Earthquake velocity modeling method and device based on deep learning, storage medium and processor

Abstract

The invention discloses a seismic velocity modeling method and device based on deep learning, a storage medium and a processor. According to the scheme, firstly, seismic data are acquired; then, the seismic data is preprocessed; the preprocessing comprises de-noising standardization operation and down-sampling operation; and finally, inputting the pre-processed seismic data into a deep neural network, performing seismic velocity prediction by the deep neural network according to the pre-processed seismic data, and outputting a seismic velocity model. The deep neural network is obtained based on sample data training. Different from an earthquake velocity model in the prior art that low-frequency signal features are not fully captured and are easily interfered by noise, and the accuracy of the earthquake velocity model is influenced, the accuracy of the earthquake velocity model is improved by introducing a frequency enhancement reasoning module and a frequency fusion layer into a deep neural network. Compared with the prior art, the method has obvious advantages.

Description

Technical Field

[0001] The present application relates to the field of seismic exploration technology, and in particular to a deep learning-based seismic velocity modeling method, device, storage medium and processor. Background Art

[0002] Seismic velocity modeling is an important field in geophysics. It mainly involves using mathematical methods to describe the velocity distribution of seismic waves propagating in the crust and mantle. This type of modeling is of great significance for understanding the internal structure of the earth, predicting earthquake disasters and resource exploration.

[0003] In the field of seismic exploration, accurate seismic velocity models are crucial to revealing underground structures, and they play a core role in resource exploration and development. Existing seismic velocity modeling technologies do not adequately capture the characteristics of medium and low-frequency signals, which affects the accuracy of seismic velocity models. For example, full waveform inversion and deep learning methods focus on extracting high-frequency features during the modeling process, and often do not adequately capture low-frequency signals. However, these low-frequency signals are crucial for depicting large-scale underground structures and will affect the accuracy of the model.

[0004] In addition, in practical applications, seismic data is often interfered by noise, and current technologies are often unable to resist the impact of high noise. In noisy datasets, existing deep learning-based models such as InversionNet and MFNSR have poor robustness, especially because noise can easily mask key low-frequency signals, further affecting the prediction accuracy of seismic velocity models.

[0005] Therefore, how to improve the accuracy of seismic velocity models is an urgent problem to be solved. Summary of the invention

[0006] Based on the above problems, the present application provides a deep learning-based seismic velocity modeling method, device, storage medium and processor, the purpose of which is to improve the accuracy of the seismic velocity model.

[0007] The embodiments of the present application disclose the following technical solutions:

[0008] The first aspect of the present application provides a seismic velocity modeling method based on deep learning, the method comprising:

[0009] Acquire seismic data; the seismic data is real seismic data or simulated seismic data for which seismic velocity modeling is to be performed;

[0010] Preprocessing the seismic data; the preprocessing includes denoising and normalization operations and downsampling operations;

[0011] The preprocessed seismic data is input into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data and outputs a seismic velocity model; the deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output the enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector; the sample data includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

[0012] In a possible implementation, the method for obtaining the simulated seismic data includes:

[0013] In response to the model parameter definition operation, determining the parameters of the seismic velocity model for generating the simulated seismic data; the parameters of the seismic velocity model for generating the simulated seismic data include P-wave velocity, S-wave velocity and density;

[0014] In response to a selection operation on a numerical simulation method, determining a selected numerical simulation method;

[0015] In response to the setting operation of boundary conditions and initial conditions of the seismic velocity model for generating the simulated seismic data, the set boundary conditions and initial conditions are determined; the boundary conditions include a free surface and an absorbing boundary; the initial conditions include the location, type and strength of the earthquake source;

[0016] Based on the seismic velocity model for generating the simulated seismic data, the numerical simulation method and the set boundary conditions and initial conditions, a simulation operation is performed to calculate the propagation process of seismic waves and generate the simulated seismic data.

[0017] In a possible implementation, the frequency enhancement inference module dynamically weights and enhances the multi-band feature vectors extracted by the convolution layer and the pooling layer, including:

[0018] The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is:

[0019]

[0020] Among them, F L (x) is the low frequency band feature vector, F H(x) is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices of the low frequency band and the high frequency band, respectively, which are learned through the deep neural network;

[0021] Calculate the attention matrix score of the low frequency band and the attention matrix score of the high frequency band through the self-attention mechanism, and use the Softmax function to normalize the attention matrix score of the low frequency band and the attention matrix score of the high frequency band respectively to obtain the attention weight of the low frequency band and the attention weight of the high frequency band;

[0022] Adding the product of the attention weight of the low frequency band and the weighted low frequency band feature vector and the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-frequency band feature vector with single-head attention enhancement;

[0023] A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

[0024] In a possible implementation manner, the frequency fusion layer dynamically fuses the enhanced feature vectors, including:

[0025] The frequency fusion layer adds the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and generates a fusion feature vector after activation through a nonlinear activation function;

[0026] The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively:

[0027]

[0028] Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

[0029] In a possible implementation, a method of training a deep neural network based on sample data includes:

[0030] Acquire sample data; the seismic data samples in the sample data include real seismic data and / or simulated seismic data for training a deep neural network, and the sample data is divided into a training set and a test set according to a preset ratio;

[0031] Preprocessing the sample data; the preprocessing includes: denoising the seismic data samples contained in the sample data; normalizing the seismic data samples contained in the denoised sample data to the same numerical range by a standardization operation; downsampling the seismic data samples contained in the sample data in a preset form;

[0032] Based on the preprocessed sample data, training parameters of the deep neural network are determined through training and testing.

[0033] A second aspect of the present application provides a seismic velocity modeling device based on deep learning, the device comprising:

[0034] A seismic data acquisition module, used to acquire seismic data; the seismic data is real seismic data or simulated seismic data to be subjected to seismic velocity modeling;

[0035] A preprocessing module, used for preprocessing the seismic data; the preprocessing includes a denoising and normalization operation and a downsampling operation;

[0036] A seismic velocity modeling module is used to input the preprocessed seismic data into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data and outputs a seismic velocity model; the deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output an enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector; the sample data includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

[0037] In a possible implementation, the frequency enhancement reasoning module is specifically used to:

[0038] The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is:

[0039]

[0040] Among them, F L (x) is the low frequency band feature vector, F H (x) is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices of the low frequency band and the high frequency band, respectively, which are learned through the deep neural network;

[0041] Calculate the attention score matrix of the low frequency band and the attention score matrix of the high frequency band through the self-attention mechanism, and use the Softmax function to normalize the attention score matrix of the low frequency band and the attention score matrix of the high frequency band respectively to obtain the attention weight of the low frequency band and the attention weight of the high frequency band;

[0042] Adding the product of the attention weight of the low frequency band and the weighted low frequency band feature vector and the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-frequency band feature vector with single-head attention enhancement;

[0043] A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

[0044] In a possible implementation manner, the frequency fusion layer is specifically used to:

[0045] Adding the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and activating them through a nonlinear activation function to generate a fusion feature vector;

[0046] The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively:

[0047]

[0048] Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

[0049] A third aspect of the present application provides a computer-readable storage medium, in which a computer program is stored. When the program is executed by a processor, a deep learning-based seismic velocity modeling method as provided in any implementation of the first aspect is implemented.

[0050] A fourth aspect of the present application provides a processor for running a computer program, which, when running, executes a deep learning-based seismic velocity modeling method as provided in any implementation of the first aspect.

[0051] Compared with the prior art, this application has the following beneficial effects:

[0052] The seismic velocity model is constructed based on a deep neural network, especially by introducing a frequency enhancement reasoning module and a frequency fusion layer in the deep neural network, the accuracy of the seismic velocity model is improved. Among them, the frequency enhancement reasoning module performs dynamic weighted processing on high-frequency and low-frequency signals, especially strengthens the low-frequency capture capability, ensures the accurate modeling of large-scale underground structures by the model, and thus improves the accuracy of the model as a whole; the frequency fusion layer effectively reduces the interference of noise in the process of feature fusion and extraction of multi-band signals, and improves the anti-noise capability of the model, especially when processing low-frequency signals. It shows excellent robustness. Therefore, the seismic velocity modeling method based on deep learning proposed in this application enables the constructed seismic velocity model to enhance the capture capability of low-frequency signals in seismic data, and maintains a high prediction accuracy when processing seismic data with large noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0054] Figure 1 A flowchart of a deep learning-based seismic velocity modeling method provided in an embodiment of the present application;

[0055] Figure 2 A flowchart of a method for training a deep neural network based on sample data provided in an embodiment of the present application;

[0056] Figure 3 A flowchart of another deep learning-based seismic velocity modeling method provided in an embodiment of the present application;

[0057] Figure 4 A flowchart of another deep learning-based seismic velocity modeling method provided in an embodiment of the present application;

[0058] Figure 5 A flowchart of a deep learning-based seismic velocity modeling method with added noise provided in an embodiment of the present application;

[0059] Figure 6 Part of the SEG Salt velocity model diagram provided for the embodiments of the present application;

[0060] Figure 7 A partial simulation earthquake velocity model diagram provided in an embodiment of the present application;

[0061] Figure 8 A seismic velocity model diagram obtained by inputting simulated seismic data into a deep neural network provided in an embodiment of the present application;

[0062] Fig. 9 A portion of real seismic data graphs selected from the training set provided in the embodiment of the present application;

[0063] Fig.10 A seismic velocity model diagram obtained by inputting real seismic data into a deep neural network provided in an embodiment of the present application;

[0064] Fig.11 A partial real salt velocity model diagram selected from the test set provided in the embodiment of the present application;

[0065] Fig.12 The seismic data diagram before and after noise addition provided in the embodiment of the present application;

[0066] Fig.13 A seismic velocity model diagram obtained by inputting the noisy seismic data provided in the embodiment of the present application into a deep neural network;

[0067] Fig.14 A comparison chart of the real seismic data before noise addition and the seismic velocity predicted by the deep neural network provided in the embodiment of the present application;

[0068] Fig.15 A comparison chart of the real seismic data after noise addition and the seismic velocity predicted by the deep neural network provided in the embodiment of the present application;

[0069] Fig.16 A schematic structural diagram of a deep learning-based seismic velocity modeling device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0070] As described above, current seismic velocity modeling technology does not adequately capture the characteristics of medium and low-frequency signals. For example, full waveform inversion and deep learning methods focus on the extraction of high-frequency features during the modeling process, and often do not adequately capture low-frequency signals. However, these low-frequency signals are crucial for depicting large-scale underground structures and will affect the accuracy of the model. Moreover, in practical applications, they are also susceptible to interference from noise. In noisy datasets, existing deep learning-based models such as InversionNet and MFNSR have poor robustness, especially because noise can easily mask key low-frequency signals, further affecting the prediction accuracy of the seismic velocity model.

[0071] In view of the above problems, the inventors have proposed a method, device, storage medium and processor for seismic velocity modeling based on deep learning after research. In this scheme, seismic data is first acquired; then the seismic data is preprocessed; the preprocessing includes denoising and normalization operations and downsampling operations; finally, the preprocessed seismic data is input into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data and outputs a seismic velocity model; the deep neural network is obtained after training based on sample data. This scheme improves the accuracy of the seismic velocity model by introducing a frequency enhancement reasoning module and a frequency fusion layer into the deep neural network.

[0072] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0073] See also Figure 1 , which is a flow chart of a deep learning-based seismic velocity modeling method provided in an embodiment of the present application. Figure 1 As shown, the method comprises the following steps:

[0074] S101. Obtain seismic data.

[0075] The seismic data are real seismic data or simulated seismic data for which seismic velocity modeling is to be performed.

[0076] In a possible implementation, real earthquake data is obtained through a seismic network or database, and simulated earthquake data is obtained through numerical simulation methods. Common simulation numerical methods include but are not limited to: finite difference method, finite element method, spectral element method and pseudo-spectral method.

[0077] Whether it is real seismic data or simulated seismic data, there are multiple ways and tools to choose from. Reasonable seismic data acquisition can provide important support for seismic velocity modeling.

[0078] S102: Preprocess the seismic data.

[0079] In the embodiment of the present application, the preprocessing operation performed on the seismic data includes a denoising and normalization operation and a downsampling operation.

[0080] In the embodiment of the present application, denoising is an operation to remove noise from seismic data and retain useful signals. Common denoising methods include filtering and wavelet transform.

[0081] In the embodiment of the present application, standardization is an operation of scaling seismic data to the same scale for subsequent processing and analysis. Common standardization methods include Z-score standardization and Min-Max standardization. In the embodiment of the present application, downsampling is an operation of reducing the number of data points to reduce the amount of data and computing cost. Common downsampling methods include time domain downsampling and frequency domain downsampling.

[0082] In seismic data, signals of different frequencies represent features of different scales. Low-frequency signals are usually related to large-scale changes in geological structures and are used to capture the overall stratigraphic structure. Medium-frequency signals reveal medium-scale geological structures, such as small and medium-sized faults, rock interfaces, storage boundaries, etc. High-frequency signals usually contain relatively subtle geological details, reflecting small-scale geological structures, such as cracks, geological interfaces, etc., and may also contain noise. Exemplarily, the frequency range of high-frequency signals is greater than 10Hz, the frequency range of medium-frequency signals is between 1Hz and 10Hz, and the frequency range of low-frequency signals is less than 1Hz.

[0083] By performing preprocessing operations such as denoising, standardization and downsampling on seismic data, the quality of seismic data and the accuracy of subsequent analysis can be significantly improved, providing a reliable data basis for seismic velocity modeling.

[0084] S103, inputting the preprocessed seismic data into a deep neural network, and the deep neural network performs seismic velocity prediction based on the preprocessed seismic data and outputs a seismic velocity model.

[0085] The deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output the enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector. The frequency enhancement reasoning module and the frequency fusion layer are important improvements in the seismic velocity model structure in the embodiment of the present application.

[0086] The sample data used to train the deep neural network includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

[0087] The working method of the frequency enhancement inference module introduced above is exemplarily described below.

[0088] The frequency enhancement inference module dynamically weights and enhances the multi-band feature vectors extracted by the convolution layer and the pooling layer, including:

[0089] The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is:

[0090]

[0091] Among them, F L (x) is the low frequency band feature vector, F H (x) is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices for the low frequency band and the high frequency band respectively, which are learned through the deep neural network.

[0092] The weight matrix will be adaptively adjusted to enhance the contribution of low-frequency signals and reduce the interference of high-frequency noise in practical applications.

[0093] The attention matrix scores of the low-frequency band and the attention matrix scores of the high-frequency band are calculated through the self-attention mechanism, and the attention matrix scores of the low-frequency band and the attention matrix scores of the high-frequency band are normalized respectively using the Softmax function to obtain the attention weights of the low-frequency band and the attention weights of the high-frequency band.

[0094] The product of the attention weight of the low frequency band and the weighted low frequency band feature vector is added to the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-band feature vector with single-head attention enhancement:

[0095]

[0096] Among them, α L and α H They are the attention weights of the low-frequency band and the high-frequency band, respectively, representing the relative importance of low-frequency signals and high-frequency signals.

[0097] A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

[0098] The frequency enhancement inference module can dynamically capture and process signals in different frequency bands by outputting enhanced feature vectors through weighted processing and self-attention mechanism.

[0099] The working method of the frequency fusion layer introduced above is exemplarily described below.

[0100] The frequency fusion layer dynamically fuses the enhanced feature vectors, including:

[0101] The frequency fusion layer adds the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and generates a fused feature vector after activation through a nonlinear activation function.

[0102] yfusion = ReLU(β L ·F enhanced_L (x)+β H ·F enhanced_H (x))

[0103] The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively:

[0104]

[0105] Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

[0106] The frequency fusion layer further fuses the high-frequency signal and the low-frequency signal to ensure that the signal of each frequency band can be fully expressed and avoid the loss of the low-frequency signal during the convolution process.

[0107] The following is an exemplary description of the method of training a deep neural network based on sample data introduced above.

[0108] Ways to train deep neural networks based on sample data include:

[0109] Obtain sample data; the seismic data samples in the sample data include real seismic data and / or simulated seismic data for deep neural network training, and the sample data is divided into a training set and a test set according to a preset ratio.

[0110] Real earthquake data usually comes from earthquake observation stations, seismic networks, etc., including earthquake waveform data, velocity and amplitude of earthquake waves, etc. Simulated earthquake data is generated by numerical simulation methods, including but not limited to finite difference method, finite element method and spectral element method, etc., and it is necessary to set the geological model, earthquake source parameters and receiver location, etc. during simulation.

[0111] The sample data is preprocessed; the preprocessing includes: denoising the seismic data samples contained in the sample data; normalizing the seismic data samples contained in the denoised sample data to the same numerical range by a standardization operation; and downsampling the seismic data samples contained in the sample data in a preset form.

[0112] Denoising methods include but are not limited to filters, wavelet transforms, and Fourier transforms. Common standardization methods include but are not limited to Z-score standardization and Min-Max standardization. Common downsampling methods include uniform downsampling and non-uniform downsampling. Preprocessing sample data is an important step to ensure the accuracy and efficiency of seismic velocity modeling.

[0113] Based on the preprocessed sample data, training parameters of the deep neural network are determined through training and testing.

[0114] The deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence. The deep neural network is trained using a training set of sample data by defining a loss function and an optimizer, the performance of the deep neural network is evaluated in a test set, and the training parameters of the trained and tested deep neural network are saved.

[0115] The seismic velocity model is constructed based on the deep neural network, especially by introducing the frequency enhancement reasoning module and the frequency fusion layer in the deep neural network, the accuracy of the seismic velocity model is improved. Among them, the frequency enhancement reasoning module performs dynamic weighted processing on high-frequency and low-frequency signals, especially strengthens the low-frequency capture ability, ensures the deep neural network to accurately model large-scale underground structures, and thus improves the accuracy of the seismic velocity model as a whole; the frequency fusion layer effectively reduces the interference of noise in the process of feature fusion and extraction of multi-band signals, improves the noise resistance of the model, and shows excellent robustness in processing low-frequency signals.

[0116] See also Figure 2 , which is a flow chart of a method for training a deep neural network based on sample data provided by an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0117] S201. Prepare sample data.

[0118] Among them, the seismic data samples in the sample data include real seismic data and / or simulated seismic data for training deep neural networks, and the sample data are divided into a training set and a test set according to a preset ratio.

[0119] In this embodiment, the sample data is divided into 80% training set and 20% test set. The training set includes 130 SEG Salt velocity models and 1600 simulated seismic velocity models, each of which corresponds to a seismic data. Figure 6 This is a partial SEG Salt speed model diagram. Figure 7 This is a partial simulation of the seismic velocity model. The Ground Truth in the title is the actual seismic velocity, the ordinate Depth is the depth in kilometers, and the abscissa Position is the position in kilometers. In the figure, you can see that contour lines are used to depict the relationship between different altitudes. The spacing between the contour lines represents the slope. The larger the spacing, the gentler the slope, and the smaller the slope, the more shaky the slope. The test set consists of 10 SEG Salt velocity models and 100 simulated seismic velocity models, which are used to evaluate the generalization ability of seismic velocity models. Each seismic velocity model covers an area of ​​3 km × 2 km and is discretized in 10-meter grids. The velocity values ​​range from 2000 m / s to 4500 m / s, representing the velocity changes in the geological environment.

[0120] S202: Preprocess the sample data.

[0121] The sample data is denoised and high-frequency noise is removed by low-pass filtering. Next, the denoised sample data is standardized to normalize the seismic velocity model and the seismic data to the same numerical range in order to accelerate the training convergence speed of the seismic velocity model. In order to reduce the computational complexity, the simulated seismic data is downsampled, and the downsampled data is finally represented in the form of a 29×301×360 matrix, where 29 is the number of seismic data channels, 301 is the number of receivers, and 360 is the number of time sampling points.

[0122] S203: Based on the preprocessed sample data, determine the training parameters of the deep neural network through training and testing.

[0123] The deep neural network includes: a convolution layer, a pooling layer, a frequency enhancement inference module, a frequency fusion layer and an upsampling layer.

[0124] The convolution layer is used to extract local features in seismic data and capture velocity information at different scales. Through multi-layer convolution operations, deep neural networks can effectively extract low-frequency and medium-frequency signals from seismic data.

[0125] The pooling layer is usually located after the convolutional layer. It compresses the feature map by downsampling in the local area, reducing the spatial dimension of the feature map, thereby reducing the amount of calculation and the number of parameters in the subsequent layers. Maximum pooling and average pooling are two commonly used pooling methods.

[0126] The frequency enhancement inference module interprets and enhances the mixed information of different frequencies contained in the extracted features, and adaptively assigns different weights to the model through weighting functions and attention mechanisms. The specific process is as follows:

[0127] (1) Input features: The features of the input frequency enhancement inference module are the results of processing by the convolution layer and the pooling layer, expressed as:

[0128] F(x)={F1(x),F2(x),…,F n (x)}

[0129] Among them, F(x) is an n-channel feature map, which contains multi-band features of different spatial resolutions. In the above expression, the subscript numbers represent different channels.

[0130] (2) Weighted processing: The weighted function is used to enhance the capture of low-frequency signals and suppress high-frequency noise. The weighted function formula is:

[0131]

[0132] in, is the low frequency band feature vector, is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices for the low frequency band and the high frequency band, respectively, which are learned through the deep neural network. The weight matrix will be adaptively adjusted to enhance the contribution of low frequency signals and reduce the interference of high frequency noise in practical applications.

[0133] (3) Self-attention mechanism: The self-attention mechanism is used to calculate the attention score matrix of the low-frequency band and the attention score matrix of the high-frequency band. The attention score matrix is ​​calculated as follows:

[0134]

[0135] Among them, d k It is a scaling factor to prevent the dot product value from being too large. Query (Q): It is used to represent the query information, asking which other features should be paid attention to for a certain feature in the current input. Key (K): It is used to represent the importance or identification of the feature, and is matched with the Query to determine which feature needs more attention. Both Query (Q) and Key (K) are learnable parameters.

[0136] The attention score matrix of the low frequency band and the attention score matrix of the high frequency band are normalized respectively by using the Softmax function to obtain the attention weight of the low frequency band and the attention weight of the high frequency band:

[0137] α L =Softmax(A L ),α H =Softmax(A H )

[0138] The product of the attention weight of the low frequency band and the weighted low frequency band feature vector is added to the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-band feature vector with single-head attention enhancement:

[0139]

[0140] The multi-band feature vectors enhanced by multiple single-head attentions are connected through a multi-head attention mechanism, and the enhanced feature vector is output after being activated by a nonlinear activation function:

[0141]

[0142] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )·WO

[0143]

[0144] F enhanced (x) = ReLU(F enhanced (x)

[0145] Frequency fusion layer: The product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector is added to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and a fusion feature vector is generated after activation through a nonlinear activation function:

[0146] y fusion =ReLU(β L ·F enhanced_L (x)+β H ·F enhanced_H (x)

[0147] in,

[0148]

[0149] Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

[0150] Upsampling layer: In order to ensure that the spatial resolution of the output features matches the requirements of the subsequent network layers, the frequency fusion layer uses a bilinear interpolation upsampling method, the formula is:

[0151] yupsampled=Bilinear(y fusion , scale_factor = 2)

[0152] Among them, scale_factor=2 means that the size of the feature map is doubled in each dimension.

[0153] Output layer: The final output is the fusion feature vector yfusion, which is optimized using the gradient descent method to minimize the loss function, thereby determining the training parameters of the deep neural network and building a seismic velocity model. The formula is as follows:

[0154]

[0155] Among them, V i represents the i-th data point of the target seismic velocity model, represents the i-th speed value predicted by the neural network, n is the total number of data points, |·| 2 is the Euclidean distance.

[0156] See also Figure 3 , which is a flowchart of another deep learning-based seismic velocity modeling method provided in an embodiment of the present application. This embodiment uses some simulated seismic data for testing. Figure 8 The seismic velocity model obtained by inputting simulated seismic data into a deep neural network is shown. Figure 3 As shown, the method comprises the following steps:

[0157] S301, obtaining simulated earthquake data.

[0158] In this embodiment, the simulated seismic data of the test set consists of 100 simulated seismic velocity models and their corresponding seismic data, covering a spatial area of ​​3 km × 2 km, each seismic velocity model is discretized with a 10-meter grid, and the velocity range is 2000 m / s to 4500 m / s. The staggered grid high-order finite difference method is used to simulate the seismic data, which is based on the solution of the first-order velocity stress equation.

[0159] Source design: The source is a Ricker wavelet with a peak frequency of 25 Hz. The source position is set on the surface of the model and a source point is set every 100 meters.

[0160]

[0161] Among them: v represents the velocity field, σ represents the stress tensor, is the density, λ and μ are Lame constants, describing the elastic properties of the medium.

[0162] Receiver setup: In the simulation area, the receivers are evenly distributed along the surface with an interval of 10 meters, and a total of 301 receivers are used. Each receiver is used to capture the propagation signal of seismic waves, record and generate corresponding seismic data.

[0163] Boundary conditions: To eliminate the impact of non-physical reflections on the simulation process, a 30-meter-thick perfectly matched layer (PML) absorbing boundary condition was used. PML can effectively reduce the interference of reflected waves on the simulation results and ensure that the generated seismic data can reflect the actual underground structure.

[0164] S302: preprocess the simulated seismic data.

[0165] The simulated seismic data was preprocessed, including denoising, normalization, and downsampling. Finally, the downsampled data was input into the deep neural network in the form of a 29×301×360 matrix, where 301 is the number of receivers and 360 is the number of time sampling points.

[0166] S303, inputting the preprocessed simulated seismic data into a deep neural network, and the deep neural network performs seismic velocity prediction based on the preprocessed simulated seismic data, and outputs a seismic velocity model.

[0167] After the preprocessed simulated seismic data is input into the trained deep neural network, the deep neural network dynamically weights the high-frequency, medium-frequency and low-frequency information through the frequency enhancement inference module and the frequency fusion layer to generate an accurate seismic velocity model. Figure 7 The seismic velocity model predicted by the simulated seismic velocity model is presented, indicating that the method of simulating seismic data can effectively capture the morphology, position and velocity changes of the strata. Figure 8 The seismic velocity model diagram obtained by inputting simulated seismic data into a deep neural network is shown, where the two figures above are, from left to right, the actual seismic velocity model of the simulated seismic data and the seismic velocity model predicted by the input deep neural network, and the figure below is the seismic velocity model predicted by the full waveform inversion method in the prior art. The results show that the deep learning-based seismic velocity modeling method provided in this embodiment can accurately capture the underground large-scale structure in the simulated seismic data using a deep neural network, and the seismic velocity model predicted by the deep neural network has a high consistency with the real seismic data.

[0168] The deep learning-based seismic velocity modeling method provided in this embodiment shows that, under noise-free conditions, the seismic velocity model generated by this method has high accuracy and robustness, especially when processing low-frequency signals, it can effectively reduce the interference of other signals.

[0169] See also Figure 4 , which is another seismic velocity modeling method based on deep learning provided in the embodiment of the present application. This embodiment uses some real seismic data for testing to verify the application effect of this method in actual geological environment. Fig. 9 As shown, Fig. 9 These are some real earthquake data images selected from the training set. Figure 4 As shown, the method comprises the following steps:

[0170] S401. Obtain real earthquake data.

[0171] The real seismic data include 10 SEG Salt velocity models and their corresponding seismic data. The spatial size of each SEG Salt velocity model is 3 km × 2 km, the grid size is 10 meters, and the velocity range is 1500 m / s to 4500 m / s. Fig.11 Some SEG Salt speed models selected in the test set.

[0172] S402: Preprocess the real seismic data.

[0173] The real seismic data was also preprocessed, including denoising, normalization, and downsampling. The normalized data was in the same format as the simulated seismic data and was finally input into the deep neural network in the form of a 29×301×360 matrix.

[0174] S403: input the preprocessed real seismic data into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed real seismic data and outputs a seismic velocity model.

[0175] After real seismic data is input into the trained deep neural network, the deep neural network extracts features and fuses multi-frequency band information through the frequency enhancement inference module and frequency fusion layer to accurately capture underground velocity changes. Fig.10 The seismic velocity model obtained by inputting real seismic data into a deep neural network is shown. Fig.10 The two figures in the upper middle are, from left to right, the actual seismic velocity model of real seismic data and the seismic velocity model predicted by the input deep neural network. The figure below is the seismic velocity model predicted by the full waveform inversion method in the prior art.

[0176] The prediction results of the seismic velocity model in the complex salt layer area are consistent with the actual observation depth, which proves the high accuracy and robustness of the seismic velocity model constructed by another seismic velocity modeling method based on deep learning provided in this embodiment under actual geological conditions.

[0177] See also Figure 5 , which is a flow chart of the noise-added deep learning-based seismic velocity modeling method provided in the embodiment of the present application. Since the signal-to-noise ratio of seismic data in actual situations is much lower than that of forward modeling seismic data, it is necessary to perform a noise robustness test. This embodiment adds noise to simulated seismic data and real seismic data to simulate the noise interference in actual seismic exploration, so as to evaluate the prediction effect of the seismic velocity model under noise conditions. Figure 5 As shown, the method comprises the following steps:

[0178] S501. Acquire seismic data, and add noise to the seismic data.

[0179] The mean of the noise is 0 and the variance is 5%. Noise is only introduced in the test phase and no noise is added in the training phase. This allows the ability of the seismic velocity model to handle noisy data in actual exploration to be evaluated.

[0180] S502: Preprocess the seismic data with noise added.

[0181] The noisy seismic data were normalized and downsampled just like the unnoised seismic data, while keeping the same matrix size, and were then input into the deep neural network.

[0182] S503: input the preprocessed seismic data with noise added into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data with noise added, and outputs a seismic velocity model.

[0183] The deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output the enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector.

[0184] Fig.12 The figure shows the seismic data before and after noise addition. The upper figure shows the seismic data before noise addition, and the lower figure shows the seismic data after noise addition. The vertical axis Time of the two figures is time in seconds, and the horizontal axis Distance is distance in kilometers. Fig.13 The seismic velocity model diagram obtained by inputting the simulated seismic data and the real seismic data into the deep neural network after noise addition is shown, wherein the two figures above are the actual seismic velocity model of the simulated seismic data and the seismic velocity model predicted by the input deep neural network from left to right, and the figure below is the seismic velocity model predicted by the full waveform inversion method in the prior art. The results show that under noise interference, the seismic velocity model constructed by the method provided in this embodiment maintains a high prediction accuracy. Although noise will interfere with seismic data, the seismic velocity model can still accurately capture the main features of the underground velocity structure through the weighted processing of signals in different frequency bands by the frequency enhancement module.

[0185] The frequency enhancement reasoning module in the noise-added deep learning-based seismic velocity modeling method provided in this embodiment effectively reduces the impact of noise on the prediction of the seismic velocity model, especially showing excellent noise resistance when processing medium and low frequency signals. The seismic velocity model generated after noise addition has a smaller error than the result without noise addition, and the seismic velocity model generated after noise addition has strong robustness.

[0186] In order to verify the predictive ability of the deep neural network under noise interference, this embodiment conducted a comparative analysis of the seismic velocity curves before and after noise addition. Fig.14 and Fig.15 The comparison results of the seismic velocity curves predicted by the deep neural network using real seismic data before and after noise addition are shown. Fig.14 is the variation of seismic wave velocity with depth at the 186th depth index, Fig.15The velocity of the seismic wave at the 133rd depth index changes with depth. The vertical coordinate Velocity in the two figures is the velocity, in meters per second, and the horizontal coordinate Depth is the depth, in kilometers. The solid line Prediction is the seismic velocity curve predicted by the seismic velocity model with unnoised seismic data input, the long dashed line FWI is the seismic velocity curve predicted by the full waveform inversion method, and the dashed line Ground Truth is the actual seismic velocity curve. The closer it is to the dashed line Ground Truth, the more accurate the seismic velocity model predicted by this method is. The results show that the curve of the seismic velocity model predicted by the deep neural network with unnoised seismic data input is basically consistent with the curve of the actual seismic velocity model, especially in the low and medium frequency bands, the seismic velocity model shows strong stability. In contrast, the seismic velocity curve after adding noise has deviations in the low and medium frequency bands, and the noise significantly affects the prediction accuracy of this frequency band. The seismic velocity curve in the high-frequency part is relatively less affected by noise, but the overall fluctuation in the low and medium frequency bands is more obvious.

[0187] Further analysis shows that noise has a greater impact on the mid- and low-frequency bands, causing the smoothness of the seismic velocity curve to decrease in these areas, especially in the flat area of ​​the curve, where the fluctuation caused by noise is more obvious. However, despite the influence of noise on the mid- and low-frequency bands, the seismic velocity model can still maintain the overall velocity trend well, and the seismic velocity model curve is highly consistent with the actual velocity curve.

[0188] The results of this embodiment show that although noise has a greater impact on the mid- and low-frequency bands, the seismic velocity model still has strong noise resistance and can maintain high accuracy in a noisy environment. This result further proves the robustness and anti-interference ability of the method provided in this embodiment when processing noisy seismic data.

[0189] Based on the deep learning-based seismic velocity modeling method introduced in the previous embodiment, the present application also provides a deep learning-based seismic velocity modeling device. Fig.16 Figure 1 is a schematic diagram of the structure of the device. Fig.16 As shown, the deep learning-based seismic velocity modeling device includes:

[0190] A seismic data acquisition module 601 is used to acquire seismic data;

[0191] The seismic data are real seismic data or simulated seismic data for which seismic velocity modeling is to be performed.

[0192] In a possible implementation, real earthquake data is obtained through a seismic network or database, and simulated earthquake data is obtained through numerical simulation methods. Common simulation numerical methods include but are not limited to: finite difference method, finite element method, spectral element method and pseudo-spectral method.

[0193] A preprocessing module 602, used for preprocessing the seismic data;

[0194] The preprocessing includes a denoising and normalization operation and a downsampling operation.

[0195] Denoising is the process of removing noise from seismic data and retaining useful signals. Common denoising methods include filtering and wavelet transform. Standardization is the process of scaling seismic data to the same scale for subsequent processing and analysis. Common standardization methods include Z-score standardization and Min-Max standardization. Downsampling is the process of reducing the number of data points to reduce the amount of data and computational cost. Common downsampling methods include time domain downsampling and frequency domain downsampling.

[0196] In seismic data, signals of different frequencies represent features of different scales. Low-frequency signals are usually related to large-scale changes in geological structures and are used to capture the overall stratigraphic structure. Medium-frequency signals reveal medium-scale geological structures, such as small and medium-sized faults, rock interfaces, storage boundaries, etc. High-frequency signals usually contain relatively subtle geological details, reflecting small-scale geological structures, such as cracks, geological interfaces, etc., and may also contain noise. Exemplarily, the frequency range of high-frequency signals is greater than 10Hz, the frequency range of medium-frequency signals is between 1Hz and 10Hz, and the frequency range of low-frequency signals is less than 1Hz.

[0197] The preprocessing module can significantly improve the quality of seismic data and the accuracy of subsequent analysis by denoising, standardizing and downsampling seismic data, providing a reliable data basis for seismic velocity modeling.

[0198] The seismic velocity modeling module 603 is used to input the preprocessed seismic data into a deep neural network, and the deep neural network performs seismic velocity prediction based on the preprocessed seismic data and outputs a seismic velocity model.

[0199] The deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output the enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector; the sample data includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

[0200] The working method of the frequency enhancement inference module introduced above is exemplarily described below.

[0201] The frequency enhancement inference module dynamically weights and enhances the multi-band feature vectors extracted by the convolution layer and the pooling layer, including:

[0202] The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is:

[0203]

[0204] Among them, F L (x) is the low frequency band feature vector, F H (x) is the high frequency band feature vector; W L and W H are weight matrices for the low frequency band and the high frequency band respectively, which are learned through the deep neural network.

[0205] The weight matrix will be adaptively adjusted to enhance the contribution of low-frequency signals and reduce the interference of high-frequency noise in practical applications.

[0206] The attention matrix scores of the low-frequency band and the attention matrix scores of the high-frequency band are calculated through the self-attention mechanism, and the attention matrix scores of the low-frequency band and the attention matrix scores of the high-frequency band are normalized respectively using the Softmax function to obtain the attention weights of the low-frequency band and the attention weights of the high-frequency band.

[0207] The product of the attention weight of the low frequency band and the weighted low frequency band feature vector is added to the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-band feature vector with single-head attention enhancement:

[0208]

[0209] Among them, α L and α H They are the attention weights of the low-frequency band and the high-frequency band, respectively, representing the relative importance of low-frequency signals and high-frequency signals.

[0210] A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

[0211] The frequency enhancement inference module can dynamically capture and process signals in different frequency bands by outputting enhanced feature vectors through weighted processing and self-attention mechanism.

[0212] The working method of the frequency fusion layer introduced above is exemplarily described below.

[0213] The frequency fusion layer dynamically fuses the enhanced feature vectors, including:

[0214] The frequency fusion layer adds the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and generates a fused feature vector after activation through a nonlinear activation function.

[0215] y fusion =ReLU(β L ·F enhanced_L (x)+β H ·F enhanced_H (x))

[0216] The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively:

[0217]

[0218] Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

[0219] The frequency fusion layer further fuses the high-frequency signal and the low-frequency signal to ensure that the signal of each frequency band can be fully expressed and avoid the loss of the low-frequency signal during the convolution process.

[0220] In addition, an embodiment of the present application also provides a computer-readable storage medium, in which a computer program is stored. When the program is executed by a processor, the deep learning-based seismic velocity modeling method introduced in any method embodiment is implemented.

[0221] In addition, an embodiment of the present application also provides a processor, which is used to run a computer program, and when the program is running, it executes the deep learning-based seismic velocity modeling method introduced in any implementation method of the aforementioned method embodiment.

[0222] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely schematic, in which the unit described as a separate component may or may not be physically separated, and the component prompted as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0223] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A seismic velocity modeling method based on deep learning, characterized in that: include: Acquisition of seismic data; The seismic data is real seismic data or simulated seismic data for which seismic velocity modeling is to be performed; Preprocessing the seismic data; the preprocessing includes denoising and normalization operations and downsampling operations; The preprocessed seismic data is input into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data and outputs a seismic velocity model; the deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used to dynamically weight and enhance the multi-band feature vectors extracted by the convolution layer and the pooling layer, and output the enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used to dynamically fuse the enhanced feature vector to generate a fused feature vector; The sample data includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

2. The method according to claim 1, characterized in that: The method for obtaining the simulated seismic data includes: In response to the model parameter definition operation, determining the parameters of the seismic velocity model for generating the simulated seismic data; the parameters of the seismic velocity model for generating the simulated seismic data include P-wave velocity, S-wave velocity and density; In response to a selection operation on a numerical simulation method, determining a selected numerical simulation method; In response to the setting operation of boundary conditions and initial conditions of the seismic velocity model for generating the simulated seismic data, the set boundary conditions and initial conditions are determined; the boundary conditions include a free surface and an absorbing boundary; the initial conditions include the location, type and strength of the earthquake source; Based on the seismic velocity model for generating the simulated seismic data, the numerical simulation method and the set boundary conditions and initial conditions, a simulation operation is performed to calculate the propagation process of seismic waves and generate the simulated seismic data.

3. The method according to claim 1, characterized in that: The frequency enhancement inference module dynamically weights and enhances the multi-band feature vectors extracted by the convolution layer and the pooling layer, including: The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is: Among them, F L (x) is the low frequency band feature vector, F H (x) is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices of the low frequency band and the high frequency band, respectively, which are learned through the deep neural network; Calculate the attention matrix score of the low frequency band and the attention matrix score of the high frequency band through the self-attention mechanism, and use the Softmax function to normalize the attention matrix score of the low frequency band and the attention matrix score of the high frequency band respectively to obtain the attention weight of the low frequency band and the attention weight of the high frequency band; Adding the product of the attention weight of the low frequency band and the weighted low frequency band feature vector and the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-frequency band feature vector with single-head attention enhancement; A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

4. The method according to claim 3, characterized in that The frequency fusion layer dynamically fuses the enhanced feature vectors, including: The frequency fusion layer adds the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and generates a fusion feature vector after activation through a nonlinear activation function; The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively: Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

5. The method according to claim 1, characterized in that Methods for training deep neural networks based on sample data include: Acquire sample data; the seismic data samples in the sample data include real seismic data and / or simulated seismic data for training a deep neural network, and the sample data is divided into a training set and a test set according to a preset ratio; Preprocessing the sample data; the preprocessing includes: denoising the seismic data samples contained in the sample data; normalizing the seismic data samples contained in the denoised sample data to the same numerical range by a standardization operation; downsampling the seismic data samples contained in the sample data in a preset form; Based on the preprocessed sample data, training parameters of the deep neural network are determined through training and testing.

6. A seismic velocity modeling device based on deep learning, characterized in that: include: A seismic data acquisition module, used for acquiring seismic data; The seismic data is real seismic data or simulated seismic data for which seismic velocity modeling is to be performed; A preprocessing module, used for preprocessing the seismic data; the preprocessing includes a denoising and normalization operation and a downsampling operation; A seismic velocity modeling module, used for inputting the preprocessed seismic data into a deep neural network, and the deep neural network predicts seismic velocity based on the preprocessed seismic data and outputs a seismic velocity model; the deep neural network is obtained after training based on sample data; the deep neural network includes a convolution layer, a pooling layer, a frequency enhancement reasoning module and a frequency fusion layer connected in sequence; the frequency enhancement reasoning module is used for dynamically weighting and enhancing the multi-band feature vectors extracted by the convolution layer and the pooling layer, and outputting an enhanced feature vector to the frequency fusion layer; the frequency fusion layer is used for dynamically fusing the enhanced feature vector to generate a fused feature vector; The sample data includes: seismic data samples and seismic velocity model labels corresponding to the seismic data samples.

7. The device according to claim 6, characterized in that The frequency enhancement inference module dynamically weights and enhances the multi-band feature vectors extracted by the convolution layer and the pooling layer, including: The multi-band feature vectors extracted by the convolution layer and the pooling layer are processed by a weighting function to obtain a weighted multi-band feature vector; the weighted multi-band feature vector is composed of a weighted low-band feature vector and a weighted high-band feature vector; the weighting function is: Among them, F L (x) is the low frequency band feature vector, F H (x) is the high frequency band feature vector; is the weighted low-frequency feature vector, is the weighted high frequency band feature vector; W L and W H are weight matrices of the low frequency band and the high frequency band, respectively, which are learned through the deep neural network; Calculate the attention score matrix of the low frequency band and the attention score matrix of the high frequency band through the self-attention mechanism, and use the Softmax function to normalize the attention score matrix of the low frequency band and the attention score matrix of the high frequency band respectively to obtain the attention weight of the low frequency band and the attention weight of the high frequency band; Adding the product of the attention weight of the low frequency band and the weighted low frequency band feature vector and the product of the attention weight of the high frequency band and the weighted high frequency band feature vector to obtain a multi-frequency band feature vector with single-head attention enhancement; A multi-head self-attention mechanism is used to connect multiple multi-band feature vectors enhanced by the single-head attention, and an enhanced feature vector is output after activation by a nonlinear activation function; the enhanced feature vector includes an enhanced low-frequency band feature vector and an enhanced high-frequency band feature vector.

8. The device according to claim 6, characterized in that The frequency fusion layer dynamically fuses the enhanced feature vectors, including: The frequency fusion layer adds the product of the fusion weight of the low frequency band and the enhanced low frequency band feature vector to the product of the fusion weight of the high frequency band and the enhanced high frequency band feature vector, and generates a fusion feature vector after activation through a nonlinear activation function; The calculation formulas for the fusion weight of the low frequency band and the fusion weight of the high frequency band are respectively: Among them, Z L is the weight score of the low frequency band, Z H is the weight score of the high frequency band; β L is the fusion weight of the low frequency band, β H is the fusion weight of the high frequency band; the weight score of the low frequency band and the weight score of the high frequency band are both obtained by learning the deep neural network.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the program is executed by a processor, the deep learning-based seismic velocity modeling method as described in any one of claims 1 to 5 is implemented.

10. A processor, characterized in that: Used to run a computer program, which, when running, executes the deep learning-based seismic velocity modeling method as described in any one of claims 1 to 5.

Images