A method, system and device for underdetermined blind source separation based on minimization and maximization
By combining the minimization maximization algorithm and the non-convex penalty function, the separation of underdetermined blind sources in high-end equipment is achieved, which solves the problem of large signal recovery error in traditional methods and provides a more accurate vibration noise source Identify and position to guide the vibration and noise reduction of the gas turbine.
Patent Information
- Application Number
- CN202010851255.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-08-21
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-08-21
AI Technical Summary
The existing under-determined blind source separation method is difficult to effectively separate multiple vibration noise sources in high-end equipment such as gas turbines. Especially when the number of sensors is insufficient in high-temperature and high-pressure environments, it leads to large signal recovery errors and the possibility of accurately identifying and positioning the vibration noise sources. .
Using the underdetermined blind source separation method based on minimization and maximization, the time domain signal is converted into the time frequency domain through short-time Fourier transform, single source points are identified, and the instantaneous hybrid matrix is estimated using hierarchical clustering method, and the The l1 norm objective function of the non-convex penalty function performs global optimization to achieve the recovery and separation of the source signal.
Improve the separation accuracy of the source signal, reduce the error between the recovery signal and the source signal, and can effectively guide the vibration reduction and noise reduction of high-end equipment, avoid systematic underestimation caused by l1 norm regularization in traditional methods, and ensure optimization The accuracy of convexity reconciliation of the process.
Smart Images

Figure CN112201274B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of mechanical vibration noise signal processing, and in particular relates to an underdetermined blind source separation method, system and device based on minimization and maximization. Background Art
[0002] Taking vibration and noise reduction measures for high-end equipment such as gas turbines can effectively suppress vibration and noise. High-end equipment such as gas turbines have complex and precise structures, numerous vibration and noise sources, and compact structures, which makes it particularly difficult to identify and locate vibration and noise sources, becoming a technical bottleneck for achieving vibration and noise reduction. Therefore, identifying the vibration and noise sources of high-end equipment such as gas turbines, accurately locating the vibration and noise source components, and revealing their transmission and mixing mechanisms can provide targeted guidance for vibration and noise reduction of high-end equipment such as gas turbines.
[0003] The signals collected by the internal sensors of high-end equipment are the superposition of multiple vibration noise sources transmitted to the sensors. However, due to the precise and compact internal structure of high-end equipment such as gas turbines, the high temperature and high pressure working environment, and the limited installation locations for sensors, it is impossible to install enough sensors. Underdetermined blind source separation can effectively separate the source signals in the absence of prior knowledge and when the number of sensors is less than the number of source signals. At present, the underdetermined blind source separation method mainly extracts single source points and estimates the mixing matrix by clustering the signals in the transform domain, and then uses the mixing matrix to optimize the recovery of the source signal using the l0 norm. However, the l0 discontinuity causes large errors in the recovered source signal, and it is impossible to effectively separate the internal mechanical mixed signals of high-end equipment such as shipborne gas turbines. Summary of the invention
[0004] The purpose of the present invention is to provide a method, system and device for underdetermined blind source separation based on minimization and maximization to overcome the shortcomings of the prior art.
[0005] In order to achieve the above object, the present invention adopts the following technical scheme:
[0006] An underdetermined blind source separation method based on minimization and maximization comprises the following steps:
[0007] Step 1), performing short-time Fourier transform on the collected observation signals to obtain a complex matrix in the time-frequency domain;
[0008] Step 2), single source point identification is performed on the time-frequency domain complex matrix to extract the time-frequency domain complex-valued single source point;
[0009] Step 3), according to the frequency domain complex valued single source points obtained in step 2, the obtained time-frequency domain complex valued single source points are clustered by a hierarchical clustering method and an instantaneous mixing matrix is estimated;
[0010] Step 4), according to the instantaneous mixing matrix obtained in step 3), the norm objective function based on the non-convex penalty function is globally optimized to realize the frequency domain complex valued source recovery, and finally the short-time Fourier inverse transform is used to transform the recovered frequency domain signal back to the time domain to obtain the separated signal.
[0011] Furthermore, step 1) is specifically as follows: collect M observation signals x(t)=[x1(t) x2(t) … x M (t)] T , where the superscript T represents transposition, the sampling time series t = 1, 2, ..., T, the collected observation signal is short-time Fourier transformed to obtain the complex matrix X in the time-frequency domain.
[0012] Furthermore, each time-domain observation signal x to be analyzed is transformed into i (t) Transform to time-frequency domain X i =[X i,1 ,X i,2 ,…,X i,J ], where the first dimension represents frequency and the second dimension represents time; the time-frequency domain of all observed signals is a three-dimensional matrix X = [X i,j,k ]The first dimension is the observed signal sequence, the second dimension represents time, and the third dimension represents frequency.
[0013] Furthermore, the single source point is identified for the time-frequency domain complex matrix, and the time-frequency domain complex matrix X = [X i,j,k ]The real and imaginary parts of a complex-valued vector at the same time and frequency:
[0014]
[0015] Where X j.k =[X 1,j,k X 2,j,k … X M,j,k ] T is a complex matrix in the time-frequency domain X = [X i,j,k ] At the jth moment, the sub-vector of the kth frequency band, Re(·) and Im(·) are respectively the real part operator and the real part imaginary part operator, and the vectors with cosine angles of real and imaginary parts greater than ε are extracted as single source point vectors, which satisfies: X j,k is a single source point vector, where |·| represents taking the modulus of the vector.
[0016] Furthermore, the obtained time-frequency domain complex-valued single source point is modulo-normalized to transform it into a real vector with a modulus of 1, that is: in Indicates X j,kThe rows are modulo-normalized. The superscript H indicates conjugate transpose. diag(·) indicates taking the diagonal elements of the square matrix and assigning them to new vectors of the same dimension. The normalized single-source point vectors are hierarchically clustered. All the cluster centers obtained are the columns of the corresponding mixing matrix. The matrix of all cluster center combinations is the estimate of the instantaneous mixing matrix. in Represents an estimate of .
[0017] Furthermore, using the instantaneous mixing matrix Construct the following objective function
[0018] in represents the square of the 2-norm of ·, ||·||1 represents the 1-norm of ·, λ>0, 1>γ≥0, It means to find the minimum value of the function f(·) with an M-dimensional complex-valued vector as the independent variable, and to find the minimum value of the objective function:
[0019]
[0020] The solution is obtained through gradient descent, and v and Y in the objective function are updated respectively to achieve the recovery of the source signal in the time-frequency domain.
[0021] Furthermore, the short-time inverse Fourier transform is used to transform the restored time-frequency domain source signal into the time domain to complete the separation of the source signal.
[0022] Furthermore, the performance of the separated signal is evaluated using the following formula:
[0023]
[0024] where s i is the ith source signal, is the estimated jth separated signal, E(·) represents the mean value, and the waveform correlation coefficient is obtained It is an indicator of the accuracy of the estimated separation signal, representing the correlation between the separation signal and the homologous signal.
[0025] An underdetermined blind source separation system based on minimization and maximization, comprising a data acquisition module, a data preprocessing module, a mixing matrix processing module and a separation module;
[0026] The data acquisition module is used to obtain the observation signal of the device to be analyzed, and transmit the obtained observation signal to the data preprocessing module;
[0027] The data preprocessing module performs short-time Fourier transform on the observed signal to obtain a complex matrix in the time-frequency domain, and at the same time, performs single source point identification based on the complex matrix in the time-frequency domain to extract the complex-valued single source point in the time-frequency domain;
[0028] The mixing matrix processing module clusters the obtained time-frequency domain complex-valued single source points according to the obtained frequency-domain complex-valued single source points through a hierarchical clustering method and estimates the instantaneous mixing matrix, and transmits the instantaneous mixing matrix to the separation module;
[0029] The separation module globally optimizes the norm objective function based on the non-convex penalty function according to the instantaneous mixing matrix obtained, realizes the frequency domain complex value source recovery, and finally uses the short-time Fourier inverse transform to transform the recovered frequency domain signal back to the time domain to obtain the separated signal and output it.
[0030] An underdetermined blind source separation device based on mini-maximization includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.
[0031] Compared with the prior art, the present invention has the following beneficial technical effects:
[0032] The invention discloses an underdetermined blind source separation method based on minimization and maximization. For blind separation of underdetermined instantaneous mixture, short-time Fourier transform is first used to transform a time-domain linear mixed signal into a component on each frequency band under instantaneous linear mixture in the time-frequency domain. At this time, the time-frequency domain signal is complex-valued. Based on the sparsity of mixed data samples and the assumption of homologous phase, the cosine angles of the real vector and the imaginary vector of the complex-valued mixed data of the same time-frequency component are respectively calculated, and the vectors whose angles meet the conditions are extracted as single source point vectors. The mixed matrix is estimated by clustering the extracted single source point vectors. The mixed matrix and the sparsity assumption of the source signal are used to transform the source signal recovery problem into an l1 norm optimization problem. After the source signal is restored in the time-frequency domain, the short-time Fourier inverse transform is used to transform the source signal in the time-frequency domain back to the time domain, thereby realizing underdetermined blind source separation. The present invention uses a mini-maximum algorithm to restore the source signal, constructs a non-convex penalty function to make the optimization of the objective function close to the l0 norm optimization, avoids the systematic underestimation of the traditional l1 norm regularization and the source signal restored according to the sparsity assumption. And ensures the convexity of the objective function, that is, it will not fall into the local optimal solution in the process of solving the optimal solution, and reduces the error between the restored signal and the source signal.
[0033] The present invention provides an underdetermined blind source separation system based on minimization and maximization, which has a simple structure and can quickly restore the source signal in the time-frequency domain according to the above method and then use short-time Fourier inverse transform to transform the source signal in the time-frequency domain back to the time domain to achieve underdetermined blind source separation, thereby improving the separation accuracy of the source signal and providing targeted guidance for vibration reduction and noise reduction of high-end equipment such as gas turbines. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 The source signal waveform and spectrum diagram in Example 1 of the present invention;
[0035] Figure 2 The mixed signal waveform and spectrum diagram in Example 1 of the present invention;
[0036] Figure 3 The signal waveform and its spectrum diagram separated by the method of the present invention in Example 1 of the present invention;
[0037] Figure 4 The l1 traditional norm optimization method described in Example 1 of the present invention is used to separate the signal waveform and the spectrum diagram. DETAILED DESCRIPTION
[0038] The present invention is further described in detail below in conjunction with the accompanying drawings:
[0039] An underdetermined blind source separation method based on minimization and maximization uses an underdetermined linear instantaneous mixture model to separate mixed observation signals, specifically comprising the following steps:
[0040] Step 1: Collect M observation signals x(t) = [x1(t) x2(t) … x M (t)] T , where the superscript T represents transposition, the sampling time series t=1,2,…,T, the collected observation signal is short-time Fourier transformed to obtain the time-frequency domain complex matrix X;
[0041] Step 2, identify the single source point of the complex matrix X in the time-frequency domain and extract the complex-valued single source point in the time-frequency domain;
[0042] Step 3, according to the time-frequency domain complex-valued single source points obtained in step 2, clustering the obtained time-frequency domain complex-valued single source points by a hierarchical clustering method and estimating an instantaneous mixing matrix;
[0043] Step 4: According to the estimated mixing matrix obtained in step 3, the l1 norm objective function based on the non-convex penalty function is globally optimized to realize the frequency domain complex valued source recovery. Finally, the short-time Fourier inverse transform is used to transform the recovered frequency domain signal back to the time domain to obtain the separated signal y(t).
[0044] In step 1, each time domain observation signal x to be analyzed is transformed into i (t) Transform to time-frequency domain X i =[X i,1 ,X i,2 ,…,X i,J ], where the first dimension represents frequency and the second dimension represents time; the time-frequency domain of all observed signals is a three-dimensional matrix X = [X i,j,k ]The first dimension is the observed signal sequence, the second dimension represents time, and the third dimension represents frequency.
[0045] In step 2, the single source point of the time-frequency domain complex matrix in step 1 is identified by taking the time-frequency domain complex matrix X = [X i,j,k ]The real and imaginary parts of a complex-valued vector at the same time and frequency:
[0046]
[0047] Where X j.k =[X 1,j,k X 2,j,k … X M,j,k ] T is a complex matrix in the time-frequency domain X = [X i,j,k ] At the jth moment, the sub-vector of the kth frequency band, Re(·) and Im(·) are respectively the real part operator and the real part imaginary part operator, and further extract the vectors with cosine angles of real and imaginary parts greater than ε as single source point vectors, that is, satisfying: X j,k is a single source point vector, where |·| represents taking the modulus of the vector.
[0048] In step 3, the single source point vector in step 2 is modulo-normalized to make it a real vector with a modulus of 1, that is: in Indicates X j,k The rows are modulo-normalized. The superscript H indicates conjugate transpose. diag(·) indicates taking the diagonal elements of the square matrix and assigning them to new vectors of the same dimension. The normalized single-source point vectors are hierarchically clustered. All the cluster centers obtained are the columns of the corresponding mixing matrix. The matrix of all cluster center combinations is the estimation of the mixing matrix. in Represents an estimate of .
[0049] Step 4 uses the estimated matrix from step 3 Construct the following objective function in represents the square of the 2-norm of ·, ||·||1 represents the 1-norm of ·, λ>0, 1>γ≥0, It means to find the minimum value of the function f(·) with an M-dimensional complex-valued vector as the independent variable. Finding the minimum value of the objective function is to solve the following minimization and maximization optimization problem:
[0050]
[0051] The solution is obtained by gradient descent, and v and Y in the objective function are updated respectively to achieve the recovery of the source signal in the time-frequency domain. The short-time inverse Fourier transform is used to transform the recovered time-frequency domain source signal into the time domain to complete the separation of the source signal.
[0052] The performance of the separated signal in step 4 is evaluated using the following formula:
[0053]
[0054] where s i is the ith source signal, is the estimated jth separated signal, E(·) represents the mean value, and the waveform correlation coefficient is obtained It is an indicator of the accuracy of the estimated separation signal, representing the correlation between the separation signal and the homologous signal.
[0055] The present invention is further described in detail below through a square array experiment.
[0056] Example 1: Select three typical mechanical signals as source signals, and their expressions are as follows:
[0057]
[0058] The obtained source signal waveform and its spectrum are as follows: Figure 1 As shown, the source signal is passed through the mixing matrix to obtain the observed mixed signal, the number of observations is 2, Figure 2 The waveform and spectrum of the mixed signal are shown in the following formula.
[0059]
[0060] from Figure 2 It can be seen that due to the superposition of source signals, it is difficult to tell what kind of signal product is contained in the time domain waveform of the observed signal. Although its frequency characteristics can be seen from the spectrum, due to the lack of prior knowledge, it is impossible to confirm which part of the frequency domain comes from which source. Therefore, it is impossible to directly obtain an estimate of the source signal by observing the signal alone.
[0061] The underdetermined blind separation method proposed in the present invention is used to process the observed signal. In this example, the short-time Fourier transform window factory is selected as 512, the window type is a Hanning window, and the moving step size is 256 each time. The separation signal waveform and source signal and their spectrum obtained by the method proposed in the present invention are shown in FIG. Figure 3 As shown, compared Figure 1 and Figure 3 It can be seen that the waveform of the separated signal is extremely similar to the waveform of the source signal, and from the comparison of the spectrum diagram, it can be seen that the main characteristic frequencies of the corresponding source signal are retained in the separated signal, indicating the accuracy and effectiveness of the method proposed in the present invention.
[0062] In this example, the short-time Fourier transform window of the traditional l1 norm optimization method is 512, the window is a Hanning window, the window step size is 256 each time, and the regularization parameter λ = 0.55. The separated signal waveform and source signal and their spectrum obtained by the traditional l1 norm optimization method are shown in the figure below: Figure 4 As shown, through Figure 1 and Figure 4 ,Comparing the spectrum of the separated signal and the source signal, the traditional l1 norm optimization method can also separate the main characteristic frequencies, but due to the inherent systematic underestimation of the real signal by the l1 norm optimization, the characteristic frequency of source signal 3 is incorrectly estimated into source signal 1 and source signal 2, resulting in a large error in the restored source signal.
[0063] In order to quantitatively evaluate the accuracy of the estimated recovered signal, the estimated source signal is processed using the following formula:
[0064]
[0065] where s i is the ith source signal, is the estimated jth separated signal, E(·) represents the mean value, and the waveform correlation coefficient is obtained It is an indicator of the accuracy of the estimated separation signal, representing the correlation between the separation signal and the homologous signal.
[0066] The separation signal waveform correlation coefficient matrix obtained by the method of the present invention is:
[0067]
[0068] The correlation coefficient matrix of the separated signal waveform obtained by the traditional l1 norm optimization method is:
[0069]
[0070] From the correlation coefficient matrix, it can be seen that the accuracy of the proposed method is higher than that of the traditional l1 norm method.
[0071] The present invention constructs a non-convex penalty function to make the optimization of the objective function approach norm optimization, avoids the systematic underestimation of traditional norm regularization, ensures that the optimal solution will not fall into the local optimal solution during the solution process, reduces the error between the restored signal and the source signal, and can accurately and effectively separate the internal mechanical mixed signals of high-end equipment such as shipborne gas turbines, providing important guidance for vibration and noise reduction of shipborne gas turbines, which has great engineering significance.
Claims
1. A method for underdetermined blind source separation based on mini-maximization, characterized in that: The following steps are involved: Step 1), performing short-time Fourier transform on the collected observation signals to obtain a complex matrix in the time-frequency domain; Specifically, collect M observation signals x(t) = [x1(t) x2(t) … x M (t)] T , where the superscript T represents transposition, the sampling time series t=1,2,…,T, the collected observation signal is short-time Fourier transformed to obtain the complex matrix X in the time-frequency domain; Step 2), single source point identification is performed on the time-frequency domain complex matrix to extract the time-frequency domain complex-valued single source point; Step 3), according to the time-frequency domain complex-valued single source points obtained in step 2, the obtained time-frequency domain complex-valued single source points are clustered by a hierarchical clustering method and an instantaneous mixing matrix is estimated; Step 4), according to the instantaneous mixing matrix obtained in step 3), the norm objective function based on the non-convex penalty function is globally optimized to realize the frequency domain complex value source recovery, and finally the recovered frequency domain signal is transformed back to the time domain by short-time inverse Fourier transform to obtain the separated signal; Each time domain observation signal x to be analyzed is transformed into i (t) Transform to time-frequency domain X i =[X i,1 ,X i,2 ,…,X i,J ], where the first dimension represents frequency and the second dimension represents time; the time-frequency domain of all observed signals is a three-dimensional matrix X = [X i,j,k ]The first dimension is the observed signal sequence, the second dimension represents time, and the third dimension represents frequency; The single source point is identified for the time-frequency domain complex matrix, and the time-frequency domain complex matrix X = [X i,j,k ]The real and imaginary parts of a complex-valued vector at the same time and frequency: Where X j.k =[X 1,j,k X 2,j,k … X M,j,k ] T is a complex matrix in the time-frequency domain X = [X i,j,k ] At the jth moment, the sub-vector of the kth frequency band, Re(·) and Im(·) are respectively the real part operator and the real part imaginary part operator, and the vectors with cosine angles of real and imaginary parts greater than ε are extracted as single source point vectors, which satisfies: X j,k is a single source point vector, where |·| represents taking the vector modulo; The obtained time-frequency domain complex-valued single source point is modulo-normalized to make it a real vector with a modulus of 1, that is: in Indicates X j,k The rows are modulo-normalized. The superscript H indicates conjugate transpose. diag(·) indicates taking the diagonal elements of the square matrix and assigning them to new vectors of the same dimension. The normalized single-source point vectors are hierarchically clustered. All the cluster centers obtained are the columns of the corresponding mixing matrix. The matrix of all cluster center combinations is the estimate of the instantaneous mixing matrix. in represents an estimate of ; Using the instantaneous mixing matrix Construct the following objective function in represents the square of the 2-norm of ·, ||·||1 represents the 1-norm of ·, λ>0, 1>γ≥0, It means to find the minimum value of the function f(·) with an M-dimensional complex-valued vector as the independent variable, and to find the minimum value of the objective function: The solution is obtained through gradient descent, and v and Y in the objective function are updated respectively to achieve the recovery of the source signal in the time-frequency domain.
2. The underdetermined blind source separation method based on mini-maximization according to claim 1, characterized in that: The short-time inverse Fourier transform is used to transform the restored time-frequency domain source signal into the time domain to complete the separation of the source signal.
3. The underdetermined blind source separation method based on mini-maximization according to claim 1, characterized in that: The performance of the separated signal is evaluated using the following formula: where s i is the ith source signal, is the estimated jth separated signal, E(·) represents the mean value, and the waveform correlation coefficient is obtained It is an indicator of the accuracy of the estimated separation signal, representing the correlation between the separation signal and the homologous signal.
4. An underdetermined blind source separation system based on mini-maximization, characterized in that: It includes a data acquisition module, a data preprocessing module, a mixed matrix processing module and a separation module; The data acquisition module is used to obtain the observation signal of the device to be analyzed, and transmit the obtained observation signal to the data preprocessing module; Collect M observation signals x(t) = [x1(t) x2(t) … x M (t)] T , where the superscript T represents transposition, the sampling time series t=1,2,…,T, the collected observation signal is short-time Fourier transformed to obtain the complex matrix X in the time-frequency domain; The data preprocessing module performs short-time Fourier transform on the observed signal to obtain a complex matrix in the time-frequency domain, and at the same time, performs single source point identification based on the complex matrix in the time-frequency domain to extract the complex-valued single source point in the time-frequency domain; The mixing matrix processing module clusters the obtained time-frequency domain complex-valued single source points according to the obtained frequency-domain complex-valued single source points through a hierarchical clustering method and estimates the instantaneous mixing matrix, and transmits the instantaneous mixing matrix to the separation module; The separation module performs global optimization on the norm objective function based on the non-convex penalty function according to the instantaneous mixing matrix obtained, realizes the frequency domain complex value source recovery, and finally uses the short-time Fourier inverse transform to transform the recovered frequency domain signal back to the time domain to obtain the separated signal and output it; Each time domain observation signal x to be analyzed is transformed into i (t) Transform to time-frequency domain X i =[X i,1 ,X i,2 ,…,X i,J ], where the first dimension represents frequency and the second dimension represents time; the time-frequency domain of all observed signals is a three-dimensional matrix X = [X i,j,k ]The first dimension is the observed signal sequence, the second dimension represents time, and the third dimension represents frequency; The single source point is identified for the time-frequency domain complex matrix, and the time-frequency domain complex matrix X = [X i,j,k ]The real and imaginary parts of a complex-valued vector at the same time and frequency: Where X j.k =[X 1,j,k X 2,j,k … X M,j,k ] T is a complex matrix in the time-frequency domain X = [X i,j,k ] At the jth moment, the sub-vector of the kth frequency band, Re(·) and Im(·) are respectively the real part operator and the real part imaginary part operator, and the vectors with cosine angles of real and imaginary parts greater than ε are extracted as single source point vectors, which satisfies: X j,k is a single source point vector, where |·| represents taking the vector modulo; The obtained time-frequency domain complex-valued single source point is modulo-normalized to make it a real vector with a modulus of 1, that is: in Indicates X j,k The rows are modulo-normalized. The superscript H indicates conjugate transpose. diag(·) indicates taking the diagonal elements of the square matrix and assigning them to new vectors of the same dimension. The normalized single-source point vectors are hierarchically clustered. All the cluster centers obtained are the columns of the corresponding mixing matrix. The matrix of all cluster center combinations is the estimate of the instantaneous mixing matrix. in represents an estimate of ; Using the instantaneous mixing matrix Construct the following objective function in represents the square of the 2-norm of ·, ||·||1 represents the 1-norm of ·, λ>0, 1>γ≥0, It means to find the minimum value of the function f(·) with an M-dimensional complex-valued vector as the independent variable, and to find the minimum value of the objective function: The solution is obtained through gradient descent, and v and Y in the objective function are updated respectively to achieve the recovery of the source signal in the time-frequency domain.
5. An underdetermined blind source separation device based on mini-maximization, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the method according to claim 1 are implemented when the processor executes the computer program.
Citation Information
Patent Citations
Blind separation hybrid matrix estimation method based on synchronous compression short-time Fourier transform
CN107301434A
Sparse source signal blind separation method based on grid and density clustering
CN108710917A
Method for blind source separation by noise-containing frequency domain convolution
CN109243483A
A complex value blind separation hybrid matrix estimation method based on a single source point
CN109711468A
Modal identification method for non-proportionally damped structures based on extended sparse component analysis
US20200089730A1
