Underdetermined blind source separation method and device based on structured sparse subspace clustering
The structured sparse subspace clustering method improves the accuracy and robustness of underdetermined blind source separation by employing time-frequency domain transformation and spatial projection for precise signal recovery in complex environments.
Patent Information
- Application Number
- CN202111239956.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-25
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-10-25
AI Technical Summary
Existing underdetermined blind source separation methods suffer from insufficient accuracy, weak robustness, and a narrow scope of application, particularly in scenarios where the number of source signals is unknown and greater than the number of observed signals.
An underdetermined blind source separation method based on structured sparse subspace clustering, involving time-frequency domain transformation, structured sparse subspace clustering to determine signal sources and clusters, improved potential function methods for mixing matrix estimation, and spatial projection with minimum l1 norm for signal reconstruction.
Enhances accuracy and robustness in determining the mixing matrix, allowing for precise signal recovery and wider applicability across various signal types, including speech and biomedical signals.
Smart Images

Figure CN113962265B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal processing, and in particular to an underdetermined blind source separation method and device based on structured sparse subspace clustering. Background Art
[0002] Blind source separation refers to a type of problem in which the source signal and its mixing mode are unknown, and the source signal is only separated from the mixed signal observed by the sensor. Since this problem is widely present in practical engineering applications, it has become a research hotspot in recent years. It is often used in speech signal processing, biomedical signal processing, image processing, fault diagnosis and other fields.
[0003] In practice, due to the complexity of the engineering environment and for cost control considerations, the number of observation sensors is often known, while the number of observed source signals is usually unknown and greater than the number of observed signals. Blind source separation in this case is called underdetermined blind source separation. Determining the number of source signals by observing signals plays an important role in solving the underdetermined blind source separation problem.
[0004] At present, the research on solving underdetermined blind source separation mainly adopts sparse component analysis method, which takes advantage of the sparsity of the signal, that is, only one signal plays a dominant role in a specific period of time in the frequency domain or time-frequency domain, and the amplitude of other signals is 0 or close to 0. It usually includes two key steps: mixing matrix estimation and source signal recovery. Mixing matrix estimation is the prerequisite for accurate signal recovery. Many studies determine the mixing matrix based on clustering methods, but these methods often require the number of source signals to be given as a known quantity. Sensitivity to noise and weak robustness are also challenges faced by current mixing matrix estimation methods. Since the mixing matrix is a non-square matrix and only satisfies the full rank of the columns, it cannot be directly inverted to obtain the reconstructed signal. Existing methods usually rely on seeking a solution that is as sparse as possible in the transform domain, and then restoring the signal to the time domain through inverse transformation to complete the signal estimation and reconstruction. Summary of the invention
[0005] Based on the above analysis, it can be seen that the underdetermined blind source separation problem urgently requires a signal estimation and reconstruction method with high accuracy, strong robustness and wide applicability.
[0006] A major problem solved by the present invention is that the existing underdetermined blind source separation method has insufficient accuracy, weak robustness and a narrow scope of application.
[0007] The present invention provides an underdetermined blind source separation method based on structured sparse subspace clustering, the method comprising the following steps:
[0008] The source signal is transformed into the time-frequency domain, and the time-frequency interval windowing method is used to remove low-energy points and screen out single source points;
[0009] Using structured sparse subspace clustering to explore the subspace distributed in the single source points, determine the number of signal sources and the cluster to which each single source point belongs;
[0010] An improved potential function method is used to find cluster centers, and a mixing matrix is estimated based on the cluster centers;
[0011] Based on the mixing matrix, reconstructing the source signal using a spatial projection method based on source number estimation combined with a minimum l1 norm method;
[0012] The reconstructed source signal in the time-frequency domain is restored to the time domain using the inverse short-time Fourier transform to obtain the estimated time-domain source signal.
[0013] Furthermore, structured sparse subspace clustering is used to explore the subspace distributed in the single source point to determine the number of signal sources and the cluster to which each single source point belongs, including the steps of:
[0014] In the initial state, the classification position of the single source point is unknown, and the model of structured sparse subspace clustering does not contain structured sparse norm terms. The model of structured sparse subspace clustering is expressed as:
[0015]
[0016] stX=XC+E,diag(C)=0
[0017] Among them, X represents the data set composed of single source points, C is the self-representation coefficient matrix of the single source point, E represents the noise matrix, λ>0 is the balance parameter, represents the regularization constraint on Gaussian noise, ||C||1 represents the l1 norm of the self-representation coefficient matrix of a single source point;
[0018] According to the model, the alternating direction multiplier method is used to obtain the self-representation coefficient matrix. Based on the self-representation coefficient matrix, spectral clustering is used to perform initial category division to determine the number of source signals and the cluster to which each single source point belongs.
[0019] Furthermore, the step of using spectral clustering to perform initial category division and determine the number of source signals includes:
[0020] Based on the first mathematical formula, we can find the similarity matrix:
[0021]
[0022] Among them, C T represents the transpose of the self-representation coefficient matrix C, and W is the similarity matrix;
[0023] Find the normalized Laplace matrix based on the second mathematical formula:
[0024]
[0025] Among them, L sym is a standardized Laplace matrix, I is the identity matrix, D is the degree matrix, specifically a diagonal matrix, and the elements on the diagonal can be obtained based on the fourth mathematical formula:
[0026]
[0027] For the matrix L sym Perform eigendecomposition and arrange the eigenvalues in order from small to large, and define the eigenvalue gap as:
[0028] Δλ m =λ m+1 -λ m ,
[0029] Among them, λ m+1 and λ m represents two adjacent eigenvalues in the sequentially arranged eigenvalue sequence, Δλ m Represents the difference between adjacent eigenvalues, that is, the eigenvalue gap;
[0030] The value of the subscript corresponding to the first maximum eigenvalue gap is the number of source signals, which is also the number of clusters.
[0031] Further, based on the self-representation coefficient matrix, spectral clustering is used to perform initial category division, thereby determining the number of source signals and the cluster to which each single source point belongs, including the steps of:
[0032] The Laplacian matrix is used as the input of spectral clustering based on the number of clusters, and the self-representation coefficient matrix is cut based on the canonical cut set criterion to obtain the partition of the cluster of any single source point.
[0033] Furthermore, after performing initial category division based on the self-representation coefficient matrix using spectral clustering, the method further includes the following steps:
[0034] The structured sparse subspace clustering framework and spectral clustering iteration are used to optimize the partition of the clusters of single source points. The steps of the iteration are:
[0035] Based on the initial classification results of the cluster of the single source point, the structured sparse norm term ||C is calculated Q , the calculation formula is the sixth mathematical formula:
[0036]
[0037] Among them, C ij Represents the element of the i-th row and j-th column of the self-representation coefficient matrix C of a single source point, and defines the partition matrix Q = [q1; q2; ..., qn ], reflecting the relationship between the degree of belonging of each sample data point and the cluster. The matrix Q only contains two values, 0 and 1. If the i-th single source point x i Distributed in the jth cluster S j In, q ij =1, otherwise q ij =0,q i and q j Represents the i-th and j-th row vectors of the partition matrix;
[0038] Determine the structured sparse norm term ||C|| Q After that, the new optimization model of structured sparse subspace clustering is obtained as the seventh mathematical formula:
[0039]
[0040] stX=XC+E,diag(C)=0,
[0041] Among them, α>0 is the equilibrium parameter;
[0042] The new self-representation coefficient matrix under the new optimization model is obtained by the alternating direction multiplication method, and the required input parameters of spectral clustering are calculated according to the second mathematical formula and the third mathematical formula, and then the division of the cluster of any single source point is updated through the spectral clustering loop iteration.
[0043] Furthermore, the method of using the structured sparse subspace clustering framework and the spectral clustering loop iteration further includes the steps of:
[0044] After each division of the cluster of the single source point is obtained, it is determined whether the number of iterations is not less than the set value. If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation is continued in a loop;
[0045] Alternatively, it is determined whether the error calculated by the alternating direction multiplier method is less than a set value. If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation is continued.
[0046] Furthermore, the improved potential function method is used to find the cluster center, and the mixing matrix is estimated based on the cluster center, including the steps of:
[0047] The number of signal sources is taken as the number of clusters, and the number of data in each cluster is {c1, c2, ..., c i}, c r is the number of data in cluster r, x r j is an element in cluster r. For any element in the cluster r Its local potential function is the eighth mathematical formula:
[0048]
[0049] Among them, the basis function f(x, y) is defined as a piecewise function, and the calculation formula is the ninth mathematical formula:
[0050]
[0051] Among them, θ(x, y) is the cosine value of the angle between any two vectors, that is, The smaller θ(x, y) is, the closer any two vectors are to each other, and ζ is the coefficient of the adjustment basis function;
[0052] The local potential values of all single-source points in the cluster are calculated, and the point with the maximum local potential value is the cluster center, and the combination of all cluster centers is the estimated mixing matrix.
[0053] Further, based on the mixing matrix, a spatial projection method based on source number estimation is used in combination with a minimum The source signal is reconstructed using the norm method, further comprising the steps of:
[0054] Assume that the set Λ is the estimated mixing matrix The set of 1 to m-1 order sub-matrices, let sub-matrix A k ∈Λ, calculate each time-frequency signal point x t (t, f) in submatrix A k The projection parameter on the spanned subspace is expressed as the tenth mathematical formula:
[0055]
[0056] Among them, t and f represent the time serial number and frequency serial number of the time-frequency signal point respectively. Represents the matrix A k The transpose of represents the inverse matrix;
[0057] A projection parameter threshold is set. If there are multiple projection parameters calculated that are greater than the projection parameter threshold, the mixed column vector corresponding to the largest projection parameter is selected as the mixed matrix of the time-frequency point. At this time, the calculation expression of the restored time-frequency signal is the eleventh mathematical formula:
[0058]
[0059] in, Represents the vector of signals whose values in the recovered time-frequency signal are not 0, represents the elements in a vector, Represents a vector of signals with a value of 0 in the restored time-frequency signal;
[0060] If any submatrix in the set Λ cannot obtain an observation point whose calculated projection parameter is greater than the projection parameter threshold, then assuming that the number of actual source signals is the same as the number of observed signals, the minimum The norm method calculates all the recovered time-frequency signals, and the calculation formula is the twelfth mathematical formula:
[0061]
[0062] Among the solutions obtained by the twelfth mathematical formula, the solution satisfying the thirteenth mathematical formula is considered to be the restored time-frequency signal. The thirteenth mathematical formula is:
[0063]
[0064] Among them, the source signal combination is the estimate of the source signal;
[0065] Determine projection parameters for all observed signals, and calculate based on the eleventh mathematical formula or the twelfth mathematical formula to obtain a complete time-frequency recovery signal;
[0066] Perform short-time inverse Fourier transform on the complete time-frequency restored signal to obtain the reconstructed source signal.
[0067] According to another aspect of the present invention, an underdetermined blind source separation device based on structured sparse subspace clustering is also disclosed. The underdetermined blind source separation device includes a memory, a processor, and an underdetermined blind source separation program based on structured sparse subspace clustering stored in the memory and executable on the processor. When the underdetermined blind source separation program based on structured sparse subspace clustering is executed by the processor, the steps of the underdetermined blind source separation method based on structured sparse subspace clustering as described in any of the preceding items are implemented.
[0068] The technical solution of the present invention can achieve the following beneficial effects:
[0069] (1) The present invention utilizes the short-term stationary characteristics of the signal and uses the windowing method to split the time-frequency domain into subintervals for high-energy point screening and single-source point screening. This method can reduce the amount of computation of the subsequent optimization algorithm, can handle the time-frequency points obtained by a variety of time-frequency transformation methods, and has a wider range of applications.
[0070] (2) The present invention uses structured sparse subspace clustering and improved potential function methods to solve the mixing matrix. Structured sparse subspace clustering can discover low-dimensional subspaces from high-dimensional observation signals, determine the number of source signals, and obtain accurate data division. The improved potential function method overcomes the disadvantage that the traditional potential function method cannot be applied to high-dimensional observation signals, and can achieve higher accuracy in determining the column vector of the mixing matrix.
[0071] (3) The present invention utilizes the fusion space projection method and the minimum The norm method reconstructs the source signal and changes the number of active source signals at each time-frequency point from a fixed value with an upper limit of m-1 to 1 to m, thereby improving the accuracy of signal recovery. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] The accompanying drawings, which constitute a part of the specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.
[0073] Figure 1 A schematic flow chart of an underdetermined blind source separation method based on structured sparse subspace clustering provided by an embodiment of the present invention;
[0074] Figure 2 A time-frequency point scatter plot after single source point screening and low energy removal provided in an embodiment of the present invention;
[0075] Figure 3 A characteristic value gap sequence diagram provided by an embodiment of the present invention;
[0076] Figure 4 A time-frequency point classification scatter plot obtained through structured sparse subspace clustering provided in an embodiment of the present invention;
[0077] Figure 5 A schematic diagram of cluster centers determined by a local potential function method provided in an embodiment of the present invention;
[0078] Figure 6 A comparison diagram of a set of source signals and reconstructed signals provided by an embodiment of the present invention; DETAILED DESCRIPTION
[0079] Various exemplary embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention.
[0080] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.
[0081] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.
[0082] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.
[0083] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.
[0084] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.
[0085] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0086] Embodiment 1, as Figure 1 As shown, it is a flow chart of an underdetermined blind source separation method based on structured sparse subspace clustering, which specifically includes the following steps:
[0087] The source signal is transformed into the time-frequency domain, and the time-frequency interval windowing method is used to remove low-energy points and screen out single source points;
[0088] Structured sparse subspace clustering is used to explore the subspace distributed in the single source point, determine the number of signal sources and the cluster to which each single source point belongs, use the improved potential function method to find the cluster center, and estimate the mixing matrix based on the cluster center;
[0089] Based on the mixing matrix, a spatial projection method based on source number estimation is used in combination with the minimum The norm method is used to reconstruct the source signal;
[0090] The reconstructed source signal in the time-frequency domain is restored to the time domain using the inverse short-time Fourier transform to obtain the estimated time-domain source signal.
[0091] In some embodiments, structured sparse subspace clustering is used to explore the subspace distributed in the single source point, and the steps of determining the number of signal sources and the cluster to which each single source point belongs are: in the initial state, the classification position of the single source point is unknown, and the model of structured sparse subspace clustering at this time does not contain a structured sparse norm term, so the model of structured sparse subspace clustering can be expressed as:
[0092]
[0093] Among them, X represents a data set composed of single-source time-frequency points, C is the self-representation coefficient matrix of the single-source point, E represents the noise matrix, and λ>0 is the balance parameter. represents the regularization constraint on Gaussian noise, and ||C||1 represents the l1 norm of the self-representation coefficient matrix of a single source point.
[0094] According to the optimization model, the alternating direction multiplier method is used to obtain the self-representation coefficient matrix C. Based on the obtained self-representation coefficient matrix, spectral clustering is used to perform initial classification of single source points to determine the number of source signals and the cluster to which each single source point belongs.
[0095] In some embodiments, the above steps use spectral clustering to perform initial category division, and the step of determining the number of source signals is specifically as follows:
[0096] First find the similarity matrix:
[0097]
[0098] Among them, C T represents the transpose of the self-representation coefficient matrix C, and W is the similarity matrix;
[0099] Then find the standardized Laplace matrix:
[0100]
[0101] Among them, L sym is a standardized Laplace matrix, I is the identity matrix, D is the degree matrix, and is specifically a diagonal matrix. The elements on the diagonal can be obtained based on the following formula:
[0102]
[0103] For the matrix L sym Perform eigendecomposition and arrange the eigenvalues in order from small to large, and define the eigenvalue gap as:
[0104] Δλ m =λ m+1 -λ m (5)
[0105] Among them, λ m+1 and λ m represents two adjacent eigenvalues in the sequentially arranged eigenvalue sequence, Δλ m Represents the difference between adjacent eigenvalues, that is, the eigenvalue gap;
[0106] The value of the subscript corresponding to the first maximum eigenvalue gap is the number of source signals, that is, the number of clusters.
[0107] In some embodiments, based on the self-representation coefficient matrix, spectral clustering is used to perform initial classification, so as to determine the number of source signals and the cluster to which each single source point belongs.
[0108] For the Laplacian matrix, the number of clusters is taken as a known quantity and as the input of spectral clustering. The self-representation coefficient matrix is cut based on the canonical cut set criterion so that the connection weight within the subgraph is as large as possible and the connection weight between subgraphs is as small as possible, thereby achieving the division of data categories and obtaining the division of clusters of all single-source points.
[0109] In some embodiments, after the single source points are initially classified using spectral clustering, the method further includes the following steps:
[0110] The structured sparse subspace clustering framework and spectral clustering iteration are used to optimize the partition of clusters of single source points.
[0111] In some embodiments, the specific steps of the loop iteration in the above steps are:
[0112] Based on the initial classification results of the clusters of single source points, the structured sparse norm term ||C|| is calculated. Q , the structured sparse norm term is used to measure the inconsistency between the self-representation matrix and the segmentation matrix Q. The specific calculation formula is:
[0113]
[0114] Among them, C ij Represents the element of the i-th row and j-th column of the self-representation coefficient matrix C of a single source point, and defines the partition matrix Q = [q1; q2; ..., q n ], reflecting the relationship between the degree of belonging of each sample data point and the cluster. The matrix Q only contains two values, 0 and 1. If the i-th single source point x i Distributed in the jth cluster S j In, q ij =1, otherwise q ij =0,q i and q j Represents the i-th and j-th row vectors of the partition matrix.
[0115] Determine the structured sparse norm term ||C|| Q After that, the new optimization model of structured sparse subspace clustering is obtained:
[0116]
[0117] Among them, α>0 is the balance parameter; the new self-representation coefficient matrix under the new optimization model is obtained by the alternating direction multiplier method and the required input parameters of spectral clustering are calculated according to formula (2) and formula (3), and then the division of the cluster of any single source point is updated through the spectral clustering loop iteration.
[0118] In some embodiments, using a structured sparse subspace clustering framework and a spectral clustering loop iteration further includes the steps of:
[0119] After each division of the cluster of the single source point is obtained, it is determined whether the number of iterations is not less than the set value. If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation is continued in a loop;
[0120] Or determine whether the error output by the alternating direction multiplier method is less than the set value (the alternating direction multiplier method requires a domain value to be set manually. If the error value of the calculated result is less than the set value, the algorithm is considered to be over. If it is greater than the set value, the algorithm continues to execute). If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation continues.
[0121] In some embodiments, the improved potential function method is used to find the cluster centers, and the mixing matrix is estimated based on the cluster centers, further comprising the steps of:
[0122] The number of signal sources is taken as the number of clusters, and the number of data in each cluster is {c1, c2, ..., c i},c r is the number of data in cluster r. For any element in cluster r Its local potential function is:
[0123]
[0124] Among them, the basis function f(x, y) is defined as a piecewise function, and the calculation formula is:
[0125]
[0126] Among them, θ(x, y) is the cosine value of the angle between any two vectors, that is, The smaller θ(x, y) is, the closer any two vectors are to each other, and ζ is the coefficient of the adjustment basis function;
[0127] The local potential values of all single-source points in the cluster are calculated. For a certain vector, the more vectors have smaller deviations from its direction, the larger its local potential value is, and the point has more potential to become the cluster center. Therefore, in this embodiment, the point with the largest local potential value is taken as the cluster center, and all cluster centers are combined into an estimated mixing matrix.
[0128] In some embodiments, based on the estimated mixing matrix, a spatial projection method based on the source number estimation is used in combination with a minimum The steps of reconstructing the source signal using the norm method specifically include:
[0129] Assume that the set Λ is the estimated mixing matrix The set of 1 to m-1 order sub-matrices, let sub-matrix A k ∈Λ, calculate each time-frequency signal point x t (t, f) in submatrix Ak The projection parameter on the spanned subspace is expressed as:
[0130]
[0131] Among them, t and f represent the time serial number and frequency serial number of the time-frequency signal point respectively. Represents the matrix A k The transpose of represents the inverse matrix;
[0132] The projection parameter threshold μ is set. To ensure the effect, 0.9≤μ<1 is generally taken. The projection parameters are calculated for each combination of mixed column vectors in the set Λ. If there are multiple calculated projection parameters greater than the projection parameter threshold, the mixed column vector corresponding to the largest projection parameter is selected as the mixing matrix of the time-frequency point. At this time, the calculation expression of the restored time-frequency signal is:
[0133]
[0134] in, Represents the vector of signals whose values in the recovered time-frequency signal are not 0, represents the elements in a vector, Represents a vector of signals with a value of 0 in the restored time-frequency signal;
[0135] If any submatrix in the set Λ cannot obtain an observation point whose calculated projection parameter is greater than the projection parameter threshold, then it is assumed that the number of actual source signals is the same as the number of observed signals and the minimum The norm method calculates all possible restored time-frequency signals, and the calculation formula is:
[0136]
[0137] Among the solutions obtained by formula (12), the solution satisfying formula (13) is considered to be the recovered time-frequency signal:
[0138]
[0139] Among them, the source signal combination is the estimate of the source signal;
[0140] Determine the projection parameters for all observed signals, and calculate the complete time-frequency recovery signal based on formula (11) or (12), that is, obtain the observation values of all source signals in the time-frequency domain;
[0141] Perform short-time inverse Fourier transform on the complete time-frequency restored signal to obtain the reconstructed signal in the time domain.
[0142] In some embodiments, any of the aforementioned underdetermined blind source separation methods may be applicable to signals including but not limited to speech signals, mechanical vibration signals, or biomedical signals.
[0143] Embodiment 2: A speech signal is selected as a source signal to specifically describe an underdetermined blind source separation method based on structured sparse subspace clustering.
[0144] Step 1: Select the speech signal. The blind source separation model for underdetermined instantaneous mixture in the time domain can be expressed as:
[0145]
[0146] Where, s(t) = [s1(t), s2(t), …, s n (t)] T represents n source signals, x(t)=[x1(t), x2(t),…, x m (t)] T represents m-path observation signals. A=[a1,a2,...,a n ]∈R M×n (m<n) represents the unknown mixing matrix. Due to the poor signal sparsity in the time domain, the time domain signal x(t) is converted to the sparse time-frequency domain through short-time Fourier transform to obtain X(t, f). When sparsity is guaranteed, if only one source plays a dominant role in a certain time-frequency interval, and the values of the other source signals are 0 or close to 0, the observation point at this time is called a single source point. The single source point exhibits linear clustering characteristics in the time-frequency domain, and because the speech signal usually has a stable characteristic in a short time, the time-frequency domain of the observed signal is cut into sub-intervals by the windowing method to screen high-energy single source points.
[0147] The time-frequency window with a time interval length of a and a frequency interval length of b is used to cut the time-frequency domain of m observation signals. i , time window [t m , t m+a The time-frequency blocks within the range of ] can be recorded as Ω1, Ω2, ..., Ω n , calculate the energy of each time-frequency block:
[0148]
[0149] Ω i represents the i-th time-frequency block, a and b are the parameters of the time-frequency interval window, the energy values of the time-frequency blocks in the same time window are arranged in descending order, and the first δ required are taken as the candidate energy intervals, which are recorded as Π(X i ), X i (t α , f β ) is the i-th time-frequency point.m The intersection of the energy candidate intervals is calculated, and the points in the intersection interval are used as the high energy points after screening:
[0150]
[0151] The set of all high energy points is denoted as Π. For the time-frequency point Single source points are screened by judging the angle between the real and imaginary parts. Ideally, the projection value between the real vector and the imaginary vector should be 1. Considering the existence of noise and interference in actual situations, in order to enhance the robustness of the single source point screening method, the parameter Relax the restrictions on the real and imaginary projection values:
[0152]
[0153] The parameters a and b of the time-frequency interval window are generally integers less than 10; δ is adjusted according to the noise intensity contained in the signal. When the noise is strong, its value takes a smaller integer; The value range is: The time-frequency point scatter plot after single source point screening and low energy removal is as follows Figure 2 shown.
[0154] Step 2: For the signal after step 1, the retained time-frequency points are used as input data for structured sparse subspace clustering.
[0155] Structured sparse subspace clustering combines the sparse representation process of separation with the spectral clustering process to form a joint optimization framework. Since the number of source signals in the underdetermined blind source separation problem is greater than the number of mixed signals, in the scatter plot, a single source point is represented as a point distributed on a one-dimensional straight line in a high-dimensional space with the observed signal as the coordinate axis, and the slope of the line corresponds to the column vector of the mixing matrix. Since structured sparse subspace clustering can explore the potential low-dimensional data structure in the high-dimensional observed signal, it can obtain better clustering performance and accurately determine the number of source signals. Therefore, it can achieve better performance in dealing with the mixing matrix estimation problem of blind source separation.
[0156] The optimization model of structured sparse subspace clustering is shown in formula (18):
[0157]
[0158]
[0159] Among them, X represents a data set composed of single-source time-frequency points, C is the self-representation coefficient matrix of the single-source point, and E represents the noise matrix. In the underdetermined blind source separation problem, the regularization constraint term of the noise is Choose as the Frobenius norm, ||C|| Qrepresents the structured sparse norm, which is used to measure the inconsistency between the self-representation matrix and the segmentation matrix Q, and can be obtained by formula (19):
[0160]
[0161] Q=[q1;q2;...,q n ] is a binary matrix reflecting the relationship between the degree of belonging of each sample data point and the cluster center. If the sample data x i Distributed in subspace S j In, q ij =1, otherwise q ij =0.
[0162] The optimization model is solved by means of the alternating direction multiplier method, which will not be described in detail here. The use of structured sparse subspace clustering to classify data includes the following steps:
[0163] Step 2.1: Initialization, obtain the number of source signals and calculate the initial structured sparse norm term;
[0164] In the initial state, the data partitioning method is unknown. In this case, the structured sparse norm term is omitted, and the optimization model can be expressed as:
[0165]
[0166] After the self-representation matrix C is obtained according to the optimization model, spectral clustering is used to perform the initial classification of categories and determine the number of source signals. First, the similarity matrix W and the standardized Laplace matrix L are obtained. sym :
[0167]
[0168]
[0169] Where D is the degree matrix,
[0170] For the matrix L sym Perform eigendecomposition and arrange the eigenvalues in order from small to large λ1≤λ2≤...≤λ n When the number of clusters is i, the first i eigenvalues are close to 0, and the i+1th eigenvalue increases significantly. Therefore, the number of clusters is determined by searching for the characteristic gaps between the eigenvalues, that is, the number of source signals in the underdetermined blind source separation problem.
[0171] Determine the structured sparse norm term ||C|| QFinally, a new optimization model of structured sparse subspace clustering is obtained. Then, the new self-representation coefficient matrix under the new optimization model is obtained by the alternating direction multiplication method, and the similarity matrix and Laplace matrix of the required input parameters of spectral clustering are calculated according to the above formula. Then, the division of the cluster of any single source point is updated through the spectral clustering loop iteration.
[0172] The gap between eigenvalues is defined as: Δλ m =λ m+1 -λ m , the subscript corresponding to the first maximum eigenvalue gap is the number of clusters. The schematic diagram of determining the number of source signals by eigenvalue gap is as follows Figure 3 shown.
[0173]
[0174] Then, the number of clusters is taken as a known quantity, and the single source point self-representation coefficient matrix C is cut and classified according to the canonical cut set criterion, so that the connection weight within the subgraph is as large as possible, and the connection weight between subgraphs is as small as possible, thereby realizing the division of data categories. Then, the initial partition matrix is generated, and the structured sparse norm term is calculated according to formula (19).
[0175] Step 2.2: Fix the partition matrix Q and solve the above optimization model (18) to obtain the self-representation coefficient matrix C;
[0176] Step 2.3: Fix the self-representation coefficient matrix C and use spectral clustering to find the category division and segmentation matrix Q;
[0177] Step 2.4: Determine whether the convergence condition is met or whether the number of iterations reaches the set maximum value. If one of them is met, the algorithm ends. If not, return to step 2.2. After iteration, the time-frequency point classification diagram obtained by structured sparse subspace clustering is as follows: Figure 4 shown.
[0178] Through structured sparse subspace clustering, we can get the number of source signals and the cluster to which each observation point belongs. Assume that the number of data in each cluster is {c1, c2, ..., c i}. Assume that for a cluster r, the number of data in it is c r , then for any element in the cluster Its local potential function can be calculated as follows:
[0179]
[0180] The basis function f(x, y) selects a piecewise function, and the calculation formula is as shown in formula (25):
[0181]
[0182] θ(x, y) is the cosine of the angle between any two vectors, that is The smaller θ(x, y) is, the closer the two vectors are. ζ is the coefficient of adjusting the basis function.
[0183] For a certain vector, the more vectors with small deviations from its direction, the greater its local potential is, and the point has more potential to become a cluster center. The point with the largest local potential is considered to be the cluster center, that is, the column vector of the mixing matrix. The combination of all cluster centers is the estimated mixing matrix The cluster centers determined by the local potential function method are as follows Figure 5 shown.
[0184] Step 3: After step 2, As a known quantity, it participates in the source signal recovery process and uses the spatial projection method based on source number estimation combined with the minimum The norm method reconstructs the source signal:
[0185] Assume that the set Λ is the estimated mixing matrix The set of 1 to m-1 order sub-matrices, let sub-matrix A k ∈Λ, any observation point x t (t, f) in A k The size of the projection of the spanned subspace is defined by the parameter D k :
[0186]
[0187] Set the threshold μ. To ensure the effect, generally take 0.9≤μ<1. Calculate the projection parameter D for each combination of mixed column vectors in the set Λ. If there are multiple cases greater than the threshold μ, select the column vector corresponding to the maximum D as the mixing matrix of the time-frequency point:
[0188] max(D k )≥μk∈{1, 2,...,l} (27)
[0189] The mixing matrix at this time is recorded as A k , the recovered time-frequency signal can be obtained as shown in formula (28):
[0190]
[0191] For observation points where any set of column vectors in the set Λ cannot satisfy the projection greater than μ, it is assumed that the number of source signals actually acting is the same as the number of observed signals, and the minimum The norm method is used to find the restored time-frequency signal, and all possible combinations of source signals are calculated using equation (29):
[0192]
[0193] The source signal combination that satisfies equation (30) This is the estimate of the source signal:
[0194]
[0195] The above operations are performed on all time-frequency points to obtain the observation values of all source signals in the time-frequency domain.
[0196] Step 4: Perform inverse short-time Fourier transform on the source signal in the time-frequency domain to obtain the reconstructed signal in the time domain like Figure 6 As shown, there is a comparison diagram of a set of source signals and reconstructed signals.
[0197] When the source signal is a mechanical vibration signal or a biomedical signal, the corresponding underdetermined blind source processing flow is similar to that of the present embodiment and will not be described in detail herein.
[0198] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
[0199] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
Claims
1. An underdetermined blind source separation method based on structured sparse subspace clustering, It is characterized in that The method comprises the following steps: The source signal is transformed into the time-frequency domain, and the time-frequency interval windowing method is used to remove low-energy points and screen out single source points; Using structured sparse subspace clustering to explore the subspace distributed in the single source points, determine the number of signal sources and the cluster to which each single source point belongs; An improved potential function method is used to find cluster centers, and a mixing matrix is estimated based on the cluster centers; Based on the mixing matrix, a spatial projection method based on source number estimation is used in combination with the minimum l 1 Reconstructing the source signal using a norm method; The reconstructed source signal in the time-frequency domain is restored to the time domain using the inverse short-time Fourier transform to obtain an estimated time-domain source signal; the structured sparse subspace clustering is used to explore the subspace distributed in the single source point to determine the number of signal sources and the cluster to which each single source point belongs, including the steps of: In the initial state, the classification position of the single source point is unknown, and the model of structured sparse subspace clustering does not contain structured sparse norm terms. The model of structured sparse subspace clustering is expressed as: stX=XC+E,diag(C)=0 Among them, X represents a data set composed of single source points, C is the self-representation coefficient matrix of the single source point, E represents the noise matrix, and λ>0 is the balance parameter. represents the regularization constraint on Gaussian noise, ||C|| 1 l represents the self-representation coefficient matrix of a single source point 1 norm; According to the model, the alternating direction multiplication method is used to obtain the self-representation coefficient matrix, and based on the self-representation coefficient matrix, spectral clustering is used to perform initial category division, so as to determine the number of source signals and the cluster to which each single source point belongs; The step of using spectral clustering to perform initial category division and determine the number of source signals includes: Based on the first mathematical formula, we can find the similarity matrix: Among them, C T represents the transpose of the self-representation coefficient matrix C, and W is the similarity matrix; Find the normalized Laplace matrix based on the second mathematical formula: Among them, L sym is a standardized Laplace matrix, I is the identity matrix, D is the degree matrix, specifically a diagonal matrix, and the elements on the diagonal are obtained based on the fourth mathematical formula: For the matrix L sym Perform eigendecomposition and arrange the eigenvalues in order from small to large, and define the eigenvalue gap as: Dl m =λ m+1 -l m , Among them, λ m+1 and λ m represents two adjacent eigenvalues in the sequentially arranged eigenvalue sequence, Δλ m Represents the difference between adjacent eigenvalues, that is, the eigenvalue gap; The value of the subscript corresponding to the first maximum eigenvalue gap is the number of source signals, which is also the number of clusters; After performing initial category division based on the self-representation coefficient matrix using spectral clustering, the method further includes the following steps: The structured sparse subspace clustering framework and spectral clustering iteration are used to optimize the partition of the clusters of single source points. The steps of the iteration are: Based on the initial classification results of the clusters of single source points, the structured sparse norm term ||C|| is calculated. Q , the calculation formula is the sixth mathematical formula: Among them, C ij The element of the i-th row and j-th column of the self-representation coefficient matrix C of a single source point is defined as the partition matrix Q = [q 1 ;q 2 ;...,q n ], reflecting the relationship between the degree of belonging of each sample data point and the cluster. The matrix Q only contains two values, 0 and 1. If the i-th single source point x i Distributed in the jth cluster S j In, q ij =1, otherwise q ij =0,q i and q j Represents the i-th and j-th row vectors of the partition matrix; Determine the structured sparse norm term ||C|| Q After that, the new optimization model of structured sparse subspace clustering is obtained as the seventh mathematical formula: stX=XC+E,diag(C)=0, Among them, α>0 is the equilibrium parameter; The new self-representation coefficient matrix under the new optimization model is obtained by the alternating direction multiplication method, and the required input parameters of spectral clustering are calculated according to the second mathematical formula and the fourth mathematical formula, and then the division of the cluster of any single source point is updated through the spectral clustering loop iteration.
2. The underdetermined blind source separation method based on structured sparse subspace clustering according to claim 1, It is characterized in that Based on the self-representation coefficient matrix, spectral clustering is used to perform initial classification, thereby determining the number of source signals and the cluster to which each single source point belongs, including the steps of: The Laplacian matrix is used as the input of spectral clustering based on the number of clusters, and the self-representation coefficient matrix is cut based on the canonical cut set criterion to obtain the partition of the cluster of any single source point.
3. The underdetermined blind source separation method based on structured sparse subspace clustering as claimed in claim 2, It is characterized in that The method uses a structured sparse subspace clustering framework and a spectral clustering loop iteration, and further includes the steps of: After each division of the cluster of the single source point is obtained, it is determined whether the number of iterations is not less than the set value. If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation is continued in a loop; Or it is determined whether the error output by the alternating direction multiplier method is less than a set value. If so, the final division of the cluster of the single source point is obtained. If not, the iterative calculation is continued.
4. The underdetermined blind source separation method based on structured sparse subspace clustering according to claim 1, It is characterized in that The method of finding cluster centers by using the improved potential function method and estimating the mixing matrix based on the cluster centers comprises the following steps: The number of signal sources is taken as the number of clusters, and the number of data in each cluster is {c 1 ,c 2 ,...,c i }, c r is the number of data in cluster r, x r j is an element in cluster r. For any element in the cluster r Its local potential function is the eighth mathematical formula: Among them, the basis function f(x,y) is defined as a piecewise function, and the calculation formula is the ninth mathematical formula: Among them, θ(x,y) is the cosine value of the angle between any two vectors, that is, The smaller θ(x,y) is, the closer any two vectors are to each other, and ζ is the coefficient of the adjustment basis function; The local potential values of all single-source points in the cluster are calculated, and the point with the maximum local potential value is the cluster center, and the combination of all cluster centers is the estimated mixing matrix.
5. The underdetermined blind source separation method based on structured sparse subspace clustering according to claim 1, It is characterized in that Based on the mixing matrix, a spatial projection method based on source number estimation is used in combination with the minimum l 1 The source signal is reconstructed using the norm method, further comprising the steps of: Assume that the set Λ is the estimated mixing matrix The set of 1 to m-1 order sub-matrices, let sub-matrix A k ∈Λ, calculate each time-frequency signal point x t (t,f) in submatrix A k The projection parameter on the spanned subspace is expressed as the tenth mathematical formula: Among them, t and f represent the time serial number and frequency serial number of the time-frequency signal point respectively. Represents the matrix A k The transpose of represents the inverse matrix; A projection parameter threshold is set. If there are multiple projection parameters calculated that are greater than the projection parameter threshold, the mixed column vector corresponding to the largest projection parameter is selected as the mixed matrix of the time-frequency point. At this time, the calculation expression of the restored time-frequency signal is the eleventh mathematical formula: in, Represents the vector of signals whose values in the recovered time-frequency signal are not 0, represents the elements in a vector, Represents a vector of signals with a value of 0 in the restored time-frequency signal; If any submatrix in the set Λ cannot obtain an observation point whose calculated projection parameter is greater than the projection parameter threshold, then assuming that the number of actual source signals is the same as the number of observed signals, the minimum l is used. 1 The norm method calculates all the recovered time-frequency signals, and the calculation formula is the twelfth mathematical formula: Among the solutions obtained by the twelfth mathematical formula, the solution satisfying the thirteenth mathematical formula is considered to be the recovered time-frequency signal. The thirteenth mathematical formula is: Among them, the source signal combination is the estimate of the source signal; Determine projection parameters for all observed signals, and calculate based on the eleventh mathematical formula or the twelfth mathematical formula to obtain a complete time-frequency recovery signal; Perform short-time inverse Fourier transform on the complete time-frequency restored signal to obtain the reconstructed source signal.
6. An underdetermined blind source separation device based on structured sparse subspace clustering, It is characterized in that The underdetermined blind source separation device includes a memory, a processor, and an underdetermined blind source separation program based on structured sparse subspace clustering stored in the memory and executable on the processor. When the underdetermined blind source separation program based on structured sparse subspace clustering is executed by the processor, the steps of the underdetermined blind source separation method based on structured sparse subspace clustering as described in any one of claims 1 to 5 are implemented.
Citation Information
Patent Citations
SAR (Synthetic Aperture Radar) image spot-inhibiting method based on spare domain noise distribution constraint
CN103971346A
Space-spectrum fused hyperspectral image mixed pixel low-rank sparse decomposition method
CN111161199A
