Shallow sea sound velocity profile matching field inversion method based on gate feedback gate recurrent unit

Abstract

The application discloses a shallow sea sound velocity profile matching field inversion method based on GFGRU. The method comprises the following steps: generating a group of dictionary base functions used to represent the sound velocity profile of a position to be inverted based on a K singular value decomposition dictionary learning algorithm, so that the accuracy is higher and the parameter quantity during searching is effectively reduced; a multi-label classification task of finding the base function representing the target sound velocity profile is completed by training a GFGRU network, a multi-label loss function is used in the training to make the network converge, in view of the problem of insufficient network training data, an artificial synthetic data set is used to pre-train the GFGRU network, and the accuracy and robustness of the GFGRU network classification can be effectively improved; a genetic algorithm is used to search the coefficients of the base function representing the target sound velocity profile output by the GFGRU classification, and a loss function is calculated by using a matching field method; iterative searching is performed until the loss function is lower than a set threshold value or the number of iteration steps reaches a set maximum value, and finally the sound velocity profile of the region to be inverted is calculated and obtained by using the optimal individual output by the searching.

Description

Technical Field

[0001] This invention relates to a method for inverting a matched field of sound velocity profile in shallow seas, specifically a method based on a gated feedback gated cyclic unit (GFGRU). Background Technology

[0002] Marine acoustic tomography is an important direction in the field of marine environmental monitoring and research. It infers the ocean conditions through which a sound field passes by precisely measuring sound propagation time or other sound propagation properties such as sound pressure amplitude and phase. Sound velocity, as one of the most important environmental parameters in underwater acoustics research, is fundamental to understanding and studying underwater acoustic channels and propagation, and is an indispensable environmental parameter in all fields of underwater acoustics research. In shallow waters, the sound velocity profile varies significantly with time and space. This results in insufficient coverage and resolution of the sound velocity profile samples obtained by directly measuring with a sound velocity meter deployed from a ship or by calculating with empirical formulas using a fixed-point temperature-depth (TDT) meter. In contrast, the inversion method requires only a single measurement to obtain the sound velocity profile at the transmit-receive cross-section. A sound velocity field for a region can be obtained through a limited number of measurements, reducing costs and saving resources.

[0003] Currently, the mainstream method for inverting shallow-sea sound velocity profiles is the matched-field method. Traditional matched-field inversion methods suffer from a certain loss of accuracy due to the use of a finite number of empirical orthogonal functions to represent the sound velocity profile. Furthermore, the inversion accuracy is not high due to limitations in the nonlinearity of the model and the tendency of the search method to get trapped in local minima. While dictionary learning and deep learning methods have been proposed in recent years to represent sound velocity profiles and have shown higher accuracy, the basis functions used to represent the sound velocity profile are too sparse in the dictionary, making it difficult to determine the basis functions used to represent unknown sound velocity profiles to be inverted, thus preventing them from reaching practical application.

[0004] In summary, existing matched-field inversion methods use empirical orthogonal functions to characterize the sound velocity profile, which easily leads to low inversion accuracy. Existing sound velocity profile characterization methods with higher accuracy, such as dictionary learning methods, have not been widely applied in matched-field inversion methods because it is difficult to determine the basis functions used to characterize the unknown sound velocity profile to be inverted. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for inverting the matched field of shallow-sea acoustic velocity profiles based on a Gated Feedback Gate Recurrent Unit (GFGRU). This method first utilizes a K-SVD (K-singular value decomposition) dictionary learning algorithm to generate basis functions, resulting in higher accuracy in representing the acoustic velocity profile while effectively reducing the number of parameters during the search. To address the difficulty of using dictionary learning methods to determine the basis functions for representing the unknown acoustic velocity profile to be inverted, this method employs a GFGRU network, suitable for sparse classification problems, to estimate the basis functions. Furthermore, to address the issue of insufficient training data for the GFGRU network, this method uses artificially synthesized datasets to pre-train the GFGRU network, effectively improving its classification accuracy and robustness while reducing the impact of insufficient experimental data on GFGRU network training. Finally, after obtaining the basis functions using the GFGRU network, a genetic algorithm is employed to complete the inverse problem of inferring the acoustic velocity profile from the sound field.

[0006] The specific technical solution of the present invention is as follows:

[0007] A method for inverting the matched field of shallow sea acoustic velocity profiles based on GFGRU, the specific steps of which are as follows:

[0008] S1: Based on the existing actual sound velocity profile data of the region to be inverted, use the K-SVD algorithm to generate a dictionary of sound velocity profiles of the region to be inverted and calculate the basis functions that constitute each sound velocity profile.

[0009] S2: Generate a synthetic sound velocity profile dataset based on the basis function distribution of the sound velocity profile in the region to be inverted. Based on the synthetic sound velocity profile dataset and the actual sound velocity profile data of the region to be inverted in step S1, calculate the sound field using the Kraken model and preprocess the sound field to obtain the synthetic dataset and the real dataset. Label each sound velocity profile in the synthetic dataset and the real dataset.

[0010] S3: The GFGRU network is pre-trained using the synthetic dataset to obtain a pre-trained model; then the pre-trained model is trained using the real dataset to complete the multi-label classification task; a multi-label loss function is used during training to help convergence;

[0011] S4: Input the measured sound field data of the region to be inverted into the trained GFGRU network to obtain the basis functions of the sound velocity profile of the region to be inverted;

[0012] S5: Use a genetic algorithm to search for the basis function coefficients of the sound velocity profile of the region to be inverted, use the matching field method to calculate the loss function, iterate until the loss function value is lower than the set threshold or the number of iterations reaches the set maximum number of steps, and use the optimal individual to calculate and obtain the sound velocity profile of the region to be inverted.

[0013] In the above technical solution, S1 is further defined as follows:

[0014] S11: Set the error ε, maximum number of iterations I, and sparsity S. p Based on the actual sound velocity profile data of the region to be inverted, a dictionary is generated using the K-SVD algorithm; the average sound velocity profile of the actual sound velocity profile data of the region to be inverted is calculated. For a given set of M sound velocity profiles discretized into K dimensions in the vertical direction, its mean sound velocity profile... The disturbance on is denoted as Let Y represent the actual sound velocity profile dataset, and the dictionary corresponding to Y is denoted as... N is the number of basis functions in the dictionary. Using the OMP algorithm, with a sparsity of S... p And the coefficient matrix reconstructed from dictionary Q by Y is denoted as The initial dictionary is obtained by randomly initializing the dictionary. Set the initial value of the loop in the K-SVD algorithm to i = 0;

[0015] S12: i = i + 1, sparse coding: for a known Q i-1 The orthogonal matching pursuit (OMP) algorithm is used to solve the constrained optimization problem, which is:

[0016]

[0017] In the formula, S p Let x represent the vector to be solved. m sparsity, and Let L2 and L0 norms be represented respectively;

[0018] The OMP algorithm is as follows:

[0019] (1) Initialize the solution vector Then residual Support set of the solution The initial value of the loop in the OMP algorithm is k = 1.

[0020] (2) Calculate the error for each basis function in the dictionary. in

[0021] (3) Find the smallest And update S k =S k-1 ∪{j k};

[0022] (4) Based on the support set S k Calculation makes Minimum solution vector

[0023] (5) Update residuals

[0024] (6) If ||r k ||2<ε or k>S p If k = k + 1, stop the algorithm; otherwise, return to (2) for the next iteration.

[0025] S13: Dictionary update.

[0026] Q i-1 basis functions in and the row vectors in X x j Update each element ∈X,j=1,…,N sequentially, with the following update steps:

[0027] (1) When updating dictionary Q i-1 The j-th basis function At that time, calculate Reconstruction error matrix of external dictionary extract x j The non-zero terms in the vector are arranged in order to form a new column vector x. R , x j The non-zero terms in E correspond to the column vectors in E, forming a new matrix E. R ;

[0028] (2) For E R Perform SVD decomposition: E R =USV T U and V represent the values ​​of E respectively. R The left and right singular matrices after SVD decomposition, where S represents the diagonal matrix composed of singular values ​​of each order;

[0029] (3) Update x R =V(:,1)S(1,1), which means that... Update x to the first left singular vector in U. R Updated to the product of the first right singular vector in V and the first-order singular value in S; and using x R The coefficients of the non-zero terms in the pairx j Update the corresponding element in the middle;

[0030] (4) Repeat (1)-(3) until all column vectors in the dictionary have been updated;

[0031] S14: Repeat S12 and S13 until i≥I, then output the dictionary Q. I And coefficient matrix X, x m Q corresponding to the zero term in the middle and non-zero terms I The basis functions in the equation are the basis functions of the m-th sound velocity profile, x m Q corresponding to zero term I The basis functions in the equation are not the basis functions that constitute the m-th sound velocity profile.

[0032] Furthermore, S2 specifically refers to:

[0033] S21: Vectorize the coefficient matrix X, find the non-zero terms, and construct a new vector. According to the law of large numbers Approximately Gaussian distribution, solve The sampling mean μ and sampling variance σ 2 Set the number of sound velocity profile samples N to be synthesized. s N s >> N; based on the sample mean μ and sample variance σ 2 and sparsity S p A new synthesis sequence matrix is ​​generated by synthesizing coefficient vectors one by one. for Any coefficient vector Its synthesis method is as follows: initialization After that S is randomly selected from N coefficients. p A set of zero values ​​is given, and then a random non-zero value is set to be obtained, which satisfies the following condition: Furthermore, the distribution of non-zero values ​​follows N(μ,σ) 2 Gaussian distribution; by Generate Ns synthetic sound velocity profiles

[0034] S22: For synthetic sound velocity profile data The sound field was calculated using the Kraken model, along with the actual sound velocity profile data Y. and The obtained sound field is then preprocessed using the following method:

[0035] For any vector in the sound field

[0036]

[0037] in, The superscript (·) represents the frequency domain sound pressure data in the real dataset, and (~) represents the frequency domain sound pressure data in the synthetic dataset; L is the number of elements in the vertical array; the covariance matrix K is calculated by normalizing the frequency domain sound pressure data:

[0038]

[0039] in, For relative complex sound pressure; take the upper triangular and diagonal elements of matrix K, and separate the real and imaginary parts to obtain the sequence f. prep :

[0040]

[0041] in, and These represent taking the real part and the imaginary part, respectively. This indicates that a new matrix is ​​formed by taking the elements from the upper triangle and the diagonal. This represents matrix serialization; taking into account the structural characteristics of GFGRU, the input data is represented as:

[0042]

[0043] in, T b This represents the number of time blocks in GFGRU;

[0044] S23: Generate labels for each sound velocity profile in the synthetic and real datasets, and obtain... and Both generate labels in the same way, for any sound velocity profile data The corresponding frequency domain sound pressure data samples are The corresponding coefficient vector is In this context, the superscript (·) represents the real dataset, and (~) represents the synthetic dataset; The corresponding reference tags are any element coefficient vector The i-th element in.

[0045] Specifically, S3 involves: first, pre-training the GFGRU network using a synthetic dataset and a multi-label classification loss function to obtain a pre-trained model; then, continuing to train the pre-trained model using a real dataset and the multi-label classification loss function to complete the training of the GFGRU network.

[0046] Before pre-training begins, set the training parameters for the GFGRU network: set the input batch size (batch_size), hidden layer size (hidden_size), and weight decay factor (weight_decay); during training, use Adam as the optimizer, set the initial learning rate (learning_rate), and set the learning rate per e... t The wheel decays to half its original value;

[0047] For sound pressure data in any frequency domain and its labels The multi-label classification loss function is calculated as follows:

[0048]

[0049] In this context, the superscript (·) represents the real dataset, and (~) represents the synthetic dataset; For the output of GFGRU; Ω neg Q represents I The set of basis functions for the sound speed profile in Central Africa, Ω pos Q represents I The middle composition The basis function set of the sound velocity profile; s i With s j As a reference measure for assessing the consistency between the network output and the label values, and and They represent and The i-th element in and They represent and The j-th element in.

[0050] Furthermore, S4 specifically includes:

[0051] The trained GFGRU network is used to estimate the basis functions constituting the sound velocity profile to be inverted. For the measured sound field data p = [p1, p2, ..., p...], ... L ] T Let the corresponding sound velocity profile of the region to be inverted be Input frequency domain acoustic field data f of GFGRU network in The data is obtained by the data preprocessing method described in step S2; the input data f is calculated by the GFGRU network. in Obtain the output result The element with a value of 1 is located in the dictionary Q. I The corresponding basis functions are the basis functions of the sound velocity profile that constitutes the region to be inverted, estimated by GFGRU.

[0052] Furthermore, S5 specifically includes:

[0053] S51: Initialize the population. Set the population size for the genetic algorithm to N. p The crossover ratio is P c The variation ratio is P v The maximum number of iterations is I. ga ; Determine the search range [min(X), max(X)] of each element in the coefficient vector x of the basis function to be searched based on the maximum and minimum values ​​of the estimated coefficient matrix X; Based on the set population size N p The population is generated randomly within the search range. Including individual samples any element For individual samples Its corresponding estimated coefficient vector Where ⊙ represents the Hadamard product, and its corresponding sound speed profile is calculated:

[0054]

[0055] in, Indicates the nth p The sound velocity profile estimation results for each individual sample; the search set is generated from the estimated coefficient vector. Initialize the number of iterations i ga =1;

[0056] S52: Calculate the loss function for each individual sequentially and sort them in ascending order; for sound velocity profiles Calculate the copy field sound field data using the Kraken model. The loss function between the measured sound field data and the copied sound field data is defined as:

[0057]

[0058] in, The cross-power spectrum is the normalized result of the signals received by any two array elements in the measured sound field data, and p m ∈p; To copy the acoustic field data, the normalized cross-power spectrum between the received signals of any two array elements is given, and T L Normalization coefficients:

[0059]

[0060] S53: Crossover operation: Initialize t c =1, initialize the child set

[0061] (1) Based on the cross ratio P c The top N population loss functions, sorted in ascending order. p ×P c Two individuals are randomly selected from the individuals. and i≠j, and randomly select the intersection position c∈[1,N], and swap them. and The coefficient vector preceding the corresponding position in the vector is used to obtain the swapped vector. and New offspring individuals and Add to child set middle;

[0062] (2)t c =t c +1, if t c ≥N p ×P c , end the loop; otherwise repeat (1);

[0063] S54: Mutation operation: Initialize t v =1;

[0064] (1) Randomly select offspring individuals with replacement from the offspring set. And in Randomly select a non-zero coefficient in the middle and modify it to a random value in [min(X), max(X)].

[0065] (2)t v =t v +1, if t v ≥N p ×P v , end the loop; otherwise repeat (1);

[0066] S55: Generating a new generation of sets: The specific steps are as follows:

[0067] (1) Merge child sets and search sets for

[0068] (2) According to step S52, calculate the loss function of each child individual in the child set. in Represents offspring individuals The corresponding estimated sound velocity profile;

[0069] (3) For sets The loss functions of all individuals are sorted in ascending order, and the top N are selected. P The search set to be formed in the next iteration

[0070] S56:i ga =i ga +1 if i ga ≥I ga Stop the iteration and output the individual with the minimum loss function value. Otherwise, repeat steps S52 to S55;

[0071] S57: Calculate the estimated sound velocity profile of the region to be inverted

[0072] The beneficial effects of this invention are as follows:

[0073] 1) This invention utilizes the K-SVD dictionary learning algorithm to generate basis functions to characterize the sound velocity profile. Traditional matched-field inversion methods, due to the use of a finite number of empirical orthogonal functions to represent the sound velocity profile, cannot accurately reflect the temporal and spatial perturbations of the sound velocity profile, resulting in a certain loss of accuracy. Using the K-SVD dictionary learning algorithm to generate basis functions can achieve higher accuracy in the representation of the sound velocity profile, require fewer basis functions, and also improve the efficiency of subsequent searches.

[0074] 2) This invention utilizes the sparse classification capability of the GFGRU network to address the problem that dictionary learning methods struggle to determine the basis functions used to characterize unknown sound velocity profiles to be inverted. The GFGRU network exhibits good performance in sparse classification problems. By training the GFGRU network using the known sound field data and basis functions corresponding to the sound velocity profiles, the basis functions for estimating unknown sound velocity profiles in the region to be inverted can be achieved.

[0075] 3) This invention uses artificially synthesized datasets to pre-train the GFGRU network, effectively improving the accuracy and robustness of GFGRU network classification while reducing the impact of insufficient experimental data samples on GFGRU network training. Simultaneously, a multi-label classification loss function is introduced, which effectively avoids the underfitting problem that traditional cross-entropy functions are prone to cause in multi-label classification tasks. Attached Figure Description

[0076] Figure 1 This is a flowchart of the shallow sea acoustic velocity profile matching field inversion method based on gated feedback loop unit of the present invention.

[0077] Figure 2 This is a diagram of the GFGRU network structure used in this invention. (a) is the GFGRU structure diagram, and (b) is the structure diagram of the GRU unit within the GFGRU.

[0078] Figure 3 This is a flowchart of the genetic algorithm used in this invention.

[0079] Figure 4 This is a simulation diagram of the classification results of GFGRU for four different sound field samples to be inverted in this invention. The solid black line represents the true labels, and the dashed red line represents the classification results of GFGRU.

[0080] Figure 5 This is a performance comparison chart of the inversion results of the shallow sea acoustic velocity profile matching field inversion method based on gated feedback gated cyclic unit of the present invention and the traditional matching field inversion results. Among them, (a) is the result of traditional matching field inversion, and (b) is the inversion result of the shallow sea acoustic velocity profile matching field inversion method based on gated feedback gated cyclic unit of the present invention. Detailed Implementation

[0081] The following simulation experiment of shallow sea sound velocity profile inversion is used as an example to further illustrate the present invention.

[0082] In this embodiment, the environmental parameters of the shallow sea, the vertical array, and the relevant parameters of the sound source in the Kraken simulation environment are as follows:

[0083] The water depth is 80m, the signal source depth is 40m, and the signal frequency is 400Hz; the vertical array is a 16-element array, uniformly distributed at depths of 15m-75m, and the horizontal distance from the signal source is 1km; the sound velocity in the seabed medium is 1800m / s, and the density is 1.65g / cm³. 3 The attenuation is 0.3dB / λ.

[0084] The flowchart of the embodiment is as follows Figure 1 As shown, the specific steps are as follows:

[0085] S1: Based on the existing actual sound velocity profile data of the region to be inverted, use the K-SVD algorithm to generate a dictionary of sound velocity profiles of the region to be inverted and calculate the basis functions that constitute each sound velocity profile.

[0086] S1 specifically refers to:

[0087] S11: Set the error ε, maximum number of iterations I, and sparsity S. p In this embodiment, ε = 1 × 10 -3 I = 200, S p =2; Based on the actual sound velocity profile data of the region to be inverted, a dictionary is generated using the K-SVD algorithm; the average sound velocity profile of the actual sound velocity profile data of the region to be inverted is calculated. For a given set of M sound velocity profiles discretized into K dimensions in the vertical direction, its mean sound velocity profile... The disturbance on is denoted as Let Y represent the actual sound velocity profile dataset, and the dictionary corresponding to Y is denoted as... N is the number of basis functions in the dictionary. Using the OMP algorithm, with a sparsity of S... p And the coefficient matrix reconstructed from dictionary Q by Y is denoted as The initial dictionary is obtained by randomly initializing the dictionary. Set the initial value of the loop in the K-SVD algorithm to i = 0;

[0088] S12: i = i + 1, sparse coding: for a known Q i-1 For constrained optimization problems

[0089]

[0090] The solution is obtained using the orthogonal matching pursuit (OMP) algorithm. Where S... p Let x represent the vector to be solved. m sparsity, and Let L2 and L0 represent the norms, respectively.

[0091] The OMP algorithm is as follows:

[0092] (1) Initialize the solution vector Then residual Support set of the solution The initial value of the loop in the OMP algorithm is k = 1;

[0093] (2) Calculate the error for each basis function in the dictionary. in

[0094] (3) Find the smallest And update S k =S k-1 ∪{j k};

[0095] (4) Based on the support set S k Calculation makes Minimum solution vector

[0096] (5) Update residuals

[0097] (6) If ||r k ||2<ε or k>S p If k = k + 1, stop the algorithm; otherwise, return to (2) for the next iteration.

[0098] S13: Dictionary update. (Regarding Q) i-1 basis functions in Row vectors in j = 1, ..., N and Xx j Update each element ∈X,j=1,…,N sequentially, with the following update steps:

[0099] (1) When updating dictionary Q i-1 The j-th basis function At that time, calculate Reconstruction error matrix of external dictionary extract x j The non-zero terms in the vector are arranged in order to form a new column vector x. R , x j The non-zero terms in E correspond to the column vectors in E, forming a new matrix E. R ;

[0100] (2) For E R Perform SVD decomposition: E R =USV T U and V represent the values ​​of E respectively. R The left and right singular matrices after SVD decomposition, where S represents the diagonal matrix composed of singular values ​​of each order;

[0101] (3) Update x R =V(:,1)S(1,1), which means that... Update x to the first left singular vector in U. R Updated to the product of the first right singular vector in V and the first-order singular value in S; and using x R The coefficients of the non-zero terms in the pair x j Update the corresponding element in the middle;

[0102] (4) Repeat (1)-(3) until all column vectors in the dictionary have been updated;

[0103] S14: Repeat S12 and S13 until i≥I, then output the dictionary Q. I And coefficient matrix X, x m Q corresponding to the zero term in the middle and non-zero terms I The basis functions in the equation are the basis functions of the m-th sound velocity profile, x m Q corresponding to zero I The basis functions in the equation are those that do not constitute the m-th sound velocity profile.

[0104] S2: Generate a synthetic sound velocity profile dataset based on the basis function distribution of the sound velocity profile in the region to be inverted. Based on the synthetic sound velocity profile dataset and the actual sound velocity profile data of the region to be inverted in step S1, calculate the sound field using the Kraken model and preprocess the sound field to obtain the synthetic dataset and the real dataset. Label each sound velocity profile in the synthetic dataset and the real dataset.

[0105] S2 specifically refers to:

[0106] S21: Vectorize the coefficient matrix X, find the non-zero terms, and construct a new vector. According to the law of large numbers Approximately Gaussian distribution, solve The sampling mean μ and sampling variance σ 2 In this embodiment, the obtained sample mean μ = 0.88 and the sample variance σ 2 =265.42. Set the number of sound velocity profile samples N to be synthesized. s N s >>N, in this embodiment, N s =12000. Based on the sample mean μ and sample variance σ 2 and sparsity S p A new synthesis sequence matrix is ​​generated by synthesizing coefficient vectors one by one. for Any coefficient vector Its synthesis method is as follows: initialization After that S is randomly selected from N coefficients. p A set of zero values ​​is set to a random non-zero value, i.e., satisfying the condition... Furthermore, the distribution of non-zero values ​​follows N(μ,σ) 2 Gaussian distribution; by Generate N s Artificially synthesized sound velocity profile data

[0107] S22: For synthetic sound velocity profile data The sound field was calculated using the Kraken model, along with the actual sound velocity profile data Y. and The obtained sound field is then preprocessed using the following method:

[0108] Sound field data samples corresponding to the synthesized sound velocity profile data For example:

[0109]

[0110] Where L is the number of elements in the vertical array, in this embodiment, L = 16. The covariance matrix K is calculated by normalizing the frequency domain sound pressure data:

[0111]

[0112] in, For relative complex sound pressure, take the upper triangular and diagonal elements of matrix K, and separate the real and imaginary parts to obtain the sequence f. prep :

[0113]

[0114] in, and These represent taking the real part and the imaginary part, respectively. This indicates that a new matrix is ​​formed by taking the elements from the upper triangle and the diagonal. This represents matrix serialization; taking into account the structural characteristics of GFGRU, the input data is represented as:

[0115]

[0116] in, T b This represents the number of time blocks in GFGRU;

[0117] S23: Generate labels for each sound velocity profile in the synthetic and real datasets, obtaining the labels. and Both generate tags in the same way, by synthesizing sound speed profiles. For example, for any sound speed profile Corresponding frequency domain sound pressure data samples Its reference label any element x i It is the i-th element in the coefficient vector.

[0118] S3: The GFGRU network is pre-trained using the synthetic dataset to obtain a pre-trained model; then the pre-trained model is trained using the real dataset to complete the multi-label classification task; a multi-label loss function is used during training to help convergence;

[0119] S3 specifically refers to:

[0120] First, construct the GFGRU network structure and set the number of GFGRU stacking layers N. h and number of time blocks T b See attached for details. Figure 2 (a) and appendix Figure 2 (b), in which attached Figure 2(b) shows the specific structure of the GRU unit; in this embodiment, N h =2,T b =10;

[0121] Set the training parameters for the GFGRU network and train it. Set the input batch size (batch_size), hidden layer size (hidden_size), and weight decay factor (weight_decay). Use Adam as the optimizer during training, and set the initial learning rate (learning_rate) per e... t The noise level is reduced to half of its original value in each training round. In this embodiment, during the first 50 training rounds, the training data contains random additive Gaussian noise ranging from -15dB to 15dB. To prevent the model from overlearning noise features, the random additive Gaussian noise in the training data is adjusted to -5dB to 15dB in the training rounds after the 50th round. In this embodiment, the input batch size is 2, the hidden layer size is 256, the number of GFGRU stacking layers is 2, and the number of horizontal time blocks is 10. In this embodiment, the optimizer used is Adam, and the initial learning rate is 4×10⁻⁶. -4 Furthermore, the weights decay to half their original value every 20 rounds, resulting in a network weight decay coefficient of 4 × 10⁻⁶. -4 .

[0122] Pre-training is performed in the same way as training. Taking pre-training using a synthetic dataset as an example, for any sample... The multi-label classification loss function used during training is as follows:

[0123]

[0124] in, For reference label values, This is the output of GFGRU. Ω neg express The set of basis functions for the sound speed profile in Central Africa, Ω pos express The set of basis functions that constitute the sound speed profile; s i With s j As a reference measure for assessing the consistency between the network output and the label values, and They represent and The i-th element in and They represent and The j-th element in;

[0125] If pre-training is performed, subsequent training using real datasets will be conducted based on the network generated by the pre-training.

[0126] S4: Input the measured sound field data of the region to be inverted into the trained GFGRU network to obtain the basis functions of the sound velocity profile of the region to be inverted;

[0127] S4 specifically refers to:

[0128] The trained GFGRU network is used to estimate the basis functions constituting the sound velocity profile to be inverted. For the measured sound field data p = [p1, p2, ..., p...], ... L ] T Let the corresponding sound velocity profile of the region to be inverted be The frequency domain sound field data is processed in the same way as S22, resulting in the final input data f to the GFGRU. in The output results are obtained by calculating the input data using the GFGRU network. The basis functions corresponding to the positions of elements with a label of 1 are the basis functions of the sound velocity profile that constitutes the region to be inverted, estimated by GFGRU.

[0129] S5: Use a genetic algorithm to search for the basis function coefficients of the sound velocity profile of the region to be inverted, use the matching field method to calculate the loss function, iterate until the loss function value is lower than the set threshold or the number of iterations reaches the set maximum number of steps, and use the optimal individual to calculate and obtain the sound velocity profile of the region to be inverted.

[0130] S5 specifically refers to:

[0131] S51: Initialize the population. Set the population size for the genetic algorithm to N. p The crossover ratio is P c The variation ratio is P v The maximum number of iterations is I. ga In this embodiment, N p =200, P c =0.5, P v =0.05, I ga =200; Since the coefficients of the basis functions obtained by dictionary learning are relatively evenly distributed in terms of energy, the search range [min(X), max(X)] of each element in the coefficient vector x of the basis function to be searched can be determined based on the maximum and minimum values ​​of the known estimated coefficient matrix X; Based on the set population size N p The population is generated randomly within the search range. Including individual samples any element For individual samples Its corresponding estimated coefficient vector Where ⊙ represents the Hadamard product, and the corresponding sound speed profile is calculated based on the basis function coefficient vector:

[0132]

[0133] in, Indicates the nth p The sound velocity profile estimation results for each individual sample; the search set is generated from the estimated coefficient vector. Initialize the number of iterations i ga =1;

[0134] S52: Calculate the loss function for each individual sequentially and sort them in ascending order. For sound velocity profiles... Calculate the copy field sound field data using the Kraken model. The loss function between the measured sound field data and the copied sound field data is defined as:

[0135]

[0136] in, The cross-power spectrum is the normalized result of the signals received by any two array elements in the measured sound field data, and p m ∈p; To copy the acoustic field data, the normalized cross-power spectrum between the received signals of any two array elements is given, and T L Normalization coefficients:

[0137]

[0138] S53: Interleaving operation. Initialize t c =1, initialize the child set

[0139] (1) Based on the cross ratio P c The top N population loss functions, sorted in ascending order. p ×P c Two individuals are randomly selected from the individuals. and i≠j, and randomly select the intersection position c∈[1,N], and swap them. and The vectors of coefficients preceding the corresponding positions are used to obtain the swapped vectors. and New offspring individuals and Add to child set middle;

[0140] (2)t c =t c +1, if tc ≥N p ×P c , end the loop; otherwise repeat (1);

[0141] S54: Mutation operation. Initialize t v =1;

[0142] (1) Randomly select offspring individuals with replacement from the offspring set. and Randomly select a non-zero coefficient in the middle and modify it to a random value in [min(X), max(X)].

[0143] (2)t v =t v +1, if t v ≥N p ×P v , end the loop; otherwise repeat (1);

[0144] S55: Generate a new generation of sets. The specific steps are as follows:

[0145] (1) Merge child sets and search sets for

[0146] (2) According to step S52, calculate the loss function of each child individual in the child set. in Represents offspring individuals The corresponding estimated sound velocity profile;

[0147] (3) For sets The loss functions of all individuals are sorted in ascending order, and the top N are selected. P The search set to be formed in the next iteration

[0148] S56:i ga =i ga +1 if i ga ≥I ga Stop the iteration and output the individual with the minimum loss function value. Otherwise, repeat steps S52 to S55;

[0149] S57: Calculate the estimated sound velocity profile of the region to be inverted

[0150] Figure 4This is a schematic diagram of the classification results using the GFGRU network. In this embodiment, the classification accuracy of GFGRU under different training sets of different data sizes and different training methods is shown in Table 1. Under training method A, GFGRU is not pre-trained, and the number of samples used during training is only 1% of the historical sample size, i.e., 109 samples. Under training method B, GFGRU is not pre-trained, and the number of samples used during training is 10% of the historical sample size, i.e., 1090 samples. Under training method C, GFGRU will use artificially synthesized datasets for pre-training, and the number of samples used during training is 10% of the historical sample size, i.e., 1090 samples. It can be seen that group A achieved good classification results on the test set even with missing training samples, indicating that under the training strategy used in this invention, the network itself learns the features of different basis functions, thus having a certain generalization ability. Group C has higher accuracy in both positive and negative examples than group B, and also has a smaller standard deviation, indicating that using synthetic datasets for pre-training can improve the problem of insufficient training samples, and also improve the accuracy and robustness of the network model.

[0151] Table 1 shows the classification accuracy of GFGRU under different training methods.

[0152] A 99.9990％ 0.9971 <![CDATA[0.2816×10 -2 ]]> <![CDATA[0.3146×10 -5 ]]> <![CDATA[0.4736×10 -4 ]]> B 99.9990％ 0.9972 <![CDATA[0.2699×10 -2 ]]> <![CDATA[0.2836×10 -5 ]]> <![CDATA[0.4449×10 -4 ]]> C 99.9991％ 0.9976 <![CDATA[0.2553×10 -2 ]]> <![CDATA[0.2021×10 -5 ]]> <![CDATA[0.3764×10 -4 ]]>

[0153] Figure 5 This is a comparison between the inversion results using the algorithm proposed in this invention and the inversion results using traditional matched-field methods. Figure 5 (a) is the inversion result of the traditional matched field method. Figure 5 (b) shows the inversion result of the algorithm in this invention, where the red dashed line is the estimated sound speed profile and the black solid line is the measured sound speed profile. It can be seen that under the same search parameters, the algorithm proposed in this invention can effectively improve the inversion accuracy.

[0154] The basic principles and main features of this invention have been described in detail above with reference to the accompanying drawings. This method utilizes the K-SVD dictionary learning algorithm to generate basis functions with higher representation accuracy, while simultaneously using a GFGRU network to estimate the basis functions of the sound velocity profile to be inverted. Furthermore, to address the problem of insufficient training data for the GFGRU network, this method uses artificially synthesized datasets to pre-train the GFGRU network, effectively improving the accuracy and robustness of the GFGRU network's classification, while reducing the impact of insufficient experimental data samples on the performance of the GFGRU network. This invention has good engineering practical value.

[0155] This invention is not limited to the embodiments described above. Those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for inverting a matched field of shallow sea acoustic velocity profiles based on GFGRU, characterized in that, The specific steps are as follows: S1: Based on the existing actual sound velocity profile data of the region to be inverted, use the K-SVD algorithm to generate a dictionary of sound velocity profiles of the region to be inverted and calculate the basis functions that constitute each sound velocity profile. S2: Generate a synthetic sound velocity profile dataset based on the basis function distribution of the sound velocity profile in the region to be inverted. Based on the synthetic sound velocity profile dataset and the actual sound velocity profile data of the region to be inverted in step S1, calculate the sound field using the Kraken model and preprocess the sound field to obtain the synthetic dataset and the real dataset. Label each sound velocity profile in the synthetic dataset and the real dataset. S3: The GFGRU network is pre-trained using the synthetic dataset to obtain a pre-trained model; then the pre-trained model is trained using the real dataset to complete the multi-label classification task; a multi-label loss function is used during training to help convergence; S4: Input the measured sound field data of the region to be inverted into the trained GFGRU network to obtain the basis functions of the sound velocity profile of the region to be inverted; S5: Use a genetic algorithm to search for the basis function coefficients of the sound velocity profile of the region to be inverted, use the matching field method to calculate the loss function, iterate until the loss function value is lower than the set threshold or the number of iterations reaches the set maximum number of steps, and use the optimal individual to calculate and obtain the sound velocity profile of the region to be inverted.

2. The method for shallow sea acoustic profile matching field inversion based on GFGRU according to claim 1, characterized in that, Specifically, S1 is: S11: Setting Error Maximum number of iterations sparsity ; A dictionary is generated using the K-SVD algorithm based on the actual sound velocity profile data of the region to be inverted. Calculate the average sound velocity profile of the actual sound velocity profile data of the region to be inverted. For a known set of M sound velocity profiles discretized into K dimensions in the vertical direction, its mean sound velocity profile... The disturbance on is denoted as ,use To refer to the actual sound speed profile dataset, The corresponding dictionary is denoted as N is the number of basis functions in the dictionary. The OMP algorithm is used to calculate the sparsity of the basis functions. and dictionaries right The reconstructed coefficient matrix is ​​denoted as The initial dictionary is obtained by randomly initializing the dictionary. Set the initial values ​​for the K-SVD algorithm loop. ; S12: Sparse coding: for known The OMP algorithm is used to solve the constrained optimization problem, which is: In the formula, S p Represents the vector to be solved sparsity, and They represent and Norm; The OMP algorithm is as follows: (1) Initialize the solution vector Then the residual The support set of the solution OMP algorithm loop initialization value ; (2) Calculate the error for each basis function in the dictionary. ,in ; (3) Find the smallest and update ; (4) Based on the support set Calculation makes Minimum solution vector ; (5) Update residuals ; (6) If or Stop the algorithm; otherwise Return to (2) for the next iteration; S13: Dictionary update: (for...) basis functions in as well as row vectors in Update them one by one, the update steps are as follows: (1) When updating the dictionary The basis functions At that time, calculate Reconstruction error matrix of external dictionary ,extract The non-zero terms in the vector are arranged in order to form a new column vector. , The non-zero terms in the text correspond to The column vectors in the matrix form a new matrix. ; (2) To Perform SVD decomposition: , and They represent respectively to The left and right singular matrices after SVD decomposition Represents a diagonal matrix composed of singular values ​​of various orders; (3) Update , Soon Updated to The first left singular vector in the middle, will Updated to The first right singular vector in the middle and The product of the first-order singular values; and using The coefficients of the non-zero terms in the pair Update the corresponding element in the middle; (4) Repeat (1) - (3) until all column vectors in the dictionary have been updated; S14: Repeat S12 and S13 until... Output dictionary and coefficient matrix X, The zero term corresponding to China and Africa The basis functions in the equation are the basis functions for the m-th sound velocity profile. The zero term corresponds to The basis functions in the equation are not the basis functions that constitute the m-th sound velocity profile.

3. The shallow sea acoustic profile matching field inversion method based on GFGRU according to claim 2, characterized in that, Specifically, S2 is: S21: Vectorize the coefficient matrix X, find the non-zero terms, and construct a new vector. According to the law of large numbers, Approximately Gaussian distribution, solve Sampling mean and sampling variance Set the number of sound velocity profile samples to be synthesized. Based on the sample mean Sampling variance sparsity A new synthesis sequence matrix is ​​generated by synthesizing coefficient vectors one by one. ,for Any coefficient vector Its synthesis method is as follows: initialization After that of Randomly select from the coefficients A set of zero values ​​is set to a random non-zero value, i.e., satisfying the condition... And the distribution of non-zero values ​​therein follows Gaussian distribution; by generate Artificially synthesized sound velocity profile data ; S22: For synthetic sound velocity profile data and actual sound velocity profile data The sound field was calculated using the Kraken model. and The obtained sound field is then preprocessed using the following method: For any vector in the sound field : in, superscript for Representing frequency domain sound pressure data from real datasets, for L represents the frequency domain sound pressure data in the synthetic dataset; L is the number of elements in the vertical matrix; the covariance matrix K is calculated by normalizing the frequency domain sound pressure data. in, For relative complex sound pressure; take the upper triangular and diagonal elements of matrix K, and separate the real and imaginary parts to obtain the sequence. : in, and These represent taking the real part and the imaginary part, respectively. This indicates that a new matrix is ​​formed by taking the elements from the upper triangle and the diagonal. This represents matrix serialization; taking into account the structural characteristics of GFGRU, the input data is represented as: in, , This represents the number of time blocks in GFGRU; S23: Generate labels for each sound velocity profile in the synthetic and real datasets, and obtain... and Both generate labels in the same way, for any sound velocity profile data. The corresponding frequency domain sound pressure data samples are The corresponding coefficient vector is , where superscript for Represents a real dataset, for Represents a synthetic dataset; The corresponding reference tags are , any element , coefficient vector The i-th element in.

4. The shallow sea acoustic velocity profile matching field inversion method based on GFGRU according to claim 3, characterized in that, Specifically, S3 is: First, the GFGRU network is pre-trained using a synthetic dataset and a multi-label classification loss function to obtain a pre-trained model. Then, the pre-trained model is trained again using a real dataset and the multi-label classification loss function to complete the training of the GFGRU network. Before pre-training begins, set the training parameters for the GFGRU network: set the input batch size (batch_size), hidden layer size (hidden_size), and weight decay coefficient (weight_decay); during training, use Adam as the optimizer, set the initial learning rate (learning_rate), and... The wheel decays to half its original value; For sound pressure data in any frequency domain and its labels The multi-label classification loss function is calculated as follows: Among them, superscript for Represents a real dataset, for Represents a synthetic dataset; Output the results for GFGRU; express The set of basis functions for the sound speed profile in Central and African countries. express The middle composition The set of basis functions for the sound velocity profile; and As a reference measure for assessing the consistency between the network output and the label values, and , , and They represent and The i-th element in and They represent and The j-th element in.

5. The shallow sea acoustic velocity profile matching field inversion method based on GFGRU according to claim 4, characterized in that, Specifically, S4 is: For measured sound field data Let the corresponding sound velocity profile of the region to be inverted be Input frequency domain acoustic field data of GFGRU network The data is obtained by the data preprocessing method described in step S2; the input data is calculated by the GFGRU network. Obtain the output result , The position of the element with a value of 1 in the dictionary The corresponding basis functions are the basis functions of the sound velocity profile that constitutes the region to be inverted, estimated by GFGRU.

6. The method for shallow sea acoustic profile matching field inversion based on GFGRU according to claim 5, characterized in that, Specifically, S5 is: S51: Initialize the population, setting the population size for the genetic algorithm to be... The crossover ratio is P c The variation ratio is P v The maximum number of iterations is Based on the estimated coefficient matrix The maximum and minimum values ​​determine the coefficient vector of the basis function to be searched. Search range of each element Based on the set population size The population is generated randomly within the search range. Including individual samples any element For individual samples The corresponding estimated coefficient vector ,in Represent the Hadamard product and calculate its corresponding sound speed profile: in, Indicates the first The sound velocity profile estimation results for each individual sample; the search set is generated from the estimated coefficient vector. Initialize the number of iterations. ; S52: Calculate the loss function for each individual sequentially and sort them in ascending order; for sound velocity profiles The copy field sound field data was calculated using the Kraken model. The loss function between the measured sound field data and the copied sound field data is defined as: in, The cross-power spectrum is the normalized result of the signals received by any two array elements in the measured sound field data, and... ; To copy the acoustic field data, the normalized cross-power spectrum between the received signals of any two array elements is given, and ;T L Normalization coefficients: S53: Cross operation: Initialization Initialize the child set ; (1) Based on the cross ratio P c The top-ranked populations in ascending order of their population loss function Two individuals are randomly selected from the individuals. and And randomly select the intersection position ,exchange and The coefficient vector preceding the corresponding position in the vector is used to obtain the swapped vector. and The new offspring individuals and Add to child set middle; (2) ,like , end the loop; otherwise repeat (1); S54: Mutation Operation: Initialization ; (1) Randomly select offspring individuals with replacement from the offspring set. and in Randomly select a non-zero coefficient and modify it to Random values ​​in; (2) ,like , end the loop; otherwise repeat (1); S55: Generating a new generation of sets: The specific steps are as follows: (1) Merge child sets and search sets for ; (2) According to step S52, calculate the loss function of each child individual in the child set. ,in Represents offspring individuals The corresponding estimated sound velocity profile; (3) For sets The loss functions of all individuals are sorted in ascending order, and the top ones are selected. The search set to be formed in the next iteration ; S56: ,like Stop the iteration and output the individual with the minimum loss function value. Otherwise, repeat steps S52 to S55. S57: Calculate the estimated sound velocity profile of the region to be inverted .

Images