Shallow-sea low-frequency near-field layered seabed earth sound inversion method and device

Abstract

The invention discloses a shallow sea low-frequency near-field layered seabed earth sound inversion method, and the method comprises the steps: respectively obtaining the pulse of a direct wave and the pulse of a seabed reflected wave; converting the pulse time domain signal into a frequency domain signal, and calculating an actual reflection coefficient; on the basis of the reflection coefficient matrix of the frequency domain signal, a neural network model based on a cross-self-attention mechanism is constructed, the input of the neural network is an analog reflection coefficient matrix, and the output of the neural network is a predicted value of an earth sound parameter; wherein the simulation reflection coefficient matrix is obtained by simulation of a forward model; the model is trained in a multi-task learning mode, the trained model is used for predicting an actual reflection coefficient matrix, and the actual reflection coefficient matrix is input of the trained model. According to the method, efficient and accurate multi-parameter inversion can be realized, and the inversion precision and stability of near-field seabed layering and earth sound parameters thereof are effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of acoustic inversion, and in particular to a shallow-water low-frequency near-field layered seabed geoacoustic inversion method and device. Background Art

[0002] The acoustic properties of seabed sediments have a significant impact on sound propagation, especially in shallow sea environments, where sound waves are reflected multiple times by the seabed. Sediment geoacoustic parameters (such as sound velocity, density, attenuation coefficient, sediment thickness, etc.) are of great significance for acoustic field modeling, target detection, and sonar performance evaluation. There are two main methods for obtaining sediment geoacoustic parameters: direct measurement and geoacoustic inversion. Direct measurement technology is expensive, and the sampling depth and range are limited. Geoacoustic inversion technology uses acoustic signals to penetrate the seabed, and uses the depth information and geoacoustic characteristics it contains to obtain acoustic parameters in deeper layers and wider areas. However, in the currently known geoacoustic inversion technology, it is usually impossible to solve the coupling problem of sound velocity and thickness in each layer of the seabed. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the embodiments of the present application provide a method, apparatus, computing device, computer storage medium and product containing a computer program for shallow water low-frequency near-field layered seabed geoacoustic inversion, which can achieve efficient and accurate multi-parameter inversion and effectively improve the inversion accuracy and stability of near-field seabed stratification and its geoacoustic parameters.

[0004] In the first aspect, an embodiment of the present application provides a shallow-water low-frequency near-field layered seabed geoacoustic inversion method, comprising: respectively obtaining direct wave pulses and seabed reflection wave pulses; converting the pulse time domain signal into a frequency domain signal, and calculating the actual reflection coefficient; based on the reflection coefficient matrix of the frequency domain signal, constructing a neural network model based on a cross-self-attention mechanism, the input of the neural network is a simulated reflection coefficient matrix, and the output is a predicted value of the geoacoustic parameter; wherein the simulated reflection coefficient matrix is ​​obtained by simulation of a forward model; the model is trained by multi-task learning, and the trained model is used to predict the actual reflection coefficient matrix, wherein the actual reflection coefficient matrix is ​​the input of the trained model.

[0005] In some possible implementations, building a neural network model based on the cross-self-attention mechanism includes: building a forward model, the forward model input is the seabed stratification and geoacoustic parameters, and the output is the reflection coefficient matrix; splitting the reflection coefficient matrix into two branches in terms of frequency and angle, and embedding a label vector; adding a feature label vector to each branch, and then performing position encoding; inputting all label vectors into L-layer multi-channel self-attention mechanism modules, using residual connections between layers, and calculating self-attention weights in the lth layer; fusing the reflection coefficient features in the angle domain and the frequency domain to obtain global features, and mapping the global features to the geoacoustic parameter vector.

[0006] In some possible implementations, the model is trained by multi-task learning, including: using a multi-task learning method based on gradient normalization, taking the prediction of each geoacoustic parameter as a separate subtask, and assigning a separate task weight to each subtask; using an automatic differentiation method to calculate the gradient norm of each subtask loss on the neural network weight parameters, and calculating the task gradient loss based on the training rate of each subtask; updating the gradient norm of all subtasks to make the gradient norm close to the standard gradient, updating the neural network weight parameters and normalizing the subtask weights.

[0007] In some possible implementations, when assigning a separate task weight to each subtask, the total loss function of the kth training cycle in the neural network model is calculated according to the following formula:

[0008]

[0009] Where, L sum (k) represents the total multi-task loss, N t Characterizes the number of subtasks, ω i (k) represents the task weight of the i-th subtask, L i (k) represents the loss of the i-th subtask, and k represents the k-th frequency point.

[0010] In some possible implementations, the gradient norm is calculated according to the following formula:

[0011]

[0012] In the formula, Represents the gradient norm of the subtask loss on the neural network weight parameter W, where W represents the neural network weight parameter, represents the partial differential operator, and i represents the i-th subtask.

[0013] In some possible implementations, the task gradient loss is calculated according to the following formula:

[0014]

[0015] Where, L grad Characterize the gradient loss, N t Represents the number of subtasks, represents the average gradient of all subtasks, and has r i (k) represents the relative training rate of the i-th subtask, and has α represents the strength parameter used to control the normalization of subtask gradients.

[0016] In some possible implementations, the processing within the angle branch may be performed according to the following formula:

[0017]

[0018] In the formula, Q h Characterize the angle branch query matrix, K h Characterizing the angular branch bond matrix, V h represents the angle branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix that represents the query matrix of the angle branch, A learnable projection matrix that represents the angular branch bond matrix, The learnable projection matrix representing the angle branch value matrix, Atten represents the self-attention weight, d θ The dimension of the feature vector representing the angle branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

[0019] In some possible implementations, the processing within the frequency branch may be performed according to the following formula:

[0020]

[0021] In the formula, Q h ′ represents the frequency branch query matrix, K h ′ represents the frequency branch bond matrix, V h ′ represents the frequency branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix representing the frequency branch query matrix, A learnable projection matrix that represents the frequency branch bond matrix, The learnable projection matrix representing the frequency branch value matrix, Atten represents the self-attention weight, d F The dimension of the feature vector representing the frequency branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

[0022] In some possible implementations, the fusion of the reflection coefficient features in the angle domain and the frequency domain may be performed according to the following formula:

[0023]

[0024] Where q represents the cross query vector, W q Represents the linear mapping matrix, Representation of CLS calculated by self-attention mechanism θ , Kh ′ represents the frequency branch characteristic bond matrix, The frequency label vector representing the L layer, The projection matrix representing the bond matrix within the frequency branch, V h ′ represents the projection matrix of the frequency branch internal value matrix, V h ' represents the frequency branch value matrix, CrossAtten represents the cross attention weights of the angle branch and the frequency branch.

[0025] In the second aspect, an embodiment of the present application provides a shallow-water low-frequency near-field layered seabed geoacoustic inversion device, including: an acquisition module, used to respectively acquire pulses of direct waves and pulses of seabed reflection waves; a processing module, used to convert the time domain signal of the pulse into a frequency domain signal, and calculate the actual reflection coefficient; the processing module is also used to construct a neural network model based on the cross-self-attention mechanism based on the reflection coefficient matrix of the frequency domain signal, the input of the neural network is a simulated reflection coefficient matrix, and the output is a predicted value of the geoacoustic parameter; wherein the simulated reflection coefficient matrix is ​​obtained by simulation of the forward model; the processing module is also used to train the model by multi-task learning, and use the trained model to predict the actual reflection coefficient matrix, wherein the actual reflection coefficient matrix is ​​the input of the trained model.

[0026] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, comprising computer-readable instructions. When a computer reads and executes the computer-readable instructions, the computer executes the method described in any one of the first aspects.

[0027] In a fourth aspect, an embodiment of the present application provides a computing device, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method described in any one of the first aspects is executed.

[0028] In a fifth aspect, an embodiment of the present application provides a product comprising a computer program. When the computer program product runs on a processor, the processor executes the method as described in any one of the first aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0030] Figure 1 It is a schematic diagram of an application scenario of a shallow-water low-frequency near-field layered seabed geoacoustic inversion method provided in an embodiment of the present application;

[0031] Figure 2 It is a flow chart of a shallow-water low-frequency near-field layered seabed geoacoustic inversion method provided in an embodiment of the present application;

[0032] Figure 3 It is a schematic diagram of a near-field time domain signal provided in an embodiment of the present application;

[0033] Figure 4 It is a schematic diagram of the pulse compression result of the acoustic signal received by all hydrophones in a vertical array provided in an embodiment of the present application;

[0034] Figure 5 is a schematic diagram of a correction result of a hydrophone offset position on a vertical array provided in an embodiment of the present application;

[0035] Figure 6 It is a frequency domain amplitude diagram of a direct wave and a seabed reflection wave provided in an embodiment of the present application;

[0036] Figure 7 is a schematic diagram of average energy at each center frequency of a direct wave and a seabed reflection wave provided in an embodiment of the present application;

[0037] Figure 8 This is a schematic diagram of actual reflection coefficient of a near-field layered seabed provided in an embodiment of the present application;

[0038] Fig. 9 It is a schematic diagram of a cross-self-attention mechanism neural network model provided in an embodiment of the present application;

[0039] Fig.10 It is a schematic diagram of the seabed reflection coefficient of a near-field layered seabed environment provided by an embodiment of the present application;

[0040] Fig.11 is a schematic diagram of global sensitivity distribution of an objective function near an optimal solution provided in an embodiment of the present application;

[0041] Fig.12 It is a schematic diagram of a normalized global attention weight distribution diagram of a cross-self-attention mechanism model provided in an embodiment of the present application;

[0042] Fig.13 It is a structural schematic diagram of a shallow-water low-frequency near-field layered seabed geoacoustic inversion device provided in an embodiment of the present application;

[0043] Fig.14 It is a structural diagram of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The term "and / or" in this article is a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The symbol " / " in this article indicates that the associated objects are in an or relationship, for example, A / B means A or B.

[0046] The terms "first" and "second" in the specification and claims herein are used to distinguish different objects rather than to describe a specific order of the objects. For example, a first response message and a second response message are used to distinguish different response messages rather than to describe a specific order of the response messages.

[0047] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0048] In the description of the embodiments of the present application, unless otherwise specified, "multiple" means two or more than two. For example, multiple processing units refer to two or more processing units, etc.; multiple elements refer to two or more elements, etc.

[0049] To facilitate the understanding of the embodiments of the present application, the following will further explain and illustrate specific embodiments in conjunction with the accompanying drawings. The embodiments do not constitute a limitation on the embodiments of the present invention.

[0050] Geoacoustic inversion can be divided into near-field and mid-field according to the distance of the sound source. Mid-field and far-field acoustic inversion can obtain seabed acoustic parameters in a large sea area, but the sound source and receiver need to be located in different places. This multi-base inversion method is relatively complicated in actual operation. Near-field acoustic inversion can arrange sound sources and receiving equipment on a single base to achieve motorized real-time inversion, which is highly practical. In addition, the near-field acoustic signal has a larger angle of incidence on the seabed and penetrates deeper into the seabed, so it can obtain more detailed seabed layered structure and its geoacoustic parameters. Traditional near-field seabed layer acquisition methods mainly include layer peeling method and matching field method. The layer peeling method requires that the time interval of the reflection signal of each layer of the seabed is greater than the signal pulse width, so a broadband high-frequency strong pulse sound source is required. The matching field method has the problem of multiple extreme values, is prone to fall into local extreme points, and has low computational efficiency. In addition, neither the layer peeling method nor the matching field method can solve the coupling problem of sound velocity and thickness in each layer of the seabed.

[0051] In view of this, in order to solve the problem of multi-parameter coupling in the traditional near-field inversion method, the embodiment of the present application provides a shallow low-frequency near-field layered seabed geoacoustic inversion method, which is implemented based on deep learning technology. Deep learning technology can train neural network models and adaptively extract features from acoustic data. According to the sensitivity differences of different geoacoustic parameters, a suitable neural network model is constructed, and an effective optimization algorithm is adopted, which is expected to solve the multi-parameter coupling problem in the inversion of seabed stratification and its geoacoustic parameters. The acoustic reflection coefficient of the layered seabed presents an oscillating structure with frequency and grazing angle, and its oscillation characteristics are related to the seabed stratification and its geoacoustic parameters. By analyzing the near-field pulse signal, the seabed reflection coefficients at different angles and frequencies are obtained. A multi-task learning method is used to train a neural network model based on the cross-self-attention mechanism, and the characteristics of the reflection coefficient in the frequency domain and the angle domain are extracted respectively, and the relationship between these characteristics and the geoacoustic parameters is learned, thereby realizing the geoacoustic parameter inversion of the near-field layered seabed, and providing a theoretical and technical basis for realizing the mobile real-time inversion of the seabed layered structure and its geoacoustic parameters.

[0052] For example, Figure 1 The following is a schematic diagram showing an application scenario of a shallow water low-frequency near-field layered seabed geoacoustic inversion method provided by an embodiment of the present application. Figure 1 As shown, a plurality of hydrophones are distributed on the vertical receiving array, and the hydrophones are used to receive acoustic signals. The ship stops near the receiving array and the transmitting transducer is suspended. In this embodiment, the transducer can use the UW350 low-frequency sound source transducer to transmit a 30-1000Hz linear frequency modulation signal. The transducer is used as a signal source to invert the geoacoustic parameters of the near-field shallow seabed, so the influence of the sedimentary layer attenuation coefficient can be ignored. The parameters to be inverted are the seabed surface sound velocity, the seabed second layer sound velocity and the thickness of the seabed surface.

[0053] The inversion method provided in the embodiment of the present application can be applied to this scenario. For example, Figure 2 FIG. 1 is a flow chart of a shallow-water low-frequency near-field layered seafloor geoacoustic inversion method provided by an embodiment of the present application. Figure 2 As shown, the method may include the following steps:

[0054] S21: Obtain the pulse of the direct wave and the pulse of the seabed reflection wave respectively.

[0055] In this embodiment, the transmitting transducer transmits an acoustic wave signal to the seabed, and a plurality of hydrophones are used to receive the return signal. On the basis of considering the emission spectrum level correction of the transmitting transducer and the Doppler effect, the received signal is correlated with the sound source signal to obtain a broadband pulse signal, which is recorded as P(t). The hydrophone array is calibrated according to the position of the sound source transducer, the position of the hydrophone, the local seawater sound velocity profile, and the corresponding relationship between the propagation distance and arrival time of the direct wave sound line at each receiving depth of the hydrophone array. The near-field multi-path arrival structure is analyzed, and the pulse signals of the direct wave and the seabed reflection wave are picked up. Among them, the ray method can be used to analyze the near-field multi-path arrival structure.

[0056] by Figure 1 Taking the application scenario shown as an example, there are 17 hydrophones on the vertical array buoy for receiving sound signals. The experimental ship stops near the buoy and drops the UW350 transmitting transducer to transmit 30-1000Hz linear frequency modulation signals. Figure 3 shows the near-field time domain signal after pulse compression. Figure 3 As shown in the figure, the near-field pulse signal after the received signal is compressed. The solid line is the pulse compression result after considering the correction of the transmitting sound source level, and the dotted line is the direct pulse compression result. It can be seen that the pulse compression after considering the correction of the sound source level can suppress the side lobes. The near-field multi-path arrival structure is analyzed using the ray model to pick up the pulse peaks of the direct wave and the seabed reflection wave. Figure 4 The figure shows the pulse compression results of the acoustic signals received by all hydrophones in the vertical array. Figure 4 As shown in the figure, the circles mark the arrival times of the direct waves, seabed reflected waves, and sea surface transmitted waves in each hydrophone signal calculated by Bellhop. Among them, the seabed reflected waves, direct waves, and sea surface transmitted waves from the 1st to the 12th seabed hydrophones are clearly distinguished. Figure 5 The results of locating the offset position of each hydrophone on the vertical array using information such as the arrival time difference of the direct wave of each hydrophone, the depth position of the sound source, the seawater depth at the vertical array position, and the seawater sound speed are shown.

[0057] S22: Convert the pulse time domain signal into a frequency domain signal, and calculate the actual reflection coefficient.

[0058] In this embodiment, the pulse signal of the direct wave and the pulse signal of the seabed reflection wave are intercepted respectively, and the fast Fourier transform is performed to convert the time domain signal into the frequency domain signal. The entire frequency range is divided according to a certain center frequency and a one-third octave bandwidth to obtain a number of frequency bands. The average energy of each frequency point in each frequency band is calculated as the energy at the center frequency, and the amplitudes of the direct wave and the seabed reflection wave are calculated respectively. After compensating for the expansion loss, the actual reflection coefficient is calculated.

[0059] In some possible embodiments, the pulse signal of the direct wave is recorded as P d (t), the pulse signal of the seabed reflection wave is denoted as P r (t), the time domain signal is converted into the frequency domain signal according to the following formula:

[0060]

[0061] In the formula, P(f k ) represents the time domain signal at the kth frequency point, f k Represents the kth frequency point, and there is f k = k·f s / N,f s represents the sampling rate, N represents the number of time domain sampling points, P(n) represents the time domain signal at the nth time point, and j represents the imaginary unit.

[0062] In some possible embodiments, the actual reflection coefficient is calculated according to the following formula:

[0063]

[0064] In the formula, R obs (f,θ) represents the actual reflection coefficient, P r (f) represents the amplitude of the reflected wave with frequency f, P d (f) represents the amplitude of the direct wave with frequency f, D r (θ) represents the propagation distance of the reflected wave, D d (θ) represents the propagation distance of the direct wave. Propagation distance D r (θ) and D d (θ) can be calculated from the depth of the sound source transducer and the hydrophone, the horizontal distance between the two, and the sea depth.

[0065] Please continue to refer to Figure 1-5 , respectively intercept Figure 4 The direct wave pulse signal P of hydrophone No. 1-13 d (t) and the pulse signal P of the seabed reflection wave r (t), and then perform fast Fourier transform to convert the time domain signal into frequency domain signal. Sampling rate f sis 16000Hz. Taking the actual signal with a sound source depth of 73.6m, a hydrophone depth of 45.58m, and a horizontal distance of 221.98m as an example, the frequency domain amplitudes of the direct wave and the seabed reflection wave are as follows: Figure 6 For a frequency domain signal of 30Hz-1kHz, the center frequency is selected every 50Hz from 150Hz to 1kHz, and the bandwidth is one-third octave, resulting in 18 frequency bands. The average energy of each frequency point in each frequency band is calculated as the energy at the center frequency, as shown in Figure 7 The amplitudes of the direct wave and the seabed reflected wave are calculated respectively, and the actual reflection coefficient is calculated after compensating for the expansion loss. Figure 8 is the actual seabed reflection coefficient at a grazing angle of 31.73 degrees.

[0066] S23: Based on the reflection coefficient matrix of the frequency domain signal, a neural network model based on the cross-self-attention mechanism is constructed. The input of the neural network is the simulated reflection coefficient matrix, and the output is the predicted value of the geoacoustic parameter.

[0067] In this embodiment, the wave number integration method is first used as the forward model for calculating the seabed reflection coefficient in the near-field layered seabed environment. All parameter combinations are traversed according to a certain step size within the range of geoacoustic parameter values, and the simulated reflection coefficient matrix is ​​calculated for each set of geoacoustic parameters. The reflection coefficient matrix is ​​recorded as R cal (f,θ). All samples together constitute a data set for training deep neural networks. The neural network model can choose a neural network model based on the cross-self-attention mechanism, and the input of the model is the reflection coefficient matrix R of all seabed models. cal (f,θ), the output is the predicted values ​​of multiple geoacoustic parameters.

[0068] Specifically, the forward model uses the wave number integration method to simulate the seabed reflection coefficient. It calculates the simulated reflection coefficient matrix based on known geoacoustic parameters (such as the nature of the seabed, the sound velocity profile of the seawater, etc.). These simulated reflection coefficient matrices are used as data sets to train the neural network model. For example, Fig. 9 A schematic diagram of a cross-self-attention mechanism neural network model provided in an embodiment of the present application is shown. Fig. 9 As shown, the neural network model converts R cal (f,θ) is split into two branches according to the frequency and angle directions and embedded with the label vector, and the feature label vector CLS is added to each branch. θ and CLS fAfter that, position encoding is performed. All the label vectors are input into the L-layer multi-channel self-attention mechanism module respectively. Residual connections are used between the layers, and the self-attention weights are calculated in the lth layer. Among them, the angle branch contains the frequency domain reflection coefficient vectors at each angle, and the frequency branch contains the angle domain reflection coefficient vectors at each frequency. The global features are obtained by processing and fusing the reflection coefficient features in the angle domain and the frequency domain, and then the global features are mapped to the ground sound parameter vector Here, i represents the number of predicted geoacoustic parameters.

[0069] In some possible embodiments, the processing within the angle branch may be performed according to the following formula:

[0070]

[0071] In the formula, Q h Characterize the angle branch query matrix, K h Characterizing the angular branch bond matrix, V h represents the angle branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix that represents the query matrix of the angle branch, A learnable projection matrix that represents the angular branch bond matrix, The learnable projection matrix representing the angle branch value matrix, Atten represents the self-attention weight, d θ The dimension of the feature vector representing the angle branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

[0072] The feature vector dimension of the branch is converted into a normalized weight coefficient through softmax processing, and then processed by the L-layer multi-channel self-attention mechanism module to obtain the sequence representation of the angle branch. in, This is the CLS calculated by the self-attention mechanism. θ Due to CLS θ and the frequency label vectors of all angles to calculate self-attention, thus including R cal Global features of (f,θ) in the angular dimension.

[0073] In some possible embodiments, the processing within the frequency branch may be performed according to the following formula:

[0074]

[0075] In the formula, Q h ′ represents the frequency branch query matrix, K h ′ represents the frequency branch bond matrix, V h′ represents the frequency branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix representing the frequency branch query matrix, A learnable projection matrix that represents the frequency branch bond matrix, The learnable projection matrix representing the frequency branch value matrix, Atten represents the self-attention weight, d F The dimension of the feature vector representing the frequency branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

[0076] The processing flow of the frequency branch is the same as that of the angle branch. After passing through the L-layer multi-channel self-attention mechanism module, the sequence representation of the frequency branch is output. Then the angle domain CLS θ is mapped to the feature space of the frequency branch and the cross attention weights are calculated.

[0077] In some possible embodiments, the reflection coefficient features in the angle domain and the frequency domain may be integrated according to the following formula:

[0078]

[0079] Where q represents the cross query vector, W q Represents the linear mapping matrix, Representation of CLS calculated by self-attention mechanism θ , K h ′ represents the frequency branch characteristic bond matrix, The frequency label vector representing the L layer, The projection matrix representing the bond matrix within the frequency branch, V h ′ represents the projection matrix of the frequency branch internal value matrix, V h ' represents the frequency branch value matrix, CrossAtten represents the cross attention weights of the angle branch and the frequency branch.

[0080] because Contains the feature label vector CLS of the frequency branch F The extracted frequency domain feature information, so CrossAtten in the above formula realizes the fusion of angle domain and frequency domain features.

[0081] By reference Figure 1-9In this example, the wave number integration method OASR is used as the forward model for calculating the seabed reflection coefficient in the near-field layered seabed environment. Since the sound source level of the experimental transmission signal is small, the energy of the transmission waves in each layer below the seabed is very small. Here, a double-layer model is considered first. All parameter combinations are traversed according to a certain step size within the range of geoacoustic parameter values, and the simulated reflection coefficient matrix R is calculated for each set of geoacoustic parameters. cal (f,θ), all samples together constitute the data set for training deep neural networks. In this example, the parameter settings of the data set are shown in the following table:

[0082] Parameters to be inverted Value range Step Length unit Surface sound velocity [1500,1700] 10 m / s Second layer of sound speed [1500,1700] 10 m / s Surface thickness [0,10] 0.5 m

[0083] Construct a neural network model based on the cross-self-attention mechanism. The input of the model is R cal (f,θ), the output is the predicted values ​​of the seafloor surface sound velocity, the seafloor second layer sound velocity and the seafloor surface thickness. cal (f,θ) is split into two branches according to the frequency and angle directions and embedded with the label vector. The angle branch contains the frequency domain reflection coefficient vector at each angle, and the frequency branch contains the angle domain reflection coefficient vector at each frequency. The feature label vector CLS is added to each branch. θ and CLS f Then position encoding is performed. All the tag vectors are input into the 5-layer (i.e. L = 5) multi-channel self-attention mechanism module, and residual connections are used between each layer. When calculating, the feature vector dimension d of the angle branch is taken separately θ = 192 and the frequency branch vector dimension d f = 96 for calculation. By processing the angle branch and the frequency branch, the reflection coefficient features in the angle domain and the frequency domain are integrated to obtain the global features, and then the global features are mapped to the geoacoustic parameter vector i=3 represents the number of predicted geoacoustic parameters, thereby realizing geoacoustic parameter prediction based on the cross-self-attention mechanism.

[0084] S24: Train the model through multi-task learning, and use the trained model to predict the actual reflection coefficient matrix.

[0085] In this embodiment, a multi-task learning method based on gradient normalization is used, and the prediction of each geoacoustic parameter is regarded as a separate subtask, and a separate task weight is assigned to each subtask. By using the automatic differentiation method of the deep learning framework, the gradient norm of each subtask loss on the neural network weight parameter W can be efficiently calculated, and then the task gradient loss is calculated according to the training rate of each subtask to adjust the gradient of each subtask to make it close to the standard gradient. The gradient of the gradient on the subtask weight is calculated by automatic differentiation, and the gradient norm of all subtasks is updated to make it close to the standard gradient. Then, the neural network weight parameters are updated and the weights of the subtasks are normalized. This process is repeated until the model training is completed. After the model training is completed, the actual reflection coefficient matrix R is input. obs (f,θ), the predicted values ​​of i geoacoustic parameters can be output, thus realizing the simultaneous inversion of multiple geoacoustic parameters.

[0086] In some possible embodiments, when assigning a separate task weight to each subtask, the total loss function of the kth training cycle in the neural network model is calculated according to the following formula:

[0087]

[0088] Where, L sum (k) represents the total multi-task loss, N t Characterizes the number of subtasks, ω i (k) represents the task weight of the i-th subtask, L i (k) represents the loss of the i-th subtask, and k represents the k-th frequency point.

[0089] In some possible embodiments, the gradient norm may be calculated according to the following formula:

[0090]

[0091] In the formula, Represents the gradient norm of the subtask loss on the neural network weight parameter W, where W represents the neural network weight parameter, represents the partial differential operator, and i represents the i-th subtask.

[0092] In some possible embodiments, the task gradient loss may be calculated according to the following formula:

[0093]

[0094] Where, L grad Characterize the gradient loss, N t Represents the number of subtasks, represents the average gradient of all subtasks, and has r i(k) represents the relative training rate of the i-th subtask, and has α represents the strength parameter used to control the normalization of subtask gradients.

[0095] After the model optimization converges, input the actual reflection coefficient matrix R obs (f,θ), the output can be the predicted values ​​of the three geoacoustic parameters, thereby realizing the simultaneous inversion of the seafloor surface velocity, the second layer velocity and the surface thickness. Figure 1 In the application scenario shown, the inversion results of the experimental sea area are that the seabed surface sound velocity is 1600-1610m / s, the second layer sound velocity is 1540-1550m / s, and the surface thickness is 2.5-3m. This prediction result is consistent with the Figure 1 The seafloor sampling results shown are basically consistent.

[0096] In order to further verify the practicability of this method, Figure 1 The experiment was conducted in the application scenario shown in the figure, and the feature extraction ability of the neural network model of the cross-autocorrelation mechanism was verified using a simulation environment. In the simulation environment, the sound speed of the seabed surface is 1550m / s, the sound speed of the second layer of the seabed is 1650m / s, the surface thickness is 5m, and the emission coefficient of the layered seabed is calculated based on OASR. The angle range is 20-60 degrees and the frequency range is 600Hz-1kHz. The results are as follows Fig.10 Near the optimal solution (i.e., true value) of the geoacoustic parameters, the partial derivatives of the reflection coefficients at each angle point and frequency point with respect to the three geoacoustic parameters are calculated, and their root mean square is used as the global sensitivity of the objective function near the optimal solution, as shown in Fig.11 As shown. Using the data set with the same angle and frequency range as the simulation environment, the cross-self-attention mechanism model of the present invention is trained. After the model optimization converges, the model normalizes the global attention weight as shown in Fig.12 By comparison Fig.10 and Fig.11 It can be found that the areas with higher sensitivity of the objective function are consistent with the areas where the seabed reflection coefficient changes dramatically, indicating that the seabed reflection coefficient in these areas is more important to the inverted geoacoustic parameters. Fig.11 and Fig.12 It can be found that the distribution of global attention weights is consistent with the global sensitivity of the objective function, indicating that the cross-self-attention mechanism model proposed in the embodiment of the present application can effectively focus on more critical local features in the reflection coefficient matrix, thereby improving the accuracy of inversion.

[0097] The above is the shallow sea low-frequency near-field layered seabed geoacoustic inversion method provided in the embodiment of the present application. The oscillation characteristics of the layered seabed reflection coefficient in the angle domain and frequency domain are used to train the neural network model using a multi-task learning method to achieve accurate and efficient acquisition of layered seabed acoustic parameters. A transducer is used to transmit a low-frequency broadband signal, and a near-field hydrophone array is used to receive the acoustic signal. Considering the spectral level and Doppler effect of the transmitting transducer sound source, the source signal is compensated and then the received signal is pulse compressed to extract the seabed reflection wave and the direct wave, calculate the seabed reflection coefficient at each frequency and angle, and construct a neural network model based on the cross-self-attention mechanism. The characteristics of the reflection coefficient matrix in the angle domain and frequency domain are extracted, and the model is trained using a multi-task learning method based on gradient normalization to achieve efficient and accurate multi-parameter inversion, effectively improving the inversion accuracy and stability of near-field seabed stratification and its geoacoustic parameters.

[0098] It is understandable that the size of the sequence number of each step in the above-mentioned embodiments does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. In addition, in some possible implementations, the steps in the above-mentioned embodiments can be selectively executed according to actual conditions, and can be partially executed or fully executed, which is not limited here. All or part of any feature of any embodiment of the present application can be freely and arbitrarily combined without contradiction. The combined technical solution is also within the scope of the present application.

[0099] The present invention proposes a shallow sea low-frequency near-field layered seabed geoacoustic inversion method based on deep learning, and the system is configured as a transmitting transducer and a hydrophone array. The present invention places the transmitting transducer and the hydrophone array in the water at a close distance, and the hydrophone receives the broadband linear frequency modulation sound signal radiated by the transducer sound source, removes the Doppler effect and the sound source spectrum level, and performs pulse compression to obtain a broadband pulse signal; the receiving array attitude is corrected using the transmitting transducer depth and relative position, the hydrophone depth, and the direct wave arrival time information; the direct wave, the sea surface reflection wave, and the seabed reflection wave arrival time are obtained using the seawater depth and sound speed information, and the channel signal that can be completely separated by the three propagation paths is selected to intercept the direct wave and the seabed reflection wave signal; the grazing angle corresponding to the seabed reflection wave is calculated according to the relative position of the transmitting transducer and the hydrophone; finally, the seabed reflection coefficient matrix is ​​calculated, and the angle and frequency oscillation structure of the seabed reflection coefficient are input into the trained neural network model to obtain the seabed layered structure and its geoacoustic parameters.

[0100] In the process of pulse compression of the received original signal, this method not only removes the Doppler effect caused by the movement of the sound source, but also corrects the source signal by considering the sound source level of the transmitting transducer, so that the main lobe of the compressed pulse signal is strong and the side lobe is weak, which is conducive to the resolution and interception of the direct wave and the reflected wave. A neural network model based on the cross-self-attention mechanism is constructed, and the self-attention weights are calculated separately in the angle branch and the frequency branch. The cross-attention mechanism is used to fuse the reflection coefficient features in the angle domain and the frequency domain to obtain the global features, and then the global features are mapped to the geoacoustic parameter vector. Through this feature extraction mechanism of separation first and then fusion, it is expected to improve the sensitivity of the parameters to be inverted and reduce the coupling between the parameters. The multi-task learning method based on gradient normalization regards the prediction of each geoacoustic parameter as a separate subtask, and assigns a separate task weight to each subtask, further eliminating the problem of multi-parameter coupling.

[0101] It is worth noting that, due to the low sound source level in the example experiment used in the introduction of the method, the seabed reflection signal only includes the reflection waves of the surface layer and the second layer, so only the geoacoustic parameters of the seabed surface layer and the second layer are obtained. The numerical simulation results show that this method is suitable for obtaining more layers of seabed stratification and their geoacoustic parameters. In practical applications, increasing the transmitting power of the transmitting transducer can obtain deeper seabed stratification structures and geoacoustic parameters. Therefore, the above examples are only used to illustrate the technical solution of the present invention and are not limiting. This near-field inversion method is also applicable to the operation mode of a ship towing a transmitting transducer and a hydrophone array at the same time, and this method can better reflect the practicality of this inversion method in the mobile acquisition of seabed stratification and its geoacoustic parameters.

[0102] Based on the method in the above embodiment, the embodiment of the present application also provides a shallow sea low-frequency near-field layered seabed geoacoustic inversion device. For example, Fig.13 A shallow water low frequency near field layered seabed geoacoustic inversion device is shown. Fig.13 As shown, the device 1300 includes an acquisition module 1301 and a processing module 1302 .

[0103] The acquisition module 1301 is used to respectively acquire the pulse of the direct wave and the pulse of the seabed reflection wave;

[0104] A processing module 1302 is used to convert the time domain signal of the pulse into a frequency domain signal and calculate an actual reflection coefficient;

[0105] The processing module 1302 is further used to construct a neural network model based on a cross-self-attention mechanism based on the reflection coefficient matrix of the frequency domain signal, the input of the neural network is a simulated reflection coefficient matrix, and the output is a predicted value of the geoacoustic parameter; wherein the simulated reflection coefficient matrix is ​​obtained by forward modeling;

[0106] The processing module 1302 is further used to train the model by multi-task learning, and use the trained model to predict the actual reflection coefficient matrix, wherein the actual reflection coefficient matrix is ​​the input of the trained model.

[0107] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method, which will not be repeated here.

[0108] The present application also provides a computing device 1400. Fig.14 As shown, computing device 1400 includes: bus 1402, processor 1404, memory 1406 and communication interface 1408. Processor 1404, memory 1406 and communication interface 1408 communicate through bus 1402. Computing device 1400 can be a server or a terminal device. It should be understood that the present application does not limit the number of processors and memories in computing device 1400.

[0109] The bus 1402 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.14 The bus 1404 may include a path for transmitting information between various components of the computing device 1400 (eg, the memory 1406, the processor 1404, and the communication interface 1408).

[0110] The processor 1404 may include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0111] The memory 1406 may include a volatile memory, such as a random access memory (RAM). The processor 1404 may also include a non-volatile memory, such as a read-only memory (ROM), a flash memory, a hard disk drive (HDD), or a solid state drive (SSD).

[0112] The memory 1406 stores executable program codes, and the processor 1404 executes the executable program codes to respectively implement the functions of the acquisition module 1301 and the processing module 1302, thereby implementing all or part of the steps of the method in the above embodiment. That is, the memory 1406 stores instructions for executing all or part of the steps in the method in the above embodiment.

[0113] Alternatively, the memory 1406 stores executable codes, and the processor 1404 executes the executable codes to respectively implement the functions of the shallow water low-frequency near-field layered seabed geoacoustic inversion device 1300, thereby implementing all or part of the steps in the above-mentioned embodiment method. That is, the memory 1406 stores instructions for executing all or part of the steps in the above-mentioned embodiment method.

[0114] The communication interface 1408 uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the computing device 1400 and other devices or communication networks.

[0115] Based on the methods in the above embodiments, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the methods in the above embodiments.

[0116] Based on the methods in the above embodiments, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor executes the methods in the above embodiments.

[0117] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0118] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0119] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions may be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state disk (SSD)), etc.

[0120] It should be understood that the various numerical numbers involved in the embodiments of the present application are only used for the convenience of description and are not used to limit the scope of the embodiments of the present application.

Claims

1. A shallow water low-frequency near-field layered seabed geoacoustic inversion method, characterized in that: The method comprises: Obtain direct wave pulses and seabed reflected wave pulses respectively; Convert the pulse time domain signal into a frequency domain signal and calculate the actual reflection coefficient; Based on the reflection coefficient matrix of the frequency domain signal, a neural network model based on a cross-self-attention mechanism is constructed, the input of the neural network is a simulated reflection coefficient matrix, and the output is a predicted value of the geoacoustic parameter; wherein the simulated reflection coefficient matrix is ​​obtained by forward modeling; The model is trained by multi-task learning, and the trained model is used to predict the actual reflection coefficient matrix, wherein the actual reflection coefficient matrix is ​​the input of the trained model.

2. The method according to claim 1, characterized in that The construction of a neural network model based on a cross-self-attention mechanism includes: Constructing a forward model, wherein the input of the forward model is seabed stratification and geoacoustic parameters, and the output is the reflection coefficient matrix; The reflection coefficient matrix is ​​divided into two branches according to the frequency and angle directions, and the label vector is embedded; after adding the feature label vector to each branch, the position encoding is performed; all the label vectors are input into the L-layer multi-channel self-attention mechanism module respectively, and the residual connection is used between the layers, and the self-attention weight is calculated in the lth layer; The reflection coefficient features in the angle domain and the frequency domain are fused to obtain the global features, which are then mapped to the geoacoustic parameter vector.

3. The method according to claim 1, characterized in that The model is trained by multi-task learning, including: A multi-task learning method based on gradient normalization is adopted to treat the prediction of each geoacoustic parameter as a separate subtask, and a separate task weight is assigned to each subtask; Using an automatic differentiation method, the gradient norm of each subtask loss on the neural network weight parameter is calculated, and according to the training rate of each subtask, the task gradient loss is calculated; Update the gradient norms of all subtasks to make them close to the standard gradient, update the neural network weight parameters and normalize the weights of the subtasks.

4. The method according to claim 3, characterized in that When assigning a separate task weight to each subtask, the total loss function for the kth training cycle in the neural network model is calculated according to the following formula: In the formula, Lsum(k) represents the total loss of multi-task, N t represents the number of subtasks, ωi(k) represents the task weight of the i-th subtask, and L i (k) represents the loss of the i-th subtask, and k represents the k-th frequency point.

5. The method according to claim 3, characterized in that: The gradient norm is calculated according to the following formula: In the formula, Represents the gradient norm of the subtask loss on the neural network weight parameter W, where W represents the neural network weight parameter, represents the partial differential operator, and i represents the i-th subtask.

6. The method according to claim 1, characterized in that The calculation task gradient loss is calculated according to the following formula: Where, L grad Characterize the gradient loss, N t Represents the number of subtasks, represents the average gradient of all subtasks, and has r i (k) represents the relative training rate of the i-th subtask, and has α represents the strength parameter used to control the normalization of subtask gradients.

7. The method according to claim 2, characterized in that: The processing in the angle branch can be performed according to the following formula: In the formula, Q h Characterize the angle branch query matrix, K h Characterizing the angular branch bond matrix, V h represents the angle branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix that represents the query matrix of the angle branch, A learnable projection matrix that represents the angular branch bond matrix, The learnable projection matrix representing the angle branch value matrix, Atten represents the self-attention weight, d θ The dimension of the feature vector representing the angle branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

8. The method according to claim 2, characterized in that: The processing within the frequency branch can be performed according to the following formula: In the formula, Q h ′ represents the frequency branch query matrix, K h ′ represents the frequency branch bond matrix, V h ′ represents the frequency branch value matrix, h represents the number of channels of the multi-channel self-attention mechanism, and each channel can learn different feature representations. A learnable projection matrix representing the frequency branch query matrix, A learnable projection matrix that represents the frequency branch bond matrix, The learnable projection matrix representing the frequency branch value matrix, Atten represents the self-attention weight, d F The dimension of the feature vector representing the frequency branch, represents the frequency label vector of the l-1 layer, and T represents the transpose.

9. The method according to claim 1, characterized in that: The reflection coefficient characteristics of the angular domain and the frequency domain can be integrated according to the following formula: Where q represents the cross query vector, W q Represents the linear mapping matrix, Representation of CLS calculated by self-attention mechanism θ , K h ′ represents the frequency branch characteristic bond matrix, The frequency label vector representing the L layer, The projection matrix representing the bond matrix within the frequency branch, V h ′ represents the projection matrix of the frequency branch internal value matrix, V h ' represents the frequency branch value matrix, CrossAtten represents the cross attention weights of the angle branch and the frequency branch.

10. A shallow sea low-frequency near-field layered seabed geoacoustic inversion device, characterized in that: The device comprises: An acquisition module, used for respectively acquiring pulses of direct waves and pulses of seabed reflected waves; A processing module, used for converting the time domain signal of the pulse into a frequency domain signal and calculating an actual reflection coefficient; The processing module is further used to construct a neural network model based on a cross-self-attention mechanism based on the reflection coefficient matrix of the frequency domain signal, the input of the neural network is a simulated reflection coefficient matrix, and the output is a predicted value of the geoacoustic parameter; wherein the simulated reflection coefficient matrix is ​​obtained by forward modeling; The processing module is also used to train the model by multi-task learning, and use the trained model to predict the actual reflection coefficient matrix, wherein the actual reflection coefficient matrix is ​​the input of the trained model.

Images