Low-complexity cascaded generalized sidelobe cancellation beam forming method and device

Abstract

The invention discloses a low-complexity cascaded generalized sidelobe cancellation beam forming method and device, and the method comprises the steps: decomposing a high-dimensional generalized sidelobe cancellation device into two low-dimensional generalized sidelobe cancellation units which are cascaded front and back through constructing a beam former of a cascaded generalized sidelobe cancellation structure; according to the method, the algorithm complexity can be remarkably reduced, the robustness is improved, the real-time performance of the algorithm is improved by utilizing a front-back stage feedback iteration processing mode based on the frame, and the method has higher practical value. According to the method, an interference noise cancellation module is constructed by further utilizing the idea of multi-channel Wiener filtering, filtering coefficients are directly updated between frames, additional noise estimation is not needed, the method is easy to implement and high in robustness, and interference noise suppression and target voice distortion can be balanced. In addition, through a multi-branch fusion and interleaving strategy, the degree of freedom of beam weight coefficient search can be improved, the algorithm can find an optimal solution more easily, and the target enhancement and interference noise suppression performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of microphone array signal processing, and in particular to a low-complexity cascaded generalized sidelobe cancellation beamforming method and device. Background Art

[0002] For large microphone arrays with a large number of elements, commonly used beamforming technologies include fixed beamforming technology and adaptive beamforming technology. Among them, fixed beamforming technology is simple to implement and has high robustness, but its disadvantage is that it has insufficient interference and noise suppression capabilities. Common adaptive beamforming technologies include minimum variance distortionless response (MVDR) beamforming, linear constrained minimum variance (LCMV) beamforming, generalized sidelobe cancellation (GSC) beamforming, etc., all of which have strong interference suppression capabilities, but when the number of elements is large, there are problems of large computational complexity and poor robustness, especially the mismatch between elements and the estimation deviation of the noise covariance matrix may even lead to performance degradation. People have also proposed Kronecker product beamforming technology to reduce the dimension of the covariance matrix to improve robustness, but it sacrifices the degree of freedom and may cause performance degradation. In practical applications, most schemes require the additional use of noise estimation algorithms, and the estimation accuracy deteriorates at low signal-to-noise ratios, which also reduces robustness. In addition, the existing Kronecker beamforming does not involve a scheme for frame-based processing that adapts to arbitrary array formations. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention aims to provide a low-complexity cascaded generalized sidelobe cancellation beamforming method and device.

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] A low-complexity cascaded generalized sidelobe cancellation beamforming method comprises the following steps:

[0006] Step 1: Using the coordinate information of each microphone array element in the microphone array and the direction parameters of the sound source wave to obtain a constraint matrix, the time domain signal collected and received by the microphone array is processed by framing, windowing and short-time Fourier transform to obtain an array frequency domain frame signal;

[0007] Step 2: interleave the array frequency domain frame signal and constraint matrix obtained in step 1 in multiple parallel branches to obtain the reordered array frequency domain frame signal and constraint matrix of each branch;

[0008] Step 3: setting a plurality of cascaded generalized sidelobe cancellers, each of which includes two low-dimensional front-stage generalized sidelobe cancellers and a rear-stage generalized sidelobe canceller that are cascaded front and back, and updating parameters through frame-based iterative operations to obtain output frequency domain information of each branch;

[0009] Step 4, the beam outputs of the cascaded generalized sidelobe cancellers of each branch are integrated to obtain the optimal solution of the beam outputs of the multi-path cascaded generalized sidelobe cancellers;

[0010] The frequency domain output signal finally outputted in step 5 and step 4 is subjected to short-time Fourier inverse transform, windowing and framing to obtain a time domain signal after noise interference suppression and target speech enhancement.

[0011] Furthermore, the specific process of step 1 is:

[0012] The microphone array consists of M microphone elements, which are numbered 1, 2, ...m, ..., M. The spatial position coordinates of the mth microphone element are expressed as r in the three-dimensional Cartesian coordinate system. m =[r x,m ,r y,m ,r z,m ] T , T represents transposition, and the far-field incident angle azimuth parameters of the known target sound source are where θ d and is the pitch angle and azimuth of the target sound source d; suppose there are J interference sound sources, and the far-field incident angle azimuth parameter of the jth interference sound source is θ j and are the pitch angle and azimuth angle of the jth interference sound source, respectively, and the constraint matrix is ​​obtained:

[0013] C(k)=[v(k,Ω d ),v(k,Ω1),…,v(k,Ω J )] T , k=0,1,2,…N fft / twenty one)

[0014] k is N fft The frequency point number of the Fourier transform of the point, v(k,Ω d ) and v(k,Ω j )(j=1,2,…J) are the guidance vectors of the target sound source and the interference sound source respectively, and the following information is obtained by using the coordinates of the microphone array element and the direction of the sound source wave:

[0015]

[0016] Where i represents the imaginary unit, κ d and κ j are the sound source wave numbers of the target sound source and the interference sound source, c is the propagation speed of sound in space, f k is the frequency value corresponding to frequency point k;

[0017] Based on the signal distortion-free criterion, the array beam weight vector w is required to satisfy CH w=g,g=[1,0,…0] T is the J+1-dimensional constraint target vector, C represents the constraint matrix, and H represents the conjugate transpose;

[0018] The time domain signal collected by the microphone array is framed, windowed and short-time Fourier transformed to obtain the array frequency domain signal of the lth frame and the kth frequency point: x(k,l)=[x1(k,l),…,x m (k,l),…,x M (k,l)] T .

[0019] Furthermore, the specific process of step 2 is:

[0020] P parallel interleaving processors are used for signal processing. The interleaving processors interleave and reorder the array frequency domain frame signal x(k, l) and the constraint matrix C(k). The interleaving table corresponding to the interleaving processor of the pth branch is m=I p (m'), then the array frequency domain frame signal interleaved and output by the interleaving processor of the pth branch is:

[0021]

[0022] Interleaved output array frequency domain frame signal vector x p The m'th element of (k,l) is taken from the m'th element of the input array frequency domain frame signal vector x(k,l); the reordered interleaved output constraint matrix is:

[0023]

[0024] is the spatial position coordinate of microphone array element m;

[0025] The interleaving operation reorders the rows of the constraint matrix, that is, the steering vectors of the target sound source and the interference sound source are reordered using the corresponding interleaving table.

[0026] Furthermore, the specific process of step 3 is:

[0027] 3.1) The array frequency domain frame signal x output by branch p p (k,l) and the constraint matrix C p (k) is fed into the cascaded generalized sidelobe canceller of the corresponding branch, the input of the preceding generalized sidelobe canceller is of M1 dimension, and the input of the succeeding generalized sidelobe canceller is of M2 dimension, M=M1×M2;

[0028] 3.2) Initialize the equivalent weight vector w2(k,l)=w2(k,0) of the subsequent generalized sidelobe canceller. This is an M2-dimensional complex coefficient vector. l=0 represents the zeroth frame, i.e., the initial value. The weight of the delayed-sum beamforming is used as the initial value:

[0029]

[0030] 3.3) Use w2(k,l) to calculate and generate the pre-stage preprocessing matrix The symbol represents the Kronecker product operation, is an M1-dimensional identity matrix, so Q1(k,l) is an M×M1-dimensional preprocessing matrix, and the interleaved output array frequency domain frame signal x in step 2 p (k,l) is transformed by the projection matrix of the previous stage preprocessing to obtain the M1-dimensional frequency domain signal vector:

[0031]

[0032] H represents conjugate transpose;

[0033] At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the front-stage generalized sidelobe canceller:

[0034]

[0035] 3.4) The front-stage generalized sidelobe canceller performs weight optimization and signal enhancement on the M1-dimensional frequency domain signal vector obtained by preprocessing;

[0036] The generalized sidelobe canceller consists of three parts: a fixed beamformer on the main path, a blocking matrix and an adaptive noise canceller on the auxiliary path. The fixed beamformer on the main path allows the target signal to pass through and be enhanced to a certain extent. The fixed beam weight vector Located in the constrained subspace, fixed beam processing is performed on the input signal vector on the main path:

[0037]

[0038] Among them, Y f,p1 (k,l) is the fixed beam output on the main path of the previous generalized sidelobe canceller on branch p, and ||·||2 represents the L2-norm of the vector;

[0039] The previous blocking matrix B1 is located in the minimum variance subspace, which prevents the target signal from entering the auxiliary path. When the M1-dimensional frequency domain signal vector is sent to the auxiliary path, the vector u containing noise and interference information is obtained through the blocking matrix. p1 (k,l):

[0040]

[0041] The fixed beam output Y f,p1 (k,l) and blocking matrix output u p1 (k,l) form an M1+1 dimensional input vector The weight vector of the adaptive noise canceller is sent to the adaptive noise canceller. It is updated frame by frame, and the fixed beam output Y on the main road f,p1 (k, l) and the adaptive noise canceller output Y on the auxiliary line e,p1 The difference between (k,l) is:

[0042]

[0043] The weight coefficients are updated frame by frame using a multi-channel Wiener filtering method. The iterative calculation formula can be expressed as:

[0044]

[0045] Among them, μ is an adjustment coefficient, Φ S,p1 (k,l),Φ N,p1 (k,l) and Φ SN,p1 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the previous generalized sidelobe canceller, which are updated in each frame:

[0046]

[0047] Among them, α is the smoothing coefficient, M1+1 dimensional signal vector h S,p1 (k,l) is given by Y f,p1 (k, l) and M1 extremely small constants ε close to 0, denoted as h S,p1 (k,l)=[Y f,p1 , ε, …ε] T ; M1+1 dimensional noise vector h N,p1 (k,l) is composed of the output vector and input vector of the adaptive noise canceller, expressed as The subscript symbol * indicates taking the conjugate of the complex number; for any frequency point, the equivalent weight vector of the previous stage is calculated after each frame iteration:

[0048]

[0049] For vector The first element in Representation vector The vector composed of the second to the last elements in the previous stage equivalent weight vector is sent to the subsequent stage generalized sidelobe canceller;

[0050] 3.5) The post-stage generalized sidelobe canceller uses w1(k,l) to calculate and generate the post-stage preprocessing matrix in is an M2-dimensional identity matrix, so Q2(k,l) is an M×M2-dimensional preprocessing matrix, interleaving the output array frequency domain frame signal x p (k,l) is projected and transformed by the post-stage preprocessing matrix to obtain the M2-dimensional frequency domain signal vector:

[0051]

[0052] At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the post-stage generalized sidelobe canceller:

[0053]

[0054] 3.6) The fixed beam former on the main path of the post-stage generalized sidelobe canceller performs fixed beam processing on the input signal vector:

[0055]

[0056] Among them, Y f,p2 (k,l) is the fixed beam output on the main path of the post-stage generalized sidelobe canceller on branch p;

[0057] In the auxiliary circuit of the post-stage generalized sidelobe canceller, the post-stage blocking matrix B2 is used to process the M2-dimensional frequency domain signal vector to obtain the noise interference information vector u p2 (k,l) is expressed as:

[0058]

[0059] The fixed beam output Y f,p2 (k,l) and blocking matrix output u p2 (k,l) composed of M2+1 dimensional input vector And sent to the adaptive noise canceller, the weight vector of the adaptive noise canceller is The multi-channel Wiener filtering method is used to update the weight coefficients frame by frame:

[0060]

[0061] Φ S,p2 (k,l),Φ N,p2 (k,l) and Φ SN,p2 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the subsequent generalized sidelobe canceller, which are updated in each frame:

[0062]

[0063] Where α = 0.02, h S,p2 (k,l)=[Y f,p2 , ε, …ε] T , For any frequency point, the equivalent weight vector of the next stage is calculated after each frame iteration:

[0064]

[0065] in, For vector The first element in Representation vector A vector consisting of the second to last elements in ;

[0066] 3.7) The beam output of the cascaded generalized sidelobe canceller can be expressed as Y f,p2 (k,l) and Y e,p2 The difference between (k,l):

[0067]

[0068] After traversing all frequency points k, all beam outputs on branch p are obtained.

[0069] Furthermore, the specific process of step 4 is:

[0070] The linear constrained minimum variance criterion is used to calculate the beam output Y of each branch. p (k, l), p = 1, ..., P for fusion. Since the target signal is distortion-free, the branch with the minimum output power in the target direction is considered to be the best. Each frame of signal is fused to obtain the branch number corresponding to the optimal solution. for:

[0071]

[0072] In the formula, the objective function at the frequency point corresponding to each frame signal is:

[0073] Γ p (k,l)=(1-β)Γ p (k,l-1)+β|Y p (k,l)| 2 (twenty four)

[0074] Among them, β is the smoothing coefficient, take β = 0.015, then determine It is the optimal solution of the beamforming output at the lth frame and the kth frequency point.

[0075] Furthermore, in steps 3.4) and 3.6, μ=1.2, α=0.02.

[0076] The present invention also provides a low-complexity cascaded generalized sidelobe cancellation beamforming device for implementing the above method, comprising a microphone array, a microphone array signal processor, a signal fusion processor, a signal enhancement processor, a plurality of interleaving processors and a plurality of cascaded generalized sidelobe cancellation processors, each cascaded generalized sidelobe cancellation processor comprising a front-stage generalized sidelobe cancellation processor and a rear-stage generalized sidelobe cancellation processor cascaded front and rear; the front-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is respectively communicatively connected to an interleaving processor, each interleaving processor is respectively communicatively connected to a microphone array signal processor, and the microphone array is communicatively connected to the microphone array signal processor; the rear-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is communicatively connected to the signal fusion processor, and the signal fusion processor is communicatively connected to the signal enhancement processor;

[0077] The microphone array signal processor is used to process the time domain signal collected and received by the microphone array to obtain an array frequency domain frame signal;

[0078] The interleaving processor is used for performing interleaving processing on the array frequency domain frame signal;

[0079] The front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor are both used to perform generalized sidelobe cancellation processing on the array frequency domain frame signal output by the interleaving processor; the front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor mutually transmit equivalent weight vector information for iterative update;

[0080] The signal fusion processor is used to fuse the beams output by all subsequent generalized sidelobe cancellation processors;

[0081] The signal enhancement processor is used to enhance the signal output by the signal fusion processor.

[0082] Furthermore, the front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor both include a main path and an auxiliary path, the main path includes a fixed beamformer, and the auxiliary path includes a blocking matrix and an adaptive noise canceller; the blocking matrix is ​​composed of an interference filter.

[0083] The beneficial effects of the present invention are:

[0084] 1. The present invention constructs a beamformer with a cascaded generalized sidelobe cancellation structure, decomposing a high-dimensional generalized sidelobe canceller into two low-dimensional generalized sidelobe cancellation units cascaded front and back, which can significantly reduce the computational complexity and improve the robustness. The frame-based front-end and back-end feedback iterative processing method is used to improve the real-time performance of the algorithm, which has more practical value.

[0085] 2. The present invention utilizes the idea of ​​multi-channel Wiener filtering to construct an interference noise cancellation module, directly updates the filter coefficients between frames, and does not require additional noise estimation. It is simple to implement and highly robust, and can also balance interference noise suppression and target speech distortion.

[0086] 3. The present invention can improve the degree of freedom of beam weight coefficient search through multi-branch fusion and interleaving strategies, which is more conducive to the algorithm to find the optimal solution and improve the performance of target enhancement and interference noise suppression.

[0087] In summary, the present invention solves the problems of poor performance, low robustness and large amount of computation in practical application of existing adaptive beamforming technology methods for large microphone arrays through a cascaded generalized sidelobe cancellation structure based on frame iteration processing, multi-channel Wiener filtering interference noise elimination, and multi-branch fusion and interleaving strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 is an overall flow chart of the method in Example 1 of the present invention;

[0089] Figure 2 Array topology and signal model diagram in the method of embodiment 1 of the present invention;

[0090] Figure 3 This is a flow chart of the cascaded generalized sidelobe cancellation method in the method of embodiment 1 of the present invention;

[0091] Figure 4 : is the signal output power pattern of each beamforming method in the experiment of Example 2 of the present invention;

[0092] Figure 5 This is a comparison diagram of output gains of various beamforming methods under different signal-to-interference ratios in the experiment of Example 2 of the present invention;

[0093] Figure 6 The simulation comparison of the operation time of each beamforming method in the experiment of Example 2 of the present invention is shown;

[0094] Figure 7 This is a schematic diagram of the structure of the device in Example 3 of the present invention. DETAILED DESCRIPTION

[0095] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0096] Example 1

[0097] This embodiment provides a low-complexity cascaded generalized sidelobe cancellation beamforming method, such as Figure 1 As shown, the following steps are included:

[0098] Step 1: The constraint matrix is ​​obtained by using the coordinate information of each microphone array element in the microphone array and the direction parameters of the sound source wave. The time domain signal collected by the microphone array is processed by framing, windowing and short-time Fourier transform to obtain the array frequency domain frame signal.

[0099] The method of this embodiment is applicable to microphone arrays of any shape, including linear arrays, planar arrays, and spherical arrays. This embodiment is described by taking a three-dimensional microphone array in which microphone array elements are arbitrarily distributed in space. The array topology and signal model are as follows: Figure 2 As shown, the microphone array consists of M omnidirectional microphone array elements, which are numbered 1, 2, ...m, ..., M in sequence. The spatial position coordinates of the mth omnidirectional microphone array element are represented by r in the three-dimensional Cartesian coordinate system. m =[r x,m ,r y,m ,r z,m ] T , T represents transposition, and the far-field incident angle azimuth parameters of the known target sound source are where θ d and is the pitch angle and azimuth of the target sound source d. Assume that there are J interference sound sources, and the far-field incident angle azimuth parameter of the jth interference sound source is θ j and are the pitch angle and azimuth angle of the jth interference sound source, respectively, and the constraint matrix is ​​obtained:

[0100] C(k)=[v(k,Ω d ),v(k,Ω1),…,v(k,Ω J )] T ,k=0,1,2,…N fft / twenty one)

[0101] k is N fft The frequency point number of the Fourier transform of the point, v(k,Ω d ) and v(k,Ω j )(j=1,2,…J) are the guidance vectors of the target sound source and the interference sound source respectively, and the following information is obtained by using the coordinates of the microphone array element and the direction of the sound source wave:

[0102]

[0103] Where i represents the imaginary unit, κ d and κ j are the sound source wave numbers of the target sound source and the interference sound source, c is the propagation speed of sound in space, f k is the frequency value corresponding to frequency point k.

[0104] Based on the signal distortion-free criterion, the array beam weight vector w is generally required to satisfy C H w=g,g=[1,0,…0] T is the J+1-dimensional constraint target vector, C represents the constraint matrix, and H represents the conjugate transpose.

[0105] The time domain signal collected by the microphone array is framed, windowed and short-time Fourier transformed to obtain the array frequency domain frame signal of the lth frame and the kth frequency point: x(k,l)=[x1(k,l),…,x m (k,l),…,x M (k,l)] T .

[0106] Step 2: interleave the array frequency domain frame signal and constraint matrix obtained in step 1 in multiple parallel branches to obtain the reordered array frequency domain frame signal and constraint matrix of each branch.

[0107] The method of this embodiment uses P parallel interleaving processors to process signals, providing more degrees of freedom for subsequent beam weight optimization. The interleaving processor interleaves and reorders the array frequency domain frame signal x(k, l) and the constraint matrix C(k). The interleaving table corresponding to the interleaving processor of the pth branch is m=I p (m'), then the array frequency domain frame signal interleaved and output by the interleaving processor of the pth branch is:

[0108]

[0109] Interleaved output array frequency domain frame signal vector x p The m'th element of (k,l) is taken from the mth element of the input array frequency domain frame signal vector x(k,l). The reordered interleaved output constraint matrix is:

[0110]

[0111] is the spatial position coordinate of microphone array element m;

[0112] The interleaving operation reorders the rows of the constraint matrix, that is, the steering vectors of the target sound source and the interference sound source are reordered using the corresponding interleaving table.

[0113] Step 3: The method of this embodiment decomposes the high-dimensional generalized sidelobe canceller into a cascaded generalized sidelobe canceller, wherein the cascaded generalized sidelobe canceller includes two low-dimensional front-stage generalized sidelobe cancellers and a rear-stage generalized sidelobe canceller, and updates the parameters through frame-based iterative operations to obtain the output frequency domain information of each branch. Figure 3 As shown, the specific process is:

[0114] 3.1) The array frequency domain frame signal x output by branch p p (k,l) and the constraint matrix C p (k) is sent to the cascaded generalized sidelobe canceller of the corresponding branch, the input of the front-stage generalized sidelobe canceller is of M1 dimension, and the input of the rear-stage generalized sidelobe canceller is of M2 dimension, M=M1×M2.

[0115] 3.2) Initialize the equivalent weight vector w2(k,l)=w2(k,0) of the subsequent generalized sidelobe canceller. This is an M2-dimensional complex coefficient vector. l=0 represents the zeroth frame, i.e., the initial value. To ensure the reliability of the algorithm, the weight of the delayed-sum beamforming is used as the initial value:

[0116]

[0117] It can be seen that the initialization value is only related to the target sound source steering vector information of the first M2 microphone array elements after interleaving.

[0118] 3.3) Use w2(k,l) to calculate and generate the pre-stage preprocessing matrix The symbol represents the Kronecker product operation, is an M1-dimensional identity matrix, so Q1(k,l) is an M×M1-dimensional preprocessing matrix, and the interleaved output array frequency domain frame signal x in step 2 p (k,l) is transformed by the projection matrix of the previous stage preprocessing to obtain the M1-dimensional frequency domain signal vector:

[0119]

[0120] H represents conjugate transpose;

[0121] At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the front-stage generalized sidelobe canceller:

[0122]

[0123] 3.4) The front-stage generalized sidelobe canceller performs weight optimization and signal enhancement on the M1-dimensional frequency domain signal vector obtained by preprocessing.

[0124] The generalized sidelobe canceller consists of three parts: a fixed beamformer on the main path, a blocking matrix and an adaptive noise canceller on the auxiliary path. The fixed beamformer on the main path allows the target signal to pass through and be enhanced to a certain extent. The fixed beam weight vector Located in the constrained subspace, fixed beam processing is performed on the input signal vector on the main path:

[0125]

[0126] Among them, Y f,p1 (k,l) is the fixed beam output on the main path of the preceding generalized sidelobe canceller on branch p, and ||·||2 represents the L2-norm of the vector.

[0127] The previous blocking matrix B1 is located in the minimum variance subspace, which prevents the target signal from entering the auxiliary path. When the M1-dimensional frequency domain signal vector is sent to the auxiliary path, the vector u containing noise and interference information is obtained through the blocking matrix. p1 (k,l):

[0128]

[0129] The fixed beam output Y f,p1 (k,l) and blocking matrix output u p1 (k,l) form an M1+1 dimensional input vector The weight vector of the adaptive noise canceller is sent to the adaptive noise canceller. It is updated frame by frame, and the fixed beam output Y on the main road f,p1 (k, l) and the adaptive noise canceller output Y on the auxiliary line e,p1 The difference between (k,l) is:

[0130]

[0131] The weight coefficients are updated frame by frame using a multi-channel Wiener filtering method. The iterative calculation formula can be expressed as:

[0132]

[0133] Among them, μ is an adjustment coefficient, here we set μ = 1.2; Φ S,p1 (k,l),Φ N,p1 (k,l) and Φ SN,p1 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the previous generalized sidelobe canceller, which are updated in each frame:

[0134]

[0135] Where α is the smoothing coefficient, and α is taken as 0.02. M1+1-dimensional signal vector h S,p1 (k,l) is given by Y f,p1 (k, l) and M1 very small constants ε close to 0 (such as ε = 10 -10 ) is composed of S,p1 (k,l)=[Y f,p1 , ε, …ε] T ; M1+1 dimensional noise vector h N,p1(k,l) is composed of the output vector and input vector of the adaptive noise canceller, expressed as The subscript symbol * indicates taking the conjugate of a complex number. For any frequency point, the equivalent weight vector of the previous stage is calculated after each frame iteration:

[0136]

[0137] For vector The first element in Representation vector The vector composed of the second to the last elements in , and the equivalent weight vector of the previous stage is sent to the generalized sidelobe canceller of the subsequent stage.

[0138] 3.5) The post-stage generalized sidelobe canceller uses w1(k,l) to calculate and generate the post-stage preprocessing matrix in is an M2-dimensional identity matrix, so Q2(k,l) is an M×M2-dimensional preprocessing matrix, interleaving the output array frequency domain frame signal x p (k,l) is projected and transformed by the post-stage preprocessing matrix to obtain the M2-dimensional frequency domain signal vector:

[0139]

[0140] At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the post-stage generalized sidelobe canceller:

[0141]

[0142] 3.6) The fixed beam former on the main path of the post-stage generalized sidelobe canceller performs fixed beam processing on the input signal vector:

[0143]

[0144] Among them, Y f,p2 (k,l) is the fixed beam output on the main path of the post-stage generalized sidelobe canceller on branch p;

[0145] In the auxiliary circuit of the post-stage generalized sidelobe canceller, the post-stage blocking matrix B2 is used to process the M2-dimensional frequency domain signal vector to obtain the noise interference information vector u p2 (k,l) is expressed as:

[0146]

[0147] The fixed beam output Y f,p2 (k,l) and blocking matrix output u p2 (k,l) composed of M2+1 dimensional input vector And sent to the adaptive noise canceller, the weight vector of the adaptive noise canceller is The multi-channel Wiener filtering method is used to update the weight coefficients frame by frame:

[0148]

[0149] Similarly, μ=1.2, Φ S,p2 (k,l),Φ N,p2 (k,l) and Φ SN,p2 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the subsequent generalized sidelobe canceller, which are updated in each frame:

[0150]

[0151] Where α = 0.02, h S,p2 (k,l)=[Y f,p2 , ε, …ε] T , For any frequency point, the equivalent weight vector of the next stage is calculated after each frame iteration:

[0152]

[0153] in, For vector The first element in Representation vector A vector consisting of the second to last elements in .

[0154] 3.7) The beam output of the cascaded generalized sidelobe canceller can be expressed as Y f,p2 (k,l) and Y e,p2 The difference between (k,l):

[0155]

[0156] After traversing all frequency points k, all beam outputs on branch p are obtained.

[0157] Step 4: The beam outputs of the cascaded generalized sidelobe cancellers of each branch are integrated to obtain the optimal solution for the beam outputs of the multi-path cascaded generalized sidelobe cancellers.

[0158] The linear constrained minimum variance criterion is used to calculate the beam output Y of each branch. p(k, l), p = 1, ..., P for fusion. Although the constraint matrices in each branch are interleaved, they are essentially the same. Since the cascaded generalized sidelobe canceller structure and the adjustment coefficient μ of each branch are the same, it can be considered that the linear constraints achieved by each branch on the target are very close. The main reason for the inconsistency is the difference in the amount of noise interference suppression. Since the target signal is distortion-free, it can be considered that the branch with the minimum output power in the target direction is the optimal one. Each frame of signal is fused to obtain the branch number corresponding to the optimal solution. for:

[0159]

[0160] In the formula, the objective function at the frequency point corresponding to each frame signal is:

[0161] Γ p (k,l)=(1-β)Γ p (k,l-1)+β|Y p (k,l)| 2 (twenty four)

[0162] Among them, β is the smoothing coefficient, taking β = 0.015, it is determined as Output the optimal solution for beamforming at the lth frame and the kth frequency point.

[0163] Step 5: The frequency domain output signal is subjected to short-time Fourier inverse transform, windowing and framing to obtain a time domain signal after noise interference suppression and target speech enhancement.

[0164] Frequency domain output signal The target sound time domain output signal can be restored through short-time Fourier inverse transform, windowing and framing operations. When the direction of the target sound source or the interference sound source changes, it is necessary to return to step 1 to generate a new constraint matrix, initialize and restart the entire calculation process.

[0165] Example 2

[0166] This example aims to verify the effectiveness of the method of Example 1 through experiments.

[0167] In the experiment, the number of microphone array elements is M=20, the cascaded generalized sidelobe cancellers are M1=5, M2=4, the parameters are defined as adjustment coefficient μ=1.2, smoothing factor α=0.02, β=0.015, and there is a scene with a target sound source and an interference sound source in the simulation environment.

[0168] Figure 4The figure is a comparison of the broadband signal output power pattern of the method of Example 1 and the common beamforming method as the incident angle changes. It can be seen that the traditional fixed beamforming technology can effectively enhance the target signal while obtaining sufficiently low side lobes through methods such as windowing, but it cannot eliminate the directional interference sound source. The adaptive beamforming method, Kronecker product minimum variance distortionless response beamforming KPMVDR method and generalized sidelobe cancellation GSC method can form a null in the direction of the interference sound source, effectively suppress the signal from the interference direction, but at the same time it will raise the side lobe, mainly in the low frequency band. The directivity will deteriorate, resulting in the introduction of directional interference noise. It can be seen that while the method of Example 1 suppresses the interference sound source, the side lobe is much lower than the KPMVDR and GSC methods, which is close to the sidelobe suppression fixed beamforming method. The sidelobe suppression outside the main lobe can reach about 20dB. The multi-channel Wiener filtering interference noise elimination and multi-branch fusion strategy can significantly improve the interference and noise suppression and robustness, and improve the target voice quality.

[0169] Figure 5 The output gain diagram of the beamforming method under different signal-to-interference ratios is compared. There are three groups of interference sound sources in the environment, including white noise, indoor noisy noise and street noise. As the interference intensity changes, the broadband signal gain of the traditional delay-sum fixed beam method (DS), MVDR method, KPMVDR method, GSC method and Example 1 method is compared. Obviously, under low signal-to-interference ratio conditions, the performance of the delay-sum beamformer is poor, and the performance gradually improves with the increase of the signal-to-interference ratio. This shows that under high interference and noise conditions, the fixed beamformer has great limitations. At the same time, whether under low or high interference levels, the performance of the KPMVDR method, GSC method and Example 1 method beamformers is far better than the minimum variance distortionless response beamformer. At the same time, because the method of Example 1 makes full use of the array degrees of freedom and the potential adaptive multi-channel Wiener filter canceller to accurately estimate the statistical characteristics of noise, its processing gain is slightly better than the KPMVDR method and the GSC method, which verifies the robustness and effectiveness of the method of Example 1 in different interference and noise scenarios.

[0170] Figure 6The algorithm operation time of the generalized sidelobe cancellation method and the cascaded generalized sidelobe cancellation method in Example 1 is compared. The simulation uses an array with an array element number of M (M1=M2=3, 4, ... 10). The time required for running the algorithm to process a sound signal of 30 seconds in length is simulated on Matlab. It can be seen that the operation time of the GSC method increases significantly as the number of array elements increases. This is because when the number of array elements increases, high-dimensional matrix operations are required, and the complexity becomes increasingly high, and it cannot be used in scenarios with high real-time requirements. The method in Example 1 is not sensitive to the increase in the number of array elements, especially for large arrays with a large number of array elements. The method in Example 1 decomposes the high-dimensional generalized sidelobe canceller into two low-dimensional generalized sidelobe cancellers in a cascade manner, which greatly reduces the operation time and complexity, and is more conducive to real-time implementation and application deployment.

[0171] Example 3

[0172] This embodiment provides a low-complexity cascaded generalized sidelobe cancellation beamforming device, such as Figure 7 As shown, it includes a microphone array, a microphone array signal processor, a signal fusion processor, a signal enhancement processor, a plurality of interleaving processors and a plurality of cascaded generalized sidelobe cancellation processors, each cascaded generalized sidelobe cancellation processor includes a front-stage generalized sidelobe cancellation processor and a rear-stage generalized sidelobe cancellation processor cascaded front and rear; the front-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is respectively communicatively connected to an interleaving processor, each interleaving processor is respectively communicatively connected to the microphone array signal processor, and the microphone array is communicatively connected to the microphone array signal processor; the rear-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is communicatively connected to the signal fusion processor, and the signal fusion processor is communicatively connected to the signal enhancement processor;

[0173] The microphone array signal processor is used to process the time domain signal collected and received by the microphone array to obtain an array frequency domain frame signal;

[0174] The interleaving processor is used for performing interleaving processing on the array frequency domain frame signal;

[0175] The front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor are both used to perform generalized sidelobe cancellation processing on the array frequency domain frame signal output by the interleaving processor; the front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor mutually transmit equivalent weight vector information for iterative update;

[0176] The signal fusion processor is used to fuse the beams output by all subsequent generalized sidelobe cancellation processors;

[0177] The signal enhancement processor is used to enhance the signal output by the signal fusion processor.

[0178] In this embodiment, the front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor both include a main path and an auxiliary path, the main path includes a fixed beamformer, and the auxiliary path includes a blocking matrix and an adaptive noise canceller; the blocking matrix is ​​composed of an interference filter.

[0179] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A low-complexity cascaded generalized sidelobe cancellation beamforming method, characterized in that: The steps include: Step 1: Using the coordinate information of each microphone array element in the microphone array and the direction parameters of the sound source wave to obtain a constraint matrix, the time domain signal collected and received by the microphone array is processed by framing, windowing and short-time Fourier transform to obtain an array frequency domain frame signal; Step 2: interleave the array frequency domain frame signal and constraint matrix obtained in step 1 in multiple parallel branches to obtain the reordered array frequency domain frame signal and constraint matrix of each branch; Step 3: setting a plurality of cascaded generalized sidelobe cancellers, each of which includes two low-dimensional front-stage generalized sidelobe cancellers and a rear-stage generalized sidelobe canceller that are cascaded front and back, and updating parameters through frame-based iterative operations to obtain output frequency domain information of each branch; Step 4, the beam outputs of the cascaded generalized sidelobe cancellers of each branch are integrated to obtain the optimal solution of the beam outputs of the multi-path cascaded generalized sidelobe cancellers; The frequency domain output signal finally outputted in step 5 and step 4 is subjected to short-time Fourier inverse transform, windowing and framing to obtain a time domain signal after noise interference suppression and target speech enhancement.

2. The method according to claim 1, characterized in that The specific process of step 1 is: The microphone array consists of M microphone elements, which are numbered 1, 2, ...m, ..., M. The spatial position coordinates of the mth microphone element are expressed as r in the three-dimensional Cartesian coordinate system. m =[r x,m ,r y,m ,r z,m ] T , T represents transposition, and the far-field incident angle azimuth parameters of the known target sound source are where θ d and is the pitch angle and azimuth of the target sound source d; suppose there are J interference sound sources, and the far-field incident angle azimuth parameter of the jth interference sound source is j=1,2,…J,θ j and are the pitch angle and azimuth angle of the jth interference sound source, respectively, and the constraint matrix is ​​obtained: C(k)=[v(k,Ω d ),v(k,Ω1),…,v(k,Ω J )] T , k=0,1,2,…N fft / 2 (1) k is N fft The frequency point number of the Fourier transform of the point, v(k,Ω d ) and v(k,Ω j )(j=1,2,…J) are the guidance vectors of the target sound source and the interference sound source respectively, and the following information is obtained by using the coordinates of the microphone array element and the direction of the sound source wave: Where i represents the imaginary unit, κ d and κ j are the sound source wave numbers of the target sound source and the interference sound source, c is the propagation speed of sound in space, f k is the frequency value corresponding to frequency point k; Based on the signal distortion-free criterion, the array beam weight vector w is required to satisfy C H w=g,g=[1,0,…0] T is the J+1-dimensional constraint target vector, C represents the constraint matrix, and H represents the conjugate transpose; The time domain signal collected by the microphone array is framed, windowed and short-time Fourier transformed to obtain the array frequency domain signal of the lth frame and the kth frequency point: x(k,l)=[x1(k,l),…,x m (k,l),…,x M (k,l)] T .

3. The method according to claim 2, characterized in that The specific process of step 2 is: P parallel interleaving processors are used for signal processing. The interleaving processors interleave and reorder the array frequency domain frame signal x(k, l) and the constraint matrix C(k). The interleaving table corresponding to the interleaving processor of the pth branch is m=I p (m'), then the array frequency domain frame signal interleaved and output by the interleaving processor of the pth branch is: Interleaved output array frequency domain frame signal vector x p The m'th element of (k,l) is taken from the m'th element of the input array frequency domain frame signal vector x(k,l); the reordered interleaved output constraint matrix is: is the spatial position coordinate of microphone array element m; The interleaving operation reorders the rows of the constraint matrix, that is, the steering vectors of the target sound source and the interference sound source are reordered using the corresponding interleaving table.

4. The method according to claim 3, characterized in that The specific process of step 3 is: 3.1) The array frequency domain frame signal x output by branch p p (k,l) and the constraint matrix C p (k) is fed into the cascaded generalized sidelobe canceller of the corresponding branch, the input of the preceding generalized sidelobe canceller is of M1 dimension, and the input of the succeeding generalized sidelobe canceller is of M2 dimension, M=M1×M2; 3.2) Initialize the equivalent weight vector w2(k,l)=w2(k,0) of the subsequent generalized sidelobe canceller. This is an M2-dimensional complex coefficient vector. l=0 represents the zeroth frame, i.e., the initial value. The weight of the delayed-sum beamforming is used as the initial value: 3.3) Use w2(k,l) to calculate and generate the pre-stage preprocessing matrix The symbol represents the Kronecker product operation, is an M1-dimensional identity matrix, so Q1(k,l) is an M×M1-dimensional preprocessing matrix, and the interleaved output array frequency domain frame signal x in step 2 p (k,l) is transformed by the projection matrix of the previous stage preprocessing to obtain the M1-dimensional frequency domain signal vector: H represents conjugate transpose; At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the front-stage generalized sidelobe canceller: 3.4) The front-stage generalized sidelobe canceller performs weight optimization and signal enhancement on the M1-dimensional frequency domain signal vector obtained by preprocessing; The generalized sidelobe canceller consists of three parts: a fixed beamformer on the main path, a blocking matrix and an adaptive noise canceller on the auxiliary path. The fixed beamformer on the main path allows the target signal to pass through and be enhanced to a certain extent. The fixed beam weight vector Located in the constrained subspace, fixed beam processing is performed on the input signal vector on the main path: Among them, Y f,p1 (k,l) is the fixed beam output on the main path of the previous generalized sidelobe canceller on branch p, and ||·||2 represents the L2-norm of the vector; The previous blocking matrix B1 is located in the minimum variance subspace, which prevents the target signal from entering the auxiliary path. When the M1-dimensional frequency domain signal vector is sent to the auxiliary path, the vector u containing noise and interference information is obtained through the blocking matrix. p1 (k,l): The fixed beam output Y f,p1 (k,l) and blocking matrix output u p1 (k,l) form an M1+1 dimensional input vector The weight vector of the adaptive noise canceller is sent to the adaptive noise canceller. It is updated frame by frame, and the fixed beam output Y on the main road f,p1 (k, l) and the adaptive noise canceller output Y on the auxiliary line e,p1 The difference between (k,l) is: The weight coefficients are updated frame by frame using a multi-channel Wiener filtering method. The iterative calculation formula can be expressed as: Among them, μ is an adjustment coefficient, Φ S,p1 (k,l),Φ N,p1 (k,l) and Φ SN,p1 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the previous generalized sidelobe canceller, which are updated in each frame: Among them, α is the smoothing coefficient, M1+1 dimensional signal vector h S,p1 (k,l) is given by Y f,p1 (k, l) and M1 extremely small constants ε close to 0, denoted as h S,p1 (k,l)=[Y f,p1 , ε, …ε] T ; M1+1 dimensional noise vector h N,p1 (k,l) is composed of the output vector and input vector of the adaptive noise canceller, expressed as The subscript symbol * indicates taking the conjugate of the complex number; for any frequency point, the equivalent weight vector of the previous stage is calculated after each frame iteration: For vector The first element in Representation vector The vector composed of the second to the last elements in the previous stage equivalent weight vector is sent to the subsequent stage generalized sidelobe canceller; 3.5) The post-stage generalized sidelobe canceller uses w1(k,l) to calculate and generate the post-stage preprocessing matrix in is an M2-dimensional identity matrix, so Q2(k,l) is an M×M2-dimensional preprocessing matrix, the interleaved output array frequency domain frame signal x p (k,l) is projected and transformed by the post-stage preprocessing matrix to obtain the M2-dimensional frequency domain signal vector: At the same time, the constraint matrix of the interleaved output is preprocessed to obtain the constraint matrix of the post-stage generalized sidelobe canceller: 3.6) The fixed beam former on the main path of the post-stage generalized sidelobe canceller performs fixed beam processing on the input signal vector: Among them, Y f,p2 (k,l) is the fixed beam output on the main path of the post-stage generalized sidelobe canceller on branch p; In the auxiliary circuit of the post-stage generalized sidelobe canceller, the post-stage blocking matrix B2 is used to process the M2-dimensional frequency domain signal vector to obtain the noise interference information vector u p2 (k,l) is expressed as: The fixed beam output Y f,p2 (k,l) and blocking matrix output u p2 (k,l) composed of M2+1 dimensional input vector And sent to the adaptive noise canceller, the weight vector of the adaptive noise canceller is The multi-channel Wiener filtering method is used to update the weight coefficients frame by frame: Φ S,p2 (k,l),Φ N,p2 (k,l) and Φ SN,p2 (k, l) are the signal covariance matrix, noise covariance matrix and signal-noise cross-covariance vector of the subsequent generalized sidelobe canceller, which are updated in each frame: Where α = 0.02, h S,p2 (k,l)=[Y f,p2 , ε, …ε] T , For any frequency point, the equivalent weight vector of the next stage is calculated after each frame iteration: in, For vector The first element in Representation vector A vector consisting of the second to last elements in ; 3.7) The beam output of the cascaded generalized sidelobe canceller can be expressed as Y f,p2 (k,l) and Y e,p2 The difference between (k,l): After traversing all frequency points k, all beam outputs on branch p are obtained.

5. The method according to claim 4, characterized in that The specific process of step 4 is: The linear constrained minimum variance criterion is used to calculate the beam output Y of each branch. p (k, l), p = 1, ..., P are fused. Since the target signal is distortion-free, the branch with the minimum output power in the target direction is considered to be the best. Each frame of signal is fused to obtain the branch number corresponding to the optimal solution. for: In the formula, the objective function at the frequency point corresponding to each frame signal is: Γ p (k,l)=(1-β)Γ p (k,l-1)+βY p (k,l) 2 (24) Among them, β is the smoothing coefficient, take β = 0.015, then determine It is the optimal solution of the beamforming output at the lth frame and the kth frequency point.

6. The method according to claim 4, characterized in that In steps 3.4) and 3.6, μ=1.2, α=0.

02.

7. A low-complexity cascaded generalized sidelobe cancellation beamforming device for implementing the method according to any one of claims 1 to 6, characterized in that: It includes a microphone array, a microphone array signal processor, a signal fusion processor, a signal enhancement processor, a plurality of interleaving processors and a plurality of cascaded generalized sidelobe cancellation processors, each of which includes a front-stage generalized sidelobe cancellation processor and a rear-stage generalized sidelobe cancellation processor cascaded front and rear; the front-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is respectively communicatively connected to an interleaving processor, each interleaving processor is respectively communicatively connected to the microphone array signal processor, and the microphone array is communicatively connected to the microphone array signal processor; the rear-stage generalized sidelobe cancellation processor of each cascaded generalized sidelobe cancellation processor is communicatively connected to the signal fusion processor, and the signal fusion processor is communicatively connected to the signal enhancement processor; The microphone array signal processor is used to process the time domain signal collected and received by the microphone array to obtain an array frequency domain frame signal; The interleaving processor is used for performing interleaving processing on the array frequency domain frame signal; The front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor are both used to perform generalized sidelobe cancellation processing on the array frequency domain frame signal output by the interleaving processor; the front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor mutually transmit equivalent weight vector information for iterative update; The signal fusion processor is used to fuse the beams output by all subsequent generalized sidelobe cancellation processors; The signal enhancement processor is used to enhance the signal output by the signal fusion processor.

8. The device according to claim 7, characterized in that The front-stage generalized sidelobe cancellation processor and the rear-stage generalized sidelobe cancellation processor both include a main path and an auxiliary path, the main path includes a fixed beamformer, and the auxiliary path includes a blocking matrix and an adaptive noise canceller; the blocking matrix is ​​composed of an interference filter.

Images