Seismic wave velocity double-driven layer-by-layer inversion network training method, inversion method and system

Abstract

The present application belongs to the field of geophysical exploration, and provides a seismic wave velocity double-drive layer-by-layer inversion network training method, an inversion method and a system. The seismic wave velocity double-drive layer-by-layer inversion network training method comprises the following steps: constructing a data-model domain data set for tunnel advanced geological prediction; performing step-by-step masking operation according to a layer-by-layer inversion strategy and a layered model; inputting the seismic data sample after the masking operation and the time information of the seismic data sample into a seismic wave velocity double-drive inversion network together to estimate the weighted features of the seismic data influence area and predict a velocity model; updating the seismic wave velocity double-drive inversion network; judging the updated focus area, and continuing to perform step-by-step masking operation according to the layer-by-layer inversion strategy and the layered model to train the seismic wave velocity double-drive inversion network until a set stop condition is reached, and a trained seismic wave velocity double-drive inversion network is obtained.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration, and particularly relates to a training method, inversion method and system for a seismic wave velocity dual-driven layer-by-layer inversion network. 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] During tunnel construction in complex geological areas such as high altitudes and deep burial sites, adverse geological conditions such as fault fracture zones, karst, and water-bearing bodies are easily encountered, which can affect the safe construction of tunnels. Tunnel seismic wave early detection methods, with their advantages of sensitivity to structural surfaces and long detection distances, have become one of the commonly used methods for tunnel early detection. Among these methods, the accurate acquisition of tunnel seismic wave velocity, as a key step in accurate imaging using tunnel seismic wave early detection, has always been a focus of research.

[0004] However, due to the limited observation space within tunnels, the effective information obtained from seismic data acquired using tunnel seismic wave detection is limited, posing a greater challenge to the task of inverting wave velocity ahead of the tunnel. In particular, the technique of using direct wave velocity to represent the wave velocity distribution ahead of the tunnel affects the location and accuracy of the imaging interface. Therefore, velocity analysis methods and tomographic imaging methods have been developed and applied, initially alleviating the dependence of imaging on wave velocity. The full waveform inversion method offers higher inversion accuracy and has also been applied in tunnels. However, the noise interference and small offset challenges brought about by the special observation environment of tunnels lead to significant ill-posedness in tunnel wave velocity inversion, affecting the application of the full waveform inversion method in tunnel seismic advance detection.

[0005] Deep learning-based wave velocity inversion methods can effectively obtain relatively accurate wave velocity distributions and are beginning to be applied to wave velocity acquisition in tunnel seismic wave detection. For tunnel seismic wave detection, existing methods are starting to utilize inversion networks with data, background wave velocities, and migration imaging as inputs, while also introducing forward modeling as a physical law-driven approach, jointly improving performance and generalization. However, during network training, existing methods primarily aim to directly recover the entire wave velocity model, lacking effective guidance based on physical laws. This results in high training difficulty and makes it hard to truly grasp the effective mapping relationship between input and output, affecting generalization and practical application effectiveness.

[0006] The inventors discovered that the current method for training tunnel seismic wave velocity dual-drive inversion networks still has problems:

[0007] First, during the training process, how can we analyze the mapping relationship between input and output using the principles of seismic exploration to improve generalization? Existing deep learning tunnel wave velocity inversion methods mostly use seismic data as network input and simultaneously invert the entire wave velocity model. The mapping relationship between the data and the wave velocity model is difficult to learn. The wave velocity variation trend and structural information in the data are implicitly reconstructed through data mining, making it difficult to grasp the effective mapping relationship between input and output. This makes network learning difficult and affects the results and generalization of the wave velocity inversion network.

[0008] Second, after finding the relationship mapping, how to apply this relationship in the network. Summary of the Invention

[0009] To address the technical problems mentioned above, this invention provides a training method, inversion method, and system for a seismic wave velocity dual-drive layer-by-layer inversion network. Based on the characteristics of seismic inversion, which involves the correspondence between previous data and near-distance wave velocities, and the layer-by-layer inversion from near to far, this invention proposes a wave velocity layer-by-layer inversion strategy that focuses on the stratigraphic structure, thereby reducing the learning difficulty of the inversion network.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first aspect of the present invention provides a method for training a seismic wave velocity dual-drive layer-by-layer inversion network.

[0012] A method for training a layer-by-layer inversion network with dual-driven seismic wave velocity includes:

[0013] Constructing a data-model domain dataset for advanced geological prediction of tunnels;

[0014] Seismic data samples from the data-model domain dataset are retrieved and progressively masked according to the layer-by-layer inversion strategy and the layered model before being input into the seismic wave velocity dual-drive inversion network.

[0015] The masked seismic data samples, along with their time information, are input into the seismic wave velocity dual-drive inversion network to estimate the weighted characteristics of the seismic data's affected area and predict the velocity model.

[0016] The weights in the loss function are adjusted based on the weighted characteristics of the affected area of ​​earthquake data to update the dual-drive inversion network for seismic wave velocity.

[0017] The focus area for updating is determined, and the seismic wave velocity dual-drive inversion network is trained by performing step-by-step masking operations according to the layer-by-layer inversion strategy and layered model until the set stopping condition is met, resulting in the trained seismic wave velocity dual-drive inversion network.

[0018] As one implementation method, the process of performing stepwise masking operations on data samples according to the layer-by-layer inversion strategy and layered model is as follows:

[0019] For the wave velocity of n types of media in the data-model domain, calculate the rounded-up value of i as n / 5;

[0020] In the first m rounds, focus on the first layer of the model medium and the subsequent wave velocities and the first reflection signal in the seismic data to perform masking;

[0021] In the m1 to m2 rounds of focusing, the reflection signals of channel (1-2i) in the model medium of layer (1-2i) and the subsequent wave velocity and seismic data are masked;

[0022] In the m2 to m3 rounds of focusing, the reflection signals of channel (1-3i) in the model medium of layer (1-3i) and the subsequent wave velocity and seismic data are masked;

[0023] Masking is performed on the entire wave velocity model and seismic data in rounds m3 to m4; where n, m, and m1 to m4 are all positive integers, and m <m1<m2<m3<m4。

[0024] As one implementation, the data-model domain dataset includes a seismic wave velocity model, seismic acquisition data, a background model, and imaging results.

[0025] In one implementation, the seismic wave velocity dual-drive inversion network includes a data domain encoder, a spatial domain encoder, and a wave velocity decoder; the data domain encoder is used to extract seismic data features.

[0026] The spatial domain encoder is used to extract background wave velocity and offset imaging features;

[0027] The wave velocity decoder is used to integrate seismic data features with background wave velocity and migration imaging features to reconstruct the true wave velocity model.

[0028] As one implementation method, the loss function includes a model loss function and a data loss function. During network training, dynamic masks are applied to the model loss function and the data loss function respectively to guide them, so as to realize a tunnel seismic wave velocity dual-drive inversion network with interface focusing layer by layer.

[0029] A second aspect of the present invention provides a seismic wave velocity dual-drive layer-by-layer inversion network training system.

[0030] A seismic wave velocity dual-driven layer-by-layer inversion network training system includes:

[0031] The dataset building module is used to build a data-model domain dataset for tunnel advanced geological prediction.

[0032] The masking module is used to retrieve seismic data samples from the data-model domain dataset and perform stepwise masking operations according to the layer-by-layer inversion strategy and the layered model before inputting them into the seismic wave velocity dual-drive inversion network.

[0033] The velocity model prediction module is used to input the masked seismic data samples along with the time information of the seismic data samples into the seismic wave velocity dual-drive inversion network, estimate the weighted characteristics of the seismic data influence area, and predict the velocity model.

[0034] The network parameter update module is used to adjust the weights in the loss function based on the weighted characteristics of the affected area of ​​the seismic data in order to update the seismic wave velocity dual-drive inversion network.

[0035] The focus area update module is used to determine the focus area to be updated and to continue training the seismic wave velocity dual-drive inversion network by performing step-by-step masking operations according to the layer-by-layer inversion strategy and the layered model until the set stopping condition is met, and the trained seismic wave velocity dual-drive inversion network is obtained.

[0036] A third aspect of the present invention provides a seismic inversion method.

[0037] A seismic inversion method includes:

[0038] A dual-drive seismic wave velocity inversion network is used to construct a mapping between seismic data and velocity models;

[0039] Based on the mapping between seismic data and velocity models, the velocity model corresponding to the seismic data is obtained, so as to realize the inversion of seismic data;

[0040] The seismic wave velocity dual-drive inversion network is obtained through the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0041] A fourth aspect of the present invention provides a seismic inversion system.

[0042] A seismic inversion system, comprising:

[0043] The mapping building block is used to construct a mapping between seismic data and velocity models using a seismic wave velocity dual-drive inversion network.

[0044] The data inversion module is used to obtain the velocity model corresponding to the seismic data based on the mapping between seismic data and velocity model, so as to realize the inversion of seismic data.

[0045] The seismic wave velocity dual-drive inversion network is obtained through the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0046] A fifth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0047] Alternatively, when the program is executed by the processor, it may implement the steps in the seismic inversion method described above.

[0048] A sixth aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0049] Alternatively, the processor may execute the steps in the seismic inversion method described above when executing the program.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] (1) In view of the problem that traditional inversion networks invert the entire model at the same time, making it difficult to learn the data-to-model mapping relationship, this invention proposes a tunnel seismic wave velocity inversion network training strategy with interface layer-by-layer focus. This strategy is based on the characteristics of tunnel seismic inversion, which has the characteristics of early data corresponding to near-distance wave velocity and inversion from near to far layer by layer, thus reducing the nonlinearity and difficulty of the inversion task.

[0052] (2) In view of the problem that traditional tunnel seismic inversion networks cannot incorporate an effective mapping of the relationship between seismic data and velocity models in tunnel inversion methods, this invention proposes to introduce a dynamic mask loss function during network training, which helps the wave velocity inversion network find the correlation between data and wave velocity, and realizes a tunnel seismic wave velocity dual-drive inversion network training method with interface layer-by-layer focusing, providing a feasible means for more accurate inversion of measured data.

[0053] (3) In view of the problem that the spatial relationship between data and model is weak in traditional inversion networks, especially the narrow space of tunnel seismic observation and the lack of effective information, resulting in strong ill-posedness and serious false anomalies in the inversion results, this invention proposes a strategy of introducing data time-to-time extraction and spatial domain information fusion in the network design, implicitly including the data-interface mapping process in the network input, realizing the layer-by-layer inversion of tunnel seismic wave velocity based on multi-modal fusion, and providing a feasible means for more accurate inversion of measured data;

[0054] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0055] 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.

[0056] Figure 1 This is a flowchart of the seismic wave velocity dual-drive layer-by-layer inversion network training method according to an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the wave velocity layer-by-layer inversion strategy for focusing on stratigraphic construction according to an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the mask network architecture for data and models in an embodiment of the present invention;

[0059] Figure 4(a) is a schematic diagram of the actual model in an embodiment of the present invention;

[0060] Figure 4(b) is a diagram showing the effect of the training set inverting layer by layer after 20 rounds of network training in an embodiment of the present invention;

[0061] Figure 4(c) is a diagram showing the layer-by-layer inversion effect of the training set after 40 rounds of network training in an embodiment of the present invention;

[0062] Figure 4(d) is a diagram showing the layer-by-layer inversion effect of the training set after 60 rounds of network training in an embodiment of the present invention;

[0063] Figure 4(e) is a diagram showing the effect of the training set inverting layer by layer after 80 rounds of network training in an embodiment of the present invention;

[0064] Figure 4(f) is a diagram showing the effect of the training set being inverted layer by layer after 100 rounds of network training in an embodiment of the present invention. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] 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.

[0067] 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.

[0068] Example 1

[0069] according to Figure 1 This embodiment provides a method for training a layer-by-layer inversion network with dual-driven seismic wave velocity, which specifically includes the following steps:

[0070] Step 1: Construct a data-model domain dataset for advanced geological prediction of tunnels;

[0071] Specifically, based on existing wave velocity modeling methods and numerical modeling methods such as forward modeling, low-frequency full waveform inversion, and migration imaging, a data-model domain dataset suitable for tunnel advanced geological prediction is constructed, including seismic wave velocity models, seismic acquisition data, background models, imaging results, and other data.

[0072] Based on the actual geological conditions, the corresponding simulated seismic observation data is calculated through numerical simulation, the corresponding large-scale wave velocity model is calculated based on low-frequency seismic data, and the migration imaging obtained by high-frequency signal calculation is used to form a data-model domain dataset for tunnel advanced geological prediction.

[0073] Specifically, numerical simulations are performed on underground geological layered wave velocity models with various undulating structures, wave velocities, and / or different numbers of layers. When constructing corresponding geological wave velocity models based on actual geological conditions, wave field simulations are performed for each geological wave velocity model with fixed source, detector positions, and observation times. Wave field data are recorded at the detector positions to obtain actual seismic data corresponding to the geological wave velocity models.

[0074] In conjunction with the target of tunnel seismic wave method advance detection, namely adverse geological structures, this example mainly focuses on training the inversion network for layered geological models. During the training process, the focus is achieved layer by layer on each interface. Therefore, the wave velocity model in front of the tunnel is simplified into multiple layered media with different wave velocities, and the layer interfaces are simplified into straight or curved interfaces, including a maximum of 5 different wave velocities, that is, a maximum of 4 interfaces.

[0075] In this example, the actual size of the inversion target is 80m × 150m, with the first 80m × 50m representing the tunnel area. A 10m × 50m low-velocity body is used to simulate the tunnel. The model area is then meshed using a 1m network to obtain 80 × 150 networks to construct the wave velocity model. Wave velocities consistent with tunnel conditions are assigned to various media. A 200Hz dominant frequency Ricker wavelet is used as the seismic source, excited at the tunnel sidewall, and receivers are set up at the sidewall to collect 0.2s of seismic data. Low-frequency data is then used for full waveform inversion, and high-frequency data is used for reverse time migration imaging to obtain other information. Finally, a dataset of 10,000 data-model domain sets is constructed.

[0076] Step 2: Retrieve seismic data samples from the data-model domain dataset and perform stepwise masking operations according to the layer-by-layer inversion strategy and layered model before inputting them into the seismic wave velocity dual-drive inversion network;

[0077] In one implementation, the seismic wave velocity dual-drive inversion network includes a data domain encoder, a spatial domain encoder, and a wave velocity decoder; the data domain encoder is used to extract seismic data features.

[0078] The spatial domain encoder is used to extract background wave velocity and offset imaging features;

[0079] The wave velocity decoder is used to integrate seismic data features with background wave velocity and migration imaging features to reconstruct the true wave velocity model.

[0080] A dual-drive seismic wave velocity inversion network is constructed to separately process and effectively fuse data domain information (potential features of seismic data) and model domain information (wave velocity and imaging) in order to simulate the seismic wave velocity inversion mapping.

[0081] Among them, such as Figure 2 As shown, the process of performing step-by-step masking on data samples according to the layer-by-layer inversion strategy and layered model is as follows:

[0082] For the wave velocity of n types of media in the data-model domain, calculate the rounded-up value of i as n / 5;

[0083] In the first m rounds, focus on the first layer of the model medium and the subsequent wave velocities and the first reflection signal in the seismic data to perform masking;

[0084] In the m1 to m2 rounds of focusing, the reflection signals of channel (1-2i) in the model medium of layer (1-2i) and the subsequent wave velocity and seismic data are masked;

[0085] In the m2 to m3 rounds of focusing, the reflection signals of channel (1-3i) in the model medium of layer (1-3i) and the subsequent wave velocity and seismic data are masked;

[0086] Masking is performed on the entire wave velocity model and seismic data in rounds m3 to m4; where n, m, and m1 to m4 are all positive integers, and m <m1<m2<m3<m4。

[0087] For example, in this embodiment, a total of 100 rounds of wave velocity inversion network training are conducted. Based on the layer-by-layer inversion strategy and layered model proposed in this paper, for wave velocities in n media types, i = n / 5 (rounded up) is calculated. In the first 20 rounds, the focus is on the wave velocities of the first media layer and its subsequent components, as well as the first reflection signal, for masking. Rounds 20-40 focus on the wave velocities of the 1st-2i media layers and their subsequent components, as well as the 1st-2i reflection signals, for masking. Rounds 40-60 focus on the wave velocities of the 1st-3i media layers and their signals, for masking. Rounds 60-100 focus on the entire wave velocity model and seismic data, for masking. Masking involves multiplying the input data and model components by a matrix of equal size, with the focused portion having a value of 1 and other portions having a value of 0, to preserve the focused portion of model and data information as input.

[0088] Step 3: Input the masked seismic data samples along with the time information of the seismic data samples into the seismic wave velocity dual-drive inversion network to estimate the weighted characteristics of the seismic data influence area and predict the velocity model;

[0089] An additional channel is added to provide the time information of the seismic data along with the input masked seismic data. This encoded data is then input into the wave velocity inversion network for processing. For example... Figure 3 As shown, when fusing data features and model features using a wave velocity decoder, the affected area of ​​the seismic data is estimated based on the background wave velocity values ​​to perform targeted weighted updates of the features. The final inversion result is then obtained.

[0090] In this example, a data encoder is used to compress and extract data features of size 100×48 from seismic data of size 2×6×1600; similarly, model information (background wave velocity and imaging results) of size 2×80×100 is compressed and extracted to model features of size 100×48; when cross-weighting the data features and model features, the background wave velocity value V0 is used as the wave velocity, and the data influence range is calculated based on the arrival time of the seismic data. Weights are set during cross-weighting to guide the decoder to focus more on the effective area; ultimately, a more accurate prediction of the inversion results is achieved.

[0091] Step 4: Adjust the weights in the loss function based on the weighted characteristics of the affected area of ​​the seismic data to update the dual-drive seismic wave velocity inversion network.

[0092] In this embodiment, a data-model domain wave velocity inversion network based on a convolutional neural network is constructed to extract and fuse features from seismic data and model domain information separately, strengthening the weight relationship between the two. Based on this, an additional channel is introduced to perform temporal calibration of the seismic data, and combined with the input background wave velocity information, an estimation region weighting is constructed, ultimately achieving wave velocity inversion result prediction under human guidance.

[0093] Partial loss function calculations are performed on seismic data and wave velocity models. Areas of interest are masked, meaning non-interested areas are not calculated, to achieve layer-by-layer dual-driven wave velocity inversion. Based on iterative calculations during the wave velocity inversion network training process, the focused interface location is gradually expanded, and the wave velocity inversion network is dynamically optimized. After the wave velocity inversion network training is complete, the background model and imaging results are calculated using field data to predict the wave velocity inversion results.

[0094] The data-model domain wave velocity inversion network can be composed of convolutional network layers, fully connected network layers, and Transformer network layers, and is constructed according to the characteristics of the data being processed.

[0095] The model loss function Φ used to update the network v and data loss function Φ d Both are used together to calculate gradients and update network parameters, enabling a dual-driven training strategy that combines data-driven and physical law-driven approaches. Their weights are adjusted appropriately based on orders of magnitude difference. The model loss function Φ v It is achieved through a mask M that includes the weights of the layer interfaces. i After multiplying the labels and output results simultaneously, the mean squared error and multi-scale structural similarity are calculated separately. This mask guides the wave velocity inversion network to focus on the forward wave velocity information by increasing the weight of the layer of interest i and the positions before it, and decreasing the weight of the interface layers behind it. The calculation formula is as follows:

[0096]

[0097] Where m is the true wave velocity model, i.e., the label, m′ is the wave velocity inversion network prediction result, and MSSIM is multi-scale structural similarity. The model loss directly affects the network prediction result; therefore, the model loss function can directly obtain the gradient with respect to the inversion network parameters and update the inversion network.

[0098] According to the loss function Φ d By calculating the estimated arrival time t(i) of the i-th layer model, the arrival time is appropriately extended. The normal performance is always in progress. Seismic data was acquired using the observation matrix R() and compared with the actual observation data d. t(i) Calculate the error to determine the data loss function for the current layer. The calculation formula is as follows:

[0099]

[0100] A velocity model m is obtained by predicting wave velocity results using a wave velocity inversion network. The prediction results and corresponding data are evaluated using a loss function. The loss function is weighted and calculated. Regions that are not of interest are masked, i.e., not calculated, in order to achieve layer-by-layer wave velocity inversion.

[0101] In this example, the following data loss function and model loss function are used for calculation:

[0102]

[0103]

[0104] Among them, for the data loss function Φ d By calculating the estimated arrival time t(i) of the i-th layer model, the arrival time is appropriately extended. Forward modeling is performed at each step, and the error is calculated against the observed data to determine the data loss function for the current layer. For the model loss function Φ... v This is achieved through a mask M that includes the layer interface weights. i After multiplying the labels and output results simultaneously, the mean squared error and multi-scale structural similarity are calculated separately. This mask guides the wave velocity inversion network to focus on the forward wave velocity information by increasing the weights of the focus layer i and its preceding positions, and decreasing the weights of the subsequent interface layers. The model loss function directly affects the network's prediction results, thus directly obtaining the gradient of the inversion network parameters and updating the inversion network. The data loss function, calculated from the forward modeled prediction results, essentially first calculates the gradient of the prediction results and then updates the wave velocity inversion network parameters.

[0105] Step 5: Determine the update focus area, and continue to train the seismic wave velocity dual-drive inversion network by performing step-by-step masking operations according to the layer-by-layer inversion strategy and layered model until the set stopping condition is met, and obtain the trained seismic wave velocity dual-drive inversion network.

[0106] As the network training process updates the layer interface i that focuses on the inversion results, the network training process is finally completed and validated. The final inversion result is obtained, as shown below. Figures 4(a)-4(f) As shown, it can be seen that the network training process indeed achieves a layer-by-layer focusing inversion of the interface. The gradient-optimized wave velocity layer-by-layer inversion network constructs a mapping relationship between seismic data and the geological velocity model, which can represent the inversion process. Partial results substituted into the test set are shown below. Figures 4(a)-4(f) As shown. From Figures 4(a)-4(f)It can be seen that the first 20 iterations focused on inverting the first layer of the model, and the wave velocity value of the first layer was well inverted, with its shape being close to the real velocity model. As the network training iterated, from the 20th to the 40th iterations, the focus shifted to inverting the first and second layers. The shape and position of the first layer were corrected, and the wave velocity value of the second layer was more accurate. Then, as can be seen from Figure 4(d), from the 40th to the 60th iterations, the first three layers were mainly inverted. The inversion results of the first and second layers were further improved, and the increasing trend of the wave velocity of the third layer was also observed. After the 60th to the 80th iterations, the results of the first two layers did not change much, the shape and value of the high-speed interlayer of the third layer were well inverted, and the wave velocity value of the fourth layer was close to the real model. After 100 iterations, the overall inversion results were further improved, the inverted wave velocity values ​​of each layer were well corrected, and results that were more consistent with the real velocity model were obtained.

[0107] The main network parameters and hardware requirements in this embodiment are as follows: computation is performed on a graphics card with four 40GB VRAM chips. The network is built on the PyTorch platform, the Adam optimizer has a batch size of 16, and the learning rate for gradient optimization of network parameters is set to 5e-4.

[0108] This embodiment calculates the model loss function and the data loss function. The loss function, as the objective function for optimization, guides the direction of the deep neural network. During network training, dynamic masks are applied to the model loss function and the data loss function at different training stages to artificially guide the network, achieving a layer-by-layer focused tunnel seismic wave velocity dual-drive inversion network. By increasing the weight of the preceding strata and decreasing the weight of the following interface layers in the model loss function through the mask, the wave velocity inversion network is guided to focus on the wave velocity information of the preceding strata, achieving the effect of layer-by-layer focused inversion. For the data loss function, the region of interest is masked, i.e., the non-interested region is not calculated, to reduce the computational load and achieve the effect of layer-by-layer wave velocity inversion. The layer-by-layer focused tunnel seismic wave velocity dual-drive inversion network constructs a mapping between seismic data and velocity models, and obtains the velocity model based on the acquired seismic data, realizing the inversion of seismic data.

[0109] This embodiment proposes a layer-by-layer wave velocity inversion strategy that focuses on stratigraphic structures, based on the characteristics of seismic inversion where earlier data corresponds to near-distance wave velocities and inversion proceeds layer by layer from near to far. This reduces the learning difficulty of the inversion network. Furthermore, the network design incorporates a strategy that fuses temporal and spatial domain information, implicitly including a data-interface mapping process in the network input. A dynamic mask loss function is introduced during network training to enable the network to focus on inverted signals from earlier interfaces in the early stages and gradually recover the entire wave velocity model in later stages, thereby improving the generalization ability of the inversion network.

[0110] Example 2

[0111] This embodiment provides a seismic wave velocity dual-driven layer-by-layer inversion network training system, which specifically includes the following modules:

[0112] The dataset building module is used to build a data-model domain dataset for advanced geological prediction of tunnels;

[0113] The masking module is used to retrieve seismic data samples from the data-model domain dataset and perform stepwise masking operations according to the layer-by-layer inversion strategy and the layered model before inputting them into the seismic wave velocity dual-drive inversion network.

[0114] The velocity model prediction module is used to input the masked seismic data samples along with the time information of the seismic data samples into the seismic wave velocity dual-drive inversion network, estimate the weighted characteristics of the seismic data influence area, and predict the velocity model.

[0115] The network parameter update module is used to adjust the weights in the loss function based on the weighted characteristics of the affected area of ​​the seismic data in order to update the seismic wave velocity dual-drive inversion network.

[0116] The focus area update module is used to determine the focus area to be updated and to continue training the seismic wave velocity dual-drive inversion network by performing step-by-step masking operations according to the layer-by-layer inversion strategy and the layered model until the set stopping condition is met, and the trained seismic wave velocity dual-drive inversion network is obtained.

[0117] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0118] Example 3

[0119] This embodiment provides a seismic inversion method, including:

[0120] A dual-drive seismic wave velocity inversion network is used to construct a mapping between seismic data and velocity models;

[0121] Based on the mapping between seismic data and velocity models, the velocity model corresponding to the seismic data is obtained, so as to realize the inversion of seismic data;

[0122] The seismic wave velocity dual-drive inversion network is obtained through the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0123] Example 4

[0124] This embodiment provides a seismic inversion system, which includes:

[0125] The mapping building module is used to construct a mapping between seismic data and velocity models using a dual-drive seismic wave velocity inversion network.

[0126] The data inversion module is used to obtain the velocity model corresponding to the seismic data based on the mapping between seismic data and velocity model, so as to realize the inversion of seismic data.

[0127] The seismic wave velocity dual-drive inversion network is obtained through the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0128] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment three, and their specific implementation process is the same, so it will not be repeated here.

[0129] Example 5

[0130] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0131] Alternatively, when the program is executed by the processor, it may implement the steps in the seismic inversion method described above.

[0132] Example 6

[0133] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method described above.

[0134] Alternatively, the processor may execute the steps in the seismic inversion method described above when executing the program.

[0135] 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, as well as 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. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] 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 within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a layer-by-layer inversion network with dual-driven seismic wave velocity, characterized in that, include: Constructing a data-model domain dataset for advanced geological prediction of tunnels; Seismic data samples from the data-model domain dataset are retrieved and progressively masked according to the layer-by-layer inversion strategy and the layered model before being input into the seismic wave velocity dual-drive inversion network. The masked seismic data samples, along with their time information, are input into the seismic wave velocity dual-drive inversion network to estimate the weighted characteristics of the seismic data's affected area and predict the velocity model. The weights in the loss function are adjusted based on the weighted characteristics of the affected area of ​​earthquake data to update the dual-drive inversion network for seismic wave velocity. The focus area for updating is determined, and the seismic velocity dual-drive inversion network is trained by progressively masking the data samples according to the layer-by-layer inversion strategy and the layered model until the set stopping condition is met, resulting in the trained seismic velocity dual-drive inversion network. The process of progressively masking the seismic data samples according to the layer-by-layer inversion strategy and the layered model is as follows: For the wave velocity of n types of media in the data-model domain, calculate the rounded-up value of i as n / 5; In the first m rounds, focus on the first layer of the model medium and the subsequent wave velocities and the first reflection signal in the seismic data to perform masking; In the m1~m2 rounds of focusing, the reflection signals of the (1-2i) channel in the model medium of the (1-2i) layer and the subsequent wave velocity and seismic data are masked; In the m2~m3 round focusing, the reflection signals of channel (1-3i) in the model medium of layer (1-3i) and the subsequent wave velocity and seismic data are masked; Masking is performed on the entire wave velocity model and seismic data in rounds m3 to m4; where n, m, and m1 to m4 are all positive integers, and m <m1<m2<m3<m4。 2. The seismic wave velocity dual-drive layer-by-layer inversion network training method as described in claim 1, characterized in that, The data-model domain dataset includes seismic wave velocity models, seismic acquisition data, background models, and imaging results.

3. The seismic wave velocity dual-drive layer-by-layer inversion network training method as described in claim 1, characterized in that, The seismic wave velocity dual-drive inversion network includes a data domain encoder, a spatial domain encoder, and a wave velocity decoder; the data domain encoder is used to extract seismic data features. The spatial domain encoder is used to extract background wave velocity and offset imaging features; The wave velocity decoder is used to integrate seismic data features with background wave velocity and migration imaging features to reconstruct the true wave velocity model.

4. The seismic wave velocity dual-drive layer-by-layer inversion network training method as described in claim 1, characterized in that, The loss function includes a model loss function and a data loss function. During network training, dynamic masks are applied to the model loss function and the data loss function respectively to guide them, so as to realize a tunnel seismic wave velocity dual-drive inversion network with interface focusing layer by layer.

5. A seismic wave velocity dual-drive layer-by-layer inversion network training system, characterized in that, include: The dataset building module is used to build a data-model domain dataset for advanced geological prediction of tunnels; The masking module is used to retrieve seismic data samples from the data-model domain dataset and perform stepwise masking operations according to the layer-by-layer inversion strategy and the layered model before inputting them into the seismic wave velocity dual-drive inversion network. The velocity model prediction module is used to input the masked seismic data samples along with the time information of the seismic data samples into the seismic wave velocity dual-drive inversion network, estimate the weighted characteristics of the seismic data influence area, and predict the velocity model. The network parameter update module is used to adjust the weights in the loss function based on the weighted characteristics of the affected area of ​​the seismic data in order to update the seismic wave velocity dual-drive inversion network. The focus area update module is used to determine the focus area for updating and to continue training the seismic velocity dual-drive inversion network by performing progressive masking operations according to the layer-by-layer inversion strategy and the layered model until the set stopping condition is met, resulting in a trained seismic velocity dual-drive inversion network. The process of progressively masking the seismic data samples according to the layer-by-layer inversion strategy and the layered model is as follows: For the n types of medium wave velocities in the data-model domain data, calculate i as the rounded-up value of n / 5; in the first m rounds, focus on the first layer of model medium and its subsequent wave velocities and the first reflection signal in the seismic data for masking; in rounds m1 to m2, focus on the (1-2i)th layer of model medium and its subsequent wave velocities and the (1-2i)th reflection signal in the seismic data for masking. In rounds m2 to m3, the reflection signals of channels (1-3i) in the model medium of layer (1-3i) and subsequent wave velocity and seismic data are masked; in rounds m3 to m4, the entire wave velocity model and seismic data are masked; where n, m, and m1~m4 are all positive integers, and m <m1<m2<m3<m4。 6. A seismic inversion method, characterized in that, include: A dual-drive seismic wave velocity inversion network is used to construct a mapping between seismic data and velocity models; Based on the mapping between seismic data and velocity models, the velocity model corresponding to the seismic data is obtained, so as to realize the inversion of seismic data; The seismic wave velocity dual-drive inversion network is obtained by the seismic wave velocity dual-drive layer-by-layer inversion network training method as described in any one of claims 1-4.

7. A seismic inversion system, characterized in that, include: The mapping building module is used to construct a mapping between seismic data and velocity models using a dual-drive seismic wave velocity inversion network. The data inversion module is used to obtain the velocity model corresponding to the seismic data based on the mapping between seismic data and velocity model, so as to realize the inversion of seismic data. The seismic wave velocity dual-drive inversion network is obtained by the seismic wave velocity dual-drive layer-by-layer inversion network training method as described in any one of claims 1-4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method as described in any one of claims 1-4; Alternatively, when the program is executed by a processor, it implements the steps in the seismic inversion method as described in claim 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the seismic wave velocity dual-drive layer-by-layer inversion network training method as described in any one of claims 1-4. Alternatively, the processor may execute the steps in the seismic inversion method as described in claim 6 when executing the program.

Images