i-th segment data y s(i) The length, N g(i) is the valid data y g(i) The length, N m(i) For missing data y m(i) The length of , q is the number of segments of segmented observation data.
3. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 1, characterized in that: Step S2 includes: S21. Under the segmented observation data model, the Fourier basis of the (k+1) column in the Fourier dictionary matrix is expressed as: Among them, ω k =2πk / K, k=0,…,K-1, K is the product of the super-resolution multiple and the total length of the segmented observation data; For The corresponding Fourier basis is, is the valid data of the i-th segment, Represents a length of N g(i) A complete Fourier basis for N o(i) is the length bias based on the segmented observation data model; S22. Based on the Fourier dictionary matrix, in combination with the signal precision value and the noise precision value, the inverse matrix to be calculated is calculated by fast Fourier transform: in, The dimension is N g ×N g The identity matrix, β is the noise accuracy value, is the Fourier dictionary matrix, Λ is 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, is a q×q Hermitian-block-Toeplitz matrix; Among them, R i,j for One dimension is N g(i) ×N g(j) The sub-matrix of Among them, r m For R i,j Elements in S23, using two permutation matrices and five custom matrices, the inverse matrix to be found is represented by the inverse matrix to be found in the target form: in, and is the permutation matrix, A custom matrix.
4. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 3, characterized in that: Step S3 includes: S31, based on the inverse matrix to be determined in the target form, calculate the inverse matrix using GS decomposition: in, and is the permutation matrix, is a custom matrix; S32, calculating the displacement representation of the inverse matrix: in, Based on Custom variables for Based on Custom variables for and is the GS decomposition factor; S33. Based on the displacement representation, calculate the GS decomposition of the inverse matrix: in, is a block Toeplitz matrix, and is the GS decomposition factor, is the block displacement matrix, M = max[N g(1) ,N g(2) ,…,N g(q) ],N g(i) is the valid data y g(i) Length; S34, using the Levinson-Durbin iterative algorithm to calculate the GS decomposition factor of the GS decomposition formula.
5. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 1, characterized in that: Step S4 includes: S41. Combine the GS decomposition factors of the inverse matrix and use fast Fourier transform to solve the diagonal elements of the covariance. The vector expression composed of the covariance and its diagonal elements is: ε=diag(Σ) Among them, ε=diag(Σ) means that the vector ε is composed of the elements on the diagonal of the matrix Σ in order, and Λ is composed of 1 / γ k The diagonal matrix constructed in sequence, γ k is the signal accuracy value, β is the noise accuracy value, is the Fourier dictionary matrix, is the inverse matrix; S42, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the mean, the expression of the mean is: Among them, β is the noise accuracy value, is the Fourier dictionary matrix, is the inverse matrix, Λ is 1 / γ k The diagonal matrix constructed in sequence, The total valid data.
6. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 5, characterized in that: Step S41 includes the steps of: S411, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform to solve the diagonal matrix: in, It represents a K-point fast Fourier transform of the vector in brackets. and is a custom vector based on the piecewise observation data model, is the Fourier dictionary matrix, is the inverse matrix; S412. Calculate the diagonal elements of the covariance using the diagonal matrix, the signal precision value, and the noise precision value: Among them, ε k is the (k+1)th value of the matrix ε, which is a vector of the elements on the diagonal of the covariance matrix in order, δ k is the (k+1)th value of the diagonal matrix δ, β is the noise accuracy value, γ k is the signal accuracy value.
7. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 5, characterized in that: Step S42 includes the steps of: S421, combining the GS decomposition factors of the inverse matrix, using fast Fourier transform and inverse fast Fourier transform to solve the first vector: in, is the inverse matrix, is valid data; S422. Calculate a second vector using the first vector and the Fourier dictionary matrix: in, It means to do K-point inverse fast Fourier transform on the vector in brackets. is the Fourier dictionary matrix, is the first vector; S423. Calculate the mean value by combining the second vector and the noise precision value: Among them, β is the noise accuracy value, Λ is calculated by 1 / γ k The diagonal matrix constructed in sequence, is the second vector.
8. The method for realizing segmented observation ISAR high-resolution imaging based on fast SBL algorithm according to claim 7, characterized in that: The convergence condition is: Among them, δ is the set convergence threshold.
9. A high-resolution ISAR imaging device based on a fast SBL algorithm for segmented observation, characterized in that: include: A segmented observation data model building module is used to model the segmented observation data according to valid sampling samples and missing sampling samples in the segmented observation data to obtain a segmented observation data model; The inverse matrix calculation module is used to calculate the inverse matrix to be calculated in the process of iterating the sparse signal by the sparse Bayesian learning algorithm using fast Fourier transform based on the Fourier dictionary matrix under the segmented observation data model, combined with the signal precision value and the noise precision value; A GS decomposition factor solving module is used to solve the GS decomposition factor of the inverse matrix by using the Levinson-Durbin iterative algorithm based on the inverse matrix to be solved; A module for solving the diagonal elements and mean of the covariance, for solving the mean of the posterior distribution and the diagonal elements of the covariance in the process of iterating the sparse signal by the sparse Bayesian learning algorithm by using the fast Fourier algorithm in combination with the GS decomposition factor of the inverse matrix; A judgment module is used to calculate a new signal accuracy value and a new noise accuracy value based on the diagonal elements of the covariance and the mean when it is judged that the mean does not meet the convergence condition, and return the inverse matrix calculation module to be calculated; when it is judged that the mean meets the convergence condition, output the mean.
