Fast deconvolution sound source localization method based on array sparse and broadband comprehensive processing

Abstract

The invention discloses a fast deconvolution sound source localization method based on array sparsity and broadband comprehensive processing, which comprises the following steps of: selecting microphones to form a sparse array, and converting a regular focusing plane into an irregular plane to obtain grid coordinates; calculating a sparse array cross-spectrum matrix based on coordinates, obtaining a function beam through eigenvalue decomposition and power exponent operation, and outputting the function beam; constructing a frequency domain deconvolution model according to the convolution relationship between the frequency domain and sound source distribution and the convolution relationship between the frequency domain and a rising power point spread function, and solving a sound source intensity distribution matrix by using fast Fourier transform and iterative optimization; estimating a noise floor-to-signal-to-noise ratio matrix, traversing frequency points, taking the maximum signal-to-noise ratio, and fusing the frequency points into a broadband imaging matrix; a time smoothing factor is introduced to carry out weighted fusion on the inter-frame imaging matrix, and positioning is completed; according to the method, low-consumption, high-precision and strong-robustness sound source localization is realized through array rarefaction and fast deconvolution load reduction, precision protection by means of function beam forming and Gaussian filtering, multi-frequency point fusion adaptive imaging, inter-frame smooth flicker suppression and adaptation to complex scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of acoustic imaging technology in the field of sound source positioning, in particular to a fast deconvolution sound source positioning method based on array sparsity and broadband comprehensive processing. BACKGROUND

[0002] Today, the acoustic imaging technology has developed to the fast deconvolution algorithm based on the microphone array, which has great advantages in performance compared with the traditional beamforming algorithm and deconvolution algorithm due to its consideration of high resolution and computational efficiency. However, the algorithm generally holds under narrowband and needs to determine the sound source frequency and bandwidth in advance. In the face of actual sound sources with unknown frequency spectrum and bandwidth and low signal-to-noise ratio, if the sound source frequency and bandwidth cannot be correctly determined, the traditional algorithm will lead to one-sided imaging, positioning drift and weak source submersion. Although the function beamforming and deconvolution technology significantly improve the narrowband imaging performance, in the broadband application scenario, there is still a lack of an integrated imaging mechanism that can adaptively fuse effective information of all frequency bands without prior spectrum knowledge, and the high computational cost brought by large arrays also restricts its engineering deployment. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a fast deconvolution sound source positioning method based on array sparsity and broadband comprehensive processing, which aims to solve the problems in the background art.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a fast deconvolution sound source positioning method based on array sparsity and broadband comprehensive processing, comprising the following steps: Step S1: divide the microphone array into several clusters by k-means clustering, and select the microphones closest to the cluster centers to form a sparse array; Step S2: convert the regular focusing plane to an irregular focusing plane, and calculate the irregular grid coordinates by angle sine value equidistant sampling; Step S3: calculate the cross-spectral matrix of the sparse array based on the irregular grid coordinates and perform eigenvalue decomposition, and introduce a power index to obtain the function beamforming output; Step S4: according to the convolution relationship between the function beamforming output, the sound source distribution and the power-up point spread function, construct a frequency domain deconvolution model; Step S5: convert the convolution in the model to a wave number domain product by using fast Fourier transform, and solve the sound source intensity distribution matrix by zero padding and iterative optimization; Step S6: estimate the noise basis of the sound source intensity distribution matrix and set a threshold, and convert it to a sound source distribution signal-to-noise ratio matrix; Step S7: traverse all frequency points for the sound source distribution signal-to-noise ratio matrix, and take the maximum signal-to-noise ratio of each grid point to fuse into a broadband comprehensive imaging matrix; Step S8: Introducing a time smoothing factor, the current frame and the last frame imaging matrix are weighted and fused to obtain the final acoustic imaging result, and the sound source positioning is completed.

[0005] Further, the specific process of step S1 is: Step S1.1: First, determine the position information of all microphones to form a position matrix containing the total number of microphones, wherein the coordinates of each microphone correspond to a row of data in the matrix; Step S1.2: Using the k-means clustering algorithm, the entire microphone array is divided into several clusters, and the clustering process aims to minimize the within-cluster sum of squares; the cluster assignment of the microphones is adjusted and the cluster centroid is updated through continuous iteration until the clustering result is stable; Step S1.3: Calculate the distance from all microphones in each cluster to the cluster centroid, and select the microphone closest to the centroid as the representative of the cluster, and finally form a sparse array composed of representative microphones of all clusters.

[0006] Further, the specific process of step S2 is: Step S2.1: Determine the coordinates of the grid points in the regular focusing plane, and the vertical distance between the focusing grid plane and the microphone array plane; Step S2.2: By calculating the distance from the grid points in the focusing plane to the reference point, and the trigonometric function relationship between the distance and the horizontal and vertical coordinates, the corresponding angle sine value is obtained; Step S2.3: Within the set range of angle sine value, discrete sampling is performed in an equal interval manner to obtain a series of angle sine values of sampling points; Step S2.4: According to the angle sine values obtained by sampling, combined with the vertical distance between the focusing plane and the microphone array plane, the specific coordinates of each grid point in the irregular focusing plane are calculated, and the conversion from the regular focusing plane to the irregular focusing plane is completed.

[0007] Further, the specific process of step S3 is: Step S3.1: Perform Fourier transform on the sound signal recorded by each microphone in the sparse array to obtain the pressure amplitude data corresponding to different frequencies for each microphone, and then construct a column matrix containing the number of sparse microphones based on the pressure amplitude data; Step S3.2: Multiply the column matrix by its conjugate transpose matrix to obtain the cross-spectrum matrix of the sparse array signal; Step S3.3: Perform eigenvalue decomposition on the cross-spectrum matrix to obtain a unitary matrix composed of the eigenvectors of the cross-spectrum matrix and a diagonal matrix composed of the eigenvalues; Step S3.4: introducing the positive integer power index, combining the constructed steering vector, calculating the function beamforming output; the steering vector is calculated by the exponential function according to the microphone array position, sound speed and frequency.

[0008] Further, the specific process of step S4 is: according to the function beamforming output obtained in step S3, analyzing the relationship between the function beamforming output and the sound source intensity distribution and the raised point spread function; the function beamforming output is expressed as the spatial convolution of the sound source intensity distribution and the raised point spread function. Based on the function beamforming output expressed as the spatial convolution of the sound source intensity distribution and the raised point spread function, an inverse convolution model in the frequency domain is constructed.

[0009] Further, the specific process of step S5 is: Step S5.1: zero padding processing is performed on the function beamforming output, the sound source intensity distribution matrix and the raised point spread function; Step S5.2: the fast Fourier transform result of the zero-padded raised point spread function is calculated, and the absolute value sum of the result is calculated; Step S5.3: the zero-padded sound source intensity distribution matrix is initialized as a zero matrix, and then iterative calculation is performed through the iterative formula: first, the function beamforming estimation result of the current iteration is calculated according to the sound source intensity distribution matrix of the current iteration, the Fourier transform result of the raised point spread function and the Gaussian regularization filter function; then, the sound source intensity distribution matrix is updated in combination with the zero-padded function beamforming output. Step S5.4: after the iteration is completed, the zero-padded part is removed to obtain the finally estimated sound source intensity distribution matrix.

[0010] Further, the specific process of step S6 is: Step S6.1: the finally estimated sound source intensity distribution matrix obtained in step S5 is converted into a column vector form, each column vector containing the sound source intensity data of all grid points at the corresponding frequency; Step S6.2: each sound source intensity column vector corresponding to a frequency is sorted in ascending order; according to the length of the sorted vector, the minimum values of a preset proportion are selected, and the average value of these minimum values is calculated as the noise floor estimation value at the corresponding frequency; Step S6.3: according to the noise floor estimation value, a threshold is set to process the sound source intensity matrix at each frequency: if the sound source intensity of a grid point is less than or equal to the noise floor estimation value, the value of the point is set to 0; if it is greater than the noise floor estimation value, the sound source intensity of the point is divided by the noise floor estimation value, and finally the sound source distribution signal-to-noise ratio matrix at each frequency is obtained.

[0011] Further, the specific process of step S7 is: for the sound source distribution signal-to-noise ratio matrix at all frequencies obtained in step S6, for each spatial grid point, traversing the signal-to-noise ratio values corresponding to the grid point at all frequencies, selecting the maximum signal-to-noise ratio value as the final imaging intensity value of the grid point; integrating the maximum signal-to-noise ratio values of all grid points to form a wideband comprehensive imaging matrix.

[0012] Further, the specific process of step S8 is: introducing a time smoothing factor, performing inter-frame weighted fusion processing on the wideband comprehensive imaging matrix obtained in step S7; during fusion, multiplying the value of each grid point in the current frame wideband comprehensive imaging matrix by the time smoothing factor, multiplying the value of the corresponding grid point in the imaging matrix of the last frame by 1 minus the time smoothing factor, adding the two results to obtain the final acoustic imaging result of the current frame, thereby completing the sound source positioning.

[0013] A non-volatile computer storage medium, the computer storage medium stores computer executable instructions, the computer executable instructions execute the fast deconvolution sound source positioning method based on array sparsification and wideband comprehensive processing.

[0014] Compared with the existing technology, the present application has the following beneficial effects: (1) The method greatly reduces the operation load by k-means clustering array sparsification on the basis of preserving the array core characteristics, improves the calculation efficiency by combining the fast deconvolution algorithm, solves the problem of high deployment cost of large array engineering, uses the function beam forming to upgrade the dynamic range, uses the frequency domain deconvolution model to combine the Gaussian filter to suppress noise, and guarantees the positioning accuracy; the wideband comprehensive mechanism of multi-frequency point signal-to-noise ratio extreme value fusion does not require prior band information, adaptively synthesizes clear imaging, avoids weak source submersion and positioning drift, and further suppresses cloud flicker through inter-frame weighted smoothing, synchronously optimizes the weak source identification ability, and finally realizes low-cost, high-precision and strong-robustness sound source positioning, which is suitable for complex actual scenes with unknown spectrum and low signal-to-noise ratio. (2) The present application introduces function beam forming and combines power index to improve imaging dynamic range, making it easier to distinguish different intensity sound sources; the frequency domain deconvolution model is constructed, the convolution is converted to wave number domain product through fast Fourier transform, combined with zero padding and iterative optimization, combined with Gaussian regularization filter to suppress high wave number noise, to accurately solve the sound source intensity distribution matrix and guarantee the positioning accuracy; the wideband comprehensive mechanism selects the frequency component with the most prominent energy and the best signal-to-noise ratio for each spatial grid point through multi-frequency point signal-to-noise ratio extreme value fusion, avoids the imaging one-sidedness and weak source submersion problem caused by unknown spectrum and bandwidth in traditional algorithms, and realizes focused and clear sound source imaging with accurate intensity. (3) The present application does not need to determine the sound source frequency and bandwidth in advance, performs the fast deconvolution algorithm independently through the full frequency band, combines noise base estimation and signal-to-noise ratio calculation, adaptively fuses effective information of the full frequency band, can stably process actual sound sources with unknown spectrum and low signal-to-noise ratio, solves the positioning drift problem of the traditional algorithm, introduces a time smoothing factor for inter-frame weighted fusion, effectively suppresses the flicker of the imaging cloud picture, improves the visual stability of the video output, and enhances the application adaptability in the dynamic sound source scene; (4) The present application estimates the noise base of the sound source intensity distribution matrix and sets a threshold, converts the sound source intensity matrix into a sound source distribution signal-to-noise ratio matrix, highlights the difference between weak sources and noise, enhances the weak source identification, and ensures that the weak source can be effectively identified in the scene where multiple sound sources coexist and there is a strong and weak difference. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The method flowchart of the present application; Figure 2 The acoustic imaging picture of the present application in a signal-to-noise ratio of -30dB environment; Figure 3 The acoustic imaging picture of the present application in a signal-to-noise ratio of -25dB environment; Figure 4 The acoustic imaging picture of the present application in a signal-to-noise ratio of -20dB environment; Figure 5 The acoustic imaging picture of the present application in a signal-to-noise ratio of -15dB environment; Figure 6 The acoustic imaging picture of the present application in a signal-to-noise ratio of -10dB environment. DETAILED DESCRIPTION

[0016] As shown in Figure 1 , the present application provides a technical solution: a fast deconvolution sound source positioning method based on array sparsification and wideband comprehensive processing, comprising the following steps: Step S1: divide the microphone array into several clusters through k-means clustering, and select the microphones closest to the cluster center in each cluster to form a sparse array (while reducing the subsequent calculation amount, the spatial coverage of the array is ensured); Step S2: convert the regular focusing plane into an irregular focusing plane, and calculate the irregular grid coordinates through equidistant sampling of the angle sine value; Step S3: calculate the cross-spectrum matrix of the sparse array based on the irregular grid coordinates and perform eigenvalue decomposition, and introduce a power index to obtain the function beamforming output (to improve the imaging dynamic range); Step S4: according to the convolution relationship between the function beamforming output and the sound source distribution and the power-up point spread function, a frequency domain deconvolution model is constructed; Step S5: Convert the convolution in the model to the wave number domain product by using fast Fourier transform, and solve the sound source intensity distribution matrix by zero padding and iterative optimization; Step S6: Estimate the noise basis of the sound source intensity distribution matrix and set a threshold to convert it into a sound source distribution signal-to-noise ratio matrix (to enhance the identification of weak sources); Step S7: Traverse all frequency points for the sound source distribution signal-to-noise ratio matrix, and take the maximum signal-to-noise ratio of each grid point to fuse into a wideband comprehensive imaging matrix; Step S8: Introduce a time smoothing factor to weight and fuse the imaging matrix of the current frame and the previous frame to obtain the final acoustic imaging result, and complete the sound source positioning.

[0017] The specific process of step S1 is as follows: Step S1.1: First, determine the position information of all microphones to form a position matrix containing the total number of microphones, where the coordinates of each microphone correspond to a row of data in the matrix; the microphone position matrix is as follows: In the formula, represents the horizontal coordinate of the m-th microphone in the plane coordinate system, , represents the total number of microphones; represents the vertical coordinate of the m-th microphone in the plane coordinate system; represents a real number matrix.

[0018] Step S1.2: Use the k-means clustering algorithm to divide the entire microphone array into clusters, and the clustering process aims to minimize the within-cluster sum of squares (the within-cluster sum of squares is the sum of the squares of the distances from all microphones in each cluster to the cluster centroid, and the centroid of the cluster is the average of all microphone coordinates in the cluster). By continuously iterating the adjustment of the cluster assignment of the microphones (assigning each microphone to the cluster with the nearest centroid) and updating the centroid of the cluster, the clustering result is stable.

[0019] Where the within-cluster sum of squares is minimized by the objective function, and the cluster assignment and centroid update are iteratively performed: The objective function is represented as: In the formula, represents the m-th cluster, ; represents the position vector of the microphone; represents the number of clusters to be divided; represents the m-th cluster centroid; ​​​​​​centroid vector of the cluster; The iterative execution of cluster assignment is represented as: ; In the formula, represents the th iteration of the th cluster; represents the position vector of the th microphone; represents the th iteration of the th cluster; represents the th iteration of the th cluster; The centroid update is represented as: .

[0020] Step S1.3: Calculate the distance of all microphones in each cluster to the centroid of the cluster , select the microphone closest to the centroid as the representative of the cluster, and finally form a sparse array composed of the representative microphones of all clusters , which can be represented as: ; ; ; ; In the formula, represents an index array, which records the index of the representative microphone of each cluster in the original array; represents the operation of selecting all columns.

[0021] The specific process of step S2 is as follows: Step S2.1: Determine the coordinates of the grid points in the regular focusing plane, and the vertical distance between the focusing grid plane and the plane of the microphone array.

[0022] Step S2.2: By calculating the distance of the grid points in the focusing plane to the reference point, and the trigonometric function relationship between the distance and the horizontal and vertical coordinates, the corresponding angle sine value is obtained.

[0023] Step S2.3: Within the set range of angle sine values (including both horizontal and vertical directions), discrete sampling is performed in an equal interval manner to obtain a series of angle sine values of sampling points.

[0024] Step S2.4: Based on the sine values ​​of the angles obtained from these samples, and combined with the vertical distance between the focusing plane and the microphone array plane, calculate the specific coordinates of each grid point in the irregular focusing plane, and complete the transformation from a regular focusing plane to an irregular focusing plane.

[0025] Wherein, the coordinates of the grid points in the focal plane are The vertical distance between the focusing plane and the microphone array plane is ,pass , ,exist and Perform equidistant divergence sampling within the range: ; In the formula, This represents the distance from a grid point in the focal plane to the reference point; This represents the vertical component, which is combined with practical calculations. Sure; The sine value representing the angle in the horizontal direction; The sine value representing the angle in the vertical direction; , These represent the minimum and maximum ranges of the sine value of the horizontal angle, respectively. , These represent the minimum and maximum ranges of the sine value of the angle in the vertical direction, respectively. Indicates the horizontal direction. The sine value of the angle at each sampling point; Indicates the vertical direction of the first The sine value of the angle at each sampling point; Indicates the number of sampling points in the horizontal direction; Indicates the number of sampling points in the vertical direction; The final result is the irregular grid coordinates: ; In the formula, Indicates the first Line number Column of irregular grid points; Indicates the first Line number The vertical coordinates of the irregular grid points.

[0026] The specific process of step S3 is as follows: Step S3.1: Perform Fourier transform on the sound signal recorded by each microphone in the sparse array to obtain the pressure amplitude data corresponding to different frequencies of each microphone, and then construct a column matrix containing the number of sparsed microphones based on the pressure amplitude data.

[0027] Step S3.2: The cross-spectral matrix of the sparse array signal is obtained by multiplying the column matrix with its conjugate transpose matrix.

[0028] Step S3.3: The cross-spectral matrix is subjected to eigenvalue decomposition to obtain a unitary matrix composed of the cross-spectral matrix eigenvectors and a diagonal matrix composed of the eigenvalues.

[0029] Step S3.4: A positive integer power index is introduced, and the output result of the functional beamforming is calculated in combination with the constructed steering vector (the steering vector is calculated by a specific exponential function based on the microphone array position, sound speed, frequency, etc.), which can reflect the sound source power at different spatial positions.

[0030] wherein the cross-spectral matrix is defined as wherein represents the column matrix composed of the pressure amplitude data of each microphone after sparsification at the frequency represents the pressure amplitude of the th microphone after sparsification at the frequency , represents the number of microphones after sparsification; represents the conjugate transpose matrix of .

[0031] wherein the cross-spectral matrix is subjected to eigenvalue decomposition: , represents the unitary matrix composed of the cross-spectral matrix eigenvectors, represents the conjugate transpose matrix of .

[0032] wherein a positive integer power index is introduced to calculate the output of the functional beamforming: wherein represents a set of positive integers; represents the output result of the functional beamforming at the position (reflecting the sound source power at the position), represents the focusing plane position of the beamforming; represents the output result of the functional beamforming at the position ​​​​​​​​​weight vector (for weighting the microphone signals to focus on a position ); denotes the conjugate transpose of ; denotes the steering vector of position , determined by microphone array position, sound speed, frequency, etc., describing the propagation characteristics of the sound source to the microphone array; denotes the conjugate transpose of ; denotes the steering vector of position , denotes the sound source position; denotes the conjugate transpose of ; denotes the intensity distribution of the sound source at position ; denotes the sound speed; denotes the position vector of the th microphone; denotes the wave number; denotes the imaginary unit.

[0033] where the power estimate at position is denoted as : ; in which, denotes the pressure amplitude column matrix of the microphone array (composed of the pressure amplitudes of each microphone); denotes the conjugate transpose of .

[0034] where the cross-spectral matrix at position is denoted as .

[0035] where the power estimate at position is denoted as : .

[0036] wherein the specific process of step S4 is: according to the functional beamforming output obtained in step S3, analyzing the relationship between the functional beamforming output and the sound source intensity distribution, the raised power point spread function, and finding that the functional beamforming output can be expressed as a spatial convolution of the sound source intensity distribution and the raised power point spread function (the raised power point spread function is the power operation result of the point spread function, and is only related to the position of the observation point relative to the sound source, that is, it has spatial shift invariance, taking the point spread function at the center of the focusing plane as the spatial shift invariable point spread function). Based on this convolution relationship, an inverse convolution model in the frequency domain is constructed (the inverse convolution model relates the functional beamforming output, the sound source intensity distribution, and the raised power point spread function).

[0037] wherein the functional beamforming output obtained in step S3 is analyzed to find the relationship between the functional beamforming output and the sound source intensity distribution, and the raised power point spread function, and is expressed as: ; ; wherein, represents the position vector of the center of the focusing plane; represents the raised power point spread function; the inverse convolution model in the frequency domain is constructed: ; wherein, represents the matrix representation of the sound source intensity distribution; represents the matrix representation of ; represents the Fourier transform operator; represents the inverse Fourier transform operator; represents the point-by-point multiplication (Hadamard product) operator in the frequency domain.

[0038] wherein the specific process of step S5 is: Step S5.1: zero padding is performed on the functional beamforming output, the sound source intensity distribution matrix, and the raised power point spread function, to ensure that the dimensions of the zero-padded matrices meet the calculation requirements (at least twice the number of rows and columns of the original sound source intensity distribution matrix); the zero-padded functional beamforming output is represented as , the zero-padded sound source intensity distribution matrix is represented as , and the zero-padded raised power point spread function is represented as .

[0039] Step S5.2: the fast Fourier transform result of the zero-padded raised power point spread function is calculated, as well as the sum of the absolute values of the result.

[0040] Step S5.3: Initialize the zero-filled sound source intensity distribution matrix as a zero matrix, and then perform iterative calculations using the iterative formula: First, based on the sound source intensity distribution matrix of the current iteration, the Fourier transform result of the power-law spread function, and the Gaussian regularized filter function (used to suppress high wavenumber noise, whose parameters are related to the cutoff wavenumber, which is determined by parameters such as the grid spacing distance of the focusing plane), calculate the function beamforming estimation result of the current iteration; then, combine the zero-filled function beamforming output to update the sound source intensity distribution matrix (ensuring that the updated matrix value is non-negative).

[0041] Step S5.4: After the iteration is completed, remove the zero-filling part to obtain the final estimated sound source intensity distribution matrix.

[0042] Among them, the calculation is: ; ; In the formula, This represents the wavenumber domain result of the zero-filled power-law point spread function after Fourier transform; This represents the sum of the absolute values ​​of the spread functions at the power-law point in the wavenumber domain; Let the zero-filling sound source intensity matrix be in the 0th iteration (initial state). Zero matrix Through the iterative formula: ; ; In the formula, Indicates the first Function beamforming estimation results at the next iteration; , They represent the first sequence The sound source intensity distribution matrix after zero-filling in the next iteration; This represents the Gaussian regularization filter function; , These represent the wavenumber components in the horizontal and vertical directions (two-dimensional components in the frequency domain), respectively. Solve for the sound source intensity distribution matrix, and After removing zero-padding, the final estimated sound source intensity distribution matrix is ​​obtained.

[0043] Among them, the Gaussian regularization filter function Represented as: ; In the formula, Indicates the cutoff wavenumber. This represents parameters related to filtering characteristics. represents the distance between the grid points of the focusing plane.

[0044] The specific process of step S6 is as follows: Step S6.1: Convert the final estimated sound source intensity distribution matrix obtained in step S5 into column vector form, and each column vector contains the sound source intensity data of all grid points at the corresponding frequency.

[0045] Step S6.2: Sort the sound source intensity column vector corresponding to each frequency in ascending order; according to the length of the sorted vector, select a certain proportion (the proportion value is between 0 and 1) of the smallest values, and calculate the average of these smallest values as the noise floor estimate at the frequency.

[0046] Step S6.3: According to the noise floor estimate, set a threshold to process the sound source intensity matrix at each frequency: if the sound source intensity of a grid point is less than or equal to the noise floor estimate, set the value of the point to 0; if it is greater than the noise floor estimate, divide the sound source intensity of the point by the noise floor estimate, and finally obtain the sound source distribution signal-to-noise ratio matrix at each frequency.

[0047] Wherein, the noise floor threshold is estimated based on the final estimated sound source intensity distribution matrix obtained in step S5. First, the final estimated sound source intensity distribution matrix is vectorized as: ; In the formula, represents the column vector after the final estimated sound source intensity distribution matrix is vectorized; represents transposition; is sorted in ascending order to obtain: ; ; In the formula, represents the column vector obtained after sorting in ascending order; represents the sound source intensity value of the th element after sorting; represents the total number of grid points; Take the average of the smallest values to obtain the noise floor estimate at the frequency: ; ; In the formula, represents the noise floor estimate at the frequency ;​ This indicates the proportion of selecting the minimum value; Indicates rounding down; Indicates the sorted order of the first... The sound source intensity value of each element; Then, based on the noise floor threshold, the sound source distribution matrix is ​​converted into a sound source distribution signal-to-noise ratio matrix: ; In the formula, Represents frequency Next Line number Elements of the signal-to-noise ratio matrix representing the sound source distribution at column grid points; Represents frequency Below, the first in the focused planar mesh Line number The sound source intensity value corresponding to the grid point.

[0048] The specific process of step S7 is as follows: For the source distribution signal-to-noise ratio (SNR) matrix obtained in step S6 at all frequencies, for each spatial grid point, iterate through the SNR values ​​corresponding to that grid point at all frequencies, and select the largest SNR value as the final imaging intensity value of that grid point; integrate the maximum SNR values ​​of all grid points to form a broadband integrated imaging matrix, represented as: ; In the formula, Indicates the first The first frame in the broadband synthetic imaging matrix Line number The final image intensity value of the column grid points; Indicates the first frequency points The signal-to-noise ratio matrix of the lower sound source distribution Line number Signal-to-noise ratio values ​​for grid points; This indicates the total number of frequency points participating in broadband convergence.

[0049] The specific process of step S8 is as follows: Introduce a time smoothing factor with a value between 0 and 1. The broadband composite imaging matrix obtained in step S7 is subjected to inter-frame weighted fusion processing. During fusion, the value of each grid point in the current frame's broadband composite imaging matrix is ​​multiplied by a time smoothing factor, and the value of the corresponding grid point in the previous frame's imaging matrix is ​​multiplied by 1 minus the time smoothing factor. The two results are added together to obtain the final acoustic imaging result of the current frame, thereby completing the sound source localization, as shown below: ; In the formula, Indicates the first The first frame in the final acoustic imaging result Line number The imaging intensity values ​​of the grid points; Indicates the first The first frame in the broadband synthetic imaging matrix Line number The final imaging intensity value of the grid points.

[0050] A non-volatile computer storage medium storing computer-executable instructions that execute a fast deconvolution sound source localization method based on array sparsity and broadband integrated processing.

[0051] Figures 2-6 Acoustic imaging images were obtained in a low signal-to-noise ratio environment of -30dB to -10dB using the method of the present invention. The test environment was as follows: a 19KHz sound source was placed at the center of two noise sources, with the sound source position at the origin and a vertical distance of 1m from the array plane.

[0052] from Figure 2 As can be seen, even in an environment with extremely low signal-to-noise ratio, the core location of the sound source can still be clearly presented, with no obvious imaging blurring or positional shift. Background noise interference is effectively suppressed, and the distinction between weak sources and noise is significant, proving the anti-interference capability and positioning effectiveness of this invention under extremely poor noise conditions.

[0053] from Figure 3 As can be seen, the sound source imaging focusing is further improved, the main lobe width is narrow, the side lobe suppression effect is excellent, there is no situation where the side lobe obscures the weak source or interferes with the localization, the sound source position accurately corresponds to the actual origin position, the imaging stability is strong, and the high resolution advantage of the method is demonstrated in medium and low signal-to-noise ratio scenarios.

[0054] from Figure 4 As can be seen, the imaging clarity and contrast are improved, the sound source outline is clearer, there is no imaging distortion (such as stretching or shifting), and the background noise floor is further reduced, verifying the actual effect of this method in suppressing spatial displacement variation of the point spread function through irregular focusing plane and noise suppression by Gaussian regularization filtering.

[0055] from Figure 5 As can be seen, the sound source localization accuracy remains stable with no localization drift, the imaging cloud map shows no flickering, and the details of weak sources (if there are multiple sources or weak source components) are clearly distinguishable, indicating the effective role of the broadband integration strategy (multi-frequency signal-to-noise ratio extreme value fusion) and the inter-frame smoothing mechanism in this signal-to-noise ratio range.

[0056] from Figure 6As can be seen, the imaging effect is optimal, the sound source positioning accuracy is high, the focusing is strong, the main side lobe ratio is significant, there is no noise redundant information interference, and the engineering level positioning requirement is completely met, further proving the robustness and practicality of the method in the full low signal-to-noise ratio range (-30dB to -10dB), and the stable imaging can be realized without prior sound source frequency and bandwidth information.

[0057] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A fast deconvolution sound source localization method based on array sparsity and broadband integrated processing, characterized in that, Includes the following steps: Step S1: Divide the microphone array into several clusters using k-means clustering, and select the microphones closest to the centroid in each cluster to form a sparse array; Step S2: Convert the regular focusing plane into an irregular focusing plane, and calculate the coordinates of the irregular grid by sampling the sine values ​​of the angles at equal intervals; Step S3: Calculate the cross-spectral matrix of the sparse array based on the irregular grid coordinates and perform eigenvalue decomposition. Introduce the power exponent to obtain the function beamforming output. Step S4: Based on the convolution relationship between the function beamforming output, the sound source distribution, and the power-law spread function, construct a frequency domain deconvolution model; Step S5: Use Fast Fourier Transform to convert the convolution in the model into wavenumber domain multiplication, and solve for the sound source intensity distribution matrix through zero-padding and iterative optimization; Step S6: Estimate the noise floor of the sound source intensity distribution matrix and set a threshold, then convert it into a sound source distribution signal-to-noise ratio matrix; Step S7: For the signal-to-noise ratio matrix of the sound source distribution, traverse all frequency points and take the maximum signal-to-noise ratio of each grid point to fuse them into a broadband integrated imaging matrix; Step S8: Introduce a time smoothing factor to weighted fuse the imaging matrices of the current frame and the previous frame to obtain the final acoustic imaging result and complete the sound source localization.

2. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 1, characterized in that: The specific process of step S1 is as follows: Step S1.1: First, determine the location information of all microphones to form a location matrix containing the total number of microphones, where the coordinates of each microphone correspond to a row of data in the matrix; Step S1.2: The k-means clustering algorithm is used to divide the entire microphone array into several clusters. The clustering process aims to minimize the sum of squares within each cluster. The cluster assignment of the microphones is adjusted and the centroid of the clusters is updated iteratively until the clustering results are stable. Step S1.3: Calculate the distance from all microphones in each cluster to the cluster centroid, select the microphone closest to the centroid as the representative of the cluster, and finally form a sparse array from the representative microphones of all clusters.

3. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 2, characterized in that: The specific process of step S2 is as follows: Step S2.1: Determine the coordinates of the grid points in the regular focusing plane, and the vertical distance between the focusing grid plane and the microphone array plane; Step S2.2: By calculating the distance from the grid point in the focusing plane to the reference point, and the trigonometric function relationship between this distance and the horizontal and vertical coordinates, the corresponding angle sine value is obtained; Step S2.3: Within the set range of angle sine values, perform discretization sampling at equal intervals to obtain a series of angle sine values ​​for sampling points; Step S2.4: Based on the sine values ​​of the angles obtained from these samples, and combined with the vertical distance between the focusing plane and the microphone array plane, calculate the specific coordinates of each grid point in the irregular focusing plane, and complete the transformation from a regular focusing plane to an irregular focusing plane.

4. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 3, characterized in that: The specific process of step S3 is as follows: Step S3.1: Perform Fourier transform on the sound signal recorded by each microphone in the sparse array to obtain the pressure amplitude data corresponding to different frequencies of each microphone, and then construct a column matrix containing the number of microphones after sparse based on the pressure amplitude data; Step S3.2: Multiply the column matrix by its conjugate transpose to obtain the cross-spectral matrix of the sparse array signal; Step S3.3: Perform eigenvalue decomposition on the cross-spectral matrix to obtain a unitary matrix composed of the eigenvectors of the cross-spectral matrix and a diagonal matrix composed of the eigenvalues; Step S3.4: Introduce a positive integer power exponent and combine it with the constructed guide vector to calculate the function beamforming output; the guide vector is calculated from the microphone array position, sound velocity, and frequency through an exponential function.

5. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 4, characterized in that: The specific process of step S4 is as follows: Based on the function beamforming output obtained in step S3, analyze its relationship with the sound source intensity distribution and the power-law spread function: the function beamforming output is represented as the spatial convolution of the sound source intensity distribution and the power-law spread function; Based on the fact that the output of function beamforming is represented as the spatial convolution of the sound source intensity distribution and the power-law spread function, a deconvolution model in the frequency domain is constructed.

6. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 5, characterized in that: The specific process of step S5 is as follows: Step S5.1: Perform zero-filling processing on the function beamforming output, the sound source intensity distribution matrix, and the power-law spread function; Step S5.2: Calculate the Fast Fourier Transform of the power-up spread function after zero-filling, and the sum of the absolute values ​​of the results; Step S5.3: Initialize the zero-filled sound source intensity distribution matrix as a zero matrix, and then perform iterative calculations using the iterative formula: First, calculate the function beamforming estimation result for the current iteration based on the sound source intensity distribution matrix of the current iteration, the Fourier transform result of the power-law spread function, and the Gaussian regularized filter function; then, update the sound source intensity distribution matrix by combining the zero-filled function beamforming output. Step S5.4: After the iteration is completed, remove the zero-filling part to obtain the final estimated sound source intensity distribution matrix.

7. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 6, characterized in that: The specific process of step S6 is as follows: Step S6.1: Convert the final estimated sound source intensity distribution matrix obtained in step S5 into column vector form, where each column vector contains the sound source intensity data of all grid points at the corresponding frequency; Step S6.2: Sort the sound source intensity column vectors corresponding to each frequency in ascending order; select the minimum value of a preset ratio according to the length of the sorted vectors, calculate the average value of these minimum values, and use it as the noise floor estimate at the corresponding frequency. Step S6.3: Set a threshold based on the noise floor estimate and process the sound source intensity matrix at each frequency: if the sound source intensity at a certain grid point is less than or equal to the noise floor estimate, then set the value of that point to 0; if it is greater than the noise floor estimate, then divide the sound source intensity at that point by the noise floor estimate, and finally obtain the sound source distribution signal-to-noise ratio matrix at each frequency.

8. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 7, characterized in that: The specific process of step S7 is as follows: For the source distribution signal-to-noise ratio matrix at all frequencies obtained in step S6, for each spatial grid point, traverse the signal-to-noise ratio values ​​corresponding to the grid point at all frequencies, and select the largest signal-to-noise ratio value as the final imaging intensity value of the grid point; integrate the maximum signal-to-noise ratio values ​​of all grid points to form a broadband integrated imaging matrix.

9. The fast deconvolution sound source localization method based on array sparsity and broadband integrated processing according to claim 8, characterized in that: The specific process of step S8 is as follows: a time smoothing factor is introduced, and the broadband integrated imaging matrix obtained in step S7 is subjected to inter-frame weighted fusion processing; during fusion, the value of each grid point in the current frame broadband integrated imaging matrix is ​​multiplied by the time smoothing factor, and the value of the corresponding grid point in the previous frame imaging matrix is ​​multiplied by 1 minus the time smoothing factor. The two results are added together to obtain the final acoustic imaging result of the current frame, thereby completing the sound source localization.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform the fast deconvolution sound source localization method based on array sparse and broadband integrated processing as described in any one of claims 1-9.

Images