Shallow sea equivalent earth sound model inversion method based on sparse Bayesian learning dispersion curve estimation
By combining sparse Bayesian learning and differential evolution algorithm, the intrinsic features of the shallow sea sound field are extracted, which solves the problems of computational complexity and parameter uncertainty in the inversion of shallow sea sound propagation model, and achieves more efficient and accurate inversion results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 715TH RES INST OF CHINA SHIPBUILDING IND CORP
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies for inverting shallow sea acoustic propagation models suffer from problems such as high computational cost, large uncertainty of inversion parameters, high dependence on prior knowledge of sound source information and seabed stratification, and severe coupling effects of attenuation coefficients, making it difficult to establish accurate acoustic propagation models.
A sparse Bayesian learning dispersion curve estimation method is adopted to extract the intrinsic features of the sound field from the vertical array data. Using the concept of equivalent ground acoustic model, parameter inversion is performed through sparse Bayesian learning algorithm and differential evolution algorithm, which reduces the dependence on sound source information and prior knowledge of seabed stratification, decouples key parameters, and improves computational efficiency.
It achieves more robust inversion results, maintains inversion accuracy under model mismatch, avoids the coupling effect of attenuation coefficient, and improves computational efficiency and inversion accuracy.
Smart Images
Figure CN122087435A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of signal processing technology, specifically relating to a method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation. Background Technology
[0002] The effectiveness of sonar detection systems is severely limited by the waveguide acoustic propagation characteristics in the marine environment. Taking shallow sea acoustic propagation as an example, it exhibits obvious multipath propagation characteristics due to significant seabed interactions. Establishing an accurate acoustic propagation model is an urgent requirement for environmentally adaptable sonar signal processing solutions, and ground acoustic parameters are precisely the most important basic information for shallow sea acoustic propagation modeling. Ground acoustic parameters are difficult to measure in situ over a large scale and are often obtained through inversion. The main difficulties in ground acoustic parameter inversion are: the high-dimensional inversion parameter space and its highly nonlinear relationship with the sound field lead to high computational costs and high uncertainty of inversion parameters in most inversion techniques; furthermore, without prior knowledge of the seabed layer structure, it is impossible to invert a true ground acoustic model.
[0003] Existing technologies use raw sound pressure data or features obtained through simple preprocessing as input for inversion. Sound pressure data is the solution to the wave equation under input parameters such as ground acoustic parameters, sound velocity profiles, and the number and location of sound sources—a direct problem. Ground acoustic inversion, on the other hand, is the input sound pressure data and output ground acoustic parameters—an inverse problem. The wave equation is nonlinear, and the relationship between the raw sound pressure data and ground acoustic parameters is highly nonlinear. The inverse problem cannot be solved directly and can only be estimated through iterative optimization algorithms or Bayesian statistical inference methods. Furthermore, the real marine environment is rife with interference, making it difficult to accurately determine the location and number of sound sources, introducing uncertainty into ground acoustic inversion. One solution is to jointly invert ground acoustic parameters, sound source locations, and even the number of sound sources, but this significantly increases the parameter space dimension and computational load. On the other hand, the inherent attenuation of the medium in ground acoustic parameters is difficult to invert directly due to its nonlinear frequency correlation, while indirect acquisition often introduces interference from other physical mechanisms. When inverting the broadband attenuation coefficient of the seabed, inconsistencies between the actual layered seabed and the assumed seabed layering can lead to "model mismatch" and "distortion" of the inherent attenuation coefficient. Meanwhile, due to the coupling between physical mechanisms, the inherent decay inversion error will interfere with the inversion of other parameters.
[0004] Therefore, there is an urgent need in this field for a new method for shallow sea acoustic parameter inversion that can reduce the dependence on sound source information and prior knowledge of seabed stratification, decouple the inversion of key parameters, and has higher computational efficiency. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation. This method directly extracts the sound field eigenvalues, i.e. dispersion curves, which are independent of the sound source from the vertical array data, and uses the concept of equivalent ground acoustic models to realize a more robust ground acoustic parameter inversion scheme that is more tolerant of model mismatch and can effectively avoid the coupling effect of attenuation coefficient.
[0006] The technical solution of this invention is to provide a method for inverting shallow-sea equivalent geosonic models based on sparse Bayesian learning dispersion curve estimation, comprising the following steps:
[0007] Step 1: Perform time-frequency transformation on the broadband sound source data received by the shallow sea vertical array to obtain frequency domain vertical array data with multiple frequency points and multiple snapshots.
[0008] Step 2: Based on the water layer sound velocity profile and prior environmental information, construct the horizontal wavenumber hypothesis interval and the mode depth function hypothesis space corresponding to each frequency point;
[0009] Step 3: Based on the normal mode propagation theory, the frequency domain vertical array data of each frequency point is modeled as a weighted sum of the mode depth functions in the mode depth function hypothesis space, and the weight vector is the horizontal wavenumber spectrum;
[0010] Step 4: Use the sparse Bayesian learning algorithm to perform sparse estimation of the horizontal wavenumber spectrum, and obtain the horizontal wavenumber estimate of the actual excitation mode at each frequency point through spectral peak detection and screening.
[0011] Step 5: Repeat steps 3 and 4 for all frequency points within the broadband range to obtain the dispersion curve estimate of the horizontal wavenumber of each mode as a function of frequency, and convert it to the modal angle domain for polynomial fitting correction.
[0012] Step 6: Construct an equivalent ground acoustic model and establish an inversion model with the error between the corrected dispersion curve estimate and the copy dispersion curve calculated based on the assumed ground acoustic parameters as the objective function.
[0013] Step 7: Optimize the objective function within the preset ground acoustic parameter search space using the differential evolution algorithm, and search for the ground acoustic parameter combination that minimizes the objective function, which is then used as the inversion result of the equivalent ground acoustic model.
[0014] The present invention proposes a method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation, which includes two parts: sparse Bayesian learning dispersion curve estimation and equivalent ground acoustic model inversion.
[0015] The first part utilizes sparse Bayesian learning to estimate the dispersion curves of each mode from broadband vertical array data, mainly including the following five steps: 1. Data preprocessing: The broadband sound source data received by the vertical array is transformed from the time domain to the frequency domain using Discrete Fourier Transform (DFT). Specifically, the discrete-time samples of each array element are processed in frames, and DFT is performed on each frame of data for each array element. The complex sound pressure value at each frequency point is extracted from the spectrum to obtain multi-shot, multi-frequency vertical array data in the frequency domain. 2. Setting the horizontal wavenumber hypothesis interval and the mode depth function hypothesis space: Based on existing prior information, the range of normal mode horizontal wavenumbers at each frequency point is initially estimated, and equally spaced discrete sampling is performed, which is called the horizontal wavenumber hypothesis interval. The mode depth function corresponding to each horizontal wavenumber value in the horizontal wavenumber hypothesis interval at each frequency point is calculated using the finite difference solution method of the wave equation and combined with the easily measurable water layer sound velocity profile. The mode depth function hypothesis space is obtained by sampling with reference to the depth of each array element of the vertical receiving array. 3. Vertical array received data modeling: Based on the normal mode sound propagation theory, the frequency domain vertical array received data at each frequency point can be viewed as a weighted sum of the mode depth functions in the hypothesis space corresponding to that frequency point. The weighted value represents the complex amplitude of each mode, also known as the horizontal wavenumber spectrum. 4. Sparse Bayesian learning estimation of the horizontal wavenumber spectrum: Since the actual number of excitation modes is finite and much smaller than the total number of modes in the hypothesis space, most elements in the weighted value are close to 0, satisfying the sparsity constraint. The horizontal wavenumber spectrum is estimated using a fixed-point iterative update sparse Bayesian learning algorithm. The peak points of the horizontal wavenumber spectrum are extracted and compared with the peak detection threshold to obtain the estimated horizontal wavenumber of the excitation mode. 5. Dispersion curve estimation and correction: For broadband data, steps 3 and 4 are repeated for each frequency point to obtain the horizontal wavenumber estimate, i.e., the dispersion curve estimate, for each frequency point. Then, based on the modal horizontal wavenumber... The modal angle approximation conversion formula transforms the dispersion curve from the horizontal wavenumber domain to the modal angle domain; in the modal angle domain, polynomial fitting techniques can be used to more effectively correct the dispersion curve estimate.
[0016] The second part uses the dispersion curve estimation as input for equivalent geosound parameter inversion, including the following two steps: 1. Establish the objective function for geosound inversion. The objective function is chosen to be a weighted sum of squared errors, that is, a weighted sum of the squares of the differences between the estimated dispersion curves of each order and the dispersion curves of each order calculated according to the assumed geosound parameters. The weights are set according to the sensitivity of different order mode dispersion curves to changes in geosound parameters and the magnitude of the estimation errors of different order mode dispersion curves. Modes with high sensitivity and small errors are assigned larger weights. 2. Determine the search range of each geosound parameter. Use the differential evolution algorithm to find the combination of geosound parameters that minimizes the objective function in the geosound parameter search space, which is taken as the geosound parameter inversion result. It should be noted that, unless otherwise specified in this invention, the horizontal wavenumber refers to the real part of the horizontal wavenumber.
[0017] As a preferred option, step 1 is performed as follows: Use Matlab software to read... The length of time for receiving data from each element of the vertical array. The data is divided into time frames, with each frame having a duration of [length missing]. With an inter-frame overlap rate of 50%, a total of Frame data, in which , This is a function for rounding down. Use... middle The function performs a discrete Fourier transform on each frame of data for each array element, and sequentially extracts the data from the spectrum within a wide frequency range. The complex sound pressure levels at each frequency point form a broadband frequency domain vertical array data, with dimensions of [missing information]. .
[0018] Preferably, in step 2, the assumed range of the horizontal wavenumber is determined as follows: the horizontal wavenumber is determined based on the frequency and sound velocity profiles. The approximate range, Then, the range is sampled at equal intervals to obtain... sampling points ,in, It's frequency. It is the lowest sound speed in the water layer. It is the speed of sound underwater;
[0019] The modulus depth function hypothesis space is determined as follows: assuming the sea surface satisfies absolutely soft boundary conditions, the modulus depth function corresponding to each sampling point within the horizontal wavenumber hypothesis interval can be solved using the finite difference method for solving the wave equation.
[0020]
[0021] in It is the sound velocity profile after linear interpolation. , It is the depth difference between two difference points. This is the total number of sampling points after interpolation; obtained by sampling the modulus depth function according to the depth of each element of the vertical matrix. That is, the hypothesis space of the modulus depth function, where This represents the depth vector of the array element.
[0022] Preferably, in step 3, the frequency domain vertical matrix data at each frequency point is modeled as a weighted sum of the modulus depth functions in the modulus depth function hypothesis space, which can be expressed as:
[0023]
[0024] in It is the horizontal wavenumber spectrum, representing the complex amplitude of each mode; A dictionary matrix representing the modulus-depth function; This is the noise term.
[0025] As a preferred option, step 4 is as follows:
[0026] Step 4.1, model the multi-frame vertical array received data at the same frequency as follows: Assuming horizontal wavenumber spectrum Let be a random vector that follows a multivariate complex Gaussian distribution, where each element is independent and uncorrelated. This represents the average power of each mode;
[0027] Step 4.2, assume the noise term The noise at each array element is independent and identically distributed Gaussian noise, and the noise is considered to be a non-stationary process in time. , For the first The variance of noise in the frame data;
[0028] Step 4.3, use a fixed-point iterative update sparse Bayesian learning algorithm to... and The estimation is performed, and the updated formula is as follows:
[0029]
[0030]
[0031] Current estimate Based on the estimation results of the previous iteration The adjusted formula is as follows: It refers to sparsity, based on the current iteration result. Select the one with the largest peak value Each spectral peak Its corresponding value in the hypothesis space of the modulus depth function The column mode depth function is composed of an effective mode selection set;
[0032] Step 4.4, based on the fact that the true sparsity of the horizontal wavenumber spectrum is equal to the number of propagable modes excited by the waveguide. A rough estimate is made by referring to the relationship between horizontal wavenumber and sea depth, seabed sound velocity, and frequency under an ideal waveguide environment (a constant-velocity seawater layer and a constant-velocity seabed half-space):
[0033]
[0034] In the formula Indicates sea depth;
[0035] use estimate ( express (Estimation) instead of horizontal wavenumber spectrum It is estimated that peak detection and threshold screening will be performed, retaining those peaks greater than the detection threshold. Each spectral peak ; corresponding to the horizontal wavenumber hypothesis interval The horizontal wavenumbers are used as estimates of the actual excited modes. .
[0036] Preferably, in step 5, within the broadband range... Repeat steps 3 and 4 for the data at each frequency point to obtain a broadband horizontal wavenumber estimate. Using modal horizontal wavenumber Approximate conversion formula for modal angles Transform the dispersion curve from the horizontal wavenumber domain to the modal angle domain. ,use Middle function Polynomial fitting was performed on the dispersion curves of each modality in the modal angular domain.
[0037] Preferably, in step 6, the equivalent geosonic model structure is a sedimentary layer. In a semi-infinite space, the geosonic parameters to be inverted include the depth of the seawater layer, the depth of the sedimentary layer, the sound velocity at the top of the sedimentary layer, the sound velocity gradient of the sedimentary layer, and the density of the sedimentary layer.
[0038] Furthermore, the objective function is chosen to be a weighted sum of squared errors. The squares of the differences between the estimated modal dispersion curves and the dispersion curves calculated based on the assumed ground acoustic parameters are weighted and summed to obtain the following:
[0039]
[0040] in The estimated first First-order mode dispersion curve; Indicates based on ground acoustic parameters use The toolbox calculates the first First-order mode dispersion curve; Indicates the first Weighting coefficients for estimating the first-order mode dispersion curve.
[0041] Compared with the prior art, the present invention has the following advantages:
[0042] By adopting the technical solution of this invention, the target can still be continuously tracked even when there is a crossover during target tracking, thus avoiding the phenomena of tracking deviating from the true trajectory and erroneous tracking in traditional methods. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the process of this invention.
[0044] Figure 2 is a flowchart and a partial MATLAB code diagram of the sparse Bayesian learning level wavenumber spectrum estimation algorithm according to an embodiment of the present invention.
[0045] Figure 3 is a schematic diagram of the real environment and its corresponding theoretical dispersion curve of a specific embodiment of the equivalent ground acoustic model inversion of a shallow sea environment according to an embodiment of the present invention.
[0046] Figure 4 is an example of how the estimation results of a single-frequency horizontal wavenumber spectrum are converted into broadband dispersion curve estimation results according to an embodiment of the present invention.
[0047] Figure 5 This refers to the equivalent ground acoustic model structure and ground acoustic parameter search range in the embodiments of the present invention.
[0048] Figure 6 is a flowchart of the differential evolution algorithm and a partial MATLAB code diagram of an embodiment of the present invention.
[0049] Figure 7 shows the equivalent ground acoustic parameter inversion results and their effectiveness in an embodiment of the present invention. Detailed Implementation
[0050] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:
[0051] The present invention proposes a method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation, which includes two parts: sparse Bayesian learning dispersion curve estimation and equivalent ground acoustic model inversion.
[0052] The main process of the method of the present invention is as follows: Figure 1 As shown, it includes the following steps:
[0053] Step 1: Perform time-frequency transformation on the broadband sound source data received by the shallow sea vertical array to obtain frequency domain vertical array data with multiple frequency points and multiple snapshots.
[0054] Step 2: Based on the water layer sound velocity profile and prior environmental information, construct the horizontal wavenumber hypothesis interval and the mode depth function hypothesis space corresponding to each frequency point;
[0055] Step 3: Based on the normal mode propagation theory, the frequency domain vertical array data of each frequency point is modeled as a weighted sum of the mode depth functions in the mode depth function hypothesis space, and the weight vector is the horizontal wavenumber spectrum;
[0056] Step 4: Use the sparse Bayesian learning algorithm to perform sparse estimation of the horizontal wavenumber spectrum, and obtain the horizontal wavenumber estimate of the actual excitation mode at each frequency point through spectral peak detection and screening.
[0057] Step 5: Repeat steps 3-4 for all frequency points within the broadband range to obtain the dispersion curve estimate of the horizontal wavenumber of each mode as a function of frequency, and convert it to the modal angle domain for polynomial fitting correction.
[0058] Step 6: Construct an equivalent ground acoustic model and establish an inversion model with the error between the corrected dispersion curve estimate and the copy dispersion curve calculated based on the assumed ground acoustic parameters as the objective function.
[0059] Step 7: Optimize the objective function within the preset ground acoustic parameter search space using the differential evolution algorithm, and search for the ground acoustic parameter combination that minimizes the objective function, which is then used as the inversion result of the equivalent ground acoustic model.
[0060] Specifically:
[0061] In step 1, use Software reading The length of time for receiving data from each element of the vertical array. The data is divided into time frames, with each frame having a duration of [length missing]. With an inter-frame overlap rate of 50%, a total of Frame data, in which , This is a function for rounding down. Use... middle The function performs a discrete Fourier transform on each frame of data for each array element, and sequentially extracts the data from the spectrum within a wide frequency range. The complex sound pressure levels at each frequency point form a broadband frequency domain vertical array data, with dimensions of [missing information]. .
[0062] In step 2, the horizontal wavenumber is determined based on the frequency and sound velocity profile. The approximate range, ,in, It's frequency. It is the lowest sound speed in the water layer. It's the speed of sound underwater, because If the value is unknown, a relatively large value (1800 m / s) can be set based on experience to ensure that the actual horizontal wavenumber is within the range; then, the range is sampled at equal intervals to obtain... sampling points This is called the horizontal wavenumber hypothesis interval.
[0063] Assuming the sea surface satisfies absolutely soft boundary conditions, the modulus-depth function corresponding to each sampling point within the horizontal wavenumber assumption interval can be solved using the finite difference method for solving the wave equation:
[0064]
[0065] in It is the sound velocity profile after linear interpolation. , It is the depth difference between two difference points. This is the total number of sampling points after interpolation; obtained by sampling the modulus depth function according to the depth of each element of the vertical matrix. That is, the hypothesis space of the modulus depth function, where This represents the depth vector of the array element.
[0066] In step 3, the vertical array received data is modeled. According to the normal mode propagation theory, the frequency domain vertical array received data at each frequency point can be regarded as a weighted sum of the mode depth functions in the hypothesis space corresponding to that frequency point:
[0067]
[0068] in It is the horizontal wavenumber spectrum, representing the complex amplitude of each mode; A dictionary matrix representing the modulus-depth function; This is the noise term.
[0069] Step 4 is detailed below, as shown in Figure 2. Figure 2(a) shows the flowchart of the sparse Bayesian learning level wavenumber spectrum estimation algorithm based on the fixed-point iterative update strategy; Figure 2(b) shows part of the algorithm code written in MATLAB.
[0070] Step 4.1: Estimate the level wavenumber spectrum using the sparse Bayesian learning algorithm. Since the actual number of excited modes is finite and much smaller than the total number of modes in the hypothesis space, the weighted vector is the level wavenumber spectrum. The vast majority of elements in the array have values close to 0, satisfying the sparsity constraint. Multiple frames of vertical array received data at the same frequency can be modeled as follows: Assuming horizontal wavenumber spectrum Let be a random vector that follows a multivariate complex Gaussian distribution, where each element is independent and uncorrelated. This represents the average power of each mode;
[0071] Step 4.2, assume the noise term The noise at each array element is independent and identically distributed Gaussian noise, and the noise is considered to be a non-stationary process in time. , For the first The variance of noise in the frame data;
[0072] Step 4.3, in this embodiment, a sparse Bayesian learning algorithm with fixed-point iterative update is used to... and The estimation is performed, and the updated formula is as follows:
[0073]
[0074]
[0075] Current estimate Based on the estimation results of the previous iteration The adjusted formula is as follows: It refers to sparsity, based on the current iteration result. Select the one with the largest peak value Each spectral peak Its corresponding value in the hypothesis space of the modulus depth function The column mode depth function is composed of an effective mode selection set;
[0076] Step 4.4, based on the fact that the true sparsity of the horizontal wavenumber spectrum is equal to the number of propagable modes excited by the waveguide. A rough estimate is made by referring to the relationship between horizontal wavenumber and sea depth, seabed sound velocity, and frequency under an ideal waveguide environment (a constant-velocity seawater layer and a constant-velocity seabed half-space):
[0077]
[0078] In the formula This represents the ocean depth; when there is a large error between the estimated and actual ocean depth, a value significantly greater than the actual seabed sound velocity is selected as a substitute. (In this embodiment, 1800 m / s is used in subsequent data processing), thus obtaining a conservative sparsity estimate. Furthermore, the iterative update algorithm used has some tolerance for smaller sparsity selections. In subsequent simulations and experiments, the sparsity... Set as .
[0079] This invention uses ( express (Estimation) instead of horizontal wavenumber spectrum Estimate, perform spectral peak detection and threshold screening, i.e., use MATLAB. Function extraction Each spectral peak is compared with a peak detection threshold (set to 0.01 times the maximum peak value), and peaks greater than the detection threshold are retained. Each spectral peak ; corresponding to the horizontal wavenumber hypothesis interval The horizontal wavenumbers are used as estimates of the actual excited modes. .
[0080] Step 5, Dispersion curve estimation and correction: For broadband data, within the broadband range... Repeat steps 3 and 4 for the data at each frequency point to obtain a broadband horizontal wavenumber estimate. This refers to dispersion curve estimation, utilizing modal level wavenumbers. Approximate conversion formula for modal angles Transform the dispersion curve from the horizontal wavenumber domain to the modal angle domain. Using functions in Matlab Polynomial fitting was performed on the dispersion curves of each modality in the modal angular domain. The rule for the polynomial fitting degree of freedom parameter in the function is as follows: starting from 1, the degree of freedom is increased successively to fit the dispersion curve until the termination condition is met: reaching the residual threshold (0.05) or the highest degree of freedom (14). The purpose of polynomial fitting is to correct the dispersion curve estimate. The dispersion curve estimate after polynomial fitting is denoted as: .
[0081] Step 6: Set the equivalent ground acoustic model and the inversion objective function. The equivalent ground acoustic model structure is a sedimentary layer. In the semi-infinite space, the geoacoustic parameters to be inverted include the depth of the seawater layer, the depth of the sedimentary layer, the sound velocity at the top of the sedimentary layer, the sound velocity gradient of the sedimentary layer, and the density of the sedimentary layer. The reason for excluding the geoacoustic parameters of the semi-infinite space from the inversion parameters is that when the sound velocity in the semi-infinite space is greater than the phase velocity of the highest-order mode, the dispersion curve is not sensitive to the geoacoustic parameters of the semi-infinite space. Therefore, a relatively large sound velocity in the semi-infinite space can be set empirically; in this invention, it is fixed at 2200 m / s. This reduces the spatial dimensionality of the inversion parameters.
[0082] The objective function for ground acoustic inversion is set, and the objective function form is chosen to be a weighted sum of squared errors. The result is obtained by weighted summation of the squares of the differences between the estimated dispersion curves of each order and the dispersion curves of each order calculated according to the assumed ground acoustic parameters.
[0083]
[0084] in The estimated first First-order mode dispersion curve; Indicates based on ground acoustic parameters The first calculation using the Kraken toolkit First-order mode dispersion curve; Indicates the first Weighting coefficients for estimating the dispersion curves of different modes. These weighting coefficients are set based on the sensitivity of the dispersion curves of different modes to changes in geoacoustic parameters and the magnitude of the estimation errors for each mode. Higher-order modes can be approximated as large grazing angle acoustic rays, which are more sensitive to geoacoustic parameters. Therefore, it is assumed that the sensitivity increases linearly with the mode order. The weighting of the first mode due to its sensitivity is . No. The weighting of the first mode due to dispersion curve error is , It is the first The squared error of the first-order modal dispersion curve estimation is replaced by the square of the difference before and after polynomial fitting of the dispersion curve, since the true value of the dispersion curve is unknown in the actual environment. The calculation formula is as follows: The final weighted value is obtained by multiplying the two weighted values: .
[0085] Referring to Figure 3, Figure 3 shows a real environment schematic diagram of a specific embodiment of the equivalent ground acoustic model inversion in a shallow sea environment and its corresponding theoretical dispersion curve. Among them, Figure 3(a) shows a real environment schematic diagram of a specific embodiment of the equivalent ground acoustic model inversion in a shallow sea environment; Figure 3(b) shows the theoretical dispersion curve calculated using the Matlab version of the normal mode toolbox (Kraken toolbox) based on the environmental parameters shown in Figure 3(a).
[0086] Step 7: Perform equivalent ground acoustic model inversion using the differential evolution algorithm. First, determine the search range for the ground acoustic parameters to be inverted. Then, use the differential evolution algorithm to find the combination of ground acoustic parameters that minimizes the objective function in the ground acoustic parameter search space, which is taken as the ground acoustic parameter inversion result. The initial parameters of the differential evolution algorithm are set as follows: maximum number of iterations 1000; main scaling factor... Secondary scaling factor Secondary scaling factor Crossover probability 0.8; Optimized variable dimension 5; Population size 25; Variable boundaries consistent with the search range of ground acoustic parameters.
[0087] Figure 4 illustrates an example of the process of estimating the horizontal wavenumber spectrum from a single frequency point to the broadband dispersion curve estimation result in this invention. In Figure 4(a), the blue solid line represents the horizontal wavenumber spectrum estimation result of the sparse Bayesian learning of the vertical receiver array data at a frequency of 50 Hz; the findpeaks function in MATLAB is used to find all spectral peaks on the horizontal wavenumber spectrum, as shown in red in the figure. As indicated by the markings; the spectral peaks after being filtered by the peak detection threshold (0.01 times the maximum peak value) are shown in the red circles in the figure. A total of 4 spectral peaks were retained, indicating that the 4th mode was actually excited at 50 Hz. The peak points are used as the level wavenumber estimates of the 4th mode at 50 Hz. Figure 4(b) shows the broadband ( The result of estimating the horizontal wavenumber of the data received by the vertical array as a function of frequency, i.e., the dispersion curve estimation. Figure 4(c) shows the result based on the horizontal wavenumber. The modal angle approximation conversion formula transforms the dispersion curve estimation results in Figure 4(b) from the horizontal wavenumber domain to the modal angle domain, allowing the 10th-order mode and its corresponding dispersion curve to be distinguished. Figure 4(d) is based on Figure 4(c), using the polyfit function in MATLAB to perform polynomial fitting correction on the dispersion curves of each order mode.
[0088] Figure 5 The equivalent ground acoustic model structure used in this invention is given, which is a sedimentary layer. Semi-infinite space. The geosonic parameters to be inverted include: water layer depth. Deposition layer thickness Sound velocity at the top of the sedimentary layer Sound velocity gradient of sedimentary layer and sediment density The inversion range for each parameter to be inverted is given. Among the sedimentary geoacoustic parameters, the medium attenuation coefficient and the semi-infinite space geoacoustic parameters are given and not involved in the inversion; the given values are mismatched with the true values.
[0089] Figure 6 shows the flowchart of the differential evolution algorithm and part of the MATLAB code. Figure 6(a) shows the flowchart of the differential evolution algorithm; Figure 6(b) shows part of the differential evolution algorithm code written in MATLAB.
[0090] Referring to Figure 7, Figure 7(a) shows the equivalent ground acoustic model parameters retrieved using the differential evolution algorithm shown in Figure 6; Figure 7(b) shows the dispersion curve calculated from the retrieved equivalent ground acoustic model (as shown by the red circle in the figure). The calculation was performed using the Kraken toolbox in Matlab and compared with the theoretical dispersion curve calculated according to Figure 3(a) (as shown by the black solid line in the figure) and the SBL dispersion curve estimation results after polynomial fitting shown in Figure 4(d). It was found that the calculated values based on the inversion results are very close to the theoretical dispersion curve. From the enlarged view in the upper left corner, it can be seen that the difference between the calculated values of the inversion results and the theoretical values is smaller than that of the SBL dispersion curve estimation after polynomial fitting. In summary, this shows that the retrieved equivalent ground acoustic model is effective.
[0091] The above description only illustrates preferred embodiments of the present invention and should not be construed as limiting the scope of the claims. Any equivalent procedural modifications made using this specification are included within the patent protection scope of this invention.
Claims
1. A method for inverting shallow-sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation, characterized in that: Includes the following steps, Step 1: Perform time-frequency transformation on the broadband sound source data received by the shallow sea vertical array to obtain frequency domain vertical array data with multiple frequency points and multiple snapshots. Step 2: Based on the water layer sound velocity profile and prior environmental information, construct the horizontal wavenumber hypothesis interval and the mode depth function hypothesis space corresponding to each frequency point; Step 3: Based on the normal mode propagation theory, the frequency domain vertical array data of each frequency point is modeled as a weighted sum of the mode depth functions in the mode depth function hypothesis space, and the weight vector is the horizontal wavenumber spectrum; Step 4: Use the sparse Bayesian learning algorithm to perform sparse estimation of the horizontal wavenumber spectrum, and obtain the horizontal wavenumber estimate of the actual excitation mode at each frequency point through spectral peak detection and screening. Step 5: Repeat steps 3-4 for all frequency points within the broadband range to obtain the dispersion curve estimate of the horizontal wavenumber of each mode as a function of frequency, and convert it to the modal angle domain for polynomial fitting correction. Step 6: Construct an equivalent ground acoustic model and establish an inversion model with the error between the corrected dispersion curve estimate and the copy dispersion curve calculated based on the assumed ground acoustic parameters as the objective function. Step 7: Optimize the objective function within the preset ground acoustic parameter search space using the differential evolution algorithm, and search for the ground acoustic parameter combination that minimizes the objective function, which is then used as the inversion result of the equivalent ground acoustic model.
2. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 1, characterized in that: Step 1 is performed as follows: Use Matlab software to read... The length of time for receiving data from each element of the vertical array. The data is divided into time frames, with each frame having a duration of [length missing]. With an inter-frame overlap rate of 50%, a total of Frame data, in which , To perform a floor function, the `fft` function in MATLAB is used to perform a Discrete Fourier Transform on each frame of data for each array element, and the data within the wide frequency range is extracted sequentially from the spectrum. The complex sound pressure levels at each frequency point form a broadband frequency domain vertical array data, with dimensions of [missing information]. .
3. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 1, characterized in that: In step 2, the assumed range of horizontal wavenumbers is determined as follows: the horizontal wavenumber is determined based on the frequency and sound velocity profiles. The approximate range, Then, the range is sampled at equal intervals to obtain M sampling points. ,in, It's frequency. It is the lowest sound speed in the water layer. It is the speed of sound underwater; The modulus depth function hypothesis space is determined as follows: assuming the sea surface satisfies absolutely soft boundary conditions, the modulus depth function corresponding to each sampling point within the horizontal wavenumber hypothesis interval can be solved using the finite difference method for solving the wave equation. in It is the sound velocity profile after linear interpolation. , It is the depth difference between two difference points. This is the total number of sampling points after interpolation; obtained by sampling the modulus depth function according to the depth of each element of the vertical matrix. That is, the hypothesis space of the modulus depth function, where This represents the depth vector of the array element.
4. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 3, characterized in that: In step 3, the frequency domain vertical matrix data at each frequency point is modeled as a weighted sum of the mode depth functions in the mode depth function hypothesis space, which can be expressed as: in It is the horizontal wavenumber spectrum, representing the complex amplitude of each mode; A dictionary matrix representing the modulus-depth function; This is the noise term.
5. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 4, characterized in that: Step 4 is detailed below: Step 4.1, model the multi-frame vertical array received data at the same frequency as follows: Assuming horizontal wavenumber spectrum Let be a random vector that follows a multivariate complex Gaussian distribution, where each element is independent and uncorrelated. This represents the average power of each mode; Step 4.2, assume the noise term The noise at each array element is independent and identically distributed Gaussian noise, and the noise is considered to be a non-stationary process in time. , For the first The variance of noise in the frame data; Step 4.3, use a fixed-point iterative update sparse Bayesian learning algorithm to... and The estimation is performed, and the updated formula is as follows: Current estimate Based on the estimation results of the previous iteration The adjusted formula is as follows: It refers to sparsity, based on the current iteration result. Select the one with the largest peak value Each spectral peak Its corresponding value in the hypothesis space of the modulus depth function The column mode depth function is composed of an effective mode selection set; Step 4.4, based on the fact that the true sparsity of the horizontal wavenumber spectrum is equal to the number of propagable modes excited by the waveguide. A rough estimate is made by referring to the relationship between horizontal wavenumber and sea depth, seabed sound velocity, and frequency under ideal waveguide conditions: In the formula Indicates sea depth; use Replace horizontal wavenumber spectrum It is estimated that peak detection and threshold screening will be performed, retaining those peaks greater than the detection threshold. Each spectral peak , express Estimate; corresponding to the horizontal wavenumber assumption interval The horizontal wavenumbers are used as estimates of the actual excited modes. .
6. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 5, characterized in that: In step 5, within the broadband range Repeat steps 3 and 4 for the data at each frequency point to obtain a broadband horizontal wavenumber estimate. Using the approximate conversion formula of modal horizontal wavenumber → modal angle Transform the dispersion curve from the horizontal wavenumber domain to the modal angle domain. The polyfit function in Matlab was used to perform polynomial fitting on the dispersion curves of each modality in the modal angular domain.
7. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 1, characterized in that: In step 6, the equivalent geosonic model structure is a sedimentary layer plus a semi-infinite space. The geosonic parameters to be inverted include the depth of the seawater layer, the depth of the sedimentary layer, the sound velocity at the top of the sedimentary layer, the sound velocity gradient of the sedimentary layer, and the density of the sedimentary layer.
8. The method for inverting shallow sea equivalent ground acoustic models based on sparse Bayesian learning dispersion curve estimation according to claim 7, characterized in that: The objective function is chosen to be a weighted sum of squared errors. The results are obtained by weighted summation of the squares of the differences between the estimated modal dispersion curves and the dispersion curves calculated based on the assumed ground acoustic parameters. in The estimated first First-order mode dispersion curve; Indicates based on ground acoustic parameters The first calculation using the Kraken toolkit First-order mode dispersion curve; Indicates the first Weighting coefficients for estimating the first-order mode dispersion curve.