A method and system for underdetermined blind source separation based on cumulative tensor decomposition
By using a cumulative tensor decomposition-based method, the problems of accuracy and computational complexity in mixing matrix estimation in underdetermined blind source separation are solved, achieving efficient signal separation and recovery, especially performing well in low signal-to-noise ratio environments.
Patent Information
- Application Number
- CN202411244273.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-09-06
AI Technical Summary
Under underdetermined conditions, existing blind source separation methods suffer from severe noise-induced signal sparsity in complex environments, leading to reduced accuracy in mixing matrix estimation and high computational complexity. Furthermore, the source signal recovery algorithm negatively impacts separation performance.
A cumulative tensor decomposition-based approach is adopted, which involves whitening, slicing, time delay covariance matrix calculation, regular multivariate decomposition, and weighted nonlinear least squares algorithm. Combined with the Jacobi operator and block Jacobi preprocessor, an appropriate hybrid vector estimation method is selected for signal reconstruction.
It improves the accuracy and robustness of hybrid matrix estimation, suppresses Gaussian white noise, enhances separation performance and computational efficiency under low signal-to-noise ratio, and improves the accuracy of source signal recovery.
Smart Images

Figure CN118760872B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of blind source separation, and in particular to an underdetermined blind source separation method and system based on cumulative tensor decomposition. Background Art
[0002] For the problem of estimating the mixing matrix under underdetermined conditions, the current methods mainly include using the sparsity of the signal or the high-order statistical characteristics of the signal. In a complex environment, the environmental noise is very dense, and the sparsity of the signal will be seriously affected. At this time, if the sparsity of the signal is still used to estimate the mixing matrix, the estimation accuracy of the mixing matrix will be greatly reduced. However, the high-order statistical characteristics of the signal are less affected by time-frequency aliasing and noise, and are an effective method to solve the low signal-to-noise ratio time-frequency aliasing underdetermined blind separation. The cumulant-based method first superimposes the statistical features into a tensor, and then obtains the estimation of the mixing matrix through joint diagonalization or tensor decomposition. Finally, based on the estimated mixing matrix, the signal reconstruction algorithm is used to restore the source signal. Therefore, there are three main research directions at present. The first is how to construct a tensor using the statistical characteristics of the signal, which can not only ensure high mixing matrix estimation accuracy and noise resistance, but also have low computational complexity. The second is the method of joint diagonalization or tensor decomposition. Joint diagonalization or tensor decomposition is essentially still an optimization problem, so finding an optimization algorithm with high solution accuracy and fast convergence speed is a problem that still needs to be studied. The third is the source signal recovery algorithm. After obtaining the mixing matrix estimate, it is still necessary to reconstruct the observed signal according to the mixing matrix to obtain the estimated source signal, i.e., the separated signal, so the source signal recovery algorithm also has a profound impact on the final separation performance. In summary, the present invention proposes an underdetermined blind source separation method based on cumulative tensor decomposition, which can effectively solve the above three problems. Compared with the existing algorithms based on sparsity, cumulant and segmentation, the accuracy of mixing matrix estimation is greatly improved, and the advantage is more obvious at low signal-to-noise ratio. Summary of the invention
[0003] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide an underdetermined blind source separation method and system based on cumulative tensor decomposition.
[0004] To achieve the above object, the present invention adopts the following technical solutions:
[0005] The present invention provides an underdetermined blind source separation method based on cumulative tensor decomposition, comprising the following steps:
[0006] S1. Perform whitening processing on the observed signal to obtain a whitened signal;
[0007] S2, slicing the whitened signal to obtain a sliced signal;
[0008] S3, calculating the delay covariance matrix of multiple whitened signal slice matrices;
[0009] S4, stacking a plurality of the delay covariance matrices into a third-order tensor;
[0010] S5, using weighted nonlinear least squares to perform regularized multivariate decomposition of third-order tensors;
[0011] S6. Calculate the Jacobian operator in weighted nonlinear least squares;
[0012] S7, calculating the gradient in the weighted nonlinear least squares;
[0013] S8, computing the block Jacobian preconditioner in weighted nonlinear least squares;
[0014] S9, when the signal-to-noise ratio is less than 10 dB, select the hybrid vector estimation method 1 to obtain the time-frequency points of the separated signal;
[0015] S10, when the signal-to-noise ratio is greater than 10 dB, select the hybrid vector estimation method 2 to obtain the time-frequency points of the separated signal;
[0016] S11, performing inverse short-time Fourier transform on the time-frequency points of the separated signal to obtain a separated source signal;
[0017] The S3 is specifically:
[0018] No. , Whitened signal slice matrix The delay covariance matrix of is:
[0019] (1);
[0020] in, Indicates Whitened signal slice matrix The delay covariance matrix of It means to find the mean value; represents conjugate transpose; Indicates delay; The mixing matrix representing the whitened signal; represents the source signal slice matrix; represents a slice of Gaussian white noise, that is ; Indicates the total number of signal slices;
[0021] The delay covariance matrix of the source signal slice matrix is:
[0022] (2);
[0023] Simplified to:
[0024] (3);
[0025] in, Represented as a slice of the observed signal;
[0026] In S9, the mixed vector matrix is extracted In Mixed vectors as a set of candidate mixed vector matrices ;
[0027] When the signal-to-noise ratio is less than 10 dB, the hybrid vector estimation method is based on formula (16) from Select the mixing vector matrix of time-frequency points in:
[0028] ;
[0029] (16);
[0030] in, represents the selected mixing vector matrix, represents the time-frequency point of the observed signal, represents the pseudo-inverse, represents the matrix formed by taking the main diagonal, Indicates the 2-norm; ;
[0031] In S10, when the signal-to-noise ratio is greater than 10 dB, the hybrid vector estimation method 2 is based on formula (17) from Select the mixing vector matrix of this time-frequency point:
[0032] (17);
[0033] in, Represents the identity matrix.
[0034] Further, the S4 is specifically: a total of signal slice, then we get The delay covariance matrix of the whitened signal slice matrix is:
[0035] ;
[0036] Among them, the matrix stack is a third-order tensor , represents the set of real numbers or the set of complex numbers, represents the delay covariance matrix of the observed signal slice; represents the delay covariance matrix of the source signal; is the size of each matrix.
[0037] Furthermore, for the third-order tensor Perform regular multivariate decomposition and get:
[0038] (5);
[0039] in, , and is the factor matrix of the regular multivariate decomposition, It is also the mixing matrix of the whitened signal, for The conjugate matrix of Regular multivariate decomposition can yield the mixing matrix , is the number of source signals; The matrices are , and No. column vector;
[0040] For ease of writing, the third-order tensor is decomposed using weighted nonlinear least squares decomposition Perform regular multivariate decomposition, that is:
[0041] (6);
[0042] (7);
[0043] in, and The factor matrix representing the decomposition of the tensor; represents the F-norm; represents the weight tensor;
[0044] Weighted nonlinear least squares uses the dogleg trust region in the Gauss-Newton method to solve the optimization problem, which is:
[0045] (8);
[0046] in, represents the step length;
[0047] , is a third-order tensor rank; , For tensors The dimension of ; Indicates the number of iterations; represents the Jacobi operator; is the trust domain; To vectorize the matrix.
[0048] Further, the Jacobi operator is written as:
[0049] (9);
[0050] in, , represents the sub-Jacobi operator;
[0051] Sub-Jacobi operator Defined as:
[0052] (10);
[0053] in, , , represents a permutation matrix; , represents the identity matrix, Indicates size The identity matrix of
[0054] (11);
[0055] in, For the matrix No. column vector; Representation weight rank.
[0056] Further, in S7, the gradient The calculation formula is:
[0057] (12);
[0058] in,
[0059] Written as:
[0060] (13);
[0061] in, express No. modal.
[0062] Furthermore, in S8, the block Jacobi preconditioner can reduce the number of conjugate gradient iterations and improve the overall convergence. The inverse, Written as:
[0063] (14);
[0064] in, ;
[0065] The pseudo-inverse for:
[0066] (15).
[0067] Furthermore, an underdetermined blind source separation system based on cumulative tensor decomposition is implemented by using the underdetermined blind source separation method based on cumulative tensor decomposition, and further includes:
[0068] Data matrix construction module, used to generate source signals and observation signals;
[0069] A data whitening module, used to construct a whitening matrix according to the sampling covariance matrix, and use the whitening matrix to perform whitening processing on the observation signal data set;
[0070] Signal slicing, used to slice the whitened signal;
[0071] A delay covariance matrix calculation module, used to calculate the delay covariance matrix of the slice matrix;
[0072] Tensor construction module, which stacks multiple delay covariance matrices to form a third-order tensor;
[0073] The tensor decomposition module uses a weighted nonlinear least squares third-order tensor to perform regularized multivariate decomposition to obtain an estimate of the mixing matrix;
[0074] The signal recovery module uses the estimated mixing matrix to obtain the separated signal according to the matrix diagonalization algorithm;
[0075] The performance evaluation module uses the average relative error to evaluate the estimation effect of the mixing matrix and the average similarity coefficient to evaluate the separation performance.
[0076] The beneficial effects of the present invention are: signal slicing can fuse multiple sub-segment information, improving the accuracy and robustness of the algorithm's mixed matrix estimation. The delay covariance matrix of Gaussian white noise is 0, so the third-order tensor stacked by multiple delay covariance matrices has a strong ability to suppress Gaussian white noise. Therefore, the underdetermined joint blind source separation method provided by the present invention can effectively suppress Gaussian white noise, thereby improving the performance of mixed matrix estimation under low signal-to-noise ratio;
[0077] The third-order tensor is regularized and decomposed using weighted nonlinear least squares. Compared with the traditional alternating least squares and nonlinear least squares, the decomposition accuracy of the tensor decomposition is improved without significantly increasing the computational complexity, thereby improving the accuracy of the mixing matrix.
[0078] Since the matrix diagonalization method is used to restore the source signal, this method first estimates the number of source signals existing at each time-frequency point, and then uses the column vector of the mixing matrix to restore the signal. Compared with the traditional subspace projection method and the shortest path method, it has higher reconstruction accuracy, thereby improving the separation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0079] Figure 1 A flowchart of an underdetermined blind source separation method based on cumulative tensor decomposition of the present invention;
[0080] Figure 2 It is four-way source signal;
[0081] Figure 3 are three-way observation signals;
[0082] Figure 4 It is a four-way separation signal;
[0083] Figure 5 Estimation error of the mixing matrix under different signal-to-noise ratios;
[0084] Figure 6 Estimation error of the mixing matrix for different numbers of slices. DETAILED DESCRIPTION
[0085] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0086] See also Figure 1 , an underdetermined blind source separation method based on cumulative tensor decomposition, comprising the following steps:
[0087] S1. Perform whitening processing on the observed signal to obtain a whitened signal;
[0088] S2, slicing the whitened signal to obtain a sliced signal;
[0089] S3, calculating the delay covariance matrix of multiple whitened signal slice matrices;
[0090] S4, stacking a plurality of the delay covariance matrices into a third-order tensor;
[0091] S5, using weighted nonlinear least squares to perform regularized multivariate decomposition of third-order tensors;
[0092] S6. Calculate the Jacobian operator in weighted nonlinear least squares;
[0093] S7, calculating the gradient in the weighted nonlinear least squares;
[0094] S8, computing the block Jacobian preconditioner in weighted nonlinear least squares;
[0095] S9, when the signal-to-noise ratio is less than 10 dB, select the hybrid vector estimation method 1 to obtain the time-frequency points of the separated signal;
[0096] S10, when the signal-to-noise ratio is greater than 10 dB, select the hybrid vector estimation method 2 to obtain the time-frequency points of the separated signal;
[0097] S11, performing inverse short-time Fourier transform on the time-frequency points of the separated signal to obtain a separated source signal;
[0098] The S3 is specifically:
[0099] No. , Whitened signal slice matrix The delay covariance matrix of is:
[0100] (1);
[0101] in, Indicates Whitened signal slice matrix The delay covariance matrix of It means to find the mean value; represents conjugate transpose; Indicates delay; The mixing matrix representing the whitened signal; represents the source signal slice matrix; represents a slice of Gaussian white noise, that is ; Indicates the total number of signal slices;
[0102] The delay covariance matrix of the source signal slice matrix is:
[0103] (2);
[0104] Simplified to:
[0105] (3);
[0106] in, Represented as a slice of the observed signal.
[0107] The S4 is specifically: signal slice, then we get The delay covariance matrix of the whitened signal slice matrix is:
[0108] ;
[0109] Among them, the matrix stack is a third-order tensor , represents the set of real numbers or the set of complex numbers, represents the delay covariance matrix of the observed signal slice; represents the delay covariance matrix of the source signal; is the size of each matrix.
[0110] Furthermore, for the third-order tensor Perform regular multivariate decomposition and get:
[0111] (5);
[0112] in, , and is the factor matrix of the regular multivariate decomposition, It is also the mixing matrix of the whitened signal, for The conjugate matrix of Regular multivariate decomposition can yield the mixing matrix , is the number of source signals; The matrices are , and No. column vector;
[0113] For ease of writing, the third-order tensor is decomposed using weighted nonlinear least squares decomposition Perform regular multivariate decomposition, that is:
[0114] (6);
[0115] (7);
[0116] in, and The factor matrix representing the decomposition of the tensor; represents the F-norm; represents the weight tensor;
[0117] Weighted nonlinear least squares uses the dogleg trust region in the Gauss-Newton method to solve the optimization problem, which is:
[0118] (8);
[0119] in, represents the step length;
[0120] , is a third-order tensor rank; , For tensors The dimension of ; Indicates the number of iterations; represents the Jacobi operator; is the trust domain; To vectorize the matrix.
[0121] The Jacobian operator is written as:
[0122] (9);
[0123] in, , represents the sub-Jacobi operator;
[0124] Sub-Jacobi operator Defined as:
[0125] (10);
[0126] in, , , represents a permutation matrix; , represents the identity matrix, Indicates size The identity matrix of
[0127] (11);
[0128] in, For the matrix No. column vector; Representation weight rank.
[0129] In S7, the gradient The calculation formula is:
[0130] (12);
[0131] in,
[0132] Written as:
[0133] (13);
[0134] in, express No. modal.
[0135] In S8, the block Jacobi preconditioner can reduce the number of conjugate gradient iterations and improve the overall convergence. The inverse, Written as:
[0136] (14);
[0137] in, ;
[0138] The pseudo-inverse for:
[0139] (15).
[0140] In S9, the mixed vector matrix is extracted In Mixed vectors as a set of candidate mixed vector matrices ;
[0141] When the signal-to-noise ratio is less than 10 dB, the hybrid vector estimation method is based on formula (16) from Select the mixing vector matrix of time-frequency points in:
[0142] ;
[0143] (16);
[0144] in, represents the selected mixing vector matrix, represents the time-frequency point of the observed signal, represents the pseudo-inverse, represents the matrix formed by taking the main diagonal, Indicates the 2-norm; .
[0145] In S10, when the signal-to-noise ratio is greater than 10 dB, the hybrid vector estimation method 2 is based on formula (17) from Select the mixing vector matrix of this time-frequency point:
[0146] (17);
[0147] in, Represents the identity matrix.
[0148] An underdetermined blind source separation system based on cumulative tensor decomposition is implemented by an underdetermined blind source separation method based on cumulative tensor decomposition, and also includes:
[0149] Data matrix construction module, used to generate source signals and observation signals;
[0150] A data whitening module, used to construct a whitening matrix according to the sampling covariance matrix, and use the whitening matrix to perform whitening processing on the observation signal data set;
[0151] Signal slicing, used to slice the whitened signal;
[0152] A delay covariance matrix calculation module, used to calculate the delay covariance matrix of the slice matrix;
[0153] Tensor construction module, which stacks multiple delay covariance matrices to form a third-order tensor;
[0154] The tensor decomposition module uses a weighted nonlinear least squares third-order tensor to perform regularized multivariate decomposition to obtain an estimate of the mixing matrix;
[0155] The signal recovery module uses the estimated mixing matrix to obtain the separated signal according to the matrix diagonalization algorithm;
[0156] The performance evaluation module uses the average relative error to evaluate the estimation effect of the mixing matrix and the average similarity coefficient to evaluate the separation performance.
[0157] The effect of the present invention is further described below in conjunction with simulation experiments.
[0158] Simulation experiment
[0159] This paper selects four-channel speech signals as source signals, such as Figure 2 , the sampling frequency is 16000Hz, and the sampling point length is 59904. Matrix mixing and then noise addition generate three-way observation signals, such as Figure 3 The estimated mixing matrix is used to restore the source signal according to the matrix diagonalization method to obtain the separated signal, such as Figure 4 When the signal-to-noise ratio is 20dB, the average relative error of the mixing matrix of the proposed algorithm reaches -19.65dB, and the average absolute correlation coefficient between the recovered source signal and the true source signal reaches 0.95. Figure 5 This is a comparison of the mixed matrix estimation errors of different algorithms under different signal-to-noise ratios. Compared with the traditional underdetermined blind source separation method, the proposed algorithm achieves the best estimation effect under different signal-to-noise ratios. Figure 6 is the mixing matrix estimation error under different slice numbers, indicating that when the number of slices is 160, the mixing matrix estimation error is the smallest.
[0160] The above-mentioned embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be based on the attached claims.
Claims
1. An underdetermined blind source separation method based on cumulative tensor decomposition, characterized in that: The following steps are involved: S1. Perform whitening processing on the observed signal to obtain a whitened signal; S2, slicing the whitened signal to obtain a sliced signal; S3, calculating the delay covariance matrix of multiple whitened signal slice matrices; S4, stacking a plurality of the delay covariance matrices into a third-order tensor; S5, using weighted nonlinear least squares to perform regularized multivariate decomposition of third-order tensors; S6. Calculate the Jacobian operator in weighted nonlinear least squares; S7, calculating the gradient in the weighted nonlinear least squares; S8, computing the block Jacobian preconditioner in weighted nonlinear least squares; S9, when the signal-to-noise ratio is less than 10 dB, select the hybrid vector estimation method 1 to obtain the time-frequency points of the separated signal; S10, when the signal-to-noise ratio is greater than 10 dB, select the hybrid vector estimation method 2 to obtain the time-frequency points of the separated signal; S11, performing inverse short-time Fourier transform on the time-frequency points of the separated signal to obtain a separated source signal; The S3 is specifically: No. , Whitened signal slice matrix The delay covariance matrix of is: (1); in, Indicates Whitened signal slice matrix The delay covariance matrix of It means to find the mean value; represents conjugate transpose; Indicates delay; The mixing matrix representing the whitened signal; represents the source signal slice matrix; represents a slice of Gaussian white noise, that is ; Indicates the total number of signal slices; The delay covariance matrix of the source signal slice matrix is: (2); Simplified to: (3); in, Represented as a slice of the observed signal; In S9, the mixed vector matrix is extracted In Mixed vectors as a set of candidate mixed vector matrices ; When the signal-to-noise ratio is less than 10 dB, the hybrid vector estimation method is based on formula (16) from Select the mixing vector matrix of time-frequency points in: ; (16); in, represents the selected mixing vector matrix, represents the time-frequency point of the observed signal, represents the pseudo-inverse, represents the matrix formed by taking the main diagonal, Indicates the 2-norm; ; In S10, when the signal-to-noise ratio is greater than 10 dB, the hybrid vector estimation method 2 is based on formula (17) from Select the mixing vector matrix of this time-frequency point: (17); in, Represents the identity matrix.
2. The underdetermined blind source separation method based on cumulative tensor decomposition according to claim 1, characterized in that: The S4 is specifically: signal slice, then we get The delay covariance matrix of the whitened signal slice matrix is: ; Among them, the matrix stack is a third-order tensor , represents the set of real numbers or the set of complex numbers, represents the delay covariance matrix of the observed signal slice; represents the delay covariance matrix of the source signal; is the size of each matrix.
3. The underdetermined blind source separation method based on cumulative tensor decomposition according to claim 2, characterized in that: For a third-order tensor Perform regular multivariate decomposition and get: (5); in, , and is the factor matrix of the regular multivariate decomposition, It is also the mixing matrix of the whitened signal, for The conjugate matrix of Regular multivariate decomposition can yield the mixing matrix , is the number of source signals; The matrices are , and No. column vector; For ease of writing, the third-order tensor is decomposed using weighted nonlinear least squares decomposition Perform regular multivariate decomposition, that is: (6); (7); in, and The factor matrix representing the decomposition of the tensor; represents the F-norm; represents the weight tensor; Weighted nonlinear least squares uses the dogleg trust region in the Gauss-Newton method to solve the optimization problem, which is: (8); in, represents the step length; , is a third-order tensor rank; , For tensors The dimension of ; Indicates the number of iterations; represents the Jacobi operator; is the trust domain; To vectorize the matrix.
4. The underdetermined blind source separation method based on cumulative tensor decomposition according to claim 3, characterized in that: The Jacobian operator is written as: (9); in, , represents the sub-Jacobi operator; Sub-Jacobi operator Defined as: (10); in, , , represents a permutation matrix; , represents the identity matrix, Indicates size The identity matrix of (11); in, For the matrix No. column vector; Representation weight rank.
5. The underdetermined blind source separation method based on cumulative tensor decomposition according to claim 4, characterized in that: In S7, the gradient The calculation formula is: (12); in, Written as: (13); in, express No. modal.
6. The underdetermined blind source separation method based on cumulative tensor decomposition according to claim 5, characterized in that: In S8, the block Jacobi preconditioner can reduce the number of conjugate gradient iterations and improve the overall convergence. The inverse, Written as: (14); in, ; The pseudo-inverse for: (15)。 7. An underdetermined blind source separation system based on cumulative tensor decomposition, characterized in that: The method is implemented by using an underdetermined blind source separation method based on cumulative tensor decomposition as claimed in any one of claims 1 to 6, further comprising: Data matrix construction module, used to generate source signals and observation signals; A data whitening module, used to construct a whitening matrix according to the sampling covariance matrix, and use the whitening matrix to perform whitening processing on the observation signal data set; Signal slicing, used to slice the whitened signal; A delay covariance matrix calculation module, used to calculate the delay covariance matrix of the slice matrix; Tensor construction module, which stacks multiple delay covariance matrices to form a third-order tensor; The tensor decomposition module uses a weighted nonlinear least squares third-order tensor to perform regularized multivariate decomposition to obtain an estimate of the mixing matrix; The signal recovery module uses the estimated mixing matrix to obtain the separated signal according to the matrix diagonalization algorithm; The performance evaluation module uses the average relative error to evaluate the estimation effect of the mixing matrix and the average similarity coefficient to evaluate the separation performance.
