Data-space domain fusion seismic wave velocity deep learning inversion method and system

Abstract

The application provides a data-space domain fusion seismic wave velocity depth learning inversion method and system, constructs a data-model domain fusion network based on attention weighted cross, takes seismic data, a background velocity model and abnormal body imaging results as inputs of the fusion network, makes the fusion network utilize the weights of the data domain and the model domain to influence each other through a weighted attention mechanism, obtains a predicted wave velocity model corresponding to seismic observation data, calculates a loss function according to the predicted wave velocity model and a real value of the wave velocity model, calculates a gradient and performs gradient back propagation to optimize and update network parameters, and processes seismic data by using the data-model domain fusion network based on the attention weighted cross after the network parameters are updated to obtain a wave velocity inversion result. The application realizes effective fusion of the three of seismic data, a background wave velocity and migration imaging results, thereby realizing more accurate wave velocity inversion.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration technology, specifically relating to a deep learning inversion method and system for seismic wave velocity based on data-spatial domain fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Advanced geological prediction plays a crucial role in tunnel construction. Among various advanced geological prediction methods for tunnels, the seismic wave method is widely used in practical engineering due to its advantages such as long detection distance and good identification of structural interfaces. Based on the elastic differences of the underground medium, the seismic wave method can locate and image anomalies. It is particularly sensitive to adverse geological bodies such as faults and fracture zones. The accurate determination of seismic wave velocity is a key factor affecting the accuracy of anomaly imaging and location in the seismic wave method.

[0004] The most mature wave velocity inversion method currently is the full waveform inversion method, which iteratively optimizes model parameters using all waveform information in the seismic record. Essentially, it's a local optimization algorithm for solving the seismic data fitting problem. However, the mapping between seismic data and seismic wave velocity is highly nonlinear, causing the inversion results to be highly dependent on the initial wave velocity model. This is especially problematic in tunnel observation environments, where the limited observation space and construction noise often result in insufficient effective information extracted from seismic wave pre-detection data, leading to significant nonlinearity and ill-posedness in wave velocity inversion. In recent years, with the rapid development of deep learning algorithms, deep neural networks, with their strong ability to perform high-dimensional optimization and simulate nonlinear mappings, have been used to solve geophysical inversion optimization problems, providing a new solution for tunnel seismic wave velocity inversion.

[0005] However, in a tunnel environment, achieving wave velocity inversion based on deep learning still faces the following three challenges:

[0006] First, existing deep learning wave velocity inversion methods mostly use seismic data as network input. Seismic data in tunnels consists of time-series data measured by multiple detectors over a certain period, while the geological model is a spatial model. The correspondence between these two is weak, and due to the limited observation space inside the tunnel, the seismic data carries relatively little effective information about the preceding strata, exacerbating the nonlinearity and ill-posedness of the inversion problem. Therefore, it is quite difficult to effectively map seismic data to wave velocity models using traditional deep learning methods.

[0007] Second, the wave velocity variation trend and structural information in seismic data are implicitly reconstructed through data mining. Since the correspondence between seismic data and geological models is quite complex, and the geological environment where tunnels are located usually contains multiple rock types and complex rock strata structures, the learning difficulty and computational load of the network are large, which affects the results and generalization of the wave velocity inversion network and easily leads to false anomalies in the wave velocity inversion results. Incorporating more information is an effective means to improve the wave velocity inversion capability and generalization.

[0008] Third, although some previous studies have also incorporated data and model information, most of them only used simple information fusion in the form of channels. However, deep neural networks tend to prioritize the use of more relevant information during training, that is, model domain information has a larger weight, while the impact of seismic data with weaker mapping relationships on the results is greatly reduced. This means that the previously used "data-model" input method essentially relies on the spatial model itself and does not make sufficient use of seismic data, resulting in inaccurate inversion results and false anomalies. Therefore, it is necessary to consider how to effectively combine data domain and spatial domain information. Summary of the Invention

[0009] To address the aforementioned problems, this invention proposes a data-spatial domain fusion seismic wave velocity deep learning inversion method and system. This invention achieves effective fusion of seismic data, background wave velocity, and migration imaging results, thereby enabling more accurate wave velocity inversion.

[0010] According to some embodiments, the present invention adopts the following technical solution:

[0011] A deep learning-based inversion method for seismic wave velocity based on data-spatial domain fusion includes the following steps:

[0012] A geological model database was constructed, a reasonable wave velocity model was built, and numerical simulation was performed using the forward modeling algorithm of the wave equation to obtain seismic data;

[0013] Calculate the background velocity model using earthquake data from a geological model database;

[0014] The obtained large-scale velocity model is used as the background velocity for anomalous body imaging;

[0015] A data-model domain fusion network based on attention-weighted cross-interaction is constructed. Seismic data, background velocity models, and anomaly imaging results are used as inputs to the fusion network. Through a weighted attention mechanism, the fusion network utilizes the mutual influence of the weights of the data domain and the model domain to obtain a predicted wave velocity model corresponding to the seismic observation data.

[0016] The loss function is calculated based on the predicted wave velocity model and the actual wave velocity model value. The gradient is calculated and backpropagated to optimize and update the network parameters.

[0017] The seismic data was processed using an attention-weighted cross-data-model domain fusion network with updated parameters to obtain wave velocity inversion results.

[0018] As an alternative implementation method, the specific process of numerical simulation using the forward modeling algorithm of the wave equation includes: constructing a tunnel seismic data learning database, performing geological modeling based on geological type information in front of the tunnel, and performing wavefield simulation for each wave velocity model using an observation system suitable for the tunnel exploration environment, recording wavefield data, and obtaining the wave velocity model and the corresponding seismic data.

[0019] As an alternative implementation method, the specific process of calculating the background velocity model includes performing full waveform inversion calculation using low-frequency signals from seismic observation data, calculating the gradient using low-frequency seismic data and source data, updating the initial model, and obtaining a low wavenumber velocity model.

[0020] As an alternative implementation method, the specific process of anomaly imaging includes: calculating the low wavenumber velocity model obtained by full waveform inversion as the background wave velocity, and using the high-frequency signal of the seismic data for reverse time migration imaging or diffraction superposition to obtain high wavenumber information and tectonic location.

[0021] As an alternative implementation, the attention-weighted cross-feedback-based data-model domain fusion network includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-fusion. The data domain encoder takes seismic observation data as input and extracts data features; the convolutional spatial domain encoder takes background wave velocity and imaging results stitched together in the channel dimension as input and extracts spatial features; the decoder based on data-model domain attention cross-fusion takes the above data and spatial features as input, performs deep fusion of the two types of features from the data domain and the model domain through a weighted attention mechanism, and performs upsampling through the decoder to restore the feature scale and obtain the predicted wave velocity model.

[0022] As a further limitation, the data domain encoder consists of six cascaded convolutional network layers.

[0023] As a further limitation, the convolutional spatial domain encoder is used to connect two channels that take the low-frequency wave velocity model and the high-frequency imaging result as input data, and then perform subsequent processing through nine cascaded convolutional network layers.

[0024] As a further limitation, the weighted attention mechanism of the decoder based on the cross-fusion of data-model domain attention is expressed as follows:

[0025]

[0026]

[0027] In the lower right corner, m and d represent the model features and data features used, respectively; Q, K, and V represent the query key, key key, and value key in the attention mechanism, respectively. k K represents the dimension size, and softmax is the normalization exponential function.

[0028] The first equation can be understood as calculating the weight relationship between data features and model features based on the correlation between data features and model features, updating the model feature construction so that it can obtain the weight influence relationship between data features and model features. The second equation is the opposite.

[0029] As a further limitation, the decoder consists of a series of cascaded conv-up blocks, each conv-up block containing two convolutional operations, wherein each convolutional operation contains a two-dimensional deconvolution operation, a batch normalization operation, and a ReLU activation function.

[0030] As an alternative implementation, the loss function of the attention-weighted cross-referencing data-model domain fusion network is calculated as follows:

[0031]

[0032] in M and R represent the predicted wave velocity model matrix and the original wave velocity model matrix, respectively, and R is the local similarity of the two matrices calculated at different scales. SSIM r The calculation formula is:

[0033]

[0034] H and W represent the height and width of the model, respectively. This represents a window of size r×r centered at point k in the M-model, where c1 and c2 are constant terms in the stable numerator and denominator.

[0035] A data-spatial domain fusion seismic wave velocity deep learning inversion system includes:

[0036] The wave velocity model building module is configured to build a geological model database and use the wave equation forward modeling algorithm to perform numerical simulations to obtain seismic data.

[0037] The calculation module is configured to calculate the background velocity model using seismic data from the geological model database;

[0038] The imaging module is configured to use the large-scale velocity model obtained by the computation module as the background velocity for imaging anomalous bodies.

[0039] The network prediction module is configured to construct a data-model domain fusion network based on attention-weighted cross-interaction. The inputs to the fusion network are seismic data, background velocity model and anomaly imaging results. Through the weighted attention mechanism, the fusion network utilizes the mutual influence of the weights of the data domain and the model domain to obtain a predicted wave velocity model corresponding to the seismic observation data.

[0040] The network parameter optimization module is configured to calculate the loss function based on the predicted wave velocity model and the actual wave velocity model, calculate the gradient and perform gradient backpropagation, and optimize and update the network parameters.

[0041] The inversion application module is configured to process seismic data using an attention-weighted cross-data-model domain fusion network with updated parameters to obtain wave velocity inversion results.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention addresses the problem that most existing deep learning inversion algorithms for seismic wave velocity use seismic observation data as input, making it difficult to learn the mapping relationship between the data and the wave velocity model. The wave velocity variation trend and structural information in the data are implicitly reconstructed through data mining, resulting in high network learning difficulty and computational load, which easily leads to false anomalies in the wave velocity inversion results. This invention proposes a research method that uses seismic data, background wave velocity models, and migration imaging results to jointly predict the accurate wave velocity ahead of the tunnel. The background wave velocity model fills in the signal in the low wavenumber segment, and the migration imaging results fill in the information in the high wavenumber segment. The mapping is improved from the previous data domain to spatial domain to a data-spatial domain to spatial domain mapping, incorporating more feature information and improving the wave velocity inversion capability and generalization.

[0044] To achieve deep fusion of data domain features and model domain features, this invention introduces an attention mechanism and constructs a data-model domain fusion network based on attention-weighted cross-interaction. It further improves upon the traditional attention mechanism, addressing the problem that the traditional deep learning wave velocity inversion method, which relies on the spatial model itself and fails to adequately utilize seismic data, relies on the "data-model" input method. This allows the weights of data domain features and model domain features to mutually constrain each other, achieving effective fusion of seismic data, background wave velocity, and migration imaging. This effectively maps seismic data to the geological velocity model, improving wave velocity inversion performance and generalization.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0047] Figure 1 This is a flowchart of the method in this embodiment;

[0048] Figure 2 This is a schematic diagram of the data-model domain fusion network model structure based on attention-weighted cross-referencing in this embodiment;

[0049] Figure 3 This example demonstrates a cross-fusion structure of data features and model features based on an attention mechanism.

[0050] Figure 4 This is the deep learning inversion result of seismic wave velocity based on data-spatial domain fusion in this example. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0054] Example 1

[0055] The deep learning-based inversion method for tunnel seismic wave velocity based on data-spatial domain fusion includes the following steps:

[0056] Based on the characteristics of tunnel seismic wave advance detection, a geological model database is constructed, a reasonable wave velocity model is established, and numerical simulation is performed through the wave equation forward modeling algorithm to finally obtain the corresponding seismic data and wave velocity model.

[0057] The large-scale background wave velocity can be calculated using seismic data from the above database. Methods include full waveform inversion, tomography, and migration velocity analysis, which can provide low wavenumber background wave velocity for subsequent imaging.

[0058] In this embodiment, low-frequency signals from seismic observation data are used for full waveform inversion calculation. When calculating the gradient using low-frequency seismic data and source data, the local minima problem of traditional full waveform inversion can be effectively alleviated to obtain the background wave velocity model. Because only a small number of iterations are used, the low-frequency inversion results differ significantly from the actual model, but they provide accurate wave velocity variation trends and approximate interface locations, thus filling in the signal gaps in the low wavenumber band and providing better constraints on the model domain and wave velocity values ​​for subsequent network optimization.

[0059] Using the obtained large-scale velocity model as the background velocity, imaging algorithms such as reverse time migration imaging or diffraction stacking are used to image the anomalous body, providing high wavenumber information and construction location for recovering accurate wave velocity.

[0060] A data-model domain fusion network based on attention-weighted cross-interaction is constructed. Seismic data, background velocity models, and imaging results are input into the network. Through a weighted attention mechanism, the network utilizes the mutual influence of the weights in the data domain and the model domain to fully utilize and fuse multimodal data, and finally obtains a predicted wave velocity model corresponding to the seismic observation data.

[0061] In this embodiment, the attention-weighted cross-feedback data-model domain fusion network includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-fusion. The data domain encoder takes seismic observation data as input and extracts data features; the convolutional spatial domain encoder takes background wave velocity and imaging results stitched together in the channel dimension as input and extracts spatial features; the decoder based on data-model domain attention cross-fusion takes the above data and spatial features as input, and through an improved attention mechanism, achieves deep fusion of the two types of features from the data domain and the model domain. The decoder then performs upsampling to restore the feature scale, resulting in a predicted wave velocity model.

[0062] The loss function is calculated by comparing the predicted wave velocity output from the attention-weighted cross-data-model domain fusion network with the actual wave velocity model value. The gradient is calculated and backpropagated to optimize and update the network parameters.

[0063] In this embodiment, a loss function based on an attention-weighted cross-linked data-model domain fusion network is designed, and the calculation formula is as follows:

[0064]

[0065] in M and R represent the predicted wave velocity model matrix and the original wave velocity model matrix, respectively, and R is the local similarity of the two matrices calculated at different scales. SSIM r The calculation formula is:

[0066]

[0067] H and W represent the height and width of the model, respectively. This represents a window of size r×r centered at point k in the M-model, where c1 and c2 are constant terms in the stable numerator and denominator.

[0068] The seismic data was processed using an attention-weighted cross-data-model domain fusion network with updated parameters to obtain wave velocity inversion results.

[0069] Example 2

[0070] A deep learning-based inversion method for seismic wave velocity based on data-spatial domain fusion, such as Figure 1 As shown, it includes the following steps:

[0071] Step S1: Construct a model-data database through computer numerical simulation. The database includes a label data group (containing wave velocity model-seismic data pairs).

[0072] This example considers that the soil and rock along the tunnel construction route mainly consists of discontinuous geological structures such as numerous joints and fissures, and sets the wave velocity range to 1500 m / s to 4500 m / s. Furthermore, based on on-site detection requirements, the model is designed to be 80m × 150m in size. The model extends 50m into the tunnel excavation area, with a 10m wide cavity in the middle and a wave velocity of 340 m / s assigned to it to simulate the tunnel. The wave velocity of the surrounding rock is consistent with the wave velocity of the first rock mass in front of the tunnel. The detection range is 50-150m in front, and its width is set to 80m to ensure effective reception of reflected signals. Unlike traditional modeling methods that simplify the wave velocity distribution within 100m in front of the tunnel into a model with 1-3 wave velocity variation interfaces, including inclined and curved interfaces, this method is used to test the proposed method.

[0073] In the numerical simulation, the finite difference grid size was set to 1m×1m, the Ricker wavelet was used as the earthquake source for simulation, and the constant density acoustic wave equation was simulated using the finite difference time-domain method to calculate the tunnel observation data.

[0074] Step S2: Perform full waveform inversion and reverse time migration imaging on the seismic data;

[0075] Although the full waveform inversion method also relies heavily on the initial model, it can provide relatively accurate trends in wave velocity inversion. Especially when using low-frequency signals for wave velocity inversion, considering that for the same wave velocity, the corresponding wavelength of the low-frequency signal is longer, its accurate wave velocity variation trend can cover inter-layer wave velocity variations. This example also chooses to use low-frequency signals for full waveform inversion calculations. After a few iterations, geological information in the low wavenumber bands can be obtained, providing a foundation for subsequent high-precision imaging. When calculating gradients using low-frequency seismic data and source data, the local minima problem of traditional full waveform inversion can be effectively alleviated.

[0076] Specifically, this example uses the surrounding rock wave velocity, which is relatively easy to obtain from the measured data, as the initial model. It adopts the previous full waveform inversion algorithm based on PyTorch, and takes the 0-100Hz frequency band signal for low-frequency full waveform inversion. The Adam optimizer is used to update the speed model learning rate, which is set to 20, and a total of 50 iterations are performed.

[0077] Based on the wave velocity model obtained from the low-frequency full waveform inversion as the background wave velocity, a round of reverse time migration imaging was performed at a higher frequency to obtain the imaging results.

[0078] Step S3, as follows Figure 2 We construct a data-model domain fusion network based on attention-weighted cross-interaction. The data-model domain fusion network based on attention-weighted cross-interaction includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-interaction.

[0079] The data domain encoder consists of a series of conv-down blocks. Each conv-down block contains two convolutional operations and a max-pooling operation. Each convolutional operation further includes a two-dimensional convolutional operation, a batch normalization operation, and a ReLU activation function. The data domain encoder compresses the spatial dimension of seismic data while increasing the number of channels, resulting in a feature matrix after depth feature extraction.

[0080] The convolutional spatial domain encoder uses full waveform inversion results and reverse time migration imaging results as dual-channel input networks. It employs a total of 9 convolutional layers to process the wave velocity-imaging channels. Each convolutional layer consists of a 7x7 convolutional kernel, batch normalization, and ReLU activation function. By setting the stride of the convolutional layers, the number of channels of the data is increased while compressing the spatial dimension. Finally, the model domain information is compressed to a feature matrix with the same feature size as the input seismic data.

[0081] The decoder based on data-model domain attention cross-fusion is as follows: Figure 3 The attention module and decoder shown can be represented by the following formula for their improved attention mechanism:

[0082]

[0083]

[0084] In the lower right corner, m and d represent the model features and data features used, respectively; Q, K, and V represent the query key, key key, and value key in the attention mechanism, respectively.

[0085] Specifically, seismic signals from different sidewalls are processed separately by the data domain encoder to obtain two sets of data features, which are then stitched together to obtain data domain features. The wave velocity-imaging matrix is ​​similarly processed by the model domain encoder module to obtain a spatial feature matrix of the same size. First, the weight matrix is ​​calculated using the data domain features as Q and K, and then a three-layer, four-head attention mechanism is used with the spatial domain features as V to obtain a feature matrix of invariant size. This can be understood as calculating the weight relationship between the data features and the model features based on their correlation, thereby updating the model feature construction so that it can capture the weight influence of data features on model features; conversely, a feature matrix of the same size can be obtained to extract the weight of model features on data features.

[0086] Finally, a decoder consisting of a series of conv-up blocks reconstructs the Bosu model. Each conv-up block contains two convolutional operations, each of which includes a two-dimensional deconvolution operation, a batch normalization operation, and a ReLU activation function. The decoder increases the spatial size of the feature matrix while reducing the number of channels through operations such as upsampling and convolution, ultimately restoring the dimensionality and size of the Bosu model.

[0087] Step S4: Design the loss function, use the training set in the database to train the network, and update and optimize the network parameters;

[0088] This example uses structural similarity and MSE as loss functions, as shown in the following formula:

[0089]

[0090] in M and the predicted wave velocity model matrix are the original wave velocity model matrix and the predicted wave velocity model matrix, respectively. R is the local similarity of the two matrices calculated at different scales. SSIM r The calculation formula is:

[0091]

[0092] H and W represent the height and width of the model, respectively. This represents a window of size r×r centered at point k in the M-model, where c1 and c2 are constant terms in the stable numerator and denominator.

[0093] The main network parameters and hardware conditions in this embodiment are as follows: computation is implemented using four NVIDIA TITAN XP chips. The network is built on the PyTorch platform, the Adam optimizer has a batch size of 16, and a learning rate of 0.0005 is used, which decays exponentially during the training process.

[0094] Step S5: Test the inversion performance of the trained attention-weighted cross-referencing data-model domain fusion network on the test set. Partial results on the test set are shown below. Figure 4 As shown, it can be seen that by introducing model domain information to improve network training, the results obtained are far superior to those of wave velocity inversion networks that simply use data features as input.

[0095] Example 3

[0096] A data-spatial domain fusion seismic wave velocity deep learning inversion system includes:

[0097] The wave velocity model building module is configured to build a geological model database and use the wave equation forward modeling algorithm to perform numerical simulation to obtain the wave velocity model.

[0098] The calculation module is configured to calculate the background velocity model using seismic data from the geological model database;

[0099] The imaging module is configured to use the obtained large-scale velocity model as background velocity to image anomalous bodies.

[0100] The prediction module is configured to construct a data-model domain fusion network based on attention-weighted cross-interaction. The network takes seismic data, background velocity model and anomaly imaging results as inputs. Through a weighted attention mechanism, the fusion network uses the weights of the data domain and model domain to influence each other and obtain a predicted wave velocity model corresponding to the seismic observation data.

[0101] The parameter optimization module is configured to calculate the loss function based on the predicted wave velocity model and the actual wave velocity model, calculate the gradient and perform gradient backpropagation, and optimize and update the network parameters.

[0102] The inversion module is configured to process seismic data using an attention-weighted cross-referenced data-model domain fusion network with updated parameters to obtain wave velocity inversion results.

[0103] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based inversion method for seismic wave velocity based on data-spatial domain fusion, characterized in that, Includes the following steps: A geological model database was constructed, a wave velocity model was established, and numerical simulations were performed using the forward modeling algorithm of the wave equation to ultimately obtain the corresponding seismic data. Calculate large-scale velocity models using seismic data from a geological model database; The obtained large-scale velocity model is used as the background velocity for anomalous body imaging; A data-model domain fusion network based on attention-weighted cross-linking is constructed. Seismic data, large-scale velocity models, and anomaly imaging results are used as inputs to the fusion network. Through a weighted attention mechanism, the fusion network utilizes the mutual influence of the weights of the data domain and the model domain to obtain a predicted wave velocity model corresponding to the seismic observation data. The loss function is calculated based on the predicted wave velocity model and the actual wave velocity model value. The gradient is calculated and backpropagated to optimize and update the network parameters. The seismic data was processed using an attention-weighted cross-data-model domain fusion network with updated parameters to obtain wave velocity inversion results; The specific process of calculating the large-scale velocity model includes full waveform inversion calculation using low-frequency signals from seismic observation data, and calculating the gradient using low-frequency seismic data and sources; the specific process of anomaly imaging includes: using the large-scale velocity model obtained from full waveform inversion as the background wave velocity, using high-frequency signals from seismic data for reverse time migration imaging or diffraction stacking to obtain high wavenumber information and tectonic location. The attention-weighted cross-data-model domain fusion network includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-fusion. The data domain encoder takes seismic observation data as input to extract data features; the convolutional spatial domain encoder takes background wave velocity and imaging results stitched together in the channel dimension as input to extract spatial features. The decoder based on data-model domain attention cross-fusion takes the aforementioned data features and spatial features as input, performs deep fusion of the two types of features from the data domain and the model domain through a weighted attention mechanism, and then performs upsampling through the decoder to restore the feature scale and obtain the predicted wave velocity model.

2. The deep learning inversion method for seismic wave velocity based on data-spatial domain fusion as described in claim 1, characterized in that, The specific process of numerical simulation using the forward modeling algorithm of the wave equation includes: constructing a learning database of tunnel seismic data, performing geological modeling based on geological type information in front of the tunnel, and performing wavefield simulation for each wave velocity model using an observation system suitable for tunnel exploration environment, recording wavefield data, and obtaining wave velocity model and corresponding seismic data.

3. The deep learning inversion method for seismic wave velocity based on data-spatial domain fusion as described in claim 1, characterized in that, The data domain encoder consists of six cascaded convolutional network layers. Alternatively, the convolutional spatial domain encoder is used to connect two channels that take the low-frequency wave velocity model and the high-frequency imaging result as input data, and then perform subsequent processing through nine cascaded convolutional network layers.

4. The deep learning inversion method for seismic wave velocity based on data-spatial domain fusion as described in claim 1, characterized in that, The weighted attention mechanism of the decoder based on data-model domain attention cross-fusion is expressed as follows: The lower right corner 'm' and 'd' represent the model features and data features used, respectively. Q, K, and V represent the query key, key key, and value key in the attention mechanism, respectively.

5. The deep learning inversion method for seismic wave velocity based on data-spatial domain fusion as described in claim 4, characterized in that, The decoder consists of a series of cascaded conv-up blocks. Each conv-up block contains two convolutional operations, and each convolutional operation contains a two-dimensional deconvolution operation, a batch normalization operation, and a ReLU activation function.

6. The deep learning inversion method for seismic wave velocity based on data-spatial domain fusion as described in claim 1, characterized in that, The loss function of the attention-weighted cross-referencing data-model domain fusion network is calculated as follows: in M and M are the predicted wave velocity model matrix and the original wave velocity model matrix, respectively. To calculate the local similarity of two matrices at different scales, The calculation formula is: in, and Divided into the height and width of the model, Representative at Take from the model Centered on point, size is The window, and This is a constant term that stabilizes the numerator and denominator.

7. A deep learning inversion system for seismic wave velocity based on data-spatial domain fusion, characterized in that, include: The wave velocity model building module is configured to build a geological model database and use the wave equation forward modeling algorithm to perform numerical simulations to obtain seismic data. The calculation module is configured to use seismic data from the geological model database to calculate large-scale velocity models; The imaging module is configured to use the obtained large-scale velocity model as background velocity to perform anomalous body imaging; The prediction module is configured to construct a data-model domain fusion network based on attention-weighted cross-linking. The network takes seismic data, large-scale velocity models, and anomaly imaging results as inputs. Through a weighted attention mechanism, the fusion network utilizes the mutual influence of the weights of the data domain and the model domain to obtain a predicted wave velocity model corresponding to the seismic observation data. The parameter optimization module is configured to calculate the loss function based on the predicted wave velocity model and the actual wave velocity model, calculate the gradient and perform gradient backpropagation, and optimize and update the network parameters. The inversion module is configured to process seismic data using an attention-weighted cross-data-model domain fusion network with updated parameters to obtain wave velocity inversion results. The specific process of calculating the large-scale velocity model includes full waveform inversion calculation using low-frequency signals from seismic observation data, and calculating the gradient using low-frequency seismic data and sources; the specific process of anomaly imaging includes: using the large-scale velocity model obtained from full waveform inversion as the background wave velocity, using high-frequency signals from seismic data for reverse time migration imaging or diffraction stacking to obtain high wavenumber information and tectonic location. The attention-weighted cross-data-model domain fusion network includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-fusion. The data domain encoder takes seismic observation data as input to extract data features; the convolutional spatial domain encoder takes background wave velocity and imaging results stitched together in the channel dimension as input to extract spatial features. The decoder based on data-model domain attention cross-fusion takes the aforementioned data features and spatial features as input, performs deep fusion of the two types of features from the data domain and the model domain through a weighted attention mechanism, and then performs upsampling through the decoder to restore the feature scale and obtain the predicted wave velocity model.

Images