Submarine full waveform inversion method, device, electronic equipment and storage medium

Abstract

The present application provides a subsea full waveform inversion method, apparatus, electronic device, and storage medium. The method includes: obtaining an initial subsea velocity model; determining a discrete wave equation of seismic waves under the initial subsea velocity model, and generating a synthetic seismic reflection record by forward modeling using the discrete wave equation; determining a velocity map of the initial subsea velocity model and at least one corresponding synthetic seismic reflection record; inputting the velocity map and the at least one corresponding synthetic seismic reflection record into a preset subsea full waveform depth neural network model for training to determine a mapping relationship between the velocity map and the at least one synthetic seismic record, and determining a full waveform inversion result based on the trained subsea full waveform depth neural network model. The weights of the trained subsea full waveform depth neural network model are saved as an inversion operator, reducing the inversion calculation time, accelerating the inference speed, and improving the accuracy of subsea full waveform inversion.

Description

Technical Field

[0001] The present application relates to the field of exploration geophysics, and particularly to a seabed full waveform inversion method, apparatus, electronic device, and storage medium. Background Art

[0002] Full waveform inversion (FWI) is a method in the field of geophysics for reconstructing a high-resolution velocity map from seismic data. The full waveform inversion problem is usually formulated as an optimization problem, using a partial differential equation (PDE) solver and a local optimization method to synthesize seismic records. FWI iteratively updates the underground model parameters by reducing the mismatch between the recorded data and the estimated data, and reconstructs the high-resolution underground physical property parameters. The existing traditional calculation methods for solving full waveform inversion not only have a large amount of calculation, but also produce low-resolution results due to the ill-posedness and cycle skipping problems of full waveform inversion. In recent years, data-driven deep learning methods have shown great potential in accelerating and simplifying the inversion process. The strategy of using deep learning technology to solve the seabed formation full waveform inversion problem is to apply a convolutional neural network (CNN) to directly derive the inversion operator f -1 , so as to obtain the velocity structure without knowing the forward operator f. However, this method also faces many challenges, including dataset constraints, poor generalizability, and weak physical interpretability. Summary of the Invention

[0003] In view of this, the purpose of the present application is to propose a seabed full waveform inversion method, apparatus, electronic device, and storage medium that overcome the above problems or at least partially solve the above problems.

[0004] Based on the above purpose, in the first aspect of the present application, a seabed full waveform inversion method is provided, including:

[0005] Obtain an initial seabed velocity model;

[0006] Determine the discrete wave equation of seismic waves under the initial seabed velocity model, and use the discrete wave equation to perform forward simulation to generate synthetic seismic reflection records;

[0007] Determine the velocity map of the initial seabed velocity model and at least one corresponding synthetic seismic reflection record;

[0008] Input the velocity map and at least one corresponding synthetic seismic reflection record into a preset seabed full waveform depth neural network model for training, determine the mapping relationship between the velocity map and at least one synthetic seismic record, and determine the full waveform inversion result based on the trained seabed full waveform depth neural network model.

[0009] Optionally, obtaining an initial seabed velocity model includes:

[0010] Obtain the layered model of the seabed formation;

[0011] Determine the minimum seabed formation velocity and the maximum seabed formation velocity;

[0012] Determine the slope of the layered model;

[0013] Perform elastic transformation and distortion on the layered model according to the minimum seabed formation velocity, the maximum seabed formation velocity, and the slope of the layered model to obtain the initial seabed velocity model.

[0014] Optionally, determine the discrete wave equation of seismic waves under the initial seabed velocity model, and use the discrete wave equation to perform forward modeling to generate a synthetic seismic reflection record, including:

[0015] Determine the acoustic wave displacement equation under the initial seabed velocity model;

[0016] Determine the source term fluctuation generated by the source wave;

[0017] Determine the number of shot points and their preset interval distances in the initial seabed velocity model;

[0018] Determine the number of geophone points and their preset interval distances in the initial seabed velocity model;

[0019] Use the shot points and geophone points to determine the reflected wave of the seismic wave;

[0020] Determine the discrete wave equation of seismic waves in the seabed velocity model according to the acoustic wave displacement equation and the source term fluctuation;

[0021] Use the discrete wave equation and the seismic reflection wave to perform forward modeling to generate a synthetic seismic reflection record.

[0022] Optionally, the acoustic wave displacement equation is expressed as:

[0023] u tt (x,z,t)-c 2 (x,z)Δ(u(x,z,y))=c 2 (x,z)f(x,z,t),

[0024] where u(x,z,t) represents the displacement of the acoustic wave, f(x,z,t) represents the external drive, c(x,z) represents the seismic wave propagation velocity, x represents the horizontal direction, and z represents the vertical direction;

[0025] The source term fluctuation is expressed as:

[0026]

[0027] where f0 represents the peak frequency of the source wave;

[0028] The discrete wave equation is expressed as:

[0029]

[0030] where u i,j,k = u(x i , z j , t k ), (x i , z j ), where i = 0, …, nx, j = 0, …, nz, nx = (x F - x I ) / Δ x , nz = (z F - z I ) / Δ z , t k = kΔt, k = 0, 1, … m, m = t F / m.

[0031] Optionally, the preset seabed full-waveform depth neural network model includes: an encoder and a decoder, the encoder includes a convolutional layer and a pooling layer, and the decoder includes a plurality of deconvolutional layers;

[0032] Inputting the velocity map and the corresponding synthetic seismic reflection record into the preset seabed full-waveform depth neural network model for training to determine the mapping relationship between the velocity map and the synthetic seismic record, and obtaining the full-waveform inversion result based on the trained seabed full-waveform depth neural network model, including:

[0033] Using the convolutional layer and the pooling layer to extract high-dimensional features of at least one of the synthetic seismic reflection records and compress the high-dimensional features into a single low-dimensional vector, and determining a first mapping relationship between at least one of the synthetic seismic reflection records and the single low-dimensional vector;

[0034] Using a plurality of deconvolutional layers to perform upsampling in sequence, deconvolving the single low-dimensional vector to increase the dimension of the single low-dimensional vector, and determining a second mapping relationship between the single low-dimensional vector and the high-dimensional velocity map;

[0035] According to the first mapping relationship and the second mapping relationship, determining the mapping relationship between the velocity map and at least one of the synthetic seismic reflection records.

[0036] Optionally, the convolutional layer and the pooling layer are expressed as:

[0037] x (l+1) = ReLU(BN(Conv(x (t) ))),

[0038] Conv(x)(i,j) = ∑ m ∑ n ∑ c K m,n,c · x (s-1)×i+m,(s-1)×j+n,c ,

[0039] MaxPooling(x) (i,j,c) = max m,n x(s - 1)×i + m, (s - 1)×j + n, c,

[0040]

[0041] where x is the independent variable of each function, Conv(x) represents the convolution function, ReLU(x) represents the activation function, MaxPooling(x) represents the max - pooling layer, BN(x) represents the batch normalization function, K represents the convolution kernel of m * n, c represents the dimension, s represents the stride between the sliding positions of the convolution kernel, γ and β represent two trainable parameters, μ B and represent the mean and variance calculated from all the calculations using the same feature map in a mini - batch respectively, ∈ represents a parameter used to increase numerical stability, and i, j, m, n represent spatial coordinates respectively;

[0042] The transposed convolution layer is expressed as:

[0043] y (l+1) = ReLU(BN(Deconv(y l )))

[0044] Deconv(y) (i,j) = ∑ m ∑ n ∑ c K m,n,c · y (s-1)×i+m,(s-1)×j+n,c ,

[0045] where Deconv(y) represents the transposed convolution function and K represents the transposed convolution kernel of m * n.

[0046] Optionally, it further includes:

[0047] Determine the mean absolute error function and the mean squared error function of the sub - sea full - waveform depth neural network model during the training process;

[0048] Use the mean absolute error function and the mean squared error function to determine the loss value for updating the parameters by backpropagation of the sub - sea full - waveform depth neural network model during the training process;

[0049] The mean absolute error function is expressed as:

[0050]

[0051] The mean square error function is expressed as:

[0052]

[0053] The loss value is expressed as:

[0054]

[0055] where y = {y1, …, y n} represents the actual subsurface formation velocity, z = {z1, …, z n} represents the predicted inversion result, n represents the number of all spatial positions in the full waveform depth neural network model of the seabed during the training process, and loss represents the loss value for updating the parameters by backpropagation of the full waveform depth neural network model of the seabed during the training process.

[0056] In a second aspect of the present application, a full waveform inversion device for the seabed is provided, including:

[0057] An acquisition module, configured to acquire an initial seabed velocity model;

[0058] A synthetic seismic reflection record module, configured to determine the discrete wave equation of seismic waves under the initial seabed velocity model, and generate a synthetic seismic reflection record by forward modeling using the discrete wave equation;

[0059] A determination module, configured to determine the velocity map of the initial seabed velocity model and the corresponding synthetic seismic reflection record;

[0060] An inversion module, configured to input the velocity map and the corresponding synthetic seismic reflection record into a preset full waveform depth neural network model of the seabed for training, determine the mapping relationship between the velocity map and the synthetic seismic record, and obtain a full waveform inversion result based on the trained full waveform depth neural network model of the seabed.

[0061] In a third aspect of the present application, an electronic device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the program, the method described in the first aspect is implemented.

[0062] In a fourth aspect of the present application, a non-transitory computer-readable storage medium is provided, where the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to cause a computer to execute the method described in the first aspect.

[0063] As can be seen from the above, the full waveform inversion method, device, electronic device, and storage medium for the seabed provided by the present application generate synthetic seismic reflection records by performing two-dimensional acoustic full waveform forward modeling on seismic waves using the initial seabed velocity model, use the synthetic seismic records and the velocity map of the initial velocity model as the data set, integrate physical constraint conditions into the data set, and finally train the weights of the preset full waveform depth neural network model for the seabed as the inversion operator f -1 Save it, input the synthetic seismic reflection record into the trained full waveform depth neural network model for the seabed and directly infer the seabed velocity model, that is, the velocity map, which reduces the inversion calculation time, speeds up the inference speed, and improves the accuracy of the full waveform inversion for the seabed.

[0064] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the description. And in order to make the above and other purposes, features, and advantages of the present invention more obvious and understandable, the following specifically gives the specific embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the present application or related technologies, the following will briefly introduce the drawings required for use in the embodiments or related technology descriptions. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0066] Figure 1 It is a flowchart of the full waveform inversion method for the seabed according to the embodiment of the present application;

[0067] Figure 2 It is a flowchart of obtaining the initial seabed velocity model according to the embodiment of the present application;

[0068] Figure 3 It is a flowchart of synthesizing seismic reflection records according to the embodiment of the present application;

[0069] Figure 4 It is a flowchart of determining the full waveform inversion result according to the embodiment of the present application;

[0070] Figure 5 It is a schematic diagram of the loss change of the training set and test set according to the embodiment of the present application;

[0071] Figure 6 It is a schematic diagram of the inversion effect according to the embodiment of the present application;

[0072] Figure 7 It is a schematic diagram of the full waveform depth neural network model for the seabed according to the embodiment of the present application;

[0073] Figure 8Schematic diagram of the seabed full waveform inversion device according to an embodiment of the present application;

[0074] Figure 9 Schematic diagram of the electronic device according to an embodiment of the present application. Detailed implementation manners

[0075] To make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0076] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be of the ordinary meaning understood by those of ordinary skill in the field to which the present application belongs. The "first", "second" and similar terms used in the embodiments of the present application do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this term cover the elements or objects listed after this term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0077] Referring to Figure 1 As shown, an embodiment of the present application provides a seabed full waveform inversion method, including:

[0078] S101. Obtain an initial seabed velocity model.

[0079] In this embodiment, as described in the background art, in the traditional full waveform inversion method, forward modeling is the basis of inversion. During the inversion process, at least one forward modeling operation is required for each iteration, with high computational requirements and large memory occupation during the operation process; when using a convolutional neural network, the inversion operator f can be directly derived without determining the forward operator. -1 , but there is a problem of weak physical interpretability. In the embodiments of the present application, a synthetic seismic reflection record is generated for subsequent forward modeling through the initial seabed velocity model, and the initial seabed velocity model and the synthetic seismic reflection record are used as the dataset for training the seabed full waveform depth neural network model, concentrating the physical constraint conditions into the dataset, effectively overcoming the problem of weak physical interpretability. At the same time, during the process of training the seabed full waveform depth neural network model, the corresponding forward modeling operations are no longer involved, reducing the amount of computation.

[0080] In some embodiments, referring to Figure 2 as shown, step S101 specifically includes:

[0081] S201. Obtain the underground stratified model;

[0082] S202. Determine the minimum and maximum seafloor formation velocities;

[0083] S203. Determine the slope of the stratified model;

[0084] S204. According to the minimum and maximum seafloor formation velocities and the slope of the stratified model, perform elastic transformation and distortion on the stratified model to obtain the initial seafloor velocity model.

[0085] In the embodiment of the present application, considering the actual situation, the minimum seafloor formation velocity is set to 1490 m / s and the maximum seafloor formation velocity is set to 4000 m / s; the underwater range is constrained between 625 m and 1125 m, and the slope of the stratified model is 4 to 12. According to the obtained minimum and maximum seafloor formation velocities and the slope of the stratified model, combined with the underwater range, the stratified model is elastically transformed and distorted, and a 1D velocity profile is cyclically created in the horizontal direction, realizing a relatively complex 2D initial seafloor velocity model with random initialization on a flat seafloor, so that the generated initial seafloor velocity model has more realistic underground geological characteristics. The process of cyclically creating a 1D velocity profile in the horizontal direction can be understood as establishing an OXY coordinate, and successively establishing velocity profiles in the OY direction along the OX direction.

[0086] S102. Determine the discrete wave equation of seismic waves under the initial seafloor velocity model, and use the discrete wave equation to perform forward simulation to generate a synthetic seismic reflection record.

[0087] In this embodiment, a synthetic seismic reflection record is generated by performing forward simulation through the obtained initial seafloor velocity model. The initial seafloor velocity model and the synthetic seismic reflection record are used as an effective training data set for subsequent training of the seafloor full-waveform depth neural network model, and physical constraint conditions are integrated into the data set.

[0088] In some embodiments, as shown in Figure 3 Step S102 includes:

[0089] S301. Determine the acoustic wave displacement equation under the initial seafloor velocity model;

[0090] S302. Determine the source term fluctuation generated by the seismic wave source;

[0091] S303. Determine the number of shot points on the initial seafloor velocity model and their preset interval distances;

[0092] S304. Determine the number of geophones on the initial seafloor velocity model and their preset interval distances;

[0093] S305. Determine the reflected wave of the seismic wave using the shot point and geophone;

[0094] S306. Determine the discrete wave equation of the seismic wave in the seabed velocity model according to the acoustic wave displacement equation and source term fluctuation;

[0095] S307. Use the discrete wave equation and forward modeling of the reflected wave of the seismic wave to generate a synthetic seismic reflection record.

[0096] In this embodiment, the number of shot points in the initial seabed velocity model is 11, the spacing in the OX (horizontal direction) is 499 m, and the value of the Z-axis (vertical direction) coordinate is 12.5 m; the number of geophones is 101, the OX spacing is 50 m, the Z-axis coordinate value is 220 m, and the frequency of the source wave is 10 HZ. The initial seabed velocity model is equivalent to an observation system for observing seismic waves. Using the shot point as the location for exciting the seismic wave, the geophone monitors the seismic wave excited by the shot point to determine the reflected wave of the seismic wave reflected back by the shot point. For the reflected wave of the seismic wave of each shot point, a synthetic seismic reflection record is obtained through forward modeling using the discrete wave equation. The purpose of this embodiment is to obtain a synthetic seismic reflection record through forward modeling, and use the initial seabed velocity model and the synthetic seismic reflection record to form a training set for training the seabed full-waveform depth neural network model.

[0097] Specifically, by determining that the acoustic wave displacement equation of the initial seabed velocity model satisfies:

[0098] u tt (x,z,t)-c 2 (x,z)Δ(u(x,z,t))=c 2 (x,z)f(x,z,t),

[0099] where u(x,z,t) represents the displacement of the acoustic wave, f(x,z,t) represents the external drive, c(x,z) represents the seismic wave propagation velocity, x represents the horizontal direction, and z represents the vertical direction;

[0100] Determine that the source term fluctuation generated by the source wave is:

[0101]

[0102] where f0 represents the peak frequency of the source wave;

[0103] Based on the central finite difference, the standard second-order explicit discretized wave equation is expressed as:

[0104]

[0105] where u i,j,k =u(x i ,z j ,tk ), (x i , z j ) where \(i = 0, \ldots, n_x\), \(j = 0, \ldots, n_z\), \(n_x=(x F -x I ) / \Delta x , n_z=(z F -z I ) / \Delta z , t k = k\Delta t\), \(k = 0, 1, \ldots, m\), \(m = t F / m

[0106] Optionally, the seismic source wave in the embodiments of the present application can be a Ricker wavelet.

[0107] S103. Determine the velocity map of the initial seabed velocity model and the corresponding at least one synthetic seismic reflection record.

[0108] In this embodiment, by making cycles according to the above steps S101 and S102, the 1 underground velocity model and the synthetic seismic reflection records corresponding to at least one shot point position are all represented as numpy matrix vectors, as the input data for training the preset seabed full waveform depth neural network model. Among them, the vector dimension corresponding to the velocity map of the initial seabed velocity model used by the seabed full waveform depth neural network model designed in the present application is \((1\times152\times500)\), and the vector dimension of the synthetic seismic reflection record is \((5\times1001\times101)\).

[0109] It can be understood that the velocity map of the initial seabed velocity model and the corresponding at least one synthetic seismic reflection record constitute a data set, but the specific corresponding relationship between the velocity map and the corresponding at least one synthetic seismic reflection record is not clear. That is to say, the synthetic seismic reflection record can be forward modeled through the velocity map, but the corresponding velocity map, that is, the velocity of the underground formation, cannot be directly obtained through the synthetic seismic reflection record. Therefore, in the embodiments of the present application, the velocity map obtained from 1 initial seabed velocity model and the synthetic seismic reflection records at 5 different shot point positions are all represented as numpy matrix vectors, as the input data of the preset seabed full waveform depth neural network model to train the weights of the seabed full waveform depth neural network model, and the corresponding relationship between the velocity map and the synthetic seismic reflection records at different shot point positions is obtained.

[0110] Finally, the data set is divided according to the ratio of the training set to the test set of 8:2.

[0111] S104. Input the velocity map and the corresponding at least one synthetic seismic reflection record into the preset seabed full waveform depth neural network model for training, determine the mapping relationship between the velocity map and at least one of the synthetic seismic records, and obtain the full waveform inversion result based on the trained seabed full waveform depth neural network model.

[0112] In this embodiment, one velocity map and five synthetic seismic reflection records included in the training set obtained from the above embodiment are input into a preset full waveform seafloor depth neural network model for training to obtain the mapping relationship between the velocity map and at least one seismic record. Through the mapping relationship, by inputting the corresponding synthetic seismic reflection record into the trained full waveform seafloor depth neural network model, the corresponding velocity map, that is, the velocity of the underground formation, can be obtained.

[0113] The preset full waveform seafloor depth neural network model includes: an encoder and a decoder. The encoder includes a convolutional layer and a pooling layer, and the decoder includes a plurality of deconvolutional layers.

[0114] Specifically, referring to Figure 4 as shown, step S104 includes:

[0115] S401. Extract high-dimensional features of at least one synthetic seismic reflection record using the convolutional layer and the pooling layer and compress the high-dimensional features into a single low-dimensional vector, and determine the first mapping relationship between at least one synthetic seismic reflection record and the single low-dimensional vector;

[0116] S402. Perform upsampling in sequence using a plurality of deconvolutional layers, deconvolve the single low-dimensional vector, increase the dimension of the single low-dimensional vector, and determine the second mapping relationship between the single low-dimensional vector and the velocity map;

[0117] According to the first mapping relationship and the second mapping relationship, determine the mapping relationship between the velocity map and at least one synthetic seismic reflection record.

[0118] During the process of training the full waveform seafloor depth neural network model, the convolutional layer and the pooling layer extract the high-dimensional features inherent in the input at least one synthetic seismic reflection record and convolve and compress them into a single low-dimensional vector, and learn the first mapping relationship between at least one synthetic seismic reflection record and the single low-dimensional vector during the convolution process, that is, the mapping relationship between the high-dimensional vector of at least one synthetic seismic record and the single low-dimensional vector. In addition, the velocity map is also input into the convolutional layer and the pooling layer. The single low-dimensional vector is deconvolved using the deconvolutional layer to increase the dimension of the single low-dimensional vector to generate a high-dimensional vector, calculate the loss value between the high-dimensional vector and the velocity map, and adjust the parameters of the convolutional kernel using the backpropagation of the deconvolutional layer according to the loss value until the loss value between the high-dimensional vector obtained by deconvolution and the velocity map is within a preset range, and obtain the second mapping relationship between the single low-dimensional vector and the velocity map, completing the training of the full waveform seafloor depth neural network model. Further, through the first mapping relationship and the second mapping relationship in the training process of the full waveform seafloor depth neural network model, the mapping relationship between the velocity map and at least one synthetic seismic reflection record is also determined during the training process.

[0119] The convolutional layer and the pooling layer are represented as:

[0120] x (l+1) = ReLU(BN(Conv(x (t) ))),

[0121] Conv(x) (i,j) = ∑ m ∑ n ∑ c K m,n,c ·x (s-1)×i+m,(s-1)×j+n,c ,

[0122] MaxPooling(x) (i,j,c) = max m,n x(s - 1)×i + m,(s - 1)×j + n,c,

[0123]

[0124] where x is the independent variable of each function, Conv(x) represents the convolution function, ReLU(x) represents the activation function, MaxPooling(x) represents the max pooling layer, BN(x) represents the batch normalization function, K represents the convolution kernel of m*n, c represents the dimension, s represents the stride between the sliding positions of the convolution kernel, γ and β represent two trainable parameters, μ B and respectively represent the mean and variance calculated for all ownerships using the same feature map in a mini - batch, ∈ represents a parameter used to increase numerical stability, and i, j, m, n respectively represent the spatial coordinates.

[0125] The decoder increases the dimension of a single low - dimensional vector through a transposed convolutional layer. In each transposed convolutional layer, a 4*4 transposed convolution with a stride of 2 is performed on the single low - dimensional vector to increase the resolution, and then a 3*3 regular convolution kernel is used to optimize the upsampled single low - dimensional vector, resulting in 5 channels at each position on the high - dimensional vector. From the above embodiments, it can be seen that the training set input to the encoder includes 1 velocity map and 5 corresponding synthetic seismic reflection records. Therefore, there are 5 channels at each position on the high - dimensional vector obtained by transposed convolution. Finally, a 1*1*5*1 convolution kernel slides on the high - dimensional vector to regress the velocity value for each position, and the convolution parameters are updated through backpropagation to obtain the mapping relationship between the velocity map and the 5 synthetic seismic reflection records on the high - dimensional vector.

[0126] The transposed convolutional layer is represented as:

[0127] y (l+1) = ReLU(BN(Deconv(y l ))),

[0128] Deconv(y)(i,j) = ∑ m ∑ n ∑ c K m,n,c ·y (s-1)×i+m,(s-1)×j+n,c ,

[0129] Among them, Deconv(y) represents the deconvolution function, and K represents the deconvolution kernel of m*n.

[0130] Through the above embodiments, the trained seabed full-waveform depth neural network model is obtained. The trained seabed full-waveform depth neural network model trains and determines the mapping relationship between the velocity map and multiple synthetic reflection seismic records, determines the weights of the trained seabed full-waveform depth neural network model, inputs the seismic reflection record into the trained seabed full-waveform depth neural network model, and directly the corresponding velocity map can be inversely obtained, that is, the velocity of the underground formation, which speeds up the inverse inference time and, without the need for a large amount of training data, improves the potential for solving the full-waveform inversion problem of actual data.

[0131] Optionally, the encoder in this embodiment includes 16 convolutional layers and 6 pooling layers, and the decoder includes 12 deconvolutional layers. Refer to Figure 7 As shown, the cuboid vectors in the figure are all the vector dimensions after convolution or deconvolution. It should be noted that every time convolution or deconvolution is performed, the dimension of the vector will change. Therefore, each cuboid vector in the picture represents one operation (convolution, pooling, deconvolution).

[0132] Figure 7 The left half in

[0133] Figure 7 The right half in

[0134] This application does not require a large number of iterative optimizations, reduces the calculation time, speeds up the inference speed, and significantly improves the accuracy of the underground formation velocity inversion.

[0135] In some embodiments, the seabed full-waveform inversion method of this embodiment of the application further includes:

[0136] Determine the mean absolute error function and the mean square error function of the seabed full-waveform depth neural network model during the training process;

[0137] Using the mean absolute error function and the mean square error function, determine the loss value for updating the parameters during the backpropagation of the subsea full waveform deep neural network model.

[0138] The mean absolute error function is expressed as:

[0139]

[0140] The mean square error function is expressed as:

[0141]

[0142] The loss value is expressed as:

[0143]

[0144] where y = {y1, …, y n} represents the actual subsurface formation velocity, z = {z1, …, z n} represents the predicted inversion result, n represents the number of all spatial positions in the subsea full waveform deep neural network model during the training process, and loss represents the loss value for updating the parameters during the backpropagation of the subsea full waveform deep neural network model.

[0145] It is clear from the above embodiments that the data set is divided into a training set and a test set. During the training process of the subsea full waveform deep neural network model, by inputting the synthetic seismic reflection records in the training set into the subsea full waveform deep neural network model during the training process, through the velocity map output by it, that is, the predicted subsurface formation velocity and the actual subsurface formation velocity included in the training set, using the mean absolute error function and the mean square error function, determine the loss value for updating the parameters during the backpropagation of the subsea full waveform deep neural network model, and use the loss value to adjust the parameters of the subsea full waveform deep neural network model to train the subsea full waveform deep neural network model to a better effect.

[0146] In some alternative embodiments, the above mean absolute error function and mean square error function can also be used to evaluate the trained subsea full waveform deep neural network model. By inputting the synthetic seismic reflection records in the test set into the trained subsea full waveform deep neural network model, through the velocity map output by it, that is, the predicted subsurface formation velocity and the actual subsurface formation velocity included in the test set, use the mean absolute error function and the mean square error function for evaluation. Refer to Figure 5 shown, which shows the comparison between the inversion result and the actual velocity value, where the inversion result relatively accurately reflects the actual velocity value, and the velocity for obtaining the inversion result is faster than the traditional inversion velocity.

[0147] In some alternative embodiments, the learning rate (lr) for training the subsea full-waveform depth neural network model of the present application is set to 0.001, the batch size is 128, and the number of training epochs is 200, as follows Figure 6 As shown, for the training and testing effects on the 2000 dataset, the loss can converge to 0.01 when the number of training epochs approaches 200. Therefore, the subsea full-waveform depth neural network model of the present application can be effective on a small dataset, greatly saving the training time and resource occupation. On the dataset used for testing the subsea full-waveform depth neural network model, the mean absolute error reaches the order of magnitude of 10 -2 and the mean square error reaches the order of magnitude of 10 -3 , and its inversion function is greatly improved compared with traditional inversion methods.

[0148] It should be noted that the method of the embodiments of the present application can be executed by a single device, such as a computer or a server, etc. The method of this embodiment can also be applied to a distributed scenario and completed by the cooperation of multiple devices. In such a distributed scenario, one of the multiple devices can only execute one or more steps of the method of the embodiments of the present application, and these multiple devices will interact with each other to complete the described method.

[0149] It should be noted that some embodiments of the present application have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order than in the above embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0150] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a subsea full-waveform inversion device.

[0151] Referring to Figure 8 , the subsea full-waveform inversion device includes:

[0152] An acquisition module 501, configured to acquire an initial subsea velocity model;

[0153] A synthetic seismic reflection record module 502, configured to determine the discrete wave equation of seismic waves under the initial subsea velocity model, and generate a synthetic seismic reflection record by forward modeling using the discrete wave equation;

[0154] A determination module 503, configured to determine the velocity map of the initial subsea velocity model and the corresponding at least one synthetic seismic reflection record;

[0155] An inversion module 504 is configured to input a velocity map and at least one corresponding synthetic seismic reflection record into a preset full waveform depth neural network model of the seabed for training, determine the mapping relationship between the velocity map and the at least one synthetic seismic record, and obtain a full waveform inversion result based on the trained full waveform depth neural network model of the seabed.

[0156] In some embodiments, the obtaining module 501 is specifically configured to,

[0157] Obtain a stratification model of the seabed formation;

[0158] Determine the minimum seabed formation velocity and the maximum seabed formation velocity;

[0159] Determine the slope of the stratification model;

[0160] According to the minimum seabed formation velocity, the maximum seabed formation velocity, and the slope of the stratification model, perform an elastic transformation and distortion on the stratification model to obtain an initial seabed velocity model.

[0161] In some embodiments, the synthetic seismic reflection record module 502 is specifically configured to determine an acoustic wave displacement equation in the initial seabed velocity model;

[0162] Determine the source term fluctuation generated by the seismic source wave;

[0163] Determine the number of shot points in the initial seabed velocity model and their preset interval distances;

[0164] Determine the number of geophone points in the initial seabed velocity model and their preset interval distances;

[0165] Use the shot points and geophone points to determine a seismic reflection record;

[0166] According to the acoustic wave displacement equation and the source term fluctuation, determine the discrete wave equation of seismic waves in the seabed velocity model;

[0167] Use the discrete wave equation and the seismic reflection record to perform forward simulation to generate a synthetic seismic reflection record.

[0168] In some embodiments, the acoustic wave displacement equation is expressed as:

[0169] u tt (x,z,t)-c 2 (x,z)Δ(u(x,z,t))=c 2 (x,z)f(x,z,t),

[0170] where u(x,z,t) represents the displacement of the acoustic wave, f(x,z,t) represents the external drive, c(x,z) represents the seismic wave propagation velocity, x represents the horizontal direction, and z represents the vertical direction;

[0171] The source term fluctuation is expressed as:

[0172]

[0173] where f0 represents the peak frequency of the source wave;

[0174] The discrete wave equation is expressed as:

[0175]

[0176] where u i,j,k = u(x i , z j , t k ), (x i , z j ) where i = 0, …, nx, j = 0, …, nz, nx = (x F - x I ) / Δ x , nz = (z F - z I ) / Δ z , t k = kΔt, k = 0, 1, … m, m = t F / m.

[0177] In some embodiments, the inversion module 504 is specifically configured to

[0178] extract high-dimensional features of at least one of the synthetic seismic reflection records by using the convolutional layer and the pooling layer and compress the high-dimensional features into a single low-dimensional vector, and determine a first mapping relationship between at least one of the synthetic seismic reflection records and the single low-dimensional vector;

[0179] perform upsampling sequentially by using a plurality of the transposed convolutional layers, perform transposed convolution on the single low-dimensional vector to increase the dimension of the single low-dimensional vector, and determine a second mapping relationship between the single low-dimensional vector and the high-dimensional velocity map;

[0180] determine a mapping relationship between the velocity map and at least one of the synthetic seismic reflection records according to the first mapping relationship and the second mapping relationship.

[0181] In some embodiments, the convolutional layer and the pooling layer are expressed as:

[0182] x (l+1) = ReLU(BN(Conv(x (t) ))),

[0183] Conv(x) (i,j) = ∑ m ∑ n ∑c K m,n,c ·x (s-1)×i+m,(s-1)×j+n,c ,

[0184] MaxPooling(x) (i,j,c) =max m,n x(s - 1)×i + m,(s - 1)×j + n,c,

[0185]

[0186] where x is the independent variable of each function, Conv(x) represents the convolution function, ReLU(x) represents the activation function, MaxPooling(x) represents the max pooling layer, BN(x) represents the batch normalization function, K represents the convolution kernel of m*n, c represents the dimension, s represents the stride between the sliding positions of the convolution kernel, γ and β represent two trainable parameters, μ B and respectively represent the mean and variance calculated from all the calculations using the same feature map in a mini - batch, ∈ represents a parameter used to increase numerical stability, and i, j, m, n represent spatial coordinates respectively;

[0187] The transposed convolution layer is expressed as:

[0188] y (l+1) =ReLU(BN(Deconv(y l )))

[0189] Deconv(y) (i,j) =∑ m ∑ n ∑ c K m,n,c ·y (s-1)×i+m,(s-1)×j+n,c ,

[0190] where Deconv(y) represents the transposed convolution function, and K represents the transposed convolution kernel of m*n.

[0191] In some embodiments, it further includes an evaluation module (not shown in the figure), and the evaluation module is used to

[0192] determine the mean absolute error function and the mean squared error function of the sub - sea full - waveform depth neural network model during the training process;

[0193] use the mean absolute error function and the mean squared error function to determine the loss value for updating the parameters by backpropagation of the sub - sea full - waveform depth neural network model during the training process;

[0194] The mean absolute error function is expressed as:

[0195]

[0196] The mean square error function is expressed as:

[0197]

[0198] The loss value is expressed as:

[0199]

[0200] where y = {y1, …, y n} represents the actual subsurface formation velocity, z = {z1, …, z n} represents the predicted inversion result, n represents the number of all spatial positions in the full waveform depth neural network model of the seabed during the training process, and loss represents the loss value for updating the parameters by backpropagation in the full waveform depth neural network model of the seabed during the training process.

[0201] For the convenience of description, when describing the above device, it is divided into various modules according to functions for separate description. Of course, when implementing the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0202] The device of the above embodiment is used to implement the corresponding full waveform inversion method of the seabed in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated herein.

[0203] Based on the same technical concept, corresponding to the method of any of the above embodiments, the present application further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the full waveform inversion method of the seabed described in any of the above embodiments.

[0204] Figure 9 FIG. shows a more specific schematic diagram of the hardware structure of the electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. Among them, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other inside the device through the bus 1050.

[0205] The processor 1010 may be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0206] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 1020 can store the operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1020 and called and executed by the processor 1010.

[0207] The input / output interface 1030 is used to connect to the input / output module to achieve information input and output. The input / output module can be configured as a component in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Among them, the input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator light, etc.

[0208] The communication interface 1040 is used to connect to the communication module (not shown in the figure) to achieve communication interaction between this device and other devices. Among them, the communication module can achieve communication through wired means (such as USB, network cable, etc.) or through wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0209] The bus 1050 includes a path for transmitting information between various components of the device (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0210] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device may also only include the components necessary to implement the solutions of the embodiments of this specification, and do not have to include all the components shown in the figure.

[0211] The electronic device in the above embodiment is used to implement the corresponding full waveform inversion method for the seabed in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0212] Based on the same technical concept, corresponding to the method in any of the above embodiments, the present application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the full waveform inversion method for the seabed as described in any of the above embodiments.

[0213] The computer-readable medium of this embodiment includes both permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0214] The computer instructions stored in the storage medium of the above embodiment are used to cause the computer to execute the full waveform inversion method for the seabed as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0215] Those of ordinary skill in the art should understand that: the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present application (including the claims) is limited to these examples; under the concept of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of brevity.

[0216] In addition, for simplicity of explanation and discussion, and in order not to make the embodiments of the present application difficult to understand, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. In addition, the devices may be shown in block diagram form to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present application will be implemented (i.e., these details should be completely within the understanding of those skilled in the art). In the case where specific details (such as circuits) are set forth to describe the exemplary embodiments of the present application, it will be apparent to those skilled in the art that the embodiments of the present application can be implemented without these specific details or with variations of these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0217] Although the present application has been described in connection with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0218] Embodiments of the present application are intended to cover all such alternatives, modifications, and variations that fall within the broad scope of the appended claims. Accordingly, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the embodiments of the present application shall be included within the protection scope of the present application.

Claims

1. A seafloor full waveform inversion method, characterized in that: include: Obtain the initial seabed velocity model; Determine a discrete wave equation of seismic waves under the seafloor initial velocity model, and generate synthetic seismic reflection records by forward modeling using the discrete wave equation; Determine a velocity map of the initial seafloor velocity model and a corresponding at least one synthetic seismic reflection record; Inputting the velocity map and the corresponding at least one synthetic seismic reflection record into a preset seafloor full waveform deep neural network model for training, determining a mapping relationship between the velocity map and the at least one synthetic seismic record, and determining a full waveform inversion result based on the trained seafloor full waveform deep neural network model; The preset seabed full-waveform deep neural network model includes: an encoder and a decoder, the encoder includes a convolution layer and a pooling layer, and the decoder includes a plurality of deconvolution layers; Inputting the velocity map and the corresponding synthetic seismic reflection record into a preset seafloor full waveform deep neural network model for training, determining a mapping relationship between the velocity map and the synthetic seismic record, and determining a full waveform inversion result based on the trained seafloor full waveform deep neural network model, including: Extracting high-dimensional features of at least one of the synthetic seismic reflection records by using the convolution layer and the pooling layer and compressing the high-dimensional features into a single low-dimensional vector, and determining a first mapping relationship between at least one of the synthetic seismic reflection records and the single low-dimensional vector; Using the plurality of deconvolution layers to sequentially perform upsampling, deconvolute the single low-dimensional vector, increase the dimension of the single low-dimensional vector, and determine a second mapping relationship between the single low-dimensional vector and the high-dimensional velocity map; A mapping relationship between the velocity map and at least one of the synthetic seismic reflection records is determined according to the first mapping relationship and the second mapping relationship.

2. The method according to claim 1, characterized in that Obtain the initial seafloor velocity model, including: Obtaining a layered model of the seafloor strata; Determine the minimum seafloor formation velocity and the maximum seafloor formation velocity; determining a slope of the layered model; According to the minimum seabed formation velocity and the maximum seabed formation velocity and the slope of the layered model, the layered model is elastically transformed and distorted to obtain an initial seabed velocity model.

3. The method according to claim 1, characterized in that Determining a discrete wave equation of seismic waves under the seafloor initial velocity model, and using the discrete wave equation to generate synthetic seismic reflection records by forward simulation, including: Determine the acoustic wave displacement equation under the seabed initial velocity model; Determine the source term fluctuations generated by the source wave; Determining the number of shot points and preset interval distances of the seabed initial velocity model; Determining the number of detection points and preset interval distances of the seabed initial velocity model; Determining the reflected waves of the seismic waves using the shot points and the detection points; Determining a discrete wave equation of seismic waves in a seafloor velocity model according to the acoustic wave displacement equation and the source term fluctuation; The discrete wave equation and the forward modeling of the seismic wave reflection wave are used to generate a synthetic seismic reflection record.

4. The method according to claim 3, characterized in that: The acoustic wave displacement equation is expressed as: u tt (x,z,t)-c 2 (x,z)Δ(u(x,z,t))=c 2 (x,z)f(x,z,t), Among them, u(x,z,t) represents the displacement of the sound wave, f(x,z,t) represents the external drive, c(x,z) represents the propagation velocity of the seismic wave, x represents the horizontal direction, z represents the vertical direction, t represents the wave field propagation time, Δ is the Laplace operator, which represents the sum of the second-order partial derivatives of the displacement with respect to the spatial variables x and z, and u tt (x,z,t) represents the second-order partial derivative of displacement with respect to time; The source term fluctuation is expressed as: Where f0 represents the peak frequency of the source wave; The discrete wave equation is expressed as: Among them, u i,j,k =u(x i ,z j ,t k ), (x i ,z j ,t k ), i and j represent the grid coordinate index after the spatial position coordinates x and z are discretized, k represents the time layer index after the time t is discretized, Δx represents the grid space step in the x direction, Δt represents the time step, and c i,j,k represents the wave speed that varies with position and time, f i,j,k Represents external drive that varies with position and time.

5. The method according to claim 1, characterized in that The convolutional layer and the pooling layer are expressed as: h (l+1) =ReLU(BN(Conv(h (l) )))), Among them, h is the independent variable of each function, that is, the feature map vector of the convolution operation, Conv represents the convolution operation, ReLU represents the activation function, MaxPooling represents the maximum pooling layer, BN represents the batch normalization operation, K represents the m*n convolution kernel, s represents the stride between the sliding positions of the convolution kernel, γ and β represent two trainable parameters, μ B and denotes the mean and variance of all calculations using the same feature map on a small batch, ∈ denotes a parameter used to increase numerical stability, h denotes the feature map vector of the convolution operation, and h (l+1) Represents the output feature map vector of the l+1th layer of the convolution operation, l i and l j They represent the spatial index coordinates of the feature map, m and n represent the spatial length and width of the convolution kernel, o represents the number of channels, and K m,n,o represents the weight of the convolution kernel, Represents the input feature area where the convolution kernel slides, ∑ represents the summation operation, and max represents the local maximum value within the pooling window; The deconvolution layer is expressed as: and (l+1) =ReLU(BN(Deconv(and l ))), Among them, y represents the feature map vector of the deconvolution operation, y (l+1) represents the output feature map vector of the l+1th layer of the deconvolution operation, Represents the input feature area where the deconvolution kernel slides, and Deconv represents the deconvolution operation.

6. The method according to claim 1, characterized in that Also includes: Determine the mean absolute error function and mean square error function of the seafloor full waveform deep neural network model during the training process; Using the mean absolute error function and the mean square error function, determining the loss value of the back propagation update parameters of the seabed full waveform deep neural network model during the training process; The mean absolute error function is expressed as: The mean square error function is expressed as: The loss value is expressed as: Where V = {V1,…,V n } represents the actual underground formation velocity, v={v1,…,v n } represents the predicted inversion result, n represents the number of all spatial positions in the seabed full waveform deep neural network model during the training process, loss represents the loss value of the back propagation update parameters of the seabed full waveform deep neural network model during the training process, and i represents the position index of the velocity model.

7. A seabed full waveform inversion device, characterized in that: include: An acquisition module is used to obtain an initial seabed velocity model; A synthetic seismic reflection record module is used to determine a discrete wave equation of seismic waves under the seafloor initial velocity model, and generate a synthetic seismic reflection record by forward simulation using the discrete wave equation; A determination module, used to determine a velocity map of the initial seafloor velocity model and a corresponding synthetic seismic reflection record; An inversion module is used to input the velocity map and the corresponding synthetic seismic reflection record into a preset seafloor full waveform deep neural network model for training, determine the mapping relationship between the velocity map and the synthetic seismic record, and obtain a full waveform inversion result based on the trained seafloor full waveform deep neural network model; The preset seabed full-waveform deep neural network model includes: an encoder and a decoder, the encoder includes a convolution layer and a pooling layer, and the decoder includes a plurality of deconvolution layers; Inputting the velocity map and the corresponding synthetic seismic reflection record into a preset seafloor full waveform deep neural network model for training, determining a mapping relationship between the velocity map and the synthetic seismic record, and determining a full waveform inversion result based on the trained seafloor full waveform deep neural network model, including: Extracting high-dimensional features of at least one of the synthetic seismic reflection records by using the convolution layer and the pooling layer and compressing the high-dimensional features into a single low-dimensional vector, and determining a first mapping relationship between at least one of the synthetic seismic reflection records and the single low-dimensional vector; Using the plurality of deconvolution layers to sequentially perform upsampling, deconvolute the single low-dimensional vector, increase the dimension of the single low-dimensional vector, and determine a second mapping relationship between the single low-dimensional vector and the high-dimensional velocity map; A mapping relationship between the velocity map and at least one of the synthetic seismic reflection records is determined according to the first mapping relationship and the second mapping relationship.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 6.

Images