Electromagnetic seismic joint inversion method and device based on multi-scale feature fusion network

Through the electromagnetic seismic joint inversion method of multi-scale feature fusion network, the problem of increasing convergence difficulty and computational volume of structural constraint loss setting and optimization process in traditional methods is solved, and the fine imaging and information coupling utilization of underground complex structures is realized. Improved inversion accuracy and resolution.

CN118884558APending Publication Date: 2024-11-01TSINGHUA UNIVERSITY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410158631.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-02-04
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

The traditional electromagnetic seismic joint inversion method has difficulty converging in the setting and optimization of structural constraint losses, and the numerical differences of different inversion systems lead to an increase in the calculation amount and limit its wide application.

Method used

Using a multi-scale feature fusion network method, the resistivity model and velocity model with random faults and salt hill structures are generated, and forward simulation is performed. The multi-scale feature fusion network model is used for iterative training, combined with a decoder and an encoder Data is encoded and fused, the loss value is calculated to output the resistivity and velocity model of joint inversion, and the model is optimized under the learning rate adjustment strategy.

Benefits of technology

Automatic inversion of the velocity model of complex underground structures is realized, the accuracy and resolution of underground structure imaging is improved, the multi-solvency of inversion is reduced, the information coupling utilization of electromagnetic and seismic data is improved, and the model training process is simplified.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118884558A_ABST
    Figure CN118884558A_ABST
Patent Text Reader

Abstract

The invention provides an electromagnetic seismic joint inversion method and device based on a multi-scale feature fusion network, and belongs to the technical field of geophysical inversion. The method comprises the following steps: generating a resistivity model and a velocity model with a random fault and a salt hill structure; performing forward modeling on the speed model and the resistivity model to obtain a sample data set; performing iterative training on a multi-scale feature fusion network model by using the sample data set to obtain a joint inversion network model; and performing joint inversion on the seismic electromagnetic data by using the joint inversion network model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the field of geophysical inversion, and in particular relates to an electromagnetic-seismic joint inversion method and device based on a multi-scale feature fusion network. Background Art

[0002] In geophysics, inverting the physical parameters of the earth medium based on the observed geophysical data is an important means of interpreting geological problems and developing resources. Seismic exploration is one of the main means of geophysical exploration. Using seismic data inversion, especially full waveform inversion (FWI), a high-resolution underground velocity estimate can be obtained. However, using only seismic data to invert underground velocity generally has serious multi-solution problems. Electromagnetic exploration can effectively assist seismic exploration due to its wide detection range and low cost. Although electromagnetic data has a low resolution, it is sensitive to underground anomalies and can penetrate high-resistance bodies. Therefore, electromagnetic data contains deep resistivity information. Therefore, using electromagnetic data and seismic data for joint inversion can effectively integrate the complementary information in the two data, reduce the multi-solution problem of inversion, and improve the accuracy and reliability of inversion.

[0003] The purpose of joint inversion is to fuse complementary information in geophysical data. According to different ways of fusing physical property model information, joint inversion can be divided into three categories: 1) model fusion technology based on the comparison of different physical property models; 2) sequential joint inversion technology that first performs a single inversion and then uses the respective inversion results as constraints to continue iterative inversion; 3) synchronous joint inversion technology that incorporates different physical property parameters into the same objective function for inversion. Since the degree of fusion of different types of data information is the highest in synchronous joint inversion technology, synchronous joint inversion is a research hotspot in geophysical joint inversion today.

[0004] There are many published research works based on the joint inversion of electromagnetic and seismic data. Haber et al. proposed to use model curvature information to construct similarity metrics for different models, but it is necessary to repeatedly adjust the curvature matching parameters to obtain the most appropriate value. Gallardo et al. proposed a structural constraint based on the cross-gradient function. This robust structural constraint method can better couple the information between different physical parameters without too much parameter adjustment, so it has been widely used. In 2008, Colombo et al. used electromagnetic and seismic for simultaneous joint inversion, and used the cross-gradient method as a structural constraint added to the objective function, realizing sub-shallow imaging in the Columbia River Basin in the pre-stack depth migration workflow.

[0005] In order to balance the role of structural constraint loss in the joint inversion objective function, Evan Schankee Um et al. proposed a joint inversion method that separates resistivity-velocity inversion and structural constraint inversion. By repeating single seismic inversion, single electromagnetic inversion and cross-gradient inversion, the role of structural constraint in the joint inversion objective function was balanced, and better subsalt imaging results were obtained. Subsequently, Bennington et al. and Ji Gao et al. made certain improvements to the cross-gradient constraint method. However, it is difficult for traditional methods to set the proportion of structural constraint loss in the optimization objective, which increases the difficulty of convergence. At the same time, the numerical values ​​of different inversion systems vary greatly. The combination of objective functions and cross-gradient functions of different inversion systems requires complex design, and the amount of calculation is greatly increased compared with single inversion, which limits its ability to be widely used.

[0006] With the development of deep learning technology, joint inversion research using deep learning methods has become a hot topic. Yen Sun et al. proposed a method that directly uses U-Net neural network to learn the mapping from seismic shot records and controlled source electromagnetic data (CSEM) to salt dome geometry. However, this scheme only achieves a preliminary segmentation of the underground salt dome structure. Daniele Colombo et al. proposed a deep learning iterative joint inversion method based on physical drive and data drive. This method reduces the requirements for data volume and proves the application potential of deep learning methods in the field of joint inversion. Subsequently, Y.Hu et al. proposed using the network to update the inversion model of the traditional method and use operations such as interpolation to match data of different resolutions. However, simple interpolation operations cannot match different types of data well, and this method only uses the inversion model of the traditional method as input and the real model as a label. The relevant electromagnetic and seismic data are not input, and the coupling between seismic electromagnetic information is not fully explored, which limits the generalization ability of the network. Summary of the invention

[0007] The present application provides an electromagnetic-seismic joint inversion method and device based on a multi-scale feature fusion network.

[0008] Some embodiments of the present application provide an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network, the method comprising:

[0009] Generate resistivity and velocity models with random fault and salt dome structures;

[0010] Performing forward simulation on the velocity model and the resistivity model to obtain a sample data set;

[0011] Using the sample data set, iteratively training the multi-scale feature fusion network model to obtain a joint inversion network model;

[0012] The joint inversion network model is used to perform joint inversion on seismic electromagnetic data.

[0013] Optionally, the multi-scale feature fusion network includes: a decoder and at least two encoders; the sample data set includes: seismic data and electromagnetic data;

[0014] The step of iteratively training the multi-scale feature fusion network model using the sample data set to obtain the joint inversion network model includes:

[0015] Performing migration processing on the seismic data using an initial velocity model;

[0016] Encoding the electromagnetic data using a first encoder to obtain electromagnetic encoding features, and encoding the migrated seismic data and a corresponding velocity model using a second encoder to obtain seismic encoding features;

[0017] Inputting the electromagnetic coding feature and the seismic coding feature into the decoder to obtain a fusion feature;

[0018] When the fusion feature meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0019] Optionally, the first encoder is an RDN encoder, the second encoder is a U-Net encoder, and the decoder is an MSFF shared encoder.

[0020] Optionally, after the step of encoding the electromagnetic data using the first encoder to obtain electromagnetic coding features, and encoding the migrated seismic data and the corresponding velocity model using the second encoder to obtain seismic coding features, the method further comprises:

[0021] Performing a frequency band alignment operation on the electromagnetic coding feature and the seismic coding feature to obtain an electromagnetic coding feature and a seismic coding feature with consistent frequency bands;

[0022] The electromagnetic coding features and seismic coding features with consistent frequency bands are sampled at different scales, and the electromagnetic coding features and seismic coding features obtained by sampling at different scales are fused and matched.

[0023] Optionally, when the fusion feature meets the loss value requirement, the step of outputting the resistivity model and velocity model of the joint inversion includes:

[0024] Calculate a first loss value, a second loss value and a third loss value based on the fusion feature, wherein the first loss value is used to measure the difference between the inverted resistivity model and velocity model and the true velocity model, the second loss value is used to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and the third loss value is used to constrain the smoothness of the output;

[0025] Combine the first loss value, the second loss value, and the third loss value to obtain a total loss value;

[0026] When the total loss value meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0027] Optionally, after the step of combining the first loss value, the second loss value, and the third loss value to obtain a total loss value, the method further includes:

[0028] When the total loss value does not meet the loss value requirement, a cosine learning rate adjustment strategy is used to increase the learning rate of the model, and then the resistivity model and the velocity model after the learning rate adjustment are iteratively trained.

[0029] Optionally, the step of generating a resistivity model and a velocity model having random fault and salt dome structures comprises:

[0030] Generate resistivity models with random fault and salt dome structures;

[0031] The resistivity of the resistivity model is converted into a velocity value using an empirical formula to obtain a velocity model.

[0032] Some embodiments of the present application provide an electromagnetic-seismic joint inversion device based on a multi-scale feature fusion network, the device comprising:

[0033] Dataset preparation module for generating resistivity and velocity models with random fault and salt dome structures;

[0034] Performing forward simulation on the velocity model and the resistivity model to obtain a sample data set;

[0035] A model training module, used to iteratively train the multi-scale feature fusion network model using the sample data set to obtain a joint inversion network model;

[0036] The joint inversion module is used to perform joint inversion on seismic electromagnetic data using the joint inversion network model.

[0037] Optionally, the multi-scale feature fusion network includes: a decoder and at least two encoders; the sample data set includes: seismic data and electromagnetic data;

[0038] The model training module is also used for:

[0039] Performing migration processing on the seismic data using an initial velocity model;

[0040] Encoding the electromagnetic data using a first encoder to obtain electromagnetic encoding features, and encoding the migrated seismic data and a corresponding velocity model using a second encoder to obtain seismic encoding features;

[0041] Inputting the electromagnetic coding feature and the seismic coding feature into the decoder to obtain a fusion feature;

[0042] When the fusion feature meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0043] Optionally, the first encoder is an RDN encoder, the second encoder is a U-Net encoder, and the decoder is an MSFF shared encoder.

[0044] Optionally, the model training module is further used to:

[0045] Performing a frequency band alignment operation on the electromagnetic coding feature and the seismic coding feature to obtain an electromagnetic coding feature and a seismic coding feature with consistent frequency bands;

[0046] The electromagnetic coding features and seismic coding features with consistent frequency bands are sampled at different scales, and the electromagnetic coding features and seismic coding features obtained by sampling at different scales are fused and matched.

[0047] Optionally, the model training module is further used to:

[0048] Calculate a first loss value, a second loss value and a third loss value based on the fusion feature, wherein the first loss value is used to measure the difference between the inverted resistivity model and velocity model and the true velocity model, the second loss value is used to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and the third loss value is used to constrain the smoothness of the output;

[0049] Combine the first loss value, the second loss value, and the third loss value to obtain a total loss value;

[0050] When the total loss value meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0051] Optionally, the model training module is further used to:

[0052] When the total loss value does not meet the loss value requirement, a cosine learning rate adjustment strategy is used to increase the learning rate of the model, and then the resistivity model and the velocity model after the learning rate adjustment are iteratively trained.

[0053] Optionally, the data set preparation module is further used to:

[0054] Generate resistivity models with random fault and salt dome structures;

[0055] The resistivity of the resistivity model is converted into a velocity value using an empirical formula to obtain a velocity model.

[0056] Some embodiments of the present application provide a computing processing device, including:

[0057] a memory having computer readable code stored therein;

[0058] One or more processors, when the computer readable code is executed by the one or more processors, the computing processing device executes the electromagnetic seismic joint inversion method based on the multi-scale feature fusion network as described above.

[0059] Some embodiments of the present application provide a non-transitory computer-readable medium storing a computer-readable code, which, when executed on a computing and processing device, causes the computing and processing device to execute the above-mentioned electromagnetic and seismic joint inversion method based on a multi-scale feature fusion network.

[0060] The present application provides an electromagnetic-seismic joint inversion method and device based on a multi-scale feature fusion network. By using apparent resistivity and phase as input electromagnetic data, and initial velocity model and corresponding depth-shifted seismic data as input seismic data, the resistivity and velocity models are inverted simultaneously, thereby realizing automatic inversion of velocity models of complex underground structures and providing a more accurate velocity model for fine imaging of underground structures.

[0061] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 A schematic diagram of a process flow of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application is schematically shown;

[0064] Figure 2 The schematic diagram schematically shows a principle diagram of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application;

[0065] Figure 3 A schematic diagram of the network structure of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application is schematically shown;

[0066] Figure 4 One of the effect schematic diagrams of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application is schematically shown;

[0067] Figure 5 The second schematic diagram schematically shows the effect of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application;

[0068] Figure 6 The third schematic diagram schematically shows the effect of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application;

[0069] Figure 7 The fourth schematic diagram schematically shows the effect of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network provided in some embodiments of the present application;

[0070] Figure 8 The structure diagram of an electromagnetic-seismic joint inversion device based on a multi-scale feature fusion network provided in some embodiments of the present application is schematically shown;

[0071] Fig. 9 A block diagram schematically shows a computing and processing device for executing the method according to some embodiments of the present application;

[0072] Fig.10 A storage unit for holding or carrying program codes for implementing methods according to some embodiments of the present application is schematically shown. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0074] Figure 1 The present application schematically shows a flow chart of an electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network, the method comprising:

[0075] Step 101, generating a resistivity model and a velocity model with random fault and salt dome structures;

[0076] Optionally, the step 101 includes:

[0077] Step 1011, generating a resistivity model with random fault and salt dome structures;

[0078] Step 1012: Convert the resistivity of the resistivity model into a velocity value using an empirical formula to obtain a velocity model.

[0079] In the embodiments of the present application, for example, reference is made to Figure 2 , in order to train the deep network, its network structure can refer to Figure 3 In the present embodiment, a training set is first constructed using artificial synthetic data. First, a large number of resistivity models R with random faults and salt dome structures are generated. T , and convert the resistivity into velocity value according to the empirical formula to obtain the corresponding velocity model V T The empirical formula (1) we use is:

[0080]

[0081] Where a and b are random coefficients, with value ranges of [1100, 1500] and [900, 1000] respectively, and mr represents mr=min({log 10 (R T )|log 10 (R T )≥2}).

[0082] Step 102, forward modeling the velocity model and the resistivity model to obtain a sample data set;

[0083] In the embodiment of the present application, the velocity model and the resistivity model are forward simulated to obtain the corresponding seismic shot gather data, apparent resistivity and impedance phase data

[0084] Specifically, 500 speed models are randomly generated For each velocity model V T j (x,z) uses a Ricker wavelet with a main frequency of 10Hz to generate shot data, with lateral and longitudinal sampling intervals of dx = 20m and dz = 10m. The number of lateral and longitudinal sampling points is 400, the starting coordinates of the shot points are 10, the interval is 13, and there are 30 shots in total. For each velocity model V T j (x,z) uses 1D initial velocity V0 j (x,z) model is used to perform reverse time migration to obtain seismic data S0 j (x,z). At the same time, for each velocity model V T j (x,z) is converted into the corresponding resistivity model R using the empirical formula T j (x,z), and perform TM mode forward modeling to obtain apparent resistivity and phase data

[0085] The data corresponding to the first 460 models are selected as the training set, the data corresponding to the 20 velocity models are selected as the validation set, and the data corresponding to the remaining 20 velocity models are selected as the test set.

[0086] Step 103, using the sample data set, iteratively training the multi-scale feature fusion network model to obtain a joint inversion network model;

[0087] Optionally, the multi-scale feature fusion network includes: a decoder and at least two encoders; the sample data set includes: seismic data and electromagnetic data, and step 103 includes:

[0088] Step 1031: Perform migration processing on the seismic data using the initial velocity model.

[0089] Step 1032: Encode the electromagnetic data using a first encoder to obtain electromagnetic encoding features, and encode the migrated seismic data and the corresponding velocity model using a second encoder to obtain seismic encoding features.

[0090] Step 1033, input the electromagnetic coding feature and the seismic coding feature into the decoder to obtain a fusion feature.

[0091] Step 1034, when the fusion feature meets the loss value requirement, output the jointly inverted resistivity model and velocity model.

[0092] Optionally, the first encoder is an RDN encoder, the second encoder is a U-Net encoder, and the decoder is an MSFF shared encoder.

[0093] Optionally, after step 1032, the method further includes:

[0094] Step 10321, performing a frequency band alignment operation on the electromagnetic coding feature and the seismic coding feature to obtain an electromagnetic coding feature and a seismic coding feature with consistent frequency bands.

[0095] Step 10322, sampling the electromagnetic coding features and seismic coding features with consistent frequency bands at different scales, and fusing and matching the electromagnetic coding features and seismic coding features obtained by sampling at different scales.

[0096] In an embodiment of the present application, the observed geophysical data obtained based on different detection methods have different effective frequency band ranges, and therefore it is necessary to perform frequency band alignment operations on the data within different effective frequency band ranges. Specifically, for response data with a narrow effective frequency band range and a low main frequency, it is necessary to use a spectrum extension technology, such as a super-resolution deep learning technology, to widen the response data with a narrow effective frequency band range and a low main frequency to be consistent with the data with a wider effective frequency band range, and through a multi-scale sampling technology, the data with consistent frequency bands after widening is sampled at different scales, and various types of observation data features are extracted at different scales for fusion and matching.

[0097] Furthermore, the present application provides a multi-scale feature fusion network model for electromagnetic and seismic joint inversion, which is composed of an RDN encoder E for encoding electromagnetic data. R (·), U-Net encoder E for encoding seismic data U (·) and a decoder D shared by electromagnetic coding features and seismic coding features. In the embodiment of the present application, the initial velocity model V0 j (x,z) and the corresponding RTM seismic data S0 j (x,z) are normalized and merged into two channels to be input into the U-Net encoder, and the resistivity and phase data are After normalization, the data is input to the RDN encoder. The two encoders will output features at different scales. Due to the low resolution of electromagnetic data, the features output by the RDN encoder need to be upsampled and downsampled at the same time to match the features of the input seismic data at different scales. The embodiment of the present application fuses the features of the same scale of seismic data and electromagnetic data, and upsamples them step by step, and finally outputs two-channel data as the resistivity model R that is inverted simultaneously. inv and speed model V inv .

[0098] Step 1033, input the electromagnetic coding feature and the seismic coding feature into the shared decoder to obtain a fusion feature.

[0099] Step 1034, when the fusion feature meets the loss value requirement, output the jointly inverted resistivity model and velocity model.

[0100] Optionally, the step 1034 includes:

[0101] Step 10341, calculating a first loss value, a second loss value, and a third loss value based on the fusion feature, wherein the first loss value is used to measure the difference between the inverted resistivity model and velocity model and the true velocity model, the second loss value is used to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and the third loss value is used to constrain the smoothness of the output;

[0102] Step 10342, combining the first loss value, the second loss value, and the third loss value to obtain a total loss value;

[0103] Step 10343, when the total loss value meets the loss value requirement, output the jointly inverted resistivity model and velocity model.

[0104] Optionally, after step 10342, the method further includes: when the total loss value does not meet the loss value requirement, increasing the learning rate of the model using a cosine learning rate adjustment strategy, and then iteratively training the resistivity model and velocity model after the learning rate is adjusted.

[0105] In the embodiment of the present application, we cannot use the same learning rate to update the weights during the entire training process, otherwise the optimal point cannot be reached, so the size of the learning rate needs to be adjusted during the training process. In the initial stage of training, since the weights belong to the state of random initialization, the loss function is easier to converge, so a larger learning rate can be set. In the later stage of training, since the weights are already close to the optimal, a larger learning rate cannot further find the optimal value, so a smaller learning rate is required. Specifically, the Warm-Up training method is used, and the cosine learning rate adjustment strategy is adopted. A smaller learning rate is adopted in the initial stage of training. As the number of iterations increases, the cosine adjustment strategy is adopted to gradually increase the learning rate to a preset value, which helps the convergence of the joint inversion network training algorithm.

[0106] Optionally, before step 10343, the method further includes:

[0107] Step 10344, when the total loss value converges to the target loss value, determining that the loss value meets the loss value requirement;

[0108] Step 10344, when the total loss value does not converge to the target loss, continue to iteratively train the multi-scale feature fusion network model.

[0109] In the embodiment of the present application, a 1-dimensional linear velocity model is used as the initial velocity model V init , reverse time migration is performed on the seismic data to obtain the migrated data S. The present application embodiment constructs a data pair {S, V init} as U-Net encoder E U The input of (·) As the RDN encoder E R (·) input, the encoded features are input to the shared decoder D(·), the output is two-channel data, and the resistivity R is inverted at the same time inv and speed model V inv , the above process can be summarized as the following formula (2):

[0110]

[0111] In network training, the embodiment of the present application uses L1 loss to measure the difference between the inverted resistivity model and velocity model and the true velocity model, uses cross gradient loss to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and adds total variation loss to impose smoothness constraints. The above loss functions can be summarized as the following formulas (3), (4), and (5):

[0112] Loss L1 =|R inv -R T |1+|V inv -V T |1 (3)

[0113] Loss CG =|▽R inv ×▽V inv |1 (4)

[0114]

[0115] The total loss function is as follows:

[0116] Loss=Loss L1 +Loss CG +Loss TV (6)

[0117] In the training phase, the embodiment of the present application uses the Adam optimizer to optimize the loss function. In the inference phase, the embodiment of the present application inputs electromagnetic data and seismic data to obtain the resistivity model and velocity model results of the joint inversion. The experimental results show that the introduction of seismic data will improve the resolution of the electromagnetic inversion results. At the same time, the electromagnetic data will be beneficial to the reconstruction of the underground structure of the seismic inversion, and the output velocity model can continue to be sent to the network for update and iterative inversion after depth migration.

[0118] During the network training process, the embodiment of the present application minimizes the loss function shown in formula (6). The embodiment of the present application uses the Adam optimizer and uses the warm-up training strategy to adjust the learning rate. A total of 400 epochs are trained, and the maximum learning rate is 0.002.

[0119] The network model based on the fully convolutional network can be directly applied to velocity models of any size. During reasoning, the embodiment of the present application inputs the velocity model, the corresponding RTM seismic data, the apparent resistivity and the phase data into the network to obtain the velocity model and the resistivity model after joint inversion. The velocity model can be directly applied to the subsequent FWI update, and the salt dome structure can be well restored even in the absence of low-frequency data.

[0120] Experimental Platform Description

[0121] The method of the embodiment of the present application is implemented based on Pytorch, and is trained and deployed on an Intel(R) Core(TM) i9-10900XCPU@3.70GHz, 64GB, RTX2080Ti machine.

[0122] Experimental simulation results

[0123] Figure 4 The velocity model and resistivity model synthesized by the embodiment of the present application and the forward modeling data are shown, wherein the seismic data are reverse time migrated using a linear velocity model to obtain the migrated data. Figure 5 The velocity and resistivity models jointly inverted using the embodiment of the present application are shown. By comparing the velocity inversion using the DVG method alone and the resistivity inversion model using EMRNet, it can be seen that the results of the inversion of the embodiment of the present application are significantly better than the individual inversion results and are also better than the traditional methods. After the introduction of seismic data, the resolution of the electromagnetic inversion is significantly improved, the resistivity layer interface is clearer, and the edge of the salt dome structure is more accurately portrayed. At the same time, the introduction of electromagnetic data alleviates the problem of insufficient illumination under the salt dome structure of the seismic data, the velocity salt dome structure of the seismic inversion is more complete, and the physical boundary is consistent with the electromagnetic inversion result. Figure 6It is the FWI inversion result with the network inversion velocity as the initial model. It can be seen that the velocity model inverted in the embodiment of the present application is closer to the true velocity model, and the FWI inversion result is better than the FWI result with the DVG inversion result as the initial model. Figure 7 The experimental results of removing low frequencies are shown. The components below 3 Hz of the forward shot data are filtered out, and then reverse time migration is performed. Figure 7 (c) and (d) show the resistivity model and velocity model inverted using the embodiment of the present application. It can be seen that even in the absence of low-frequency data, the embodiment of the present application can restore most of the salt dome structure. The joint inversion of seismic data and electromagnetic data promotes each other and improves the inversion accuracy of each other.

[0124]

[0125] Table (1)

[0126] Table (1) shows the comparison of the MAE index of the inversion results of the embodiment of the present application and other methods with the real model. The method combining the inversion results of the embodiment of the present application with FWI has the lowest MAE index, which is also consistent with the subjective evaluation.

[0127] Step 104: perform joint inversion on the seismic electromagnetic data using the joint inversion network model.

[0128] The frequency band complementarity of prestack seismic data and electromagnetic data can be effectively utilized. Based on the low-frequency constraints provided by electromagnetic data and the medium- and high-frequency constraints of seismic data, high-resolution resistivity models and velocity models can be inverted, thereby improving the accuracy and reliability of the inversion.

[0129] The embodiment of the present application integrates the super-resolution network model RDN and the deep generation model U-Net as sub-models based on the joint learning technology in deep learning, and uses the strategy of partial parameter sharing to simultaneously invert the velocity model and resistivity model. In the traditional iterative update of the velocity model, the processing personnel make a qualitative assessment of the quality of the velocity model by subjectively judging the matching degree between the velocity model and the main reflection surface in the seismic depth migration data body. The higher the matching degree, the higher the quality of the velocity model. The deep generation model realizes the quantification of this qualitative evaluation standard through a neural network, and the joint learning technology extends this velocity-seismic reflection surface quantitative matching to resistivity-seismic reflection surface quantitative matching, thereby assisting in the inversion of a high-resolution resistivity model. At the same time, the AI ​​super-resolution technology based on RDN broadens the medium and high frequency information of electromagnetic data, fuses electromagnetic data with different frequency band information of seismic data, reduces the multi-solution of seismic and electromagnetic inversion problems, broadens the convergence domain of the initial velocity model in the seismic inversion problem, and even with a simple one-dimensional initial velocity model, a high-precision velocity model and resistivity model can be inverted, reducing the manpower and time cost of velocity modeling.

[0130] For a neural network

[21] that learns based on raw shot collections, different numbers of shot collections and their different combinations constitute all of its 3D raw data space in the 2D case. When there are N grids horizontally, different numbers of shot collections and their different combinations have a total of And these 2 N -1 data need to be mapped to the same velocity model in theory, which causes the pathological nature of the mapping relationship and the network is prone to overfitting. The migration operator of the embodiment of the present application can be regarded as a data compression operator based on physical constraints, which compresses the shot gather data in the original 3D data space and maps it into 2D seismic image data. In the case of sufficient lateral coverage, different pre-stack shot gathers will correspond to the same migration results, which greatly reduces the size of the original data space and enhances the interpretability of the network.

[0131] The network structure proposed in the embodiment of the present application consists of three parts: an RDN encoder for encoding electromagnetic data, a U-Net encoder for encoding seismic data, and a decoder that shares electromagnetic encoding features with seismic encoding features. The apparent resistivity and phase data obtained by MT exploration are denoted as Input initial velocity model V init , using V init Perform prestack depth migration to obtain seismic data S. Input RDN decoder, {S,V init} Input the U-Net decoder to obtain the corresponding electromagnetic data features and seismic data features, then fuse the features at each scale through upsampling and downsampling operations, and simultaneously output the inverted resistivity and velocity models through a shared decoder. The loss function of the embodiment of the present application includes three items: L1 loss is used to measure the difference between the output result and the true answer, the cross-gradient loss is used to constrain the physical property boundaries of the joint inversion results to be consistent, and the total variation loss function is used to impose smoothness constraints. Experiments have shown that the SEMI-Net proposed in the embodiment of the present application can effectively integrate electromagnetic and seismic information, improve the resolution of electromagnetic and seismic inversion, and reduce the multi-solution of inversion. Even when the initial velocity model is very far from the true velocity model, it can effectively invert the underground velocity and assist in the refined imaging of the underground structure. It can also be used as a velocity model input for subsequent further iterations.

[0132] Figure 8 The structure diagram of an electromagnetic-seismic joint inversion device 20 based on a multi-scale feature fusion network provided by the present application is schematically shown, and the device comprises:

[0133] A data set preparation module 201 is used to generate a resistivity model and a velocity model with random fault and salt dome structures;

[0134] Performing forward simulation on the velocity model and the resistivity model to obtain a sample data set;

[0135] A model training module 202 is used to iteratively train the multi-scale feature fusion network model using the sample data set to obtain a joint inversion network model;

[0136] The joint inversion module 203 is used to perform joint inversion on the seismic electromagnetic data using the joint inversion network model.

[0137] Optionally, the multi-scale feature fusion network includes: a decoder and at least two encoders; the sample data set includes: seismic data and electromagnetic data;

[0138] The model training module 202 is also used for:

[0139] Performing migration processing on the seismic data using an initial velocity model;

[0140] Encoding the electromagnetic data using a first encoder to obtain electromagnetic encoding features, and encoding the migrated seismic data and a corresponding velocity model using a second encoder to obtain seismic encoding features;

[0141] Inputting the electromagnetic coding feature and the seismic coding feature into the decoder to obtain a fusion feature;

[0142] When the fusion feature meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0143] Optionally, the first encoder is an RDN encoder, the second encoder is a U-Net encoder, and the decoder is an MSFF shared encoder.

[0144] Optionally, the model training module 202 is further used to:

[0145] Performing a frequency band alignment operation on the electromagnetic coding feature and the seismic coding feature to obtain an electromagnetic coding feature and a seismic coding feature with consistent frequency bands;

[0146] The electromagnetic coding features and seismic coding features with consistent frequency bands are sampled at different scales, and the electromagnetic coding features and seismic coding features obtained by sampling at different scales are fused and matched.

[0147] Optionally, the model training module 202 is further used to:

[0148] Calculate a first loss value, a second loss value and a third loss value based on the fusion feature, wherein the first loss value is used to measure the difference between the inverted resistivity model and velocity model and the true velocity model, the second loss value is used to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and the third loss value is used to constrain the smoothness of the output;

[0149] Combine the first loss value, the second loss value, and the third loss value to obtain a total loss value;

[0150] When the total loss value meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

[0151] Optionally, the model training module 202 is further used to:

[0152] When the total loss value does not meet the loss value requirement, a cosine learning rate adjustment strategy is used to increase the learning rate of the model, and then the resistivity model and the velocity model after the learning rate adjustment are iteratively trained.

[0153] Optionally, the data set preparation module 201 is further used to:

[0154] Generate resistivity models with random fault and salt dome structures;

[0155] The resistivity of the resistivity model is converted into a velocity value using an empirical formula to obtain a velocity model.

[0156] The embodiment of the present application uses apparent resistivity and phase as input electromagnetic data, initial velocity model and corresponding depth offset seismic data as input seismic data, and simultaneously inverts resistivity and velocity models to achieve automatic inversion of velocity models of complex underground structures, thereby providing a more accurate velocity model for fine imaging of underground structures.

[0157] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0158] The various component embodiments of the present application can be implemented in hardware, or implemented in software modules running on one or more processors, or implemented in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all functions of some or all components in the computing processing device according to the embodiment of the present application. The application can also be implemented as a device or apparatus program (e.g., computer program and computer program product) for performing a part or all of the methods described herein. Such a program implementing the present application can be stored on a non-transient computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0159] For example, Fig. 9 A computing processing device that can implement the method according to the present application is shown. The computing processing device traditionally includes a processor 310 and a computer program product or a non-transitory computer-readable medium in the form of a memory 320. The memory 320 can be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read-only memory), an EPROM, a hard disk or a ROM. The memory 320 has a storage space 330 for a program code 331 for executing any method step in the above method. For example, the storage space 330 for the program code may include individual program codes 331 for implementing various steps in the above method respectively. These program codes can be read from or written to one or more computer program products. These computer program products include program code carriers such as a hard disk, a compact disk (CD), a memory card or a floppy disk. Such a computer program product is usually as shown in reference Fig.10 The portable or fixed storage unit. The storage unit may have Fig. 9 The memory 320 in the computing and processing device of the computer 320 is similarly arranged as a storage segment, a storage space, etc. The program code can be compressed, for example, in an appropriate form. Generally, the storage unit includes a computer readable code 331 ', that is, a code that can be read by a processor such as 310, which, when executed by the computing and processing device, causes the computing and processing device to perform the various steps in the method described above.

[0160] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0161] The term "one embodiment", "embodiment" or "one or more embodiments" herein means that a particular feature, structure or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present application. In addition, please note that the examples of the term "in one embodiment" here do not necessarily all refer to the same embodiment.

[0162] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0163] In the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An electromagnetic-seismic joint inversion method based on a multi-scale feature fusion network, characterized in that: The method comprises: Generate resistivity and velocity models with random fault and salt dome structures; Performing forward simulation on the velocity model and the resistivity model to obtain a sample data set; Using the sample data set, iteratively training the multi-scale feature fusion network model to obtain a joint inversion network model; The joint inversion network model is used to perform joint inversion on seismic electromagnetic data.

2. The method according to claim 1, characterized in that: The multi-scale feature fusion network includes: a decoder and at least two encoders; the sample data set includes: seismic data and electromagnetic data; The step of iteratively training the multi-scale feature fusion network model using the sample data set to obtain the joint inversion network model includes: Performing migration processing on the seismic data using an initial velocity model; Encoding the electromagnetic data using a first encoder to obtain electromagnetic encoding features, and encoding the migrated seismic data and a corresponding velocity model using a second encoder to obtain seismic encoding features; Inputting the electromagnetic coding feature and the seismic coding feature into the decoder to obtain a fusion feature; When the fusion feature meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

3. The method according to claim 2, characterized in that The first encoder is an RDN encoder, the second encoder is a U-Net encoder, and the decoder is an MSFF shared encoder.

4. The method according to claim 2, characterized in that: After the steps of encoding the electromagnetic data using the first encoder to obtain electromagnetic encoding features, and encoding the migrated seismic data and the corresponding velocity model using the second encoder to obtain seismic encoding features, the method further includes: Performing a frequency band alignment operation on the electromagnetic coding feature and the seismic coding feature to obtain an electromagnetic coding feature and a seismic coding feature with consistent frequency bands; The electromagnetic coding features and seismic coding features with consistent frequency bands are sampled at different scales, and the electromagnetic coding features and seismic coding features obtained by sampling at different scales are fused and matched.

5. The method according to claim 2, characterized in that: The step of outputting the resistivity model and velocity model of the joint inversion when the fusion feature meets the loss value requirement includes: Calculate a first loss value, a second loss value and a third loss value based on the fusion feature, wherein the first loss value is used to measure the difference between the inverted resistivity model and velocity model and the true velocity model, the second loss value is used to constrain the physical property boundaries of the inverted resistivity and velocity models to be consistent, and the third loss value is used to constrain the smoothness of the output; Combine the first loss value, the second loss value, and the third loss value to obtain a total loss value; When the total loss value meets the loss value requirement, the resistivity model and velocity model of the joint inversion are output.

6. The method according to claim 5, characterized in that After the step of combining the first loss value, the second loss value, and the third loss value to obtain a total loss value, the method further includes: When the total loss value does not meet the loss value requirement, a cosine learning rate adjustment strategy is used to increase the learning rate of the model, and then the resistivity model and the velocity model after the learning rate adjustment are iteratively trained.

7. The method according to claim 1, characterized in that The step of generating a resistivity model and a velocity model with random fault and salt dome structures comprises: Generate resistivity models with random fault and salt dome structures; The resistivity of the resistivity model is converted into a velocity value using an empirical formula to obtain a velocity model.

8. An electromagnetic-seismic joint inversion device based on a multi-scale feature fusion network, characterized in that: The device comprises: Dataset preparation module for generating resistivity and velocity models with random fault and salt dome structures; Performing forward simulation on the velocity model and the resistivity model to obtain a sample data set; A model training module, used to iteratively train the multi-scale feature fusion network model using the sample data set to obtain a joint inversion network model; The joint inversion module is used to perform joint inversion on seismic electromagnetic data using the joint inversion network.

9. A computing device, characterized in that: include: a memory having computer readable code stored therein; One or more processors, when the computer readable code is executed by the one or more processors, the computing processing device executes the electromagnetic seismic joint inversion method based on a multi-scale feature fusion network as described in any one of claims 1-7.

10. A non-transitory computer readable medium, characterized in that Computer readable codes are stored, and when the computer readable codes are run on a computing and processing device, the computing and processing device is caused to execute the electromagnetic and seismic joint inversion method based on a multi-scale feature fusion network as described in any one of claims 1 to 7.

Citation Information

Cited By

  • Joint inversion method based on multi-geophysical data feature fusion and dual joint strategy driving

    CN121637005A