CN118151236A - Seismic velocity deep learning inversion method and system based on data-space domain fusion - Google Patents

Abstract

The invention provides a data-space domain fusion seismic wave velocity deep learning inversion method and a system, which are used for constructing a data-model domain fusion network based on attention weighted intersection, taking a seismic data, a background velocity model and an abnormal imaging result as inputs of the fusion network, and enabling the fusion network to utilize the mutual influence of the weights of a data domain and a model domain through a weighted attention mechanism so as to obtain a prediction wave velocity model corresponding to seismic observation data; calculating a loss function according to the predicted wave velocity model and the true value of the wave velocity model, calculating a gradient, carrying out gradient feedback, and optimizing and updating network parameters; and processing the seismic data by using a data-model domain fusion network based on attention weighted intersection after updating parameters to obtain a wave velocity inversion result. The invention realizes the effective fusion of the seismic data, the background wave velocity and the offset imaging result, thereby realizing more accurate wave velocity inversion.

Description

Seismic velocity deep learning inversion method and system based on data-space domain fusion

Technical Field

The present invention belongs to the technical field of geophysical exploration, and in particular relates to a data-space domain fusion seismic velocity deep learning inversion method and system.

Background technique

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

Advanced geological prediction is of great significance in tunnel construction. Among various tunnel advanced geological prediction methods, the seismic wave method has been widely used in actual engineering due to its advantages such as long detection distance and good identification effect on structural interfaces. The seismic wave method is based on the elastic difference of underground media, which can locate and image abnormal bodies. It is more sensitive to adverse geological bodies such as faults and fracture zones. The accurate determination of seismic wave velocity in the seismic wave method is a key factor affecting the accuracy of abnormal body imaging and positioning.

The most mature velocity inversion method at present is the full waveform inversion method, which uses all the waveform information in the seismic record to iteratively optimize the model parameters. It is essentially a local optimization algorithm for solving the seismic data fitting problem. The mapping between seismic data and seismic velocity is strongly nonlinear, which causes its inversion results to be highly dependent on the initial velocity model. Especially for tunnel observation environments, due to the narrow observation space inside the tunnel and the construction noise, the effective information extracted from the advance detection data of the seismic wave method is often insufficient, resulting in serious nonlinearity and ill-posedness in velocity inversion. In recent years, with the rapid development of deep learning algorithms, deep neural networks have been used to solve geophysical inversion optimization problems with their strong high-dimensional optimization and simulation of nonlinear mapping capabilities, providing a new solution for tunnel seismic velocity inversion.

However, in tunnel environments, there are still three challenges in implementing velocity inversion based on deep learning:

First, existing deep learning velocity inversion mostly uses seismic data as network input. The seismic data in the tunnel is time series data measured by multiple detectors within a certain period of time, while the geological model is a spatial model. The correspondence between the two is weak, and due to the narrow observation space inside the tunnel, the seismic data carries less effective information about the front strata, which aggravates the nonlinearity and ill-posedness of the inversion problem. Therefore, it is difficult to use traditional deep learning methods to effectively map seismic data to velocity models.

Second, the velocity change trend and structural information in the seismic data are implicitly reconstructed through data mining. Since the correspondence between seismic data and geological models is relatively complex, and since the geological environment in which the tunnel is located usually contains a variety of rock types and the rock structure is complex, the network learning difficulty and the amount of calculation are large, which affects the results and generalization of the velocity inversion network and easily leads to false anomalies in the velocity inversion results. Incorporating more information is an effective means to improve the velocity inversion capability and generalization.

Third, although some previous research works have also introduced data and model information at the same time, and most of them only perform simple information fusion in a channel manner, deep neural networks tend to give priority to more relevant information during the training process, that is, model domain information occupies a larger weight, while the impact of seismic data with weaker mapping relationships on the results will be greatly discounted. This leads to the previously adopted "data-model" input method, which essentially relies on the spatial model itself, and insufficient use of seismic data, resulting in inaccurate inversion results, false anomalies and other problems. Therefore, it is necessary to consider how to effectively combine data domain and spatial domain information.

Summary of the invention

In order to solve the above problems, the present invention proposes a data-space domain fusion seismic velocity deep learning inversion method and system, which realizes the effective fusion of seismic data, background velocity and offset imaging results, thereby achieving more accurate velocity inversion.

According to some embodiments, the present invention adopts the following technical solutions:

A data-space domain fusion seismic velocity deep learning inversion method comprises the following steps:

Build a geological model database, construct a reasonable wave velocity model, use the wave equation forward algorithm to perform numerical simulation, and obtain seismic data;

Calculate the background velocity model using seismic data in the geological model database;

The obtained large-scale velocity model is used as the background velocity to perform anomaly imaging;

A data-model domain fusion network based on attention weighted cross is constructed, with seismic data, background velocity model and abnormal volume imaging results as the input of the fusion network. Through the weighted attention mechanism, the fusion network uses the weights of the data domain and the model domain to influence each other, and obtains a predicted wave velocity model corresponding to the seismic observation data.

Calculate the loss function based on the predicted wave speed model and the true value of the wave speed model, calculate the gradient and return the gradient to optimize and update the network parameters;

The seismic data are processed using the attention-weighted cross-talk-based data-model domain fusion network with updated parameters to obtain the velocity inversion results.

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

As an optional implementation, the specific process of calculating the background velocity model includes using the low-frequency signal of the seismic observation data to perform full waveform inversion calculation, using the seismic data and source of the low-frequency signal to calculate the gradient and update the initial model to obtain a low-wavenumber velocity model.

As an optional implementation, the specific process of performing anomaly imaging includes: calculating the low wave number velocity model obtained by full waveform inversion as the background wave velocity, using the high frequency band signal of the seismic data to perform reverse time migration imaging or diffraction stacking to obtain high wave number information and structural position.

As an optional implementation, the data-model domain fusion network based on weighted cross-attention includes a data domain encoder, a convolutional spatial domain encoder and a decoder based on data-model domain attention cross-fusion, wherein the data domain encoder takes seismic observation data as input to extract data features; the convolutional spatial domain encoder takes the background wave velocity and imaging results spliced in the channel dimension as input to extract 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, data domain and model domain, through a weighted attention mechanism, and performs upsampling through the decoder to restore the feature scale to obtain a predicted wave velocity model.

As a further limitation, the data domain encoder is composed of 6 convolutional network layers cascaded in sequence.

As a further limitation, the convolutional spatial domain encoder is used to connect the low-frequency wave velocity model and the high-frequency imaging result as two channels of input data, and then perform subsequent processing through 9 layers of convolutional network layers cascaded in sequence.

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

The m and d in the lower right corner represent the model features and data features respectively; Q, K and V represent the query key, key key and value key in the attention mechanism respectively, d _k is the dimension size of K, and softmax is the normalized exponential function.

The former formula can be understood as calculating the weight relationship between data features and model features by their correlation, and updating the model feature structure so that it can obtain the weight influence relationship between data features and model features. The latter formula is the opposite.

As a further limitation, the decoder consists of a series of cascaded conv-up blocks, each conv-up block contains two convolution operations, each of which contains a two-dimensional deconvolution operation, a batch normalization operation, and a ReLU activation function.

As an optional implementation, the loss function of the attention-weighted cross-data-model domain fusion network is calculated as follows:

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

Among them, H and W are the height and width of the model. It represents a window of size r×r with point k as the center in the M model, and _c1 and _c2 are constant terms that stabilize the numerator and denominator.

A data-space domain fusion seismic wave velocity deep learning inversion system, comprising:

A wave velocity model building module is configured to build a geological model database and perform numerical simulation using a wave equation forward algorithm to obtain seismic data;

A calculation module configured to calculate a background velocity model using seismic data in a geological model database;

An imaging module is configured to perform abnormal body imaging using the large-scale velocity model obtained by the calculation module as a background velocity;

The network prediction module is configured to construct a data-model domain fusion network based on attention weighted crossover, with seismic data, background velocity model and abnormal volume imaging results as inputs of the fusion network. Through the weighted attention mechanism, the fusion network uses the weights of the data domain and the model domain to influence each other, and obtains a predicted wave velocity model corresponding to the seismic observation data;

The network parameter optimization module is configured to calculate the loss function according to the predicted wave velocity model and the true value of the wave velocity model, calculate the gradient and perform gradient backpropagation, and optimize and update the network parameters;

The inversion application module is configured to process seismic data using the attention-weighted cross-talk-based data-model domain fusion network with updated parameters to obtain velocity inversion results.

Compared with the prior art, the present invention has the following beneficial effects:

In view of the fact that most of the existing deep learning inversion algorithms for seismic wave velocity use seismic observation data as input, the mapping relationship between the seismic wave velocity model and the wave velocity model is difficult to learn, the wave velocity change trend and structural information in the data are implicitly reconstructed through data mining, the network learning difficulty and the amount of calculation are large, and it is easy to cause the problem of false anomalies in the wave velocity inversion results. The present invention proposes a research method for jointly predicting the accurate wave velocity in front of the tunnel using seismic data, background wave velocity model, and offset imaging results. The background wave velocity model is used to fill the signal of the low wave number segment, and the offset imaging result is used to fill the information of the high wave number segment. The previous mapping from the data domain to the space domain is improved to the mapping from the data-space domain to the space domain, which incorporates more feature information and improves the wave velocity inversion capability and generalization.

In order to achieve the deep fusion of data domain features and model domain features, the present invention introduces an attention mechanism, constructs a data-model domain fusion network based on attention weighted cross, and makes further improvements on the traditional attention mechanism, which solves the problem that the "data-model" input method adopted by traditional deep learning velocity inversion essentially relies on the spatial model itself and insufficiently utilizes seismic data. The weights of data domain features and model domain features are mutually constrained, and the effective fusion of seismic data, background velocity and offset imaging is achieved, which effectively completes the mapping of seismic data to geological velocity model, and improves the velocity inversion effect and generalization.

In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

FIG1 is a flow chart of the method of this embodiment;

FIG2 is a schematic diagram of the structure of a data-model domain fusion network model based on attention weighted crossover of this embodiment;

FIG3 is a cross-fusion structure of data features and model features based on the attention mechanism of this example;

FIG4 is the seismic velocity deep learning inversion result based on data-space domain fusion in this example.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and/or combinations thereof.

Embodiment 1

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

According to 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 carried out through the wave equation forward algorithm to finally obtain the corresponding seismic data and wave velocity model.

The seismic data in the above database is used to calculate the large-scale background wave velocity. The methods that can be used include full waveform inversion, tomography, migration velocity analysis, etc., to provide low-wavenumber background wave velocity for subsequent imaging;

In this embodiment, the low-frequency signal of the seismic observation data is used for full waveform inversion calculation. When the seismic data of the low-frequency signal and the source are used to calculate the gradient, the local minimum problem of the traditional full waveform inversion can be effectively alleviated to obtain the background wave velocity model. Since only a small number of iterations are used, the low-frequency inversion result is quite different from the real model, but it provides an accurate wave velocity change trend and an approximate interface position, that is, it fills the signal of the low wave number segment, which provides a better constraint on the model domain and wave velocity value for the subsequent network optimization.

Using the large-scale velocity model as the background velocity, imaging algorithms such as reverse time migration imaging or diffraction stacking are used to image the anomaly, providing high-wave number information and structural position for restoring accurate wave velocity;

A data-model domain fusion network based on weighted cross-attention is constructed. The seismic data, background velocity model and imaging results are input into the network. Through the weighted attention mechanism, the network uses the weights of the data domain and the model domain to influence each other, so that the multimodal data can be fully utilized and fused, and finally a predicted wave velocity model corresponding to the seismic observation data is obtained.

In this embodiment, the data-model domain fusion network based on attention weighted cross 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 the background wave velocity and imaging results spliced in the channel dimension as input to extract spatial features; the decoder based on data-model domain attention cross fusion takes the above data and spatial features as input, and realizes the deep fusion of the two types of features in the data domain and the model domain by improving the attention mechanism, and performs upsampling through the decoder to restore the feature scale and obtain the predicted wave velocity model.

The loss function is calculated based on the predicted wave speed output by the data-model domain fusion network based on the attention weighted cross and the true value of the wave speed model, the gradient is calculated and the gradient is returned to optimize and update the network parameters;

In this embodiment, a loss function of a data-model domain fusion network based on attention weighted cross is designed, and the calculation formula is:

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

Among them, H and W are the height and width of the model. It represents a window of size r×r with point k as the center in the M model, and _c1 and _c2 are constant terms that stabilize the numerator and denominator.

The seismic data are processed using the attention-weighted cross-talk-based data-model domain fusion network with updated parameters to obtain the velocity inversion results.

Embodiment 2

The seismic velocity deep learning inversion method based on data-spatial domain fusion, as shown in FIG1 , includes the following steps:

Step S1, constructing a model-data database through computer numerical simulation, wherein the database includes a label data set (including wave velocity model-seismic data pairs).

In this example, considering that the rock and soil along the tunnel construction route are mostly non-continuous geological structures such as a large number of joints and fissures, the wave velocity range is set to 1500m/s to 4500m/s. In addition, according to the needs of on-site detection, the model is designed to be 80m×150m in size. The front of the model is a 50m tunnel excavation range, and a 10m wide cavity is arranged in the middle and a wave velocity of 340m/s is assigned to simulate the tunnel. The wave velocity of the surrounding rock around the tunnel is consistent with the wave velocity of the first rock medium in front of the tunnel; 50-150m in front is the front detection range. In order to ensure the effective reception of reflected signals, its width is set to 80m. Different from the traditional modeling method, which simplifies the wave velocity distribution within 100m in front of the tunnel into a model with 1-3 wave velocity change interfaces, the interface morphology includes inclined interfaces and curved interfaces, which jointly test the disclosed method.

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

Step S2, performing full waveform inversion and reverse time migration imaging on the seismic data;

Although the full waveform inversion method is also relatively dependent on the initial model, it can provide a relatively accurate change trend for velocity inversion. In particular, when using low-frequency signals for velocity inversion, considering that for the same velocity, the corresponding wavelength of the low-frequency signal is longer, its accurate velocity change trend can cover the velocity change between layers. This example also chooses to use low-frequency signals for full waveform inversion calculation. After a small number of iterations, geological information in the low-wave number segment can be obtained, providing a basis for subsequent high-precision imaging. When calculating the gradient with seismic data and earthquake sources of low-frequency signals, the local minimum problem of traditional full waveform inversion can be effectively alleviated.

Specifically, this example uses the surrounding rock velocity that is easier to obtain in the measured data as the initial model, adopts the previous Pytorch-based full waveform inversion algorithm, and takes the 0-100 Hz frequency band signal for low-frequency full waveform inversion. The Adam optimizer is used to update the speed model with a learning rate set to 20, and a total of 50 rounds of iterations are performed.

The velocity model obtained by low-frequency full waveform inversion is used as the background velocity, and a round of reverse time migration imaging is performed at a higher frequency to obtain the imaging results.

Step S3, as shown in FIG2, constructs a data-model domain fusion network based on attention weighted cross-talk, wherein the data-model domain fusion network based on attention weighted cross-talk includes a data domain encoder, a convolutional spatial domain encoder, and a decoder based on data-model domain attention cross-talk fusion;

The data domain encoder is composed of a series of conv-down blocks. Each conv-down block contains two convolution operations and a maximum pooling operation. Each convolution operation contains a two-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The data domain encoder compresses the spatial dimension of the seismic data while increasing the number of channels of the data to obtain the feature matrix after deep feature extraction.

The convolutional spatial domain encoder uses the full waveform inversion results and the reverse time migration imaging results as the dual-channel input network, and uses a cumulative 9-layer convolution layer to process the velocity-imaging channel. Each convolution layer consists of a convolution kernel of size 7x7, batch normalization and ReLU activation function. By setting the step size of the inter-layer convolution, the spatial dimension size is compressed while increasing the number of data channels. Finally, the model domain information is compressed to a feature matrix consistent with the feature size of the input seismic data.

The decoder based on data-model domain attention cross-fusion consists of an attention module and a decoder as shown in Figure 3. Its improved attention mechanism can be expressed as the following formula:

The m and d in the lower right corner represent the adopted model features and data features respectively; Q, K and V represent the query key, key key and value key in the attention mechanism respectively.

Specifically, the seismic signals from different side walls are processed separately by the data domain encoder to obtain two sets of data features, which are then spliced to obtain data domain features, while the wave velocity-imaging matrix is also 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 the three-layer four-head attention mechanism is used with the spatial domain features as V to obtain a feature matrix of the same size. It can be understood as calculating the weight relationship between the data feature correlation and the model feature, and then updating the model feature structure so that it can obtain the weight influence relationship between the data feature and the model feature; conversely, features of the same size can be obtained to extract the weight of the model feature to the data feature.

Finally, the decoder composed of a series of conv-up blocks restores the wave speed model. Each conv-up block contains two convolution operations, each of which contains 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, and finally restores the dimension and size of the wave speed model.

Step S4, designing a loss function, using the training set in the database to train the network, and updating and optimizing the network parameters;

This example uses structural similarity and MSE as the loss function. The specific formula is as follows:

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

Among them, H and W are the height and width of the model. It represents a window of size r×r with point k as the center in the M model, and _c1 and _c2 are constant terms that stabilize the numerator and denominator.

The main network parameters and hardware conditions in this embodiment are as follows: the calculation is implemented using 4 NVIDIA TITAN Xp chips. The network is built based on the PyTorch platform, the batch size of the Adam optimizer is 16, and the learning rate is 0.0005 with exponential decay during the training process.

Step S5, the trained attention-weighted cross-linked data-model domain fusion network is tested on the test set for inversion effect. Some results on the test set are shown in Figure 4. It can be seen that by introducing model domain information to improve network training, the wave velocity inversion network results that are far superior to those that simply use data features as input are achieved.

Embodiment 3

A data-space domain fusion seismic wave velocity deep learning inversion system, comprising:

The wave velocity model building module is configured to build a geological model database, perform numerical simulation using a wave equation forward algorithm, and obtain a wave velocity model;

A calculation module configured to calculate a background velocity model using seismic data in a geological model database;

An imaging module is configured to perform anomaly imaging using the obtained large-scale velocity model as a background velocity;

The prediction module is configured to construct a data-model domain fusion network based on attention weighted cross, with seismic data, background velocity model and abnormal volume imaging results as inputs of the fusion network. Through the weighted attention mechanism, the fusion network uses the weights of the data domain and the model domain to influence each other, and obtains a predicted wave velocity model corresponding to the seismic observation data;

The parameter optimization module is configured to calculate the loss function according to the predicted wave speed model and the true value of the wave speed model, calculate the gradient and perform gradient backpropagation, and optimize and update the network parameters;

The inversion module is configured to process the seismic data using the attention-weighted cross-talk-based data-model domain fusion network with updated parameters to obtain a velocity inversion result.

Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented 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.

The present invention is described with reference to the flowchart and/or block diagram of the method, device (system), and computer program product according to the embodiment of the present invention. It should be understood that each process and/or box in the flowchart and/or block diagram, as well as the combination of the process and/or box in the flowchart and/or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims (10)

1. A data-space domain fusion seismic wave velocity deep learning inversion method is characterized by comprising the following steps:

And constructing a geological model database, establishing a reasonable wave velocity model, performing numerical simulation through a wave equation positive algorithm, and finally obtaining corresponding seismic data and the wave velocity model.

Calculating a background velocity model using the seismic data in the geologic model database;

taking the obtained large-scale speed model as a background speed, and performing abnormal body imaging;

constructing a data-model domain fusion network based on attention weighted intersection, taking the seismic data, a background velocity model and an abnormal body imaging result as inputs of the fusion network, and enabling the fusion network to mutually influence by using the weights of a data domain and a model domain through a weighted attention mechanism to obtain a prediction wave velocity model corresponding to seismic observation data;

calculating a loss function according to the predicted wave velocity model and the true value of the wave velocity model, calculating a gradient, carrying out gradient feedback, and optimizing and updating network parameters;

and processing the seismic data by using a data-model domain fusion network based on attention weighted intersection after updating parameters to obtain a wave velocity inversion result.

2. The method for deep learning and inversion of seismic wave velocity with data-space domain fusion according to claim 1, wherein the specific process of numerical simulation by utilizing wave equation forward algorithm comprises constructing a tunnel seismic data learning database, performing geological modeling based on geological type information in front of a tunnel, performing wave field simulation on each wave velocity model by using an observation system suitable for tunnel exploration environment, and recording wave field data to obtain a wave velocity model and corresponding seismic data.

3. The method of claim 1, wherein the step of computing the background velocity model comprises performing full waveform inversion computation using low frequency signals of the seismic observation data, and computing gradients from the seismic data of the low frequency signals and the seismic source.

4. The method for deep learning inversion of seismic wave velocity for data-space domain fusion according to claim 1, wherein the specific process of performing abnormal body imaging comprises: and taking a low-resolution speed model obtained by full waveform inversion as a background wave speed, and adopting a high-frequency band signal of the seismic data to perform reverse time migration imaging or diffraction superposition to obtain high wave number information and a construction position.

5. The method of seismic wave velocity deep learning inversion of a data-space domain fusion of claim 1, wherein the data-model domain fusion network based on attention weighted intersection comprises a data domain encoder, a convolution space domain encoder and a decoder based on data-model domain attention intersection fusion, wherein the data domain encoder takes seismic observation data as input to extract data features; the convolution space domain encoder takes the background wave velocity and the imaging result as input and extracts space features in the channel dimension; the decoder based on the data-model domain attention cross fusion takes the data and the spatial characteristics as input, performs depth fusion of two types of characteristics of the data domain and the model domain through a weighted attention mechanism, performs up-sampling through the decoder, and restores the characteristic scale to obtain a prediction wave velocity model.

6. The method for deep learning and inverting of seismic wave velocity by fusion of data-space domain as claimed in claim 5, wherein the data domain encoder is composed of 6 sequentially cascaded convolutional network layers;

Or the convolution space domain coder is used for connecting the low-frequency wave velocity model with the high-frequency imaging result as two channels of input data, and then carrying out subsequent processing through 9 sequentially cascaded convolution network layers.

7. A data-space domain fusion seismic wave velocity deep learning inversion method as defined in claim 5 wherein said data-model domain attention cross fusion based decoder has weighted attention mechanisms expressed as:

Wherein the lower right angles m and d represent the model features and the data features, respectively; q, K and V represent the query key, key, and numeric key, respectively, in the attention mechanism.

8. The method of seismic wave velocity deep learning inversion of a data-space domain fusion of claim 7 wherein said decoder is comprised of a series of conv-up blocks cascaded in sequence, each conv-up block comprising two convolution operations, wherein each convolution operation in turn comprises a two-dimensional deconvolution operation, a batch normalization operation, and a ReLU activation function.

9. The method for deep learning and inversion of seismic wave velocity based on data-space domain fusion according to claim 1, wherein the loss function of the data-model domain fusion network based on attention weighted intersection has a calculation formula:

Wherein the method comprises the steps of And M is a predicted wave velocity model matrix and an original wave velocity model matrix respectively, R is the local similarity of the two matrices calculated by different scales, and a calculation formula of the SSIM ^r is as follows:

wherein H and W are divided into the height and width of the model, Represents a window with k point as center and size of r×r in the model, c ₁ and c ₂ are constant terms of stable molecular denominator.

10. A data-space domain fusion seismic wave velocity deep learning inversion system is characterized by comprising:

The wave velocity model construction module is configured to construct a geological model database, and performs numerical simulation by using a wave equation positive algorithm to obtain seismic data;

A computing module configured to compute a background velocity model using the seismic data in the geologic model database;

An imaging module configured to perform abnormal body imaging with the obtained large-scale velocity model as a background velocity;

the prediction module is configured to construct a data-model domain fusion network based on attention weighted intersection, takes the seismic data, a background velocity model and an abnormal body imaging result as the input of the fusion network, and enables the fusion network to mutually influence by using the weights of the data domain and the model domain through a weighted attention mechanism so as to obtain a prediction wave velocity model corresponding to the seismic observation data;

the parameter optimization module is configured to calculate a loss function according to the predicted wave velocity model and the actual value of the wave velocity model, calculate gradient and carry out gradient return, and optimize and update network parameters;

and the inversion module is configured to process the seismic data by using a data-model domain fusion network based on attention weighted intersection after updating parameters to obtain a wave velocity inversion result.