An adaptive blind separation method, program, and system for radar main lobe composite interference signals

Abstract

This invention belongs to the field of radar array signal blind source separation technology, specifically involving an adaptive blind separation method, program, and system for radar main lobe composite interference signals. This invention constructs a robust complex whitening matrix using a non-zero time delay covariance matrix, reducing the complete dependence on the zero time delay covariance matrix during whitening. This reduces the impact of spatial colored noise on the whitening matrix and improves the stability of blind separation processing under spatial colored noise conditions. This invention uses the number of time delay covariance matrices participating in joint diagonalization as an adaptive selection object, enabling the determination of the target time delay matrix number based on changes in the second-order correlation structure under different composite interference scenarios, avoiding scenario mismatch problems caused by fixed number settings. This invention combines robust complex whitening, adaptive time delay matrix number selection, and joint diagonalization demixing processing into a unified second-order blind identification process, enabling the separation of radar main lobe composite interference observation signals into clearer signal components.

Description

Technical Field

[0001] This invention belongs to the field of radar array signal blind source separation technology, specifically relating to an adaptive blind separation method, program and system for radar main lobe composite interference signals. Background Technology

[0002] In radar main lobe interference scenarios, different interference components have similar incident directions and overlapping time-frequency distributions. Relying solely on angle resolution, energy detection, or overall category identification makes it difficult to obtain the independent waveforms of each individual interference source. If the receiver cannot effectively separate the composite interference at the signal level, subsequent interference type identification, parameter estimation, and targeted suppression processing will all be affected. Therefore, how to separate the superimposed suppressive interference, deceptive interference, and related echo components from the mixed observation signals received by the array is a crucial problem that needs to be solved in radar main lobe anti-interference processing.

[0003] In existing radar composite jamming processing methods, one type focuses on identifying jamming patterns from time-frequency maps, image features, or tag information. While this type of method can obtain certain category judgment results, its output is usually not a separation of the source signal components. When subsequent processing requires analysis of the waveform, time delay, frequency, or modulation characteristics of each jamming component, further signal-level separation is still necessary. Another type of method employs the idea of ​​blind source separation, which can recover several source signal components from multi-channel observed signals with limited prior information. Second-order blind source separation methods utilize the second-order time correlation structure of the source signals, estimating the unmixing matrix through whitening processing and joint diagonalization of multiple time delay covariance matrices. This type of method does not rely on higher-order non-Gaussianity assumptions and is suitable for radar jamming signals with time correlation structures.

[0004] However, existing SOBI-like methods still have shortcomings in separating composite interference from radar main lobe. Firstly, SOBI typically performs whitening based on the zero-delay covariance matrix and requires the equivalent mixing matrix after whitening to maintain an approximately unitary structure so that subsequent source signal separation can be achieved through joint diagonalization of multiple delay covariance matrices. In actual array receiving channels, factors such as channel coupling and non-ideal characteristics of the RF front-end introduce spatial colored noise. The covariance matrix of this type of noise is generally not in the proportional form of an identity matrix and will enter the zero-delay covariance estimation result, affecting the whitening matrix with the noise covariance structure and thus disrupting the approximately unitary structure of the mixed matrix after whitening. The resulting off-diagonal perturbation will propagate to the subsequent joint diagonalization process of the delay covariance matrices, resulting in residual mixing in the separation result and reducing the performance of composite interference separation. Secondly, the number of delay covariance matrices participating in joint diagonalization in SOBI usually needs to be manually set. Under different main lobe composite interference scenarios, the second-order correlation structure, correlation attenuation rate, and time-frequency overlap of suppression interference and deception interference are not the same. If too few time delay covariance matrices are selected, the time delay correlation differences between different interference sources cannot be fully utilized; if too many are selected, noise-dominated or weakly correlated time delay matrices are introduced, interfering with the joint diagonalization target and increasing the residual mixing risk. Therefore, fixing or manually setting the number of time delay matrices is difficult to adapt to changes in the second-order correlation structure under different composite interference scenarios.

[0005] In summary, existing technologies for separating radar mainlobe composite interference signals still suffer from problems such as insufficient whitening reliability under spatial colored noise conditions, lack of adaptive capability in selecting the time delay covariance matrix, and decreased separation stability when the composite interference scenario changes. Therefore, it is necessary to propose an adaptive blind separation method and system for radar mainlobe composite interference signals. This method can improve the robustness of whitening processing under spatial colored noise and changing composite interference scenarios, and adaptively determine the number of time delay covariance matrices participating in joint diagonalization, thereby achieving effective separation of composite interference components. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of signal-level separation of radar main lobe composite interference in the prior art, and to provide an adaptive blind separation method, program and system for radar main lobe composite interference signals.

[0007] An adaptive blind separation method for radar main lobe composite interference signals includes the following steps:

[0008] Obtain the composite interference observation signal vector received by the radar array; arrange the composite interference observation signal vectors at all sampling times in sequence to construct the array observation matrix;

[0009] Construct a set of non-zero delays for robust whitening. Calculate the autocorrelation matrix corresponding to each non-zero delay based on the composite interference observation signal vector at each sampling time, and perform weighted summation to obtain the robust covariance matrix.

[0010] The robust covariance matrix is ​​decomposed into eigenvalues. Based on the principal eigenvector matrix and the principal eigenvalue diagonal matrix obtained from the decomposition, a robust complex whitening matrix is ​​constructed. The array observation matrix is ​​then whitened using the robust complex whitening matrix to obtain the whitened observation matrix, and the whitened observation signal vector at each sampling time is determined.

[0011] Construct a candidate non-zero delay set for joint diagonalization. Calculate the covariance matrix corresponding to each candidate non-zero delay based on the whitened observation signal vector at each sampling time and perform normalization. Arrange the normalized covariance matrices corresponding to all candidate non-zero delays in sequence to construct a candidate delay correlation matrix.

[0012] Determine the optimal delay quantity, and select a target subset from the candidate delay correlation matrix based on the optimal delay quantity; perform joint diagonalization on the target subset, and perform rotation transformation on the target subset through the unitary matrix, with the optimization objective being that the transformed matrix is ​​close to the diagonal matrix; obtain the optimal unitary matrix by solving the optimization problem.

[0013] The product of the conjugate transpose of the optimal unitary matrix and the robust complex whitening matrix is ​​used to construct the overall unmixing matrix. The product of the overall unmixing matrix and the array observation matrix is ​​then calculated to obtain the blind separation result of the radar array receiving the composite interference observation signal.

[0014] Furthermore, the construction of the non-zero latency set for robust whitening Based on the composite interference observation signal vector at each sampling time Calculate the autocorrelation matrix corresponding to each non-zero delay. ;

[0015]

[0016] in, Non-zero latency The corresponding covariance matrix, , , For the non-zero delay set used for robust whitening The total number of zero-latency connections between China and Africa; The first received by the radar array Composite interference observation signal vector at sampling time, , The total number of sampling times. This represents the conjugate transpose operation;

[0017] Perform a weighted summation to obtain the robust covariance matrix. ;

[0018]

[0019] in, The summation index is used to represent sets with non-zero delays. The Middle A non-zero latency, ; Non-zero latency The weights satisfy , .

[0020] Furthermore, the non-zero latency weight The setup method is as follows:

[0021] All non-zero delays The corresponding covariance matrix Arranged sequentially, they form a time delay correlation matrix. ;

[0022]

[0023] For the time delay correlation matrix Perform singular value decomposition and select the number of source signals to be separated. The corresponding principal left singular vectors yield the signal subspace basis. Calculate each non-zero delay The corresponding complex field projection matrix ;

[0024]

[0025] Set each non-zero delay weight , make the set Non-zero latency in each region The corresponding complex field projection matrix The combined matrix obtained by weighted summation It is a positive definite matrix. .

[0026] Furthermore, the robust covariance matrix Perform eigenvalue decomposition:

[0027]

[0028] Based on the principal eigenvector matrix obtained from the decomposition diagonal matrix with principal eigenvalues Construct a robust complex whitening matrix :

[0029]

[0030] Based on robust complex whitening matrix For array observation matrix Whitening is performed to obtain the whitening observation matrix. This allows for the determination of the whitened observation signal vector at each sampling time. :

[0031]

[0032] in, , .

[0033] Furthermore, the construction of the candidate non-zero delay set for joint diagonalization Based on the whitening observation signal vector at each sampling time Calculate the non-zero delay for each candidate The corresponding covariance matrix :

[0034]

[0035] in, , For the candidate non-zero delay set used for joint diagonalization The total number of candidate non-zero latency;

[0036] Perform normalization:

[0037]

[0038] in, This indicates the calculation of the Frobenius norm; Used to prevent the denominator from being 0;

[0039] All candidate non-zero latency The corresponding normalized covariance matrix Arranged sequentially, they form a candidate time delay correlation matrix. , .

[0040] Furthermore, the determination of the optimal delay quantity Based on the optimal number of delays, from the candidate delay correlation matrix Select target subset :

[0041]

[0042] For the target subset Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ;

[0043]

[0044] in, The unitary transformation matrix to be optimized during the joint diagonalization process. for A set of unitary matrices The number of source signals to be separated; This means setting the diagonal elements of the matrix to zero and retaining the off-diagonal elements; This represents the Frobenius norm, used to calculate the square root of the sum of the squares of the moduli of a matrix.

[0045] The optimal unitary matrix The conjugate transpose and robust complex whitening matrix The product of these components forms the overall unmixing matrix. ;

[0046]

[0047] Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ;

[0048]

[0049] in, , For the first Blind separation results of the source signal ; , For the first The source signal is at the Blind separation results at the sampling time.

[0050] Furthermore, the determination of the optimal delay quantity The method is as follows:

[0051] Based on Markov decision processes, a maximum number of iterations is set. With convergence threshold Build latency Corresponding separation quality evaluation function , This represents the number of iterations.

[0052] In each iteration, based on the amount of delay in the current iteration... From the candidate time delay correlation matrix Selecting a subset ;

[0053]

[0054] Based on the array observation matrix subset Number of delays in the current iteration Separation quality evaluation in the current iteration Historical best latency and historical best separation quality evaluation Construct decision state vector ;

[0055] Decision state vector The input is fed into a pre-trained action network to obtain actions. ;

[0056] According to the action Update latency ;

[0057]

[0058] in, Indicates will Limited to 1 to between;

[0059] Based on the updated latency Calculate the separation quality evaluation quantity ;

[0060] If the separation quality evaluation quantity Greater than the historical best separation quality evaluation metric ,Right now Then update the historical best latency number. Separation quality evaluation metric from historical best ;

[0061] Otherwise, do not update the historical best latency quantity and the historical best separation quality evaluation quantity, i.e. , ;

[0062] If the maximum number of iterations is reached, that is , or continuous The best latency in the next iteration If none of the above steps are updated, stop the iteration and output the optimal latency. .

[0063] Furthermore, the number of delays Corresponding separation quality evaluation function Specifically:

[0064] Pair subset Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ;

[0065]

[0066] The optimal unitary matrix The conjugate transpose and robust complex whitening matrix The product of these products forms the overall unmixing matrix. ;

[0067]

[0068] Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ;

[0069]

[0070] in, , To determine the amount of delay The obtained number Blind separation results of the source signal;

[0071] Blind separation results of each source signal Perform a short-time Fourier transform and take the amplitude to obtain the time-frequency amplitude spectrum of each source signal. ;

[0072] Set sorting threshold For each source signal, the time-frequency amplitude spectrum Sort the amplitude values ​​in descending order and take the first few. The amplitude value at the boundary is used as a high-energy threshold. ;

[0073] Based on high energy threshold Time-frequency amplitude spectrum of the source signal High-energy thresholding is performed to obtain a sparse mask of the source signal. ;

[0074]

[0075] in, To determine the amount of delay The obtained number Time-frequency amplitude spectrum of the source signal median coordinate The amplitude value at that point; To determine the amount of delay The obtained number Sparse mask of the source signal median coordinate The amplitude value at that point;

[0076] Based on the sparse mask of each source signal Calculate the number of delays Corresponding separation quality evaluation quantity :

[0077]

[0078] in, For the first Source signal and the first The time-frequency crosstalk coefficient of the source signal;

[0079]

[0080] in, , ; This indicates the calculation of the matrix inner product; A regularization constant is introduced to prevent the denominator from being zero or too small, satisfying the following conditions: .

[0081] A computer program product includes computer instructions that, when executed by a processor, implement the steps of the aforementioned adaptive blind separation method for radar mainlobe composite interference signals.

[0082] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described adaptive blind separation method for radar mainlobe composite interference signals.

[0083] The beneficial effects of this invention are as follows:

[0084] This invention constructs a robust complex whitening matrix using a non-zero time-delay covariance matrix, reducing the complete dependence on the zero-time-delay covariance matrix during whitening. This reduces the impact of spatial colored noise on the whitening matrix and improves the stability of blind separation processing under spatial colored noise conditions. The invention uses the number of time-delay covariance matrices participating in joint diagonalization as an adaptive selection object, determining the target time-delay matrix number based on changes in the second-order correlation structure under different composite interference scenarios, avoiding scenario mismatch problems caused by fixed number settings. This invention reduces the adverse effects of noise-dominated, weakly correlated, or statistically redundant time-delay covariance matrices on the joint diagonalization process, thereby reducing the residual mixing risk in the separation results. This invention combines robust complex whitening, adaptive time-delay matrix number selection, and joint diagonalization demixing processing into a unified second-order blind identification process, enabling the separation of radar mainlobe composite interference observation signals into clearer signal components, providing more reliable input for subsequent interference identification, interference suppression, parameter estimation, and target detection. Attached Figure Description

[0085] Figure 1 This is a diagram of the overall architecture of the present invention.

[0086] Figure 2 This is a flowchart illustrating the adaptive selection of the number of time delay covariance matrices in this invention.

[0087] Figure 3 This is a schematic diagram of an adaptive blind separation system for radar main lobe composite interference signals in an embodiment of the present invention.

[0088] Figure 4 The figure shows the simulation verification results of the selection of the number of joint diagonals of the time delay covariance matrix in different composite interference scenarios in the embodiments of the present invention.

[0089] Figure 5 This is a schematic diagram illustrating the simulation verification results of the echo construction of the radar main lobe receiving composite interference signals under different interference-to-noise ratio conditions in an embodiment of the present invention.

[0090] Figure 6 This is a comparison chart showing the effect of various algorithms in decomposing composite signals into single signals under different interference-to-noise ratios in the embodiments of the present invention.

[0091] Figure 7 This is a performance comparison chart of the method under different interference-to-noise ratios and different composite interference signal scenarios in the embodiments of the present invention.

[0092] Figure 8 This is a schematic diagram of ablation verification results in an embodiment of the present invention, showing the adaptive selection of the number of diagonals for robust whitening and time delay covariance matrix. Detailed Implementation

[0093] The present invention will now be further described with reference to the accompanying drawings.

[0094] This invention relates to the fields of blind source separation of radar array signals, second-order statistical signal processing, and adaptive intelligent optimization. The adaptive blind source separation method for radar main lobe composite interference signals provided by this invention includes the following steps:

[0095] Step 1: Obtain the composite interference observation signal vector received by the radar array at each sampling time. Construct the array observation matrix Determine the number of source signals to be separated. ;

[0096] in, This is the index of the sampling time. , This represents the total number of sampling times.

[0097] Step 2: Construct a non-zero latency set for robust whitening Based on the composite interference observation signal vector at each sampling time Calculate each non-zero delay The corresponding covariance matrix ;

[0098]

[0099] in, , For set The total number of zero-latency connections between China and Africa; Indicates conjugate transpose;

[0100] Based on each non-zero delay The corresponding covariance matrix Construct each non-zero latency The corresponding autocorrelation matrix ;

[0101]

[0102] According to the set Non-zero latency in each region The corresponding covariance matrix Construct a time delay correlation matrix ;

[0103]

[0104] Step 3: Analyze the time delay correlation matrix Perform singular value decomposition and select the number of source signals to be separated. The corresponding principal left singular vectors yield the signal subspace basis. ;

[0105] All non-zero delays The corresponding covariance matrix Projecting onto the signal subspace and taking the Hermitian portion yields the non-zero time delays. The corresponding complex field projection matrix ;

[0106]

[0107] Step 4: Construct each non-zero latency weight , make the set Non-zero latency in each region The corresponding complex field projection matrix The combined matrix obtained by weighted summation It is a positive definite matrix;

[0108] , , ,

[0109] in, express The smallest eigenvalue is greater than 0, therefore It is a positive definite matrix;

[0110] Based on each non-zero delay weight For sets Non-zero latency in each region The corresponding autocorrelation matrix Weighted summation yields the robust covariance matrix. ;

[0111]

[0112] Step 5: For the robust covariance matrix Perform eigenvalue decomposition:

[0113]

[0114] Based on the principal eigenvector matrix after eigenvalue decomposition diagonal matrix with principal eigenvalues Construct a robust complex whitening matrix ;

[0115]

[0116] Based on robust complex whitening matrix For array observation matrix Whitening is performed to obtain the whitening observation matrix. This allows for the determination of the whitened observation signal vector at each sampling time.

[0117]

[0118]

[0119] Step 6: Construct a candidate non-zero delay set for joint diagonalization Based on the whitening observation signal vector at each sampling time Calculate the non-zero delay for each candidate The corresponding covariance matrix ;

[0120]

[0121] in, , For set The total number of candidate non-zero latency;

[0122] For each candidate, there is a non-zero latency The corresponding covariance matrix Perform normalization to obtain non-zero latency for each candidate. The corresponding normalized covariance matrix ;

[0123]

[0124] in, This indicates the calculation of the Frobenius norm; Used to prevent the denominator from being 0;

[0125] Based on the non-zero delay of each candidate The corresponding normalized covariance matrix Construct candidate delay correlation matrix ;

[0126]

[0127] Step 7: Based on the Markov decision process, according to the array observation matrix Correlation matrix with candidate delay Determine the optimal delay amount ;

[0128] Step 7.1: Set the maximum number of iterations With convergence threshold ;

[0129] Initialize the number of iterations Initialization delay quantity Historical best latency Separation quality evaluation quantity Historical best separation quality evaluation metric ;

[0130] The delay amount Corresponding separation quality evaluation quantity The calculation method is as follows:

[0131] Step 7.1.1: Based on the number of delays From the candidate time delay correlation matrix Selecting a subset ;

[0132] Step 7.1.2: For subsets Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ;

[0133]

[0134] in, for A set of unitary matrices; This means setting the diagonal elements of the matrix to zero and retaining the off-diagonal elements; This indicates the calculation of the Frobenius norm;

[0135] Step 7.1.3: Calculate the optimal unitary matrix. The conjugate transpose and robust complex whitening matrix The product of these products forms the overall unmixing matrix. ;

[0136]

[0137] Step 7.1.4: Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ;

[0138]

[0139]

[0140] in, , To determine the amount of delay The obtained number Blind separation results of the source signal ; , To determine the amount of delay The obtained number The source signal is at the The value at the sampling time;

[0141] Step 7.1.5: Blind separation results of each source signal Perform a short-time Fourier transform and take the amplitude to obtain the time-frequency amplitude spectrum of each source signal. ;

[0142] Set sorting threshold For each source signal, the time-frequency amplitude spectrum Sort the amplitude values ​​in descending order and take the first few. The amplitude value at the boundary is used as a high-energy threshold. Based on the high energy threshold Time-frequency amplitude spectrum of the source signal High-energy thresholding is performed to obtain a sparse mask of the source signal. ;

[0143]

[0144] in, To determine the amount of delay The obtained number Time-frequency amplitude spectrum of the source signal median coordinate The amplitude value at that point; To determine the amount of delay The obtained number Sparse mask of the source signal median coordinate The amplitude value at that point;

[0145] Step 7.1.6: Based on the sparse mask of each source signal Calculate the number of delays Corresponding separation quality evaluation quantity :

[0146]

[0147] in, For the first Source signal and the first The time-frequency crosstalk coefficient of the source signal;

[0148]

[0149] in, , ; This indicates the calculation of the matrix inner product;

[0150] Step 7.2: Based on the array observation matrix Candidate delay correlation matrix Number of delays in the current iteration Separate quality evaluation metrics in the current iteration Historical best latency and historical best separation quality evaluation metric Construct decision state vector ;

[0151] Decision state vector The input is fed into a pre-trained action network to obtain actions. ;

[0152] According to the action Update latency ;

[0153]

[0154] in, Indicates will Limited to 1 to between;

[0155] Step 7.3: Based on the updated latency quantity Calculate the separation quality evaluation quantity ;

[0156] If the separation quality evaluation quantity Greater than the historical best separation quality evaluation metric ,Right now Then update the historical best latency number. Separation quality evaluation metric from historical best ;

[0157] Otherwise, do not update the historical best latency quantity and the historical best separation quality evaluation quantity, i.e. , ;

[0158] Step 7.4: If , or continuous The best latency in the next iteration If none of the above steps are updated, stop the iteration and output the optimal latency. ;

[0159] Otherwise, let Return to step 7.2;

[0160] Step 8: Based on the optimal latency amount Obtain the blind separation results of the radar array receiving composite interference observation signals. The method is the same as steps 7.1.1 to 7.1.4.

[0161] Example 1:

[0162] Reference Figure 1 In one embodiment, the present invention provides an adaptive blind separation method for composite interference signals on the main lobe of radar. This method takes the composite interference observation signal received by the radar array as input and sequentially completes array observation modeling, determination of the number of components to be separated, construction of a non-zero delay covariance matrix, robust complex whitening, construction of the whitened delay covariance matrix, adaptive selection of the target delay quantity, joint diagonalization, and demixing output.

[0163] Step S1: Acquire the composite interference observation signal received by the radar array. In one embodiment, the radar array includes L receiving channels, acquiring N fast-time sampling point data within the same observation window. The array received observation vector at the nth sampling time is represented as:

[0164]

[0165] In the formula, For array receiving observation vectors, It is an array of mixed matrices. Let M be the source signal vector to be separated, and M be the number of components to be separated. The noise vector takes into account environmental noise and spatial correlation noise between radar array elements. The source signal to be separated may include target echo components, suppression jamming components, deception jamming components, or combinations thereof forming the signal component to be separated.

[0166] In one alternative implementation, the array blending matrix consists of multiple array guiding vectors:

[0167]

[0168] In the formula, The incident direction corresponding to the m-th source signal to be separated The array steering vector, M represents the number of source signals to be separated.

[0169] For a uniform linear array, the array steering vector can be expressed as:

[0170]

[0171] In the formula, Let be the incident direction of the m-th source signal to be separated, and d be the element spacing. Where L is the signal wavelength, L is the number of array elements, and j is the imaginary unit.

[0172] Arrange the N snapshots column-wise to obtain the array observation matrix:

[0173] ,

[0174] The output of step S1 is the composite interference observation matrix. The technical role of this step is to unify the radar composite interference reception problem into an array instantaneous linear mixing problem, providing a normalized input for subsequent blind source separation based on second-order statistics.

[0175] Step S2: Determine the number of components to be separated based on the observation signals received by the radar array. In one embodiment, this is based on the array observation matrix. The sample covariance matrix, effective eigenvalue distribution, information criterion, or source number estimation results are used to determine the number of components M to be separated. The number of components M is used to determine the signal subspace dimension in subsequent robust complex whitening processing, and also to determine the dimensions of the joint diagonalization matrix and the unmixing matrix.

[0176] Step S3: Construct a non-zero delay covariance matrix for robust whitening from the original array observation signals. Let the set of non-zero delays used for robust whitening be:

[0177]

[0178] in, , This represents the number of non-zero delays used in constructing the whitening matrix. For each The non-zero time delay covariance matrix is ​​directly calculated from the original array observation signals:

[0179]

[0180] In the formula, For the original observed signal with a non-zero time delay The sample delay covariance matrix. This matrix is ​​derived from the array received observation vectors before whitening. A construct is used to reduce the dependence on the zero-delay covariance matrix.

[0181] Step S4, based on Construct a robust complex whitening matrix and obtain whitening observation data.

[0182] In one implementation, multiple original non-zero time delay covariance matrices are concatenated column-wise to obtain a time delay correlation matrix:

[0183]

[0184] For the time delay correlation matrix Perform singular value decomposition and select the number of source signals to be separated. The corresponding principal left singular vectors yield the signal subspace basis. .

[0185] For each Projecting it onto the signal subspace and taking the Hermitian part yields the complex domain projection matrix:

[0186]

[0187] In the formula, For the first Several projection matrices are used. Hermitianization improves the stability of the complex-domain covariance combination. Positive definite combination iterations are performed on each complex-domain projection matrix to obtain the combination matrix:

[0188]

[0189] In the formula, It is a combined matrix obtained by weighting multiple complex field projection matrices;

[0190] Based on each non-zero delay weight We obtain the robust covariance matrix by weighted summation of the autocorrelation matrices of samples with strictly non-zero time delays:

[0191]

[0192] Robust covariance matrix It can reduce the impact of spatial colored noise on the zero-delay covariance matrix;

[0193] For robust covariance matrix Perform eigenvalue decomposition:

[0194]

[0195] In the formula, The selected principal eigenvector matrix, The corresponding principal eigenvalue diagonal matrix. Construct the robust complex whitening matrix. :

[0196]

[0197] Based on robust complex whitening matrix For array observation matrix Whitening is performed to obtain the whitening observation matrix. This allows for the determination of the whitened observation signal vector at each sampling time. :

[0198]

[0199] in, .

[0200] Step S5: Construct a set of candidate delay covariance matrices for joint diagonalization based on the whitened observation data. Let the set of candidate non-zero delays for joint diagonalization be:

[0201]

[0202] in, , For the candidate non-zero delay set used for joint diagonalization The total number of candidate non-zero latency;

[0203] For each candidate delay Based on whitening observation data Construct the whitened time delay covariance matrix:

[0204]

[0205] This matrix is ​​different from the one in step S3. : Constructed from the original observation signal and used for whitening, Constructed from whitening observation data and used for joint diagonalization.

[0206] To reduce the difference in amplitude between different time delay matrices, Normalize:

[0207]

[0208] This yields the candidate set:

[0209]

[0210] The output of step S5 is a set of candidate delay covariance matrices, sorted in ascending order of non-zero delay and normalized. The technical purpose of this step is to provide a uniformly scaled set of candidate second-order statistical matrices for subsequent selection of the target delay quantity and joint diagonalization.

[0211] Step S6: Using the trained latency selection model, determine the optimal latency online. .

[0212] In one implementation, the number of whitened time-delay covariance matrices participating in joint diagonalization is denoted as... Since the second-order correlation structure differs under different composite interference scenarios, this invention does not... Instead of fixing it to an experience value, it will be... The choice is modeled as a finite-step adaptive decision process.

[0213] Set the maximum number of iterations. With convergence threshold Build latency Corresponding separation quality evaluation function , This represents the number of iterations.

[0214] In each iteration, based on the amount of delay in the current iteration... From the candidate time delay correlation matrix Selecting a subset ;

[0215]

[0216] Based on the array observation matrix subset Number of delays in the current iteration Separation quality evaluation in the current iteration Historical best latency and historical best separation quality evaluation Construct decision state vector ;

[0217] Decision state vector The input is fed into a pre-trained action network to obtain actions. ;

[0218] According to the action Update latency ;

[0219]

[0220] in, Indicates will Limited to 1 to between;

[0221] Based on the updated latency Calculate the separation quality evaluation quantity ;

[0222] If the separation quality evaluation quantity Greater than the historical best separation quality evaluation metric ,Right now Then update the historical best latency number. Separation quality evaluation metric from historical best ;

[0223] Otherwise, do not update the historical best latency quantity and the historical best separation quality evaluation quantity, i.e. , ;

[0224] If the maximum number of iterations is reached, that is , or continuous The best latency in the next iteration If none of the above steps are updated, stop the iteration and output the optimal latency. .

[0225] In one embodiment of the present invention, the action set can be represented as:

[0226]

[0227] For the updated latency A subset of time delay covariance matrices participating in the trial evaluation is determined from the set of candidate whitened time delay covariance matrices. Based on this subset, joint diagonalization and separation quality evaluation are performed to obtain the corresponding quality evaluation value. .like If the quality rating is better than the historical best, then it will be Update to the historical best latency value; otherwise, keep the historical best latency value unchanged and record the number of consecutive times there is no improvement.

[0228] When the latency selection model outputs a stop action, or when the number of consecutive times without improvement reaches a preset threshold, the search process ends, and the historical best latency is determined as the target latency. .

[0229] In one implementation, the latency selection model is obtained through offline training. (See reference...) Figure 2 During offline training, training samples are constructed with different composite interference scenarios, different noise conditions, and different numbers of candidate delays. A reward value is then constructed based on the separation results corresponding to the number of candidate delays to update the delay selection model. In training samples with a reference source signal, the reward value may further include a separation performance evaluation term calculated based on the reference source signal. The reference source signal is only used for reward construction during the offline training phase; no real source signal or reference interference component needs to be input during the online application phase.

[0230] Step S7, based on the optimal number of delays, from the candidate delay correlation matrix Select target subset :

[0231]

[0232] For the target subset Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ;

[0233]

[0234] in, for A set of unitary matrices The number of source signals to be separated; This means setting the diagonal elements of the matrix to zero and retaining the off-diagonal elements; This refers to the unitary matrix to be optimized during the joint diagonalization process;

[0235] The optimal unitary matrix The conjugate transpose and robust complex whitening matrix The product of these components forms the overall unmixing matrix. ;

[0236]

[0237] in, ;

[0238] Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ;

[0239]

[0240] in, , For the first Blind separation results of the source signal ; , For the first The source signal is at the Blind separation results at the sampling time.

[0241] Through the above steps, this invention utilizes the non-zero time delay covariance matrix of the original observed signal during the whitening stage. Construct a robust complex whitening matrix and utilize the whitening delay covariance matrix during the joint diagonalization stage. The optimal latency is determined online using a trained policy model. This enables adaptive blind separation under varying conditions of spatial colored noise and complex interference.

[0242] Reference Figure 3 This application also proposes an adaptive blind separation system for radar mainlobe composite interference signals. The system includes an observation signal acquisition module, a module for determining the number of components to be separated, a module for constructing a non-zero time delay covariance matrix, a robust complex whitening module, a candidate time delay covariance matrix construction module, an adaptive decision-making module, a joint diagonal demixing module, and a separated signal output module.

[0243] The observation signal acquisition module is used to execute step S1, acquire the radar array received observation signals, and represent the radar array received observation signals as array observation data. The component number determination module is used to execute step S2, and determine the number of components to be separated based on the array observation data. The original non-zero delay covariance matrix construction module is used to execute step S3, and construct multiple non-zero delay covariance matrices based on the array observation data to form a set of original non-zero delay covariance matrices.

[0244] The robust complex whitening module executes step S4, constructing a robust complex whitening matrix based on the original set of non-zero delay covariance matrices, and then using the robust complex whitening matrix to whiten the array observation data, obtaining whitened observation data. The candidate whitened delay covariance matrix construction module executes step S5, constructing multiple whitened delay covariance matrices based on the whitened observation data, forming a set of candidate whitened delay covariance matrices.

[0245] The adaptive delay quantity selection module performs step S6 by inputting the relevant features of the current composite interference scenario and the candidate whitened delay covariance matrices into the pre-trained delay quantity selection model. The delay quantity selection model adaptively determines the target delay quantity participating in joint diagonalization. The joint diagonalization module performs step S7 by determining the target delay covariance matrix set from the candidate whitened delay covariance matrix set based on the target delay quantity, and performing joint diagonalization on the target delay covariance matrix set to obtain the diagonalized matrix. The demixing output module performs step S8 by constructing the overall demixing matrix based on the diagonalized matrix and the robust complex whitening matrix, and using the overall demixing matrix to perform blind separation of the array observation data, outputting the separated radar interference signal components.

[0246] It should be noted that the functional modules in the system can be implemented in software, hardware, or a combination of software and hardware. The names of each module are used to represent their functional divisions, and there is no requirement that each module must be a physically independent hardware component.

[0247] Furthermore, to facilitate understanding of the technical effects of the present invention, the effectiveness of the method of the present invention can be verified without changing steps S1 to S8. The effectiveness verification can be implemented using simulation, and is only used to illustrate the technical effects of the method of the present invention, and does not constitute a limitation on radar system, array size, interference-to-noise ratio range, number of interference types, or number of candidate delays.

[0248] Reference Figure 4 By adopting the adaptive selection process of target delay quantity in step S6, the result of target delay quantity change under different composite interference scenarios can be obtained. Figure 4To clarify: the target delay quantity is not a fixed empirical parameter, but can be adjusted based on the scene description vector of the current composite interference observation signal, the candidate delay covariance matrix profile, the dynamic search state, and the separation quality feedback. When the noise intensity, composite interference structure, or finite sample estimation stability changes, the number of effective delay covariance matrices participating in the joint diagonalization also changes. Therefore, Figure 4 This can help explain the necessity of introducing the reinforcement learning decision-making process to select the target delay in step S6.

[0249] Reference Figures 5 to 8 This patent presents an adaptive blind separation method (PSOBI) for radar mainlobe composite interference signals. The results are compared with those of traditional second-order blind identification algorithms (SOBI), robust whitening second-order blind identification algorithms (CROBI), and tensor decomposition blind source separation (TD-BSS). The signal-to-interference ratio (SIR), minimum distance (MD) index, and Amari error are used as performance evaluation metrics to assess the separated signal output in step S8. The SIR characterizes the relative strength of the desired component to the crosstalk component in the estimated source signal, while the minimum distance index and Amari error characterize the deviation between the estimated unmixing matrix and the ideal separation state. Under different interference-to-noise ratios, the method can adaptively select the target delay based on the current composite interference scenario and, combined with robust whitening and joint diagonalization processing in the complex domain, obtain a separation output with good time-frequency separability. Figure 5 This is to construct the echo for receiving composite interference signals on the radar main lobe. Figure 6 To compare the performance of various algorithms in decomposing composite signals into single signals, it can be seen that the invented PSOBI method has a better separation effect than the other methods, and is most similar to the source signal. Figure 7 The performance of the proposed PSOBI method is compared and quantified under different interference-to-noise ratios and different composite interference signal scenarios. It can be seen that the proposed PSOBI method has better performance metrics compared to other methods. This figure is used to aid in understanding the beneficial effects of the invention and does not limit the invention to being verified solely through evaluation metrics.

[0250] Reference Figure 8 The effects of steps S4 and S6 can be explained using module ablation methods. These module ablation methods include: a second-order blind source separation method using standard whitening with a fixed delay (Base-SOBI); a method using standard whitening with adaptive selection of the target delay (SOBI+PPO-K); a method using complex-domain robust whitening with a fixed delay (RW-SOBI); and a method simultaneously using complex-domain robust whitening and adaptive selection of the target delay (PSOBI). Figure 6 For illustration: Overall, PSOBI achieves a high SIR under this simulation setting, while maintaining low AmariError and MD. The SIR curves show that Base-SOBI and SOBI+PPO-K offer limited improvement in the low to medium JNR range, indicating that when standard whitening exhibits mismatch, relying solely on PPO for adaptive adjustment is insufficient. It is difficult to sufficiently improve the separation results; when the whitening coordinates are already mismatched, only adjustment is possible. Unable to correct the source direction estimation, it instead selects a suboptimal delay quantity under erroneous state characteristics. In contrast, RW-SOBI outperforms Base-SOBI under most JNR conditions, indicating that robust whitening can mitigate whitening errors caused by correlated noise and provide a more reliable second-order statistical basis for joint diagonalization. Further comparison of RW-SOBI and PSOBI reveals that, based on the same robust whitening, PSOBI further improves its SIR and error metrics, indicating that when whitening quality is guaranteed, PPO-K can further select a more suitable joint diagonalization matrix quantity based on the delay-related structure of the current composite interference. In contrast, the error curve of SOBI+PPO-K is not superior to that of Base-SOBI, indicating that the effectiveness of PPO-K depends on a reliable whitening premise. Therefore, these results show that robust whitening is the foundation for performance improvement, and PPO-K is a key module for further releasing scene adaptability.

[0251] It should be noted that, Figures 4 to 8 These are all simulation effect verification diagrams of the technical effect of the method of the present invention, and are not necessary execution steps of the method of the present invention. The scope of protection of the present invention is defined by steps S1 to S8 and the system modules as specified in the claims, and is not subject to... Figures 4 to 8 The specific curve shape, numerical range, or comparison object shown are subject to limitations.

[0252] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive blind separation method for radar main lobe composite interference signals, characterized in that: Obtain the composite interference observation signal vector received by the radar array; arrange the composite interference observation signal vectors at all sampling times in sequence to construct the array observation matrix; Construct a set of non-zero delays for robust whitening. Calculate the autocorrelation matrix corresponding to each non-zero delay based on the composite interference observation signal vector at each sampling time, and perform weighted summation to obtain the robust covariance matrix. The robust covariance matrix is ​​decomposed into eigenvalues. Based on the principal eigenvector matrix and the principal eigenvalue diagonal matrix obtained from the decomposition, a robust complex whitening matrix is ​​constructed. The array observation matrix is ​​then whitened using the robust complex whitening matrix to obtain the whitened observation matrix, and the whitened observation signal vector at each sampling time is determined. Construct a candidate non-zero delay set for joint diagonalization, calculate the covariance matrix corresponding to each candidate non-zero delay based on the whitened observation signal vector at each sampling time, and perform normalization. Arrange the normalized covariance matrices corresponding to all candidate non-zero delays in sequence to construct the candidate delay correlation matrix; Determine the optimal delay quantity, and select the target subset from the candidate delay correlation matrix based on the optimal delay quantity; Perform joint diagonalization on the target subset, and rotate the target subset using a unitary matrix. The optimization objective is to make the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. The product of the conjugate transpose of the optimal unitary matrix and the robust complex whitening matrix is ​​used to construct the overall unmixing matrix. The product of the overall unmixing matrix and the array observation matrix is ​​then calculated to obtain the blind separation result of the radar array receiving the composite interference observation signal.

2. The adaptive blind separation method for radar main lobe composite interference signals according to claim 1, characterized in that: The construction of the non-zero latency set for robust whitening Based on the composite interference observation signal vector at each sampling time Calculate the autocorrelation matrix corresponding to each non-zero delay. ; in, Non-zero latency The corresponding covariance matrix, , , For the non-zero delay set used for robust whitening The total number of zero-latency connections between China and Africa; The first received by the radar array The composite interference observation signal vector at the sampling time, , The total number of sampling times. This represents the conjugate transpose operation; Perform a weighted summation to obtain the robust covariance matrix. ; in, The summation index is used to represent sets with non-zero delays. The Middle A non-zero latency, ; Non-zero latency The weights satisfy , .

3. The adaptive blind separation method for radar main lobe composite interference signals according to claim 2, characterized in that: The non-zero delay weight The setup method is as follows: All non-zero delays The corresponding covariance matrix Arranged sequentially, they form a time delay correlation matrix. ; For the time delay correlation matrix Perform singular value decomposition and select the number of source signals to be separated. The corresponding principal left singular vectors yield the signal subspace basis. Calculate each non-zero delay The corresponding complex field projection matrix ; Set each non-zero delay weight , make the set Non-zero latency in each region The corresponding complex field projection matrix The combined matrix obtained by weighted summation It is a positive definite matrix. .

4. The adaptive blind separation method for radar main lobe composite interference signals according to claim 2, characterized in that: The robust covariance matrix Perform eigenvalue decomposition: Based on the principal eigenvector matrix obtained from the decomposition diagonal matrix with principal eigenvalues Construct a robust complex whitening matrix : Based on robust complex whitening matrix For array observation matrix Whitening is performed to obtain the whitening observation matrix. This allows for the determination of the whitened observation signal vector at each sampling time. : in, , .

5. The adaptive blind separation method for radar main lobe composite interference signals according to claim 4, characterized in that: The construction of a candidate nonzero delay set for joint diagonalization Based on the whitening observation signal vector at each sampling time Calculate the non-zero delay for each candidate The corresponding covariance matrix : in, , For the candidate non-zero delay set used for joint diagonalization The total number of candidate non-zero latency; Perform normalization: in, This indicates the calculation of the Frobenius norm; Used to prevent the denominator from being 0; All candidate non-zero latency The corresponding normalized covariance matrix Arranged sequentially, they form a candidate time delay correlation matrix. , .

6. The adaptive blind separation method for radar main lobe composite interference signals according to claim 5, characterized in that: Determining the optimal delay amount Based on the optimal number of delays, from the candidate delay correlation matrix Select target subset : For the target subset Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ; in, The unitary transformation matrix to be optimized during the joint diagonalization process. for A set of unitary matrices The number of source signals to be separated; This means setting the diagonal elements of the matrix to zero and retaining the off-diagonal elements; This represents the Frobenius norm, used to calculate the square root of the sum of the squares of the moduli of a matrix. The optimal unitary matrix The conjugate transpose and robust complex whitening matrix The product of these components forms the overall unmixing matrix. ; Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ; in, , For the first Blind separation results of the source signal ; , For the first The source signal is at the Blind separation results at the sampling time.

7. The adaptive blind separation method for radar main lobe composite interference signals according to claim 6, characterized in that: Determining the optimal delay amount The method is as follows: Based on Markov decision processes, a maximum number of iterations is set. With convergence threshold Build latency Corresponding separation quality evaluation function , This represents the number of iterations. In each iteration, based on the amount of delay in the current iteration... From the candidate time delay correlation matrix Selecting a subset ; Based on the array observation matrix subset Number of delays in the current iteration Separation quality evaluation in the current iteration Historical best latency and historical best separation quality evaluation Construct decision state vector ; Decision state vector The input is fed into a pre-trained action network to obtain actions. ; According to the action Update latency ; in, Indicates will Limited to 1 to between; Based on the updated latency Calculate the separation quality evaluation quantity ; If the separation quality evaluation quantity Greater than the historical best separation quality evaluation metric ,Right now Then update the historical best latency number. Separation quality evaluation metric from historical best ; Otherwise, do not update the historical best latency quantity and the historical best separation quality evaluation quantity, i.e. , ; If the maximum number of iterations is reached, that is , or continuous The best latency in the next iteration If none of the above steps are updated, stop the iteration and output the optimal latency. .

8. The adaptive blind separation method for radar main lobe composite interference signals according to claim 7, characterized in that: The delay amount Corresponding separation quality evaluation function Specifically: Pair subset Perform joint diagonalization, rotate the target subset using a unitary matrix, and optimize the result by making the transformed matrix as close as possible to the diagonal matrix. The optimal unitary matrix is ​​obtained by solving the optimization problem. ; The optimal unitary matrix The conjugate transpose and robust complex whitening matrix The product of these products forms the overall unmixing matrix. ; Calculate the overall unmixing matrix With array observation matrix The product of these two factors yields the blind separation result of the radar array receiving the composite interference observation signal. ; in, , To determine the amount of delay The obtained number Blind separation results of the source signal; Blind separation results of each source signal Perform a short-time Fourier transform and take the amplitude to obtain the time-frequency amplitude spectrum of each source signal. ; Set sorting threshold For each source signal, the time-frequency amplitude spectrum Sort the amplitude values ​​in descending order and take the first one. The amplitude value at the boundary is used as a high-energy threshold. ; Based on high energy threshold Time-frequency amplitude spectrum of the source signal High-energy thresholding is performed to obtain a sparse mask of the source signal. ; in, To determine the amount of delay The obtained number Time-frequency amplitude spectrum of the source signal median coordinate The amplitude value at that point; To determine the amount of delay The obtained number Sparse mask of the source signal median coordinate The amplitude value at that point; Based on the sparse mask of each source signal Calculate the number of delays Corresponding separation quality evaluation quantity : in, For the first Source signal and the first The time-frequency crosstalk coefficient of the source signal; in, , ; This indicates the calculation of the matrix inner product; A regularization constant is introduced to prevent the denominator from being zero or too small, satisfying the following conditions: .

9. A computer program product comprising computer instructions, characterized in that: When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1 to 8.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

Images