Broadband signal multiband joint direction finding method based on support vector regression

By constructing array mathematical model and preprocessing of broadband signals, and building a support vector regression model after dimensionality reduction, the problem of training time growth in broadband signal direction finding is solved, and efficient and accurate multi-band joint direction finding performance is achieved.

CN119001594AActive Publication Date: 2024-11-22XIAN UNIV OF TECH

Patent Information

Application Number
CN202411284759.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-11-22
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

When existing direction finding of broadband signals using support vector regression, parallel input of multi-frequency data leads to a sudden increase in training time, especially when processing broadband or ultra-wideband signals, the training time increases exponentially, making it difficult to meet the demand for efficient direction finding.

Method used

By constructing an array mathematical model of broadband signals, performing discrete Fourier transform and data preprocessing, focusing on multi-band data to reference frequency points, building a support vector regression model after dimensionality reduction, training and application of this model for multi-band joint measurement of broadband signals Towards.

Benefits of technology

It realizes high-resolution, high-precision and high-efficiency broadband signal direction finding performance, overcomes the dependence of traditional methods on the model, and improves the robustness and training efficiency of the algorithm in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119001594A_ABST
    Figure CN119001594A_ABST
Patent Text Reader

Abstract

The invention discloses a broadband signal multiband joint direction finding method based on support vector regression. The broadband signal multiband joint direction finding method is specifically implemented according to the following steps: step 1, constructing an array mathematical model of the lth snapshot of a broadband signal; step 2, calculating array receiving data focused to a reference frequency point by each frequency point; step 3, performing data preprocessing on the reference frequency point array receiving data to obtain input features after dimension reduction; step 4, building an SVR model of broadband signal multiband joint direction finding; step 5, training the SVR model to obtain an SVR model trained by broadband signal multi-band joint direction finding; and step 6, inputting a test sample into the model trained in the step 5 to obtain a broadband signal multiband joint direction finding result. According to the invention, high-resolution, high-precision and high-efficiency broadband signal direction finding performance is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of radar technology, and in particular relates to a broadband signal multi-band joint direction finding method based on support vector regression. Background Art

[0002] With the development of modern radar communication technology, the application field of array direction finding system is constantly expanding, the signal environment is becoming more and more complex, and narrowband signals can no longer meet the actual application needs. In addition, in the process of signal transmission, since broadband signals have a relatively large bandwidth, they can carry more information, have a stronger anti-interference ability, and are more conducive to target detection, parameter estimation, and feature extraction. Therefore, broadband signals are increasingly used in radar and communication systems.

[0003] Array direction finding technology is an important technical means to achieve target positioning. A typical processing method in broadband signal direction finding technology is frequency domain processing. The broadband array received data is decomposed into narrowband components of multiple frequency points through discrete Fourier transform, and then processed using incoherent subspace algorithm or coherent subspace algorithm. Experiments have verified that the performance of the coherent subspace algorithm is better than that of the incoherent subspace algorithm, indicating that for broadband signals, multi-band joint estimation has better performance.

[0004] Support vector regression is a machine learning method, mainly a regression algorithm based on small sample statistical regression theory, suitable for constructing a functional mapping relationship between input signal features and arrival angles. It overcomes the dependence of traditional methods on models and is widely used in array direction finding. However, when using support vector regression to find the direction of broadband signals, there is a problem of parallel input of multi-frequency data, which increases the training time suddenly. Especially when processing broadband signals or ultra-wideband signals, the training time increases exponentially. Summary of the invention

[0005] The purpose of the present invention is to provide a wideband signal multi-band joint direction finding method based on support vector regression to achieve high-resolution, high-precision and high-efficiency wideband signal direction finding performance.

[0006] The technical solution adopted by the present invention is a broadband signal multi-band joint direction finding method based on support vector regression, which is specifically implemented according to the following steps:

[0007] Step 1, constructing an array mathematical model of the lth snapshot of the broadband signal;

[0008] Step 2, calculating the array receiving data of each frequency point focused on the reference frequency point;

[0009] Step 3: preprocess the reference frequency array received data to obtain input features after dimensionality reduction;

[0010] Step 4: Build an SVR model for broadband signal multi-band joint direction finding;

[0011] Step 5: training the SVR model to obtain a trained SVR model for broadband signal multi-band joint direction finding;

[0012] Step 6: Input the test sample into the model trained in step 5 to obtain the broadband signal multi-band joint direction finding result.

[0013] The present invention is also characterized in that:

[0014] Step 1 is implemented as follows:

[0015] Step 1.1: Calculate the echo signal x received by the mth array element according to the array structure of the uniform linear array and the position information of the spatial target relative to the array. m (t):

[0016]

[0017] Among them, s k (t) represents the kth broadband signal, k = 1, 2, ..., K, K represents the number of targets, τ m,k =(m-1)d sinθ k / c means signal s k (t) is the time delay difference when it reaches the mth array element relative to the reference array element, M represents the number of array elements, θ k represents the direction of the kth signal, c represents the propagation speed of electromagnetic waves, d represents the array element spacing, and n m (t) represents the Gaussian white noise on the mth array element, with a mean of 0 and a variance of

[0018] Step 1.2: Perform discrete Fourier transform on the array echo signal obtained in step 1.1 to obtain the frequency f under single snapshot j The mathematical model of the point array is:

[0019] x(f j )=A(f j ,θ)s(f j )+n(f j ),j=1,2,…,J

[0020] Among them, x(f j )=[x1(f j ),x2(f j ),…,x M (f j )] T , s(f j )=[s1(f j ),s2(f j),…,s K (f j )] T and n(f j )=[n1(f j ),n2(f j ),…,n M (f j )] T The frequencies are f j The discrete Fourier transform of the array echo signal, target signal and noise signal of the point is assumed to obey Gaussian distribution and the noise of each sub-band is independent of each other. J represents the number of sub-bands, (·) T represents the transposition operation, θ=[θ1,θ2,…,θ K ],A(f j ,θ)=[a(f j ,θ1),…,a(f j ,θ K )]∈C M×K is the frequency f j The array manifold matrix of points, is the angle θ k The steering vector of

[0021] Step 1.3: The single snapshot frequency f obtained in step 1.2 j The array mathematical model of the points is used to divide a whole period of time into L equally spaced segments, and then discrete Fourier transform is performed on each segment to obtain multiple snapshot data in the frequency domain. The array mathematical model of the lth snapshot is:

[0022] x l (f j )=A(f j ,θ)s l (f j )+n l (f j ),l=1,2,…,L

[0023] Among them, s l (f j ) is the frequency f j The signal vector of the lth snapshot, n l (f j ) is the frequency f j The noise vector of the lth snapshot.

[0024] Step 2 is implemented as follows:

[0025] Step 2.1: Based on the array mathematical model of the lth snapshot obtained in step 1, assuming that the signal is uncorrelated with the noise and that the noises of different channels are uncorrelated, the covariance matrix R corresponding to the jth frequency point is obtained:j for:

[0026]

[0027] Where E[·] represents the expected operation, is the signal covariance matrix, I represents the unit matrix, (·) H represents the conjugate transpose operation. In practice,

[0028] Step 2.2, set the lowest frequency point of the broadband signal as the reference frequency point f0;

[0029] Step 2.3: Design the focusing matrix T(f j ):

[0030] T(f j )=Q(f0)Q H (f j )

[0031] Among them, Q(f0), Q(f j ) are P(f0), P(f j ), is the covariance matrix of the array data after denoising at the reference frequency point f0, is the frequency f j The covariance matrix of the array data after denoising, the noise power and are the frequencies f0 and f j The average of the small eigenvalues ​​of the data covariance matrix at ;

[0032] Step 2.4: Use the focusing matrix T(f j ) The frequency f j The array receiving data at the reference frequency f0 is focused to obtain the focused data.

[0033]

[0034] Step 3 is implemented as follows:

[0035] Step 3.1: Calculate the covariance matrix R0 at the reference frequency point f0 after focusing based on the array data of each frequency point obtained in step 2.4:

[0036]

[0037] Step 3.2: Extract the upper triangular part of the covariance matrix R0 obtained in step 3.1 and recombine it to obtain the vector η:

[0038] η=[r 12 ,…,r 1M ,r 23 ,…,r 2M ,…,r (M-1)M ]

[0039] where r ij = R0(i,j) is the i-th row and j-th column element of R0;

[0040] Step 3.3: Normalize the vector η obtained in step 3.2 to obtain the reduced-dimensional input feature z0:

[0041]

[0042] Among them, ||·||2 represents the l2 norm of the vector, and Represents the real and imaginary part operations respectively.

[0043] Step 4 is implemented according to the following steps:

[0044] Step 4.1: Fit z0 and output angle according to the input features obtained in step 3.3 and support vector regression theory The function mapping relationship between them is:

[0045]

[0046] in, represents the estimation of θ, Φ(·) is the nonlinear mapping function, and {w,b} are the model parameters to be trained;

[0047] Step 4.2: Based on the function mapping relationship obtained in step 4.1, build an SVR model for broadband signal multi-band joint direction finding:

[0048] {

[0049] SVR1{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0050] SVR2{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0051] ···

[0052] SVR K {Kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; Penalty parameter: C}

[0053] }

[0054] Among them, the Gaussian kernel function Represents the pth sample data, p=1,2,…,P.

[0055] Step 5 is implemented according to the following steps:

[0056] Step 5.1: Generate a set of P angle labels for K targets And meet

[0057] Step 5.2: Use the angle label set Θ obtained in step 5.1 k Combine steps 1, 2, and 3 to generate a training sample set

[0058] Step 5.3: Construct optimization objective function:

[0059]

[0060] in, represents the loss function, ε is the error tolerance;

[0061] Step 5.4: The angle label set Θ obtained according to step 5.1 k and the sample set Z obtained in step 5.2 P , construct P sets of training data for the kth target in The pth group of training data represents the kth target angle, and the model training starts from k=1, p=1;

[0062] Step 5.5: Use the p-th group of training data of the k-th target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0063] Step 5.6: Use the p+1th group of training data of the kth target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0064] Step 5.7: Comparison function and Size:

[0065] like Then p=p+1, and repeat steps 5.5, 5.6 and 5.7;

[0066] like The model training of the k-th target angle is ended, and the optimal model parameter w corresponding to the k-th target angle is obtained. k ,b k ;

[0067] Step 5.8, let k = k + 1, and repeat steps 5.4, 5.5, 5.6 and 5.7 until k = K, all target models are trained, and the training is completed.

[0068] Step 6 is implemented according to the following steps:

[0069] Step 6.1, randomly generate target angle information θ, and combine steps 1, 2, and 3 to obtain the test sample z0;

[0070] Step 6.2: Input the test sample z0 obtained in step 6.1 into the model trained in step 5, and obtain the estimated angle of the kth target of the broadband signal:

[0071]

[0072] The beneficial effects of the present invention are as follows: compared with the existing broadband signal direction finding method, the broadband signal multi-band joint direction finding method based on support vector regression benefits from: 1) the method of the present invention is data-driven, which overcomes the traditional method's dependence on the model and improves the robustness of the algorithm in complex environments; 2) the method of the present invention focuses the multi-band data to a reference frequency point and performs preprocessing, which reduces the input dimension of the network model while achieving denoising and improves the training efficiency of the algorithm; 3) the method of the present invention establishes a support vector regression model for broadband signal multi-band joint estimation, achieves multi-band joint direction finding, and improves the estimation performance of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the implementation of the present invention;

[0074] FIG. 2( a ) is a diagram showing simulation results of multi-signal angle estimation using the method of the present invention;

[0075] FIG2( b ) is a diagram showing the error simulation results of multi-signal angle estimation using the method of the present invention;

[0076] FIG3(a) is a simulation result diagram of the root mean square error of multiple signals versus signal-to-noise ratio using the existing method and the method of the present invention;

[0077] FIG3( b ) is a simulation result diagram of the multi-signal root mean square error versus snapshot number curve using the existing method and the method of the present invention;

[0078] FIG4(a) is a simulation result diagram of the curve of the success resolution probability of multiple signals changing with the angular interval using the existing method and the method of the present invention;

[0079] FIG4( b ) is a simulation result diagram of the curve of the probability of successful resolution of multiple signals changing with the signal-to-noise ratio using the existing method and the method of the present invention;

[0080] Figure 5 The figure is a simulation result diagram of the curve of the running time of multi-signal training samples and the average running time of test samples changing with the number of sub-frequency bands using the existing method and the method of the present invention. DETAILED DESCRIPTION

[0081] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] The invention is based on the support vector regression broadband signal multi-band joint direction finding method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0083] Step 1: construct an array mathematical model of the lth snapshot of the broadband signal according to the array structure of the uniform linear array and the position information of the spatial target relative to the array;

[0084] Step 1 is implemented according to the following steps:

[0085] Step 1.1: Calculate the echo signal x received by the mth array element according to the array structure of the uniform linear array and the position information of the spatial target relative to the array. m (t):

[0086]

[0087] Among them, s k (t) represents the kth broadband signal, k = 1, 2, ..., K, K represents the number of targets, τ m,k =(m-1)d sinθ k / c means signal s k (t) is the time delay difference when it reaches the mth array element relative to the reference array element, M represents the number of array elements, θ k represents the direction of the kth signal, c represents the propagation speed of electromagnetic waves, d represents the array element spacing, and n m (t) represents the Gaussian white noise on the mth array element, with a mean of 0 and a variance of

[0088] Step 1.2: Perform discrete Fourier transform on the array echo signal obtained in step 1.1 to obtain the frequency f under single snapshot j The mathematical model of the point array is:

[0089] x(f j )=A(f j ,θ)s(f j )+n(f j ),j=1,2,…,J

[0090] Among them, x(f j )=[x1(f j ),x2(f j ),…,x M (f j )] T , s(f j )=[s1(f j ),s2(f j ),…,s K (f j )] T and n(f j )=[n1(f j ),n2(f j ),…,n M (f j )] T The frequencies are f j The discrete Fourier transform of the array echo signal, target signal and noise signal of the point is assumed to obey Gaussian distribution and the noise of each sub-band is independent of each other. J represents the number of sub-bands, (·) T represents the transposition operation, θ=[θ1,θ2,…,θ K ],A(f j ,θ)=[a(f j ,θ1),…,a(f j ,θ K )]∈C M×K is the frequency f j The array manifold matrix of points, is the angle θ k The steering vector of

[0091] Step 1.3: The single snapshot frequency f obtained in step 1.2 j The array mathematical model of the points is used to divide a whole period of time into L equally spaced segments, and then discrete Fourier transform is performed on each segment to obtain multiple snapshot data in the frequency domain. The array mathematical model of the lth snapshot is:

[0092] x l (f j )=A(f j ,θ)s l (f j )+n l (f j ),l=1,2,…,L

[0093] Among them, s l (f j ) is the frequency f j The signal vector of the lth snapshot, n l(f j ) is the frequency f j The noise vector of the lth snapshot.

[0094] Step 2: Calculate the array receiving data of each frequency point focused on the reference frequency point according to the array mathematical model of the lth snapshot of the broadband signal obtained in step 1;

[0095] Step 2 is implemented as follows:

[0096] Step 2.1: Based on the array mathematical model of the lth snapshot obtained in step 1, assuming that the signal is uncorrelated with the noise and that the noises of different channels are uncorrelated, obtain the covariance matrix R corresponding to the jth frequency point: j for:

[0097]

[0098] Where E[·] represents the expected operation, is the signal covariance matrix, I represents the unit matrix, (·) H represents the conjugate transpose operation. In practice,

[0099] Step 2.2, set the lowest frequency point of the broadband signal as the reference frequency point f0;

[0100] Step 2.3: Design the focusing matrix T(f j ):

[0101] T(f j )=Q(f0)Q H (f j )

[0102] Among them, Q(f0), Q(f j ) are P(f0), P(f j ), is the covariance matrix of the array data after denoising at the reference frequency point f0, is the frequency f j The covariance matrix of the array data after denoising, the noise power and are the frequencies f0 and f j The average of the small eigenvalues ​​of the data covariance matrix at ;

[0103] Step 2.4: Use the focusing matrix T(f j ) The frequency f j The array receiving data at the reference frequency f0 is focused to obtain the focused data.

[0104]

[0105] Step 3, performing data preprocessing on the reference frequency array received data obtained in step 2 to obtain input features after dimensionality reduction;

[0106] Step 3 is implemented as follows:

[0107] Step 3.1: Calculate the covariance matrix R0 at the reference frequency point f0 after focusing based on the array data of each frequency point obtained in step 2.4:

[0108]

[0109] Step 3.2: Extract the upper triangular part (excluding diagonal elements) of the covariance matrix R0 obtained in step 3.1 and recombine them to obtain the vector η:

[0110] η=[r 12 ,…,r 1M ,r 23 ,…,r 2M ,…,r (M-1)M ]

[0111] where r ij = R0(i,j) is the i-th row and j-th column element of R0;

[0112] Step 3.3: Normalize the vector η obtained in step 3.2 to obtain the reduced-dimensional input feature z0:

[0113]

[0114] Among them, ||·||2 represents the l2 norm of the vector, and Represents the real and imaginary part operations respectively.

[0115] Step 4: Based on the input features obtained in step 3 and support vector regression theory, an SVR model for multi-band joint direction finding of broadband signals is constructed;

[0116] Step 4 is implemented according to the following steps:

[0117] Step 4.1: Fit z0 and output angle according to the input features obtained in step 3.3 and support vector regression theory The function mapping relationship between them is:

[0118]

[0119] in, represents the estimation of θ, Φ(·) is the nonlinear mapping function, and {w,b} are the model parameters to be trained;

[0120] Step 4.2: Based on the function mapping relationship obtained in step 4.1, build an SVR model for broadband signal multi-band joint direction finding:

[0121] {

[0122] SVR1{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0123] SVR2{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0124] ···

[0125] SVR K {Kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; Penalty parameter: C}

[0126] }

[0127] Among them, the Gaussian kernel function Represents the pth sample data, p=1,2,…,P.

[0128] Step 5: training the SVR model constructed in step 4 to obtain a trained SVR model for broadband signal multi-band joint direction finding;

[0129] Step 5 is implemented according to the following steps:

[0130] Step 5.1: Generate a set of P angle labels for K targets And meet

[0131] Step 5.2: Use the angle label set Θ obtained in step 5.1 k Combine steps 1, 2, and 3 to generate a training sample set

[0132] Step 5.3: Construct optimization objective function:

[0133]

[0134] in, represents the loss function, ε is the error tolerance;

[0135] Step 5.4: The angle label set Θ obtained according to step 5.1 k and the sample set Z obtained in step 5.2 P , construct P sets of training data for the kth target in The pth group of training data represents the kth target angle, and the model training starts from k=1, p=1;

[0136] Step 5.5: Use the p-th group of training data of the k-th target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0137] Step 5.6: Use the p+1th group of training data of the kth target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0138] Step 5.7: Comparison function and Size:

[0139] like Then p=p+1, and repeat steps 5.5, 5.6 and 5.7;

[0140] like The model training of the k-th target angle is ended, and the optimal model parameter w corresponding to the k-th target angle is obtained. k ,b k ;

[0141] Step 5.8, let k = k + 1, and repeat steps 5.4, 5.5, 5.6 and 5.7 until k = K, all target models are trained, and the training is completed.

[0142] Step 6: Randomly generate target angle information, and combine steps 1, 2, and 3 to obtain test samples, which are then input into the model trained in step 5 to obtain the broadband signal multi-band joint direction finding results.

[0143] Step 6 is implemented according to the following steps:

[0144] Step 6.1, randomly generate target angle information θ, and combine steps 1, 2, and 3 to obtain the test sample z0;

[0145] Step 6.2: Input the test sample z0 obtained in step 6.1 into the model trained in step 5, and obtain the estimated angle of the kth target of the broadband signal:

[0146]

[0147] Example 1

[0148] The invention is based on the support vector regression broadband signal multi-band joint direction finding method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0149] Step 1: construct an array mathematical model of the lth snapshot of the broadband signal according to the array structure of the uniform linear array and the position information of the spatial target relative to the array;

[0150] Step 2: Calculate the array receiving data of each frequency point focused on the reference frequency point according to the array mathematical model of the lth snapshot of the broadband signal obtained in step 1;

[0151] Step 3, performing data preprocessing on the reference frequency array received data obtained in step 2 to obtain input features after dimensionality reduction;

[0152] Step 4: Based on the input features obtained in step 3 and support vector regression theory, an SVR model for multi-band joint direction finding of broadband signals is constructed;

[0153] Step 5: training the SVR model constructed in step 4 to obtain a trained SVR model for broadband signal multi-band joint direction finding;

[0154] Step 6: Input the test sample into the model trained in step 5 to obtain the broadband signal multi-band joint direction finding result.

[0155] Example 2

[0156] The invention is based on the support vector regression broadband signal multi-band joint direction finding method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0157] Step 1: construct an array mathematical model of the lth snapshot of the broadband signal according to the array structure of the uniform linear array and the position information of the spatial target relative to the array;

[0158] Step 1 is implemented as follows:

[0159] Step 1.1: Calculate the echo signal x received by the mth array element according to the array structure of the uniform linear array and the position information of the spatial target relative to the array. m (t):

[0160]

[0161] Among them, s k (t) represents the kth broadband signal, k = 1, 2, ..., K, K represents the number of targets, τ m,k =(m-1)d sinθ k / c means signal s k (t) is the time delay difference when it reaches the mth array element relative to the reference array element, M represents the number of array elements, θk represents the direction of the kth signal, c represents the propagation speed of electromagnetic waves, d represents the array element spacing, and n m (t) represents the Gaussian white noise on the mth array element, with a mean of 0 and a variance of σ n 2 ;

[0162] Step 1.2: Perform discrete Fourier transform on the array echo signal obtained in step 1.1 to obtain the frequency f under single snapshot j The mathematical model of the point array is:

[0163] x(f j )=A(f j ,θ)s(f j )+n(f j ),j=1,2,…,J

[0164] Among them, x(f j )=[x1(f j ),x2(f j ),…,x M (f j )] T , s(f j )=[s1(f j ),s2(f j ),…,s K (f j )] T and n(f j )=[n1(f j ),n2(f j ),…,n M (f j )] T The frequencies are f j The discrete Fourier transform of the array echo signal, target signal and noise signal of the point is assumed to obey Gaussian distribution and the noise of each sub-band is independent of each other. J represents the number of sub-bands, (·) T represents the transposition operation, θ=[θ1,θ2,…,θ K ],A(f j ,θ)=[a(f j ,θ1),…,a(f j ,θ K )]∈C M×K is the frequency f j The array manifold matrix of points, is the angle θ k The steering vector of

[0165] Step 1.3: The single snapshot frequency f obtained in step 1.2 jThe array mathematical model of the points is used to divide a whole period of time into L equally spaced segments, and then discrete Fourier transform is performed on each segment to obtain multiple snapshot data in the frequency domain. The array mathematical model of the lth snapshot is:

[0166] x l (f j )=A(f j ,θ)s l (f j )+n l (f j ),l=1,2,…,L

[0167] Among them, s l (f j ) is the frequency f j The signal vector of the lth snapshot at n l (f j ) is the frequency f j The noise vector of the lth snapshot.

[0168] Step 2: Calculate the array receiving data of each frequency point focused on the reference frequency point according to the array mathematical model of the lth snapshot of the broadband signal obtained in step 1;

[0169] Step 2 is implemented as follows:

[0170] Step 2.1: Based on the array mathematical model of the lth snapshot obtained in step 1, assuming that the signal is uncorrelated with the noise and that the noises of different channels are uncorrelated, obtain the covariance matrix R corresponding to the jth frequency point: j for:

[0171]

[0172] Where E[·] represents the expected operation, is the signal covariance matrix, I represents the unit matrix, (·) H represents the conjugate transpose operation. In practice,

[0173] Step 2.2, set the lowest frequency point of the broadband signal as the reference frequency point f0;

[0174] Step 2.3: Design the focusing matrix T(f j ):

[0175] T(f j )=Q(f0)Q H (f j )

[0176] Among them, Q(f0), Q(f j ) are P(f0), P(fj ), is the covariance matrix of the array data after denoising at the reference frequency point f0, is the frequency f j The covariance matrix of the array data after denoising, the noise power and are the frequencies f0 and f j The average of the small eigenvalues ​​of the data covariance matrix at ;

[0177] Step 2.4: Use the focusing matrix T(f j ) The frequency f j The array receiving data at the reference frequency f0 is focused to obtain the focused data.

[0178]

[0179] Step 3, performing data preprocessing on the reference frequency array received data obtained in step 2 to obtain input features after dimensionality reduction;

[0180] Step 4: Based on the input features obtained in step 3 and support vector regression theory, an SVR model for multi-band joint direction finding of broadband signals is constructed;

[0181] Step 4 is implemented according to the following steps:

[0182] Step 4.1: Fit z0 and output angle according to the input features obtained in step 3.3 and support vector regression theory The function mapping relationship between them is:

[0183]

[0184] in, represents the estimation of θ, Φ(·) is the nonlinear mapping function, and {w,b} are the model parameters to be trained;

[0185] Step 4.2: Based on the function mapping relationship obtained in step 4.1, build an SVR model for broadband signal multi-band joint direction finding:

[0186] {

[0187] SVR1{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0188] SVR2{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C}

[0189] ···

[0190] SVR K{Kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; Penalty parameter: C}

[0191] }

[0192] Among them, the Gaussian kernel function Represents the pth sample data, p=1,2,…,P.

[0193] Step 5: training the SVR model constructed in step 4 to obtain a trained SVR model for broadband signal multi-band joint direction finding;

[0194] Step 5 is implemented according to the following steps:

[0195] Step 5.1: Generate a set of P angle labels for K targets And meet

[0196] Step 5.2: Use the angle label set Θ obtained in step 5.1 k Combine steps 1, 2, and 3 to generate a training sample set

[0197] Step 5.3: Construct optimization objective function:

[0198]

[0199] in, represents the loss function, ε is the error tolerance;

[0200] Step 5.4: The angle label set Θ obtained according to step 5.1 k and the sample set Z obtained in step 5.2 P , construct P sets of training data for the kth target in The pth group of training data represents the kth target angle, and the model training starts from k=1, p=1;

[0201] Step 5.5: Use the p-th group of training data of the k-th target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0202] Step 5.6: Use the p+1th group of training data of the kth target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k, and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value

[0203] Step 5.7: Comparison function and Size:

[0204] like Then p=p+1, and repeat steps 5.5, 5.6 and 5.7;

[0205] like The model training of the k-th target angle is ended, and the optimal model parameter w corresponding to the k-th target angle is obtained. k ,b k ;

[0206] Step 5.8, let k = k + 1, and repeat steps 5.4, 5.5, 5.6 and 5.7 until k = K, all target models are trained, and the training is completed.

[0207] Step 6: Randomly generate target angle information, and combine steps 1, 2, and 3 to obtain test samples, which are then input into the model trained in step 5 to obtain the broadband signal multi-band joint direction finding results.

[0208] Step 6 is implemented according to the following steps:

[0209] Step 6.1, randomly generate target angle information θ, and combine steps 1, 2, and 3 to obtain the test sample z0;

[0210] Step 6.2: Input the test sample z0 obtained in step 6.1 into the model trained in step 5, and obtain the estimated angle of the kth target of the broadband signal:

[0211]

[0212] Example 3

[0213] The estimation performance of the target angle information of the present invention can be further verified through the following simulation experiments.

[0214] 1. Experimental scenario:

[0215] A uniform linear array is used, with the number of array elements M = 6 and the array element spacing d = 15m. It is assumed that all far-field broadband signal sources have the same bandwidth B = 4MHz, the center frequency is 8MHz, and the number of sub-bands is J = 64. The kernel function uses the Gaussian kernel function, with the Gaussian kernel function parameter γ = 0.1, the penalty parameter C = 1000, and ε = 0.1.

[0216] 2. Experimental content and analysis

[0217] Experiment 1: In the training phase, the BPSK signal is selected as the training signal. The angle range of the two signals is [-60°, 60°], the step is 0.1°, the signal angle interval is [9, 11, ..., 35], and a total of 13734 sets of training samples are generated, and the signal-to-noise ratio is 10 dB. In the test phase, the test signal is a 6dB BPSK signal with the same bandwidth. Three signal intervals [10, 20, 30] that are not within the angle interval range of the training set are selected. The angle range of the two signals is [-60°, 60°], the step is 1°, and a total of 303 sets of test samples are generated. The angle estimation result of multi-signal direction finding obtained by the method of the present invention is shown in Figure 2 (a) and the estimation error is shown in Figure 2 (b).

[0218] As can be seen from Figure 2, the method of the present invention can be used to estimate the angles of multiple sources, and the estimation results are basically consistent with the true values. Most of the angle estimation errors are concentrated in [-0.5, 0.5], which proves the effectiveness of the method of the present invention. In addition, the angle interval in the test set is not in the training set, indicating that the method of the present invention has generalization ability.

[0219] Experiment 2: Assume that there are two broadband signals in the far field of space, and the directions of the incoming waves are -9.56°+ζ and 15.72°+ζ, where ζ is randomly selected in the range of [-0.5°, 0.5°], the number of snapshots is fixed to 1024, the signal-to-noise ratio is -15dB to 15dB, and the step is 1dB. 500 Monte Carlo experiments are performed at each signal-to-noise ratio, and the root mean square error curves of the existing method and the method of the present invention with the signal-to-noise ratio are respectively statistically shown in Figure 3(a). The signal-to-noise ratio is fixed to 6dB and remains unchanged. The number of snapshots increases by 50 between 10 and 1050. 500 Monte Carlo experiments are performed at each snapshot number, and the root mean square error curves of the existing method and the method of the present invention with the number of snapshots are respectively statistically shown in Figure 3(b).

[0220] As can be seen from Figure 3, the root mean square error of the method of the present invention gradually decreases with the increase of the signal-to-noise ratio and the number of snapshots. It can also be seen that when the signal-to-noise ratio is greater than -9dB or the number of frequency domain snapshots is greater than 350, the root mean square error of the method of the present invention is lower than that of the existing method, indicating that the estimation accuracy of the method of the present invention is higher.

[0221] Experiment 3: Assume that there are two broadband far-field signals to study the resolution performance. The first signal incident angle is fixed at θ1 = 0°, and the second signal incident angle is θ2 = θ1 + Δθ, where Δθ is the interval between the two angles, Δθ∈[1°:1°:20°], the signal-to-noise ratio is 8dB, and the number of snapshots is 1024. Define that in each experiment the angle estimate satisfies and If the experiment succeeds, then it fails. Represents an estimate of θ. The probability of successful resolution is the ratio of the number of successful experiments to the total number of experiments. 200 Monte Carlo experiments were performed in each angular direction, and the statistical curves of the probability of successful resolution of the existing method and the method of the present invention with the angle interval are shown in Figure 4(a). The experimental conditions were changed, the first signal incident angle was fixed at θ1 = 0°, the second signal incident angle was fixed at θ2 = 7°, the signal-to-noise ratio was -10dB to 10dB, and the step was 2dB. 200 Monte Carlo experiments were performed at each signal-to-noise ratio, and the statistical curves of the probability of successful resolution of the existing method and the method of the present invention with the signal-to-noise ratio are shown in Figure 4(b).

[0222] It can be seen from FIG4 that with the increase of the angular interval or the signal-to-noise ratio, the resolution probability of all methods is on an increasing trend, and the successful resolution probability of the method of the present invention reaches 100% first, and at the same angular interval or signal-to-noise ratio, the successful resolution probability of the method of the present invention is significantly higher than that of the existing methods, especially at a small angular interval or a low signal-to-noise ratio, indicating that the resolution performance of the method of the present invention is higher than that of the existing methods.

[0223] Experiment 4: Assuming that there are two far-field broadband signals, the number of sub-bands ranges from 10 to 226. The running time of the training samples of the existing method and the method of the present invention and the average running time of 500 test samples are respectively counted as the variation curve of the number of sub-bands. Figure 5 shown.

[0224] Depend on Figure 5 It can be seen that compared with the existing method, the training time and test time of the method of the present invention are shorter, and when the number of sub-bands is greater than 136, the training time of the existing method increases exponentially, while the training time of the method of the present invention remains almost unchanged, indicating that when the number of sub-bands is large, the training efficiency of the method of the present invention is higher and can better meet the real-time requirements.

[0225] In summary, the method of the present invention can effectively estimate the angle of a broadband signal source, improve the accuracy and resolution performance of target angle estimation, and has high training efficiency, which can better meet the needs of practical engineering applications.

Claims

1. A multi-band joint direction finding method for broadband signals based on support vector regression, characterized in that: Follow the steps below to implement it: Step 1, constructing an array mathematical model of the lth snapshot of the broadband signal; Step 2, calculating the array receiving data of each frequency point focused on the reference frequency point; Step 3: preprocess the reference frequency array received data to obtain input features after dimensionality reduction; Step 4: Build an SVR model for broadband signal multi-band joint direction finding; Step 5: training the SVR model to obtain a trained SVR model for broadband signal multi-band joint direction finding; Step 6: Input the test sample into the model trained in step 5 to obtain the broadband signal multi-band joint direction finding result.

2. The method for multi-band joint direction finding of broadband signals based on support vector regression according to claim 1, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1: Calculate the echo signal x received by the mth array element according to the array structure of the uniform linear array and the position information of the spatial target relative to the array. m (t): Among them, s k (t) represents the kth broadband signal, k = 1, 2, ..., K, K represents the number of targets, τ m,k =(m-1)dsinθ k / c means signal s k (t) is the time delay difference when it reaches the mth array element relative to the reference array element, M represents the number of array elements, θ k represents the direction of the kth signal, c represents the propagation speed of electromagnetic waves, d represents the array element spacing, and n m (t) represents the Gaussian white noise on the mth array element, with a mean of 0 and a variance of Step 1.2: Perform discrete Fourier transform on the array echo signal obtained in step 1.1 to obtain the frequency f under single snapshot j The mathematical model of the point array is: x(f j )=A(f j ,θ)s(f j )+n(f j ),j=1,2,…,J Among them, x(f j )=[x1(f j ),x2(f j ),…,x M (f j )] T , s(f j )=[s1(f j ),s2(f j ),…,s K (f j )] T and n(f j )=[n1(f j ),n2(f j ),…,n M (f j )] T The frequencies are f j The discrete Fourier transform of the array echo signal, target signal and noise signal of the point is assumed to obey Gaussian distribution and the noise of each sub-band is independent of each other. J represents the number of sub-bands, (·) T represents the transposition operation, θ=[θ1,θ2,…,θ K ],A(f j ,θ)=[a(f j ,θ1),…,a(f j ,θ K )]∈C M×K is the frequency f j The array manifold matrix of points, is the angle θ k The steering vector of Step 1.3: The single snapshot frequency f obtained in step 1.2 j The array mathematical model of the points is used to divide a whole period of time into L equally spaced segments, and then discrete Fourier transform is performed on each segment to obtain multiple snapshot data in the frequency domain. The array mathematical model of the lth snapshot is: x l (f j )=A(f j ,θ)s l (f j )+n l (f j ),l=1,2,…,L Among them, s l (f j ) is the frequency f j The signal vector of the lth snapshot, n l (f j ) is the frequency f j The noise vector of the lth snapshot.

3. The method for multi-band joint direction finding of broadband signals based on support vector regression according to claim 2, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1: Based on the array mathematical model of the lth snapshot obtained in step 1, assuming that the signal is uncorrelated with the noise and that the noises of different channels are uncorrelated, the covariance matrix R corresponding to the jth frequency point is obtained: j for: Where E[·] represents the expected operation, is the signal covariance matrix, I represents the unit matrix, (·) H represents the conjugate transpose operation. In practice, Step 2.2, set the lowest frequency point of the broadband signal as the reference frequency point f0; Step 2.3: Design the focusing matrix T(f j ): T(f j )=Q(f0)Q H (f j ) Among them, Q(f0), Q(f j ) are P(f0), P(f j ), is the covariance matrix of the array data after denoising at the reference frequency point f0, is the frequency f j The covariance matrix of the array data after denoising, the noise power and are the frequencies f0 and f j The average of the small eigenvalues ​​of the data covariance matrix at ; Step 2.4: Use the focusing matrix T(f j ) The frequency f j The array receiving data at the reference frequency f0 is focused to obtain the focused data.

4. The method for multi-band joint direction finding of broadband signals based on support vector regression according to claim 3 is characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Calculate the covariance matrix R0 at the reference frequency point f0 after focusing based on the array data of each frequency point obtained in step 2.4: Step 3.2: Extract the upper triangular part of the covariance matrix R0 obtained in step 3.1 and recombine it to obtain the vector η: η=[r 12 ,…,r 1M ,r 23 ,…,r 2M ,…,r (M-1)M ] where r ij = R0(i,j) is the i-th row and j-th column element of R0; Step 3.3: Normalize the vector η obtained in step 3.2 to obtain the input feature z0 after dimensionality reduction: Among them, ||·||2 represents the l2 norm of the vector, and Represents the real and imaginary part operations respectively.

5. The method for multi-band joint direction finding of broadband signals based on support vector regression according to claim 4 is characterized in that: The step 4 is specifically implemented according to the following steps: Step 4.1: Fit z0 and output angle according to the input features obtained in step 3.3 and support vector regression theory The function mapping relationship between them is: in, represents the estimation of θ, Φ(·) is the nonlinear mapping function, and {w,b} are the model parameters to be trained; Step 4.2: Based on the function mapping relationship obtained in step 4.1, build an SVR model for broadband signal multi-band joint direction finding: { SVR1{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C} SVR2{kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; penalty parameter: C} ··· SVR K {Kernel function: Gaussian kernel function; Gaussian kernel function parameter: γ; Penalty parameter: C} } Among them, the Gaussian kernel function Represents the pth sample data, p=1,2,…,P.

6. The method for multi-band joint direction finding of wideband signals based on support vector regression according to claim 5, characterized in that: The step 5 is specifically implemented according to the following steps: Step 5.1: Generate a set of P angle labels for K targets And meet Step 5.2: Use the angle label set Θ obtained in step 5.1 k Combine steps 1, 2, and 3 to generate a training sample set Step 5.3: Construct optimization objective function: in, represents the loss function, ε is the error tolerance; Step 5.4: The angle label set Θ obtained according to step 5.1 k and the sample set Z obtained in step 5.2 P , construct P sets of training data for the kth target in The pth group of training data represents the kth target angle, and the model training starts from k=1, p=1; Step 5.5: Use the p-th group of training data of the k-th target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value Step 5.6: Use the p+1th group of training data of the kth target angle obtained in step 5.4 Train the SVR model obtained in step 4.2 and update the parameter w k ,b k , and substitute it into the optimization objective function constructed in step 5.3 to obtain the function value Step 5.7: Comparison function and Size: like Then p=p+1, and repeat steps 5.5, 5.6 and 5.7; like The model training of the k-th target angle is ended, and the optimal model parameter w corresponding to the k-th target angle is obtained. k ,b k ; Step 5.8, let k = k + 1, and repeat steps 5.4, 5.5, 5.6 and 5.7 until k = K, all target models are trained, and the training is completed.

7. The method for multi-band joint direction finding of wideband signals based on support vector regression according to claim 6, characterized in that: The step 6 is specifically implemented according to the following steps: Step 6.1, randomly generate target angle information θ, and combine steps 1, 2, and 3 to obtain the test sample z0; Step 6.2: Input the test sample z0 obtained in step 6.1 into the model trained in step 5, and obtain the estimated angle of the kth target of the broadband signal:

Citation Information

Patent Citations

  • Wideband signal DOA estimation method based on sparse representation of covariance matrix

    CN109901148A

  • One-dimensional DOA estimation method by utilizing secondary virtual extension of multi-frequency signal

    CN110133574A

  • Direction-of-arrival estimation method based on differential array

    CN111239680A

  • Broadband signal sparse representation direction-of-arrival estimation method based on focusing transformation

    CN113805139A

  • Direction of arrival estimation method based on broadband signal multiband joint sparse Bayesian learning

    CN118501803A

Cited By

  • Elastic wave device performance sample preprocessing method and system, apparatus, and storage medium

    WO2026184585A1