A method and system for extracting target signal based on audio blind source separation

Abstract

The present invention discloses a method and system for extracting a target signal based on audio blind source separation, the method comprising: S1 establishing an echo signal model of multi-source audio; S2 preprocessing the echo signal and constructing a covariance matrix, eigendecomposing the covariance matrix, and extracting the principal component of the echo signal based on the eigendecomposition result; S3 time-compensating the echo signal and repeating step S2 to extract the principal component of the echo signal after time compensation; S4 using a trained deep learning network to identify and mark the principal components extracted in steps S2 and S3, and building a target signal library based on the marking result; S5 removing duplicate signals in the target signal library; S6 reorganizing the signals in the target signal library to obtain the target signal of the audio signal. Based on the blind source separation method, the present invention obtains a clean target signal by adding time compensation, removing noise influence, removing duplicate signals, signal reorganization, etc., thereby ensuring the signal extraction quality.

Description

Technical Field

[0001] The present invention relates to the technical field of audio signal extraction, and in particular to a target signal extraction method and system based on audio blind source separation. Background Art

[0002] When processing audio signals, technicians must face two major challenges: noise pollution and multipath signal component interference. Noise pollution refers to the noise generated by environmental background, electronic equipment interference, communication interference and other factors. Multipath signal component interference refers to the audio signal received by different paths during the transmission process. The phase and amplitude of the audio signal change, which affects the clarity of the audio signal and causes the audio signal to be unclear. Therefore, how to extract a clear target signal has always been the research focus in the field of audio signal processing.

[0003] The existing methods for extracting target signals are mainly divided into two categories: signal separation methods based on the main components of the signal and deep learning discrimination methods based on signal features. However, the above methods still face the following problems in practical applications: when the source signal is not completely independent or there are multipath signals, the extraction process of the target signal is easily affected, which is manifested in incomplete extracted target signals and incorrect target signal discrimination. Summary of the invention

[0004] In view of this, the present invention combines the blind source separation method with the deep learning network. First, the main components of the echo signal are extracted using the feature decomposition method, and then the target signal library is constructed after judgment by the deep learning network. Finally, a clean and complete target signal is obtained by removing duplicate signals and signal recombination operations.

[0005] In order to achieve the above object, the present invention proposes a target signal extraction method based on audio blind source separation, comprising the following steps:

[0006] S1. Establishing an echo signal model of multi-source audio; including:

[0007] S11. Establish a multi-target signal source model:

[0008] ;

[0009] in, Indicates the signal source, represents the signal from the first target, represents the signal from the second target, Indicates that it comes from A target signal, Representation Matrix The transposed matrix of

[0010] S12, establish the coefficient matrix of multi-source audio:

[0011] ;

[0012] in, The coefficient matrix representing multi-source audio, represents the weight factor, represents the delay factor, , Represents the imaginary number symbol, Indicates the signal carrier frequency, Represents the unit delay, Indicates the number of receiving sources, The receiving sources are arranged at equal intervals;

[0013] S13, establishing an echo signal model of multi-source audio:

[0014] ;

[0015] in, Represents the echo signal of multi-source audio, Represents noise:

[0016] ;

[0017] S2. Preprocess the echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal component of the echo signal based on the eigendecomposition result;

[0018] S3, performing time compensation on the echo signal and repeating step S2 to extract the main component of the echo signal after time compensation;

[0019] S4, using the trained deep learning network to identify and mark the principal components extracted in steps S2 and S3, and constructing a target signal library based on the marking results;

[0020] S5, eliminating duplicate signals in the target signal library;

[0021] S6. Reorganize the signals in the target signal library to obtain the target signal of the audio signal.

[0022] Preferably, the step S2 comprises:

[0023] S21, preprocessing the echo signal of the multi-source audio:

[0024] ;

[0025] in, represents the echo signal after preprocessing, Represents the mean value of the echo signal;

[0026] S22. Constructing the covariance matrix , and perform eigendecomposition on the covariance matrix:

[0027] ;

[0028] in, Representation Matrix The transpose of is the covariance matrix The eigenvector matrix of Each column of is a feature vector, is the covariance matrix The eigenvalue diagonal matrix of ;

[0029] Preserve eigenvector matrix Center front The submatrix composed of eigenvectors constitutes the main eigenvector matrix ;

[0030] S23, extract the main component of the echo signal:

[0031] Constructing the dimensionality reduction matrix :

[0032] ;

[0033] Take the row vector of the reduced dimension matrix as the principal component of the echo signal ， in, express From the Dimensionality reduction matrix , express The first A row vector.

[0034] Preferably, step S3 comprises:

[0035] S31, performing time compensation on the echo signal:

[0036] Constructing compensation factors buchang :

[0037] ;

[0038] Echo signal of multi-source audio based on compensation factor To make time compensation:

[0039] ;

[0040] in, Represents the echo signal after time compensation;

[0041] S32, repeat steps S21, S22 and S23 to obtain the main component of the echo signal after time compensation .

[0042] Preferably, step S4 comprises:

[0043] Input the principal components extracted in steps S2 and S3 into the trained deep learning network, identify the input principal components and mark them as useful signals or noise signals respectively;

[0044] Construct a target signal library from the principal components marked as useful signals:

[0045] ;

[0046] in, represents the target signal library, Represents a signal in the target signal library.

[0047] Preferably, the training process of the deep learning network in step S4 includes:

[0048] Solve for the multipath echo signal of each target:

[0049] ;

[0050] ;

[0051] in, Represents the multipath echo signal of the target, Represents the target echo signal collected under laboratory conditions, represents the multipath factor;

[0052] The multipath echo signals of each target are collected and marked as useful signals, a training library is constructed, and the deep learning network is trained.

[0053] Preferably, step S5 comprises:

[0054] S51, target signal library Any two signals in and Perform coherence judgment. If the two are coherent, select the target signal library Reject signal or , if the two are not relevant, no processing will be done;

[0055] S52, repeat step S51, traverse the target signal library All signal pairs in .

[0056] Preferably, in step S51, the signal and The process of coherence determination includes:

[0057] Compute the dot product:

[0058] ;

[0059] in, Indicates the length of the signal, Indicates in the signal No. The signal value of the sampling point, Indicates in the signal No. The signal value of each sampling point;

[0060] if If it is equal to 0, the signal and If the signal is coherent, otherwise, and Not relevant.

[0061] Preferably, step S6 comprises:

[0062] Using the Target Signal Library The signal in is recombined to obtain the target signal of the audio signal:

[0063] ;

[0064] in, represents the target signal of the obtained audio signal, Represents the target signal library after coherence judgment The signal in .

[0065] The present invention also proposes a target signal extraction system based on audio blind source separation, comprising:

[0066] Echo signal model module: used to establish the echo signal model of multi-source audio;

[0067] Time compensation module: used to perform time compensation on the echo signal;

[0068] Echo signal principal component extraction module: used to pre-process the echo signal and the time-compensated echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal component of the echo signal based on the eigendecomposition result;

[0069] Target signal library module: used to train the deep learning network and identify and mark the main components of the input deep learning network, build the target signal library based on the marking results, and remove duplicate signals in the target signal library;

[0070] Recombination signal module: used to recombine the signals in the target signal library to obtain the target signal of the audio signal;

[0071] To realize the above-mentioned target signal extraction method based on audio blind source separation.

[0072] The technical solution of the present invention has at least the following beneficial effects:

[0073] 1. The present invention fully considers the delay time of the target signal, and uses the signal arrival time difference to construct the coefficient matrix of the multi-source audio, so that the established echo signal model of the multi-source audio is more in line with the actual situation; the present invention preprocesses the echo signal of the multi-source audio to ensure signal standardization, so as to facilitate subsequent feature decomposition; based on the blind source separation method, the present invention first performs time compensation on the echo signal of the multi-source audio, which is conducive to obtaining the main components of the echo signal with different delays, ensuring that the target signal is complete and not missed, and then extracts the echo signal and the main components of the echo signal after time compensation respectively through the feature decomposition method, effectively eliminating the noise pollution in the audio signal, and achieving high-quality echo signal main components.

[0074] 2. The present invention adopts a deep learning network to identify the main components of the echo signal, and constructs a target signal library based on the identified useful signals. It also eliminates some repeated signals based on the spatial correlation between the signals in the target signal library to ensure that the reconstructed target signal does not contain multipath signals from the same target, thereby obtaining a clean target signal and ensuring the quality of signal extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A flowchart of a target signal extraction method based on audio blind source separation provided in Example 1 of the present invention.

[0076] Figure 2 The echo signal of the multi-source audio obtained in step S13 of embodiment 1 of the present invention Schematic diagram of .

[0077] Figure 3 The recombinant target signal obtained in step S6 in Example 1 of the present invention Schematic diagram of . DETAILED DESCRIPTION

[0078] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0079] Embodiment 1:

[0080] A target signal extraction method based on audio blind source separation, the basic idea of ​​which is to obtain the principal component of the directly transmitted echo signal and the principal component of the echo signal under multipath propagation conditions, and to construct the target signal using the principal component with unique characteristics and non-redundancy, such as Figure 1 As shown, the method comprises the following steps:

[0081] S1. Establishing echo signal model of multi-source audio:

[0082] S11. Establish a multi-target signal source model:

[0083] ;

[0084] in, Indicates the signal source, represents the signal from the first target, represents the signal from the second target, Indicates that it comes from A target signal, Representation Matrix The transposed matrix of

[0085] S12, establish the coefficient matrix of multi-source audio:

[0086] ;

[0087] in, The coefficient matrix representing multi-source audio, represents the weight factor, represents the delay factor, , Represents the imaginary number symbol, Indicates the signal carrier frequency, Represents the unit delay, Represents the unit phase shift of the echo signal, Indicates the number of receiving sources, The receiving sources are arranged at equal intervals;

[0088] S13, establishing an echo signal model of multi-source audio:

[0089] ;

[0090] in, Represents the echo signal of multi-source audio, Represents noise:

[0091] .

[0092] S2. Preprocess the echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal components of the echo signal based on the eigendecomposition results:

[0093] S21, preprocessing the echo signal of the multi-source audio:

[0094] ;

[0095] in, represents the echo signal after preprocessing, Represents the mean value of the echo signal;

[0096] S22. Constructing the covariance matrix , and perform eigendecomposition on the covariance matrix:

[0097] ;

[0098] in, Representation Matrix The transpose of is the covariance matrix The eigenvector matrix of Each column of is a feature vector, is the covariance matrix The eigenvalue diagonal matrix of ;

[0099] Preserve eigenvector matrix Center front The submatrix composed of eigenvectors constitutes the main eigenvector matrix ;

[0100] S23, extract the main component of the echo signal:

[0101] Constructing the dimensionality reduction matrix :

[0102] ;

[0103] Take the row vector of the reduced dimension matrix as the principal component of the echo signal ， in, express From the Dimensionality reduction matrix , express The first A row vector.

[0104] S3, time-compensate the echo signal and repeat step S2 to extract the main component of the echo signal after time compensation:

[0105] S31, performing time compensation on the echo signal:

[0106] Constructing compensation factors buchang :

[0107] ;

[0108] Echo signal of multi-source audio based on compensation factor To make time compensation:

[0109] ;

[0110] in, Represents the echo signal after time compensation;

[0111] S32, repeat steps S21, S22 and S23 to obtain the main component of the echo signal after time compensation .

[0112] S4. Use the trained deep learning network to identify and mark the principal components extracted in steps S2 and S3, and build a target signal library based on the marking results:

[0113] Input the principal components extracted in steps S2 and S3 into the trained deep learning network, identify the input principal components and mark them as useful signals or noise signals respectively;

[0114] Construct a target signal library from the principal components marked as useful signals:

[0115] ;

[0116] in, represents the target signal library, represents a signal in a target signal library;

[0117] When training a deep learning network:

[0118] Solve for the multipath echo signal of each target:

[0119] ;

[0120] ;

[0121] in, Represents the multipath echo signal of the target, Represents the target echo signal collected under laboratory conditions, represents the multipath factor;

[0122] The multipath echo signals of each target are collected and marked as useful signals, a training library is constructed, and the deep learning network is trained.

[0123] S5. Eliminate duplicate signals in the target signal library:

[0124] S51, target signal library Any two signals in and To determine the coherence:

[0125] Compute the dot product:

[0126] ;

[0127] in, Indicates the length of the signal, Indicates in the signal No. The signal value of the sampling point, Indicates in the signal No. The signal value of each sampling point;

[0128] if If it is equal to 0, the signal and With coherence, from the target signal library Reject signal or Otherwise, the signal is judged and Not relevant, no processing;

[0129] S52, repeat step S51, traverse the target signal library All signal pairs in .

[0130] S6. Reorganize the signals in the target signal library to obtain the target signal of the audio signal:

[0131] Using the Target Signal Library The signal in is recombined to obtain the target signal of the audio signal:

[0132] ;

[0133] in, represents the target signal of the obtained audio signal, Represents the target signal library after coherence judgment The signal in .

[0134] In this embodiment:

[0135] Step S2 directly uses the characteristic decomposition method to extract the principal component of the echo signal, and obtains the target echo signal under the direct propagation condition, while step S3 first time compensates the echo signal, and then uses the characteristic decomposition method to extract the principal component, and obtains the target echo signal of multipath propagation under different time delay conditions. These two steps include the target echo signal under all path propagation conditions and are redundant.

[0136] Step S4 uses the trained deep learning network to judge the redundant principal components, marks the principal components from the target as useful signals and establishes a target signal library. This redundant input method can improve the judgment accuracy of the deep learning network, while ensuring that the principal components from the target are not missed and removing the influence of noise.

[0137] Step S5 removes duplicate data in the target signal library by performing correlation determination on the signals, thereby ensuring that the signals in the target signal library belong to different targets.

[0138] Step S6 superimposes the signals in the target signal library and uses the mean value of the echo signal to correct the signal value, thereby obtaining the target-related signal in the audio signal and realizing accurate extraction of the target signal.

[0139] The method in this embodiment is used to perform a simulation experiment, and the following simulation conditions are set:

[0140] The signal transmitted to the target is in the form of a frequency modulated interrupted continuous wave, expressed as:

[0141] ;

[0142] in, Indicates the signal transmitted to the target, Represents the imaginary number symbol, Represents the symbol for pi, Indicates the signal carrier frequency, represents the signal bandwidth, represents the frequency sweep period, express is the time domain signal, Signal transmission adopts the form of one transmitting source and multiple receiving sources. The number of receiving sources is .

[0143] The number of targets is 5, and the distance between the first target and the emission source is 10km, the distance between the second target and the emission source is 12km, the distance between the third target and the emission source is 15km, the distance between the fourth target and the emission source is 20km, and the distance between the fifth target and the emission source is 50km. The signals from each target are expressed as:

[0144] ;

[0145] ;

[0146] ;

[0147] ;

[0148] ;

[0149] in, represents the signal from the first target, represents the signal from the second target, Represents a signal from a third target, Indicates the signal from the fourth target, Indicates the signal from the fifth target, represents the audio propagation rate,

[0150] The 2.5 in the The signal amplitude,

[0151] The 1.98 in The signal amplitude,

[0152] The 10 in the The signal amplitude,

[0153] 27 in the The signal amplitude is the unit delay, and the time it takes for the audio signal to propagate 1 km .

[0154] In the above simulation experiment, the echo signal of the multi-source audio is obtained in step S13. , see the schematic diagram Figure 2 , the target signal of the audio signal obtained in step S6 , see the schematic diagram Figure 3 .contrast Figure 2 and Figure 3 It can be seen that the target signal of the audio signal reconstructed by the method of the present invention not only maintains the basic change trend of the original echo signal, but also significantly reduces noise interference, thereby achieving high-quality extraction of the target signal.

[0155] Embodiment 2:

[0156] The present invention also provides a target signal extraction system based on audio blind source separation, comprising the following modules:

[0157] Echo signal model module: used to establish the echo signal model of multi-source audio;

[0158] Time compensation module: used to perform time compensation on the echo signal;

[0159] Echo signal principal component extraction module: used to pre-process the echo signal and the time-compensated echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal component of the echo signal based on the eigendecomposition result;

[0160] Target signal library module: used to train the deep learning network and identify and mark the main components of the input deep learning network, build the target signal library based on the marking results, and remove duplicate signals in the target signal library;

[0161] Recombination signal module: used to recombine the signals in the target signal library to obtain the target signal of the audio signal;

[0162] To implement the target signal extraction method based on audio blind source separation in Example 1.

[0163] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method 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, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0164] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0165] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A target signal extraction method based on audio blind source separation, characterized in that: The following steps are involved: S1. Establishing an echo signal model of multi-source audio, including: S11. Establish a multi-target signal source model: ； in, Indicates the signal source, represents the signal from the first target, represents the signal from the second target, Indicates that it comes from A target signal, Representation Matrix The transposed matrix of S12, establish the coefficient matrix of multi-source audio: ； in, The coefficient matrix representing multi-source audio, represents the weight factor, represents the delay factor, , Represents the imaginary number symbol, Indicates the signal carrier frequency, Represents the unit delay, Indicates the number of receiving sources, The receiving sources are arranged at equal intervals; S13, establishing an echo signal model of multi-source audio: ； in, Represents the echo signal of multi-source audio, Represents noise: ； S2. Preprocess the echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal component of the echo signal based on the eigendecomposition result; S3, performing time compensation on the echo signal and repeating step S2 to extract the main component of the echo signal after time compensation; S4, using the trained deep learning network to identify and mark the principal components extracted in steps S2 and S3, and constructing a target signal library based on the marking results; S5, eliminating duplicate signals in the target signal library; S6. Reorganize the signals in the target signal library to obtain the target signal of the audio signal.

2. The target signal extraction method based on audio blind source separation according to claim 1 is characterized in that: The step S2 comprises: S21, preprocessing the echo signal of the multi-source audio: ； in, represents the echo signal after preprocessing, Represents the mean value of the echo signal; S22. Constructing the covariance matrix , and perform eigendecomposition on the covariance matrix: ； in, Representation Matrix The transpose of is the covariance matrix The eigenvector matrix of Each column of is a feature vector, is the covariance matrix The eigenvalue diagonal matrix of ; Preserve eigenvector matrix Center front The submatrix composed of eigenvectors constitutes the main eigenvector matrix ; S23, extract the main component of the echo signal: Constructing the dimensionality reduction matrix : ； Take the row vector of the reduced dimension matrix as the principal component of the echo signal ， in, express From the Dimensionality reduction matrix , express The first A row vector.

3. The target signal extraction method based on audio blind source separation according to claim 2 is characterized in that: The step S3 comprises: S31, performing time compensation on the echo signal: Constructing compensation factors buchang : ； Echo signal of multi-source audio based on compensation factor To make time compensation: ； in, Represents the echo signal after time compensation; S32, repeat steps S21, S22 and S23 to obtain the main component of the echo signal after time compensation .

4. The target signal extraction method based on audio blind source separation according to claim 3 is characterized in that: The step S4 comprises: Input the principal components extracted in steps S2 and S3 into the trained deep learning network, identify the input principal components and mark them as useful signals or noise signals respectively; Construct a target signal library from the principal components marked as useful signals: ； in, represents the target signal library, Represents a signal in the target signal library.

5. The target signal extraction method based on audio blind source separation according to claim 4 is characterized in that: The training process of the deep learning network in step S4 includes: Solve for the multipath echo signal of each target: ； ； in, Represents the multipath echo signal of the target, Represents the target echo signal collected under laboratory conditions, represents the multipath factor; The multipath echo signals of each target are collected and marked as useful signals, a training library is constructed, and the deep learning network is trained.

6. The target signal extraction method based on audio blind source separation according to claim 4 is characterized in that: The step S5 comprises: S51, target signal library Any two signals in and Perform coherence judgment. If the two are coherent, select the target signal library Reject signal or , if the two are not relevant, no processing will be done; S52, repeat step S51, traverse the target signal library All signal pairs in .

7. The target signal extraction method based on audio blind source separation according to claim 6 is characterized in that: In step S51, the signal and The process of coherence determination includes: Compute the dot product: ； in, Indicates the length of the signal, Indicates in the signal No. The signal value of the sampling point, Indicates in the signal No. The signal value of each sampling point; if If it is equal to 0, the signal and If the signal is coherent, otherwise, and Not relevant.

8. The target signal extraction method based on audio blind source separation according to claim 6 is characterized in that: The step S6 comprises: Using the Target Signal Library The signal in is recombined to obtain the target signal of the audio signal: ； in, represents the target signal of the obtained audio signal, Represents the target signal library after coherence judgment The signal in .

9. A target signal extraction system based on audio blind source separation, characterized in that: include: Echo signal model module: used to establish the echo signal model of multi-source audio; Time compensation module: used to perform time compensation on the echo signal; Echo signal principal component extraction module: used to pre-process the echo signal and the time-compensated echo signal and construct a covariance matrix, perform eigendecomposition on the covariance matrix, and extract the principal component of the echo signal based on the eigendecomposition result; Target signal library module: used to train the deep learning network and identify and mark the main components of the input deep learning network, build the target signal library based on the marking results, and remove duplicate signals in the target signal library; Recombination signal module: used to recombine the signals in the target signal library to obtain the target signal of the audio signal; To realize the target signal extraction method based on audio blind source separation as described in any one of claims 1-8.

Images