Improved direction of arrival sound source estimation method based on singular value decomposition

Abstract

The invention discloses an improved DOA (direction of arrival) sound source estimation method based on singular value decomposition, and relates to the field of acoustic imaging sound signal processing. The method is a method for improving a MUSIC algorithm based on the combination of singular value decomposition and noise subspace weighting, and comprises the following steps: S1, a receiving end receives a signal source X (t) through a microphone array; s2, preprocessing the acquired signals; s3, constructing a cross matrix Z, and constructing a new covariance matrix Rnew for the collected data; s4, performing singular value decomposition (SVD) on the new covariance matrix Rnew to obtain a noise subspace Un; s5, performing weighting processing on the noise subspace; and S6, calculating a spatial spectrogram and searching a peak value position. According to the method, the estimation performance of the MUSIC algorithm under the non-ideal conditions of low signal-to-noise ratio and few snapshots is improved, and meanwhile, the target resolution capability of the adjacent angle under the condition of low signal-to-noise ratio is improved.

Description

Technical Field

[0001] The invention relates to the field of acoustic imaging sound signal processing, in particular to an improved direction-of-arrival sound source estimation method based on singular value decomposition. Background Art

[0002] Estimation of the direction of arrival (DOA) of incident signals is a key research topic in array signal processing and has been widely used in radar, acoustic imaging, mobile communications, and medical fields. In recent decades, a large number of research results have emerged in DOA estimation, and super-resolution algorithms have become the mainstream. The most well-known of these is the Multiple Signal Classification (MUSIC) method, which constructs a spatial spectrum function through the orthogonality of the signal subspace and the noise subspace to estimate the signal DOA.

[0003] The introduction of the MUSIC algorithm has promoted the rapid development of super-resolution algorithms. In practical applications, due to the complex communication environment, the noise amplitude received by the receiver is often large, resulting in a low signal-to-noise ratio or insufficient number of snapshots. At this time, the resolution of the traditional algorithm will be reduced. When there are two or more signal sources with similar arrival angles, the spectral peaks of the two will gradually merge with each other, so that the directions of the waves of multiple targets cannot be distinguished, and the accuracy of sound source positioning will indeed decrease. This is because coherent signal sources (signal sources with similar directions) will cause the intervals between the eigenvalues ​​of the covariance matrix to become smaller, which in turn affects the separation of the noise subspace, resulting in a decrease in the resolution of the MUSIC algorithm.

[0004] Among the existing invention methods, there is a method that combines Toeplitz matrix reconstruction with a DOA estimation algorithm, which can estimate multiple coherent sound source targets, but the resolution of adjacent targets is not high; there is a method that applies a spatial smoothing method to DOA estimation, which has a higher estimation accuracy, but the estimation performance is lacking in an environment with low signal-to-noise ratio and a small number of block beats. Summary of the invention

[0005] In order to overcome the shortcomings of existing research, the present invention aims at the problem that the estimation performance of the traditional MUSIC algorithm fails under the conditions of low signal-to-noise ratio and small snapshot number, and proposes an improved direction-of-arrival sound source estimation method based on singular value decomposition. This method can make maximum use of the mutual correlation information to construct a new matrix and use singular value decomposition to obtain the signal subspace and noise subspace, and the improved MUSIC algorithm combined with weighted processing of the improved noise subspace has high resolution for close targets under the conditions of low signal-to-noise ratio and small snapshot number without basically increasing the computational complexity, and has high practical application value.

[0006] In order to achieve the above object, the technical solution specifically adopted by the present invention is as follows:

[0007] An improved direction of arrival sound source estimation method based on singular value decomposition is characterized in that: the method optimizes the MUSIC algorithm by using a method combining singular value decomposition and weighted noise subspace, and comprises the following steps:

[0008] S1, the receiving end receives the signal source signal through the microphone array;

[0009] S2, preprocessing the collected signal X(t);

[0010] S3, construct the cross matrix Z, and construct a new covariance matrix R for the collected data new ;

[0011] S4. New covariance matrix R new Singular value decomposition obtains the noise subspace U n ;

[0012] S5, noise subspace weighted processing;

[0013] S6. Calculate the spatial spectrum P MUSIC (θ), find the peak position.

[0014] Furthermore, in step S1, let the number of sensors be N, the sensor spacing be d and less than half a wavelength, and at each discrete time t, the array output signal vector X(t) = [x1(t), x2(t), ... x N (t)] T It can be expressed as:

[0015] X(t)=A(θ)*S(t)+N(t)

[0016] In the formula, N(t) means the mean is 0 and the variance is Gaussian white noise, S(t) is the matrix of the sound source signal, usually a column vector, where each column represents the time value of a different sound source signal, and A(θ)*S(t) is the combination of multiple sound source signals after passing through the array response;

[0017] A(θ)=[a(θ1)*a(θ2)*…a(θ M )]

[0018] Where a(θ) is the array steering vector, with a size of N*M, and each column corresponds to the steering vector of a signal source, where a(θ) is expressed as:

[0019]

[0020] In the formula, f represents the signal frequency, c is the speed of sound, xy is the two-dimensional coordinate of the signal (the Z axis is 0); S is the signal vector of the signal source, with a size of M*1, containing the signals of M signal sources at time t.

[0021] Furthermore, the specific steps of step S2 are:

[0022] S2-1, split the long signal into short frames by frame division and windowing, where the window function w is the Hamming window function;

[0023] X frame =X(:,start idx :end idx )*w;

[0024] In the formula, start idx :end idx is the start index and end index of the time period, indicating the selection of a certain section (i.e., a frame) of the signal, and: represents all elements of the row. This formula takes the start and end index of the array X from start idx to the end idx (excluding end idx )

[0025] S2-2, calculate the covariance matrix of each frame and accumulate it, and take the average value to get the covariance matrix R x :

[0026]

[0027] In the formula, num frames is the number of frames (number of windows) after the entire signal is framed, which is the total number of frames obtained after sliding window or framing; frame_size is the length of each frame, that is, the number of sampling points contained in each frame; X H frame For X frame The conjugate transpose of , that is, the Hermitian transpose matrix.

[0028] Furthermore, the specific steps of step S3 are:

[0029] S3-1. Construct the cross matrix Z and generate the auxiliary matrix R y , R xy ;in:

[0030]

[0031] R x_loaded =R x +δI;

[0032]

[0033] R xy =R y +ZR y Z T ;

[0034] Where, the diagonal loading factor δ is usually a small positive number, I is the unit matrix, and its size is the same as the covariance matrix. is the covariance matrix after diagonal loading, Z T is the transpose of the cross matrix Z;

[0035] S3-2. Calculate the new covariance combination R1, R2, R3, and finally combine them into a new covariance matrix R new ;in:

[0036]

[0037]

[0038] R3=R xy +ZR xy Z T ;

[0039] R new =[R1,R2,R3].

[0040] Furthermore, the specific steps of step S4 are:

[0041] S4-1. For the new covariance matrix R new Perform singular value decomposition to obtain matrix U, matrix S, and matrix V;

[0042] R new =USV H

[0043] Among them, the matrix U and the matrix V are orthogonal matrices, the matrix S is a diagonal matrix, and V H represents the Hermitian transpose of the matrix V;

[0044] S4-2. Forming the signal subspace U s and the noise subspace U n :

[0045] U s =U(:,1:M)

[0046] U n =U(:,M+1:N)

[0047] Among them, for the covariance matrix R new, the singular values ​​corresponding to the left singular vectors, i.e., the diagonal elements in the S matrix, are sorted by size, and the left singular vectors corresponding to the first M largest singular values ​​form the signal subspace U s , the remaining singular values ​​are small, usually close to zero, and the corresponding left singular vectors form the noise subspace U n .

[0048] Furthermore, the specific steps of step S5 are:

[0049] S5-1. Define the weighted matrix W: first extract the part of the singular value matrix S related to the noise subspace (i.e., from M+1 to the Nth singular value), then calculate its reciprocal, and perform power adjustment according to the weighting coefficient α, and finally obtain a diagonal matrix W, the diagonal elements of which are the reciprocals of these adjusted singular values;

[0050]

[0051] S5-2, the noise subspace matrix U n Multiply it with the weighting matrix W to get the weighted noise subspace matrix E n ;

[0052] E n =U n *W.

[0053] Furthermore, the specific steps of step S6 are:

[0054] S6-1, calculating the spatial spectrum, and normalizing the results after calculating the spatial spectrum of all angles;

[0055]

[0056] Where P MUSIC (θ) is the spatial spectrum value at the direction angle θ, representing the estimation ability of the sound source signal, a H (θ) is the conjugate transpose of the array steering vector a(θ), E n H is the noise subspace E n The conjugate transpose of is the normalized spatial spectrum value.

[0057] S6-2, finding the angle peak of the spatial spectrum function in the direction of the signal source to determine the direction of the signal source;

[0058]

[0059] In the formula, θ estimated is the estimated signal source direction, that is, the angle corresponding to the maximum spectrum value found in the normalized spatial spectrum.

[0060] Furthermore, after completing the preprocessing of the acquisition signal X(t), the method further includes setting a diagonal loading factor δ to perform diagonal loading on the covariance matrix:

[0061] R x_loaded =R x +δI

[0062] Where I is the identity matrix, which is the same size as the covariance matrix.

[0063] The present invention has the following characteristics and beneficial effects:

[0064] The improved MUSIC algorithm based on singular value decomposition and noise subspace weighted combination proposed in the present invention greatly improves the estimation performance of adjacent incident angles under low signal-to-noise ratio and small block number conditions while ensuring real-time positioning, thereby reducing positioning errors.

[0065] Compared with the traditional MUSIC algorithm and the weighted MUSIC algorithm, the method of the present invention has lower error and better positioning performance under low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0067] Figure 1 The present invention provides a flowchart of an improved method for direction of arrival sound source estimation based on singular value decomposition.

[0068] Figure 2 This is a schematic diagram of a 128-microphone spiral array used in the present invention.

[0069] Figure 3 Schematic diagram of the three algorithms' ability to resolve neighboring targets under non-ideal conditions of low signal-to-noise ratio and insufficient snapshots;

[0070] In the figure: (a) snr=2, k=512, [0,2,4]; (b) snr=2, k=512, [0,3,6].

[0071] Figure 4 The RMSE comparison chart of the three algorithms at five groups of adjacent incident angles [0,1,2], [0,2,4], [0,3,6], [0,4,8], [0,5,10] is shown;

[0072] In the figure: (a) snr=20, k=128, [0,2,4]; (b) snr=20, k=128, [0,3,6]. Figure 5This is a comparison chart of the RMSE of three MUSIC algorithms under different incident angle groups. DETAILED DESCRIPTION

[0073] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0074] In order to solve the problem that the traditional MUSIC algorithm fails to estimate performance under low signal-to-noise ratio and small number of snapshots. At the same time, when there are two sound sources with similar arrival angles, the spectral peaks of the two will gradually merge with each other, so that the directions of the waves coming from the adjacent angles cannot be distinguished. The embodiment of the present invention provides an improved wave direction sound source estimation method based on singular value decomposition, such as Figure 1 As shown, the following steps are included:

[0075] Step 1: The receiver receives the signal source signal through the microphone array. Before receiving the signal source signal, the microphone array data must be loaded and preprocessed, and the microphone array data must be converted into a suitable format to ensure the correct unit. Then, constants and parameters are defined, including angle, number of sources, sound speed, frame size, and frame overlap.

[0076] The 128-channel spiral microphone array used in the present invention is as follows: Figure 2 As shown. Here, the number of sensors is N, the sensor spacing is d and is less than half a wavelength, and at each discrete time t, the array output signal vector X(t) = [x1(t), x2(t), …x N (t)] T It can be expressed as:

[0077] X(t)=A(θ)*S(t)+N(t)

[0078] In the formula, N(t) means the mean is 0 and the variance is Gaussian white noise, S(t) is the matrix of the sound source signal, usually a column vector, where each column represents the time value of a different sound source signal, and A(θ)*S(t) is the combination of multiple sound source signals after passing through the array response;

[0079] A(θ)=[a(θ1)*a(θ2)*…a(θ M )]

[0080] Where a(θ) is the array steering vector, with a size of N*M, and each column corresponds to the steering vector of a signal source, where a(θ) is expressed as:

[0081] In the formula, f represents the signal frequency, c is the speed of sound, and xy

[0082] is the two-dimensional coordinate of the signal (Z axis is 0); S is the signal vector of the signal source, with a size of M*1, containing the signals of M signal sources at time t.

[0083] Step 2: Preprocess the collected signal X(t), divide the frame and add windows. Directly performing Fourier transform on the entire signal may cause spectrum leakage. Windowing can reduce this leakage and improve the accuracy of spectrum analysis.

[0084] Step 1: Assume that the signal X(t) is an N*K matrix, where N is the number of microphones and K is the total number of sampling points. We first divide the long signal into short frames by framing and windowing, so that the signal in each frame can be approximated as a stationary signal, which is convenient for analysis and processing. The size of each frame is frame_size, and the overlap between frames is overlap. Calculate the number of frames and mark the start and end positions of each frame as start_idx and end_idx; here, the window function w selects the Hamming window function.

[0085] X frame =X(:,start idx :end idx )*w

[0086] In the formula, start idx :end idx It is the start and end index of the time period, indicating the data of a certain section (i.e., a frame) in the selected signal.

[0087] Step 2: Calculate the covariance matrix of each frame. For the kth frame, the covariance matrix is ​​calculated as

[0088]

[0089] Step 3: After accumulating the covariance matrix, take the average value to get the covariance matrix R x :

[0090]

[0091] In the formula, num frames is the number of frames (number of windows) after the entire signal is framed, which is the total number of frames obtained after sliding window or framing; frame_size is the length of each frame, that is, the number of sampling points contained in each frame; X H frame For X frame The conjugate transpose of , that is, the Hermitian transpose matrix.

[0092] Step 3: Diagonal loading process

[0093] Step 1: To avoid ill-conditioned matrices, set a diagonal loading factor δ, which is usually a small positive number.

[0094] Step 2: Load the covariance matrix diagonally:

[0095] R x_loaded =R x +δI

[0096] Where I is the identity matrix, which is the same size as the covariance matrix.

[0097] Step 4: Construct a cross matrix and construct a new covariance matrix for the collected data. The specific steps are as follows:

[0098] In order to enhance the structural information of the covariance matrix, effectively disperse and average the noise impact, and improve the resolution and estimation accuracy, the cross matrix Z is constructed as follows:

[0099]

[0100] By using the cross matrix Z multiple times, multiple auxiliary matrices R are generated y , R xy , using their correlation, combine them to get a larger matrix R new , this new matrix contains more information of the original matrix and is more stable, which is helpful for the subsequent subspace decomposition;

[0101] R x_loaded =R x +δI

[0102]

[0103] R xy =R y +ZR y Z T

[0104] Where, the diagonal loading factor δ is usually a small positive number, I is the unit matrix, and its size is the same as the covariance matrix. is the covariance matrix after diagonal loading, Z T is the transpose of the cross matrix Z.

[0105] Calculate the new covariance combination R1, R2, R3:

[0106]

[0107]

[0108] R3=R xy +ZR xy ZT

[0109] Finally, the new covariance matrix is ​​combined:

[0110] R new =[R1,R2,R3]

[0111] Step 5: Decompose the new covariance matrix singular value to obtain the noise subspace. The specific steps are as follows:

[0112] For the new covariance matrix R new Perform singular value decomposition to obtain matrix U, matrix S, and matrix V, where matrix U and matrix V are orthogonal matrices and matrix S is a diagonal matrix. Singular value decomposition decomposes the matrix into different components, which facilitates the subsequent extraction of signal subspace and noise subspace: R new =USV H Among them, matrix U and matrix V are orthogonal matrices, matrix S is a diagonal matrix, and V H Represents the Hermitian transpose of the matrix V.

[0113] The column vectors in the matrix U are R new The left singular vector of . For the covariance matrix R new , the singular values ​​corresponding to the left singular vectors (i.e., the diagonal elements in the S matrix) are sorted by size, and the left singular vectors corresponding to the first M largest singular values ​​form the signal subspace U s The remaining singular values ​​are small, usually close to zero, and the corresponding left singular vectors form the noise subspace U n .

[0114] U s =U(:,1:M)

[0115] U n =U(:,M+1:N)

[0116] Step 6: Noise subspace weighted processing, the specific steps are as follows:

[0117] First, extract the part of the singular value matrix S related to the noise subspace (i.e., from the M+1th to the Nth singular value), then calculate its reciprocal, and adjust it according to the weighting coefficient α, and finally obtain a diagonal matrix W, where the diagonal elements are the reciprocals of these adjusted singular values. i is the i-th singular value in the singular value matrix S, obtained by singular value decomposition (SVD).

[0118]

[0119] Secondly, the noise subspace matrix U n Multiply it with the weighting matrix W to get the weighted noise subspace matrix En This weighted processing improves the stability and accuracy of subsequent direction estimation by amplifying or shrinking certain components of the noise subspace.

[0120] E n =U n *W

[0121] The present invention adjusts the contribution of the noise subspace according to the size of the singular value, thereby improving the robustness of the direction estimation algorithm.

[0122] Step 7: Calculate the spatial spectrum and find the peak position. The specific steps are as follows:

[0123] Traverse all possible angles to estimate the direction of the signal source and calculate the spatial spectrum function P MUSIC (θ):

[0124]

[0125] Where P MUSIC (θ) is the spatial spectrum value at the direction angle θ, representing the estimation ability of the sound source signal, α H (θ) is the conjugate transpose of the array steering vector α(θ), is the normalized spatial spectrum value.

[0126] After calculating the spatial spectra of all angles, the results are normalized and converted into decibel (dB) to facilitate subsequent peak detection and direction estimation;

[0127]

[0128] The spatial spectrum function has a peak at the angle in the direction of the signal source. The peak is found to determine the direction of the signal source.

[0129]

[0130] In the formula, θ estimated is the estimated signal source direction, that is, the angle corresponding to the maximum spectrum value found in the normalized spatial spectrum.

[0131] Finally, the root mean square error RMSE is calculated to view the estimation accuracy of the improved MUSIC method:

[0132]

[0133] Where M is the number of observations, is the estimated value of the ith observation, and θ is the actual value of the ith observation.

[0134] In order to verify the estimation effect of the present invention on DOA, the traditional MUSIC algorithm and the weighted MUSIC algorithm are compared on the data set constructed by adding noise to the sound source signal obtained by simulation. The incident angle directions of 0°, 2°, 4° and 0°, 3°, 6° are selected, and the traditional MUSIC algorithm and the weighted MUSIC algorithm are selected for comparison with the present invention.

[0135] When the incident angle is [0, 2, 4], under the conditions of signal-to-noise ratio of 2 and number of snapshots of 512, the three algorithms have the following ability to distinguish neighboring targets: Figure 3 As shown in (a), the spatial spectrum recognition effect of the SVD-MUSIC algorithm is much better than the other two algorithms, and the root mean square error is the smallest; when the signal-to-noise ratio is 20 and the number of snapshots is 128, the resolution ability of the three algorithms for adjacent angles is as follows: Figure 4 As shown in (a), the traditional MUSIC algorithm loses its resolution, the weighted MUSIC algorithm identifies two incident angles, and the SVD-MUSIC algorithm identifies three incident angles more accurately.

[0136] When the incident angle is [0, 3, 6], under the conditions of signal-to-noise ratio of 2 and number of snapshots of 512, the three algorithms have the following ability to distinguish neighboring targets: Figure 3 As shown in (b), only the spectrum of the SVD-MUSIC algorithm can identify three incident angles with the smallest root mean square error, while the other two algorithms can only estimate two incident angles. When the signal-to-noise ratio is 20 and the number of snapshots is 128, the resolution of the three algorithms for adjacent angles is as follows: Figure 4 As shown in (a), the weighted MUSIC algorithm loses its resolution ability, while the SVD-MUSIC algorithm can still recognize the three incident angles, and its recognition ability is much better than that of the traditional MUSIC algorithm.

[0137] It can be seen that the weighted MUSIC algorithm cannot distinguish adjacent targets under non-ideal conditions such as low signal-to-noise ratio or insufficient snapshots. The improved algorithm in this paper has an obvious depression between the two directions and successfully distinguishes two adjacent targets. The error is much lower than that of the traditional MUSIC algorithm, and the peak of the spectrum is more obvious.

[0138] In order to verify the positioning performance of the present invention for adjacent angles, under the same environment as the above experiment, the signal-to-noise ratio is set to 2, the number of snapshots is set to 512, and the RMSE of the three algorithms at five groups of adjacent incident angles [0, 1, 2], [0, 2, 4], [0, 3, 6], [0, 4, 8], [0, 5, 10] are tested and counted. The results are as follows Figure 4 As shown in the figure, it can be seen that the SVD-MUSIC algorithm has a much lower ability to distinguish adjacent sound sources than the other two algorithms in a low signal-to-noise ratio environment.

[0139] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An improved direction of arrival sound source estimation method based on singular value decomposition, characterized in that: This method optimizes the MUSIC algorithm by using the singular value decomposition and weighted noise subspace joint method, and includes the following steps: S1, the receiving end receives the signal source signal through the microphone array; S2, preprocessing the collected signal X(t); S3, construct the cross matrix Z, and construct a new covariance matrix R for the collected data new ; S4. New covariance matrix R new Singular value decomposition obtains the noise subspace U n ; S5, noise subspace weighted processing; S6. Calculate the spatial spectrum P MUSIC (θ), find the peak position.

2. The method for improving direction of arrival sound source estimation based on singular value decomposition according to claim 1, characterized in that: In step S1, let the number of sensors be N, the sensor spacing be d and less than half a wavelength, and at each discrete time t, the array output signal vector X(t)=[x1(t), x2(t), ...x N (t)] T It can be expressed as: X(t)=A(θ)*S(t)+N(t) In the formula, N(t) means the mean is 0 and the variance is Gaussian white noise, S(t) is the matrix of sound source signals, each column of which represents the time value of different sound source signals, and A(θ)*S(t) is the combination of multiple sound source signals after passing through the array response; A(θ)=[a(θ1)*a(θ2)*…a(θ M )] Where A(θ) is the array response matrix, a(θ) is the array steering vector, the size is N*M, each column corresponds to the steering vector of a signal source, and a(θ) is expressed as: In the formula, f represents the signal frequency, c is the speed of sound, xy is the two-dimensional coordinate of the signal (the Z axis is 0); S is the signal source signal vector, with a size of M*1, containing the signals of M signal sources at time t.

3. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: The specific steps of step S2 are: S2-1, split the long signal into short frames by framing and windowing. Here, the window function w is the Hamming window function. The signal segment X after framing and windowing frame It is expressed as: X frame =X(:,start idx :end idx )*w; In the formula, start idx :end idx is the start and end index of the time period, indicating the selection of a certain section (i.e., a frame) of the signal. : represents all elements of the row. This formula takes the start and end index of the time period in the array X from start idx to the end idx (excluding end idx ) all columns; S2-2, calculate the covariance matrix of each frame and accumulate it, and take the average value to get the covariance matrix R x : In the formula, num frames is the number of frames after the entire signal is framed, which is the total number of frames obtained after sliding window or framing; frame_size is the length of each frame, that is, the number of sampling points contained in each frame; X H frame For X frame The conjugate transpose of , that is, the Hermitian transpose matrix.

4. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: The specific steps of step S3 are: S3-1. Construct a cross matrix Z, which is used to mix the forward and backward signals together, mainly to eliminate the interference caused by the coherent signal and generate the auxiliary matrix R y , R xy ; in: R x_loaded =R x +δI; R xy =R y +ZR y With T ; Where δ is the diagonal loading factor and I is the unit matrix, which is the same size as the covariance matrix. is the covariance matrix after diagonal loading, Z T is the transpose of the cross matrix Z; S3-2. Calculate the new covariance combination R1, R2, R3, and finally combine them into a new covariance matrix R new ;in: R3=R xy +ZR xy With T ; R new =[R1,R2,R3]。 5. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: The specific steps of step S4 are: S4-1. For the new covariance matrix R new Perform singular value decomposition: R new =USV H Among them, the matrix U and the matrix V are orthogonal matrices, the matrix S is a diagonal matrix, and V H represents the Hermitian transpose of the matrix V; S4-2. Forming the signal subspace U s and the noise subspace U n : U s =U(:,1:M); U n =U(:,M+1:N)。 6. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: The specific steps of step S5 are: S5-1. Define the weighted matrix W: first extract the part of the singular value matrix S related to the noise subspace, then calculate its inverse, and perform power adjustment according to the weighting coefficient α, and finally obtain a diagonal matrix W, the diagonal elements of which are the inverses of these adjusted singular values; In the formula, σ i is the i-th singular value in the singular value matrix S, obtained by singular value decomposition (SVD); S5-2, the noise subspace matrix U n Multiply it with the weighting matrix W to get the weighted noise subspace matrix E n ; HAVE BEEN n =U n *W。 7. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: The specific steps of step S6 are: S6-1, calculating the spatial spectrum, and normalizing the results after calculating the spatial spectrum of all angles; Where P MUSIC (θ) is the spatial spectrum value at the direction angle θ, representing the estimation ability of the sound source signal, a H (θ) is the conjugate transpose of the array steering vector a(θ), is the normalized spatial spectrum value; E n H is the noise subspace E n The conjugate transpose of ; S6-2, finding the angle peak of the spatial spectrum function in the direction of the signal source to determine the direction of the signal source; In the formula, θ estimated is the estimated signal source direction, that is, the angle corresponding to the maximum spectrum value found in the normalized spatial spectrum.

8. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: After completing the preprocessing of the acquired signal X(t), the step of setting a diagonal loading factor δ to perform diagonal loading on the covariance matrix is ​​also included: R x_loaded =R x +δI Where I is the identity matrix, which is the same size as the covariance matrix.

9. The improved direction of arrival sound source estimation method based on singular value decomposition according to claim 1, characterized in that: For the covariance matrix R new , the singular values ​​corresponding to the left singular vectors, i.e., the diagonal elements in the S matrix, are sorted by size, and the left singular vectors corresponding to the first M largest singular values ​​form the signal subspace U s , the left singular vectors corresponding to the remaining singular values ​​form the noise subspace U n .

Images