CN111562611A - A semi-supervised deep learning seismic data inversion method driven by wave equation - Google Patents

Abstract

The present disclosure provides a semi-supervised deep learning seismic data inversion method driven by a wave equation, which can realize a deep learning inversion network in the absence of a corresponding geological model for some seismic data. A fully connected network is used to enhance the seismic data, and by extracting the feature map and finally obtaining the geological wave velocity model, the mapping relationship between the seismic data and the underground multi-layer medium model is completed. For the seismic data of geological model, the data loss function is used to replace the wave velocity loss function, physical laws are introduced, and a semi-supervised learning strategy is implemented. Through semi-supervised deep learning seismic data inversion network, the inversion effect of deep learning network is improved when there are few labeled data.

Description

A semi-supervised deep learning seismic data inversion method driven by wave equation

technical field

The present disclosure belongs to the field of geophysical exploration, and relates to a semi-supervised deep learning seismic data inversion method driven by a wave equation.

Background technique

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

As one of the most commonly used geophysical exploration methods, seismic method is widely used in petroleum exploration, coal field and metal mineral exploration, etc., and has broad application prospects. The main principle of the seismic method is based on wave field propagation. Multiple geophones are arranged on the surface, and the wave field is generated by repeatedly exciting the artificial seismic source and propagates in the underground medium. When the wave impedance changes in the underground medium, it is reflected or refracted back to the ground. The geophones located on the ground record the vibration information transmitted to the ground, and the seismic data is processed by imaging or inversion methods to obtain the distribution information of the subsurface medium. Among them, the inversion method can improve the seismic resolution, obtain more accurate subsurface structural information, and improve the evaluation ability of subsurface media, and has gradually become an inaccessible part of seismic data processing.

As far as the inventors know, the most commonly used seismic data inversion method is full waveform inversion. For pre-stack seismic data, high-precision characterization of underground structures can be accomplished by using the wave field kinematics and dynamics information contained therein. Although the full waveform inversion method has been developed for many years, there are still problems such as heavy dependence on the initial model, easy to fall into local optimum and low computational efficiency, which need to be further improved. In recent years, with the emergence of deep learning methods, the field of geophysics has gradually begun to use deep learning methods to solve data processing, inversion imaging and other problems, providing new solutions.

There are two problems in realizing seismic deep learning inversion:

First, unlike deep learning to solve traditional image processing problems, the correspondence between seismic data and geological models is more complicated. Spatial model, the corresponding relationship between the two is weak, and it is difficult to use traditional deep learning methods for effective mapping.

Second, as a data-driven algorithm, deep learning relies on massive data support and needs to extract the correspondence between data and labels based on a large number of data-label pairs, that is, supervised deep learning methods to complete the mapping between data. For seismic data, because it is difficult to obtain the information of the underground medium, the corresponding labels of all seismic data cannot be obtained, which cannot meet the conditions of supervised deep learning. In addition, this data-driven algorithm lacks the support of physical meaning, and the generalization of the method is limited.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present disclosure proposes a semi-supervised deep learning seismic data inversion method driven by the wave equation. The present disclosure constructs a trace convolution-fully connected network according to the characteristics of the seismic data. To solve the problem, the data loss function is calculated by the wave equation to replace the commonly used wave velocity loss function to obtain the subsurface wave velocity model, which improves the inversion effect and generalization of pre-stack seismic data.

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

A semi-supervised deep learning seismic data inversion method driven by wave equation, including the following steps:

Build a geological model library based on the modeling of the geological velocity model, calculate the corresponding seismic data through the wave equation, classify the data with or without labels, and build a seismic semi-supervised learning database;

Construct a deep learning-based trace convolution-full connection network, input the seismic data with labeled data sets and unlabeled data sets into the trace convolution-full connection network, and obtain the corresponding predicted wave velocity model;

The prediction model corresponding to the unlabeled data is further subjected to wave field simulation based on the wave equation to obtain the predicted seismic data corresponding to the predicted wave velocity model;

Calculate the wave velocity loss function and the data loss function, and integrate the two loss functions to optimize the convolution-fully connected network;

The mapping between the seismic data and the geological velocity model is constructed through the trace convolution-full connection network, and the velocity model map is obtained according to the acquired seismic data to realize the inversion of the seismic data.

As an optional implementation, constructing the seismic semi-supervised learning database includes:

Numerical modeling of subsurface geological layered wave velocity models with various relief structures, wave velocities and/or different layers;

For each wave velocity model, the wave field simulation is performed with a fixed source, geophone position and observation time, and the wave field data is recorded at the geophone position to obtain the seismic data corresponding to the wave velocity model;

The data is randomly classified in a certain proportion and divided into labeled data groups and unlabeled data groups, and the seismic semi-supervised learning database is obtained.

As an alternative embodiment, the labeled data set contains velocity model-seismic data pairs; the unlabeled data set contains only seismic data.

As an optional embodiment, a deep learning-based trace convolution-fully connected network is constructed, and the input seismic data is encoded and preprocessed, including feature enhancement on the three-dimensional seismic data section. For each single shot and single trace seismic data after encoding The feature generation and decoding of each feature vector are realized by the encoding network structure, the feature generating network structure and the decoding network structure, respectively.

As an optional implementation manner, the coding network structure includes global features, adjacent domain information and position coding, wherein the global feature extraction includes two sets of 6-layer convolutional structures cascaded in sequence, and the adjacent domain information extraction includes two sets of 3-layered convolutional structures. The convolutional structure is connected, and the location code is a set of vectors marking the location of the hypocenter and the receiver.

As an optional embodiment, the feature generation network structure includes 5 fully connected layers.

As an optional implementation manner, the decoding network structure includes three layers of convolutional structures cascaded in sequence, and the outputs of the third layer of convolutional structures are respectively connected to four layers of parallel convolutional structures. The output ends of the convolutional layer are superimposed, and the input segment of the last layer of convolutional structure is connected, and the final output channel convolution-full connection network result, that is, the predicted wave speed model.

As an optional embodiment, an acoustic wave field simulation is performed on the prediction model corresponding to the unlabeled data, and the calculation formula is based on the acoustic wave equation:

Among them, v is the predicted velocity model, t is the wave field propagation time, x and y are the spatial position of the model, and p is the pressure, that is, the wave field.

As an optional implementation, the specific process of optimizing the convolution-fully connected network by integrating two loss functions includes:

The wave velocity loss function is calculated for the prediction model corresponding to the labeled data set and the original wave velocity model. The calculation formula is:

Calculate the data loss function for the predicted seismic data and the input seismic data corresponding to the unlabeled data set, and the calculation formula is:

和M分别为预测波速模型矩阵和原波速模型矩阵，

和D分别为预测地震数据矩阵和原地震数据矩阵，R为以不同尺度计算两个矩阵的局部相似性，λ_r为不同尺度局部相似性的系数，SSIM^r为不同尺度下两个矩阵的局部相似性，SSIM^r的计算公式如下：in

and M are the predicted wave velocity model matrix and the original wave velocity model matrix, respectively,

and D are the predicted seismic data matrix and the original seismic data matrix, respectively, R is the local similarity of the two matrices calculated at different scales, λ _r is the coefficient of local similarity at different scales, SSIM ^r is the local similarity of the two matrices at different scales Similarity, the calculation formula of SSIM ^r is as follows:

代表在Μ模型中取k点为中心，大小为r×r的窗口，c₁和c₂为稳定分子分母的常数项。Among them, H and W are divided into the height and width of the model,

Represents a window of size r × r taking point k as the center in the M model, and c ₁ and c ₂ are constant terms of the stable numerator and denominator.

Accumulate the wave speed loss function and the data loss function in a certain proportion. The proportion of the two loss functions can be determined according to the actual situation. The two loss functions need to reach the same order of magnitude to prevent the calculation from being skewed due to an excessively large loss function. The loss function calculates the gradient of the loss function, in which the data loss function is calculated through the inverse calculation of the wave equation, and the gradient of the model is calculated, and the wave velocity loss function is calculated directly. The gradient of the model is accumulated, and finally the convolution is updated. - Parameters of fully connected neural network.

A computer-readable storage medium stores a plurality of instructions, and the instructions are adapted to be loaded by a processor of a terminal device and execute the method for inversion of seismic data based on semi-supervised deep learning driven by wave equations.

A terminal device, comprising a processor and a computer-readable storage medium, where the processor is used to implement various instructions; the computer-readable storage medium is used to store a plurality of instructions, the instructions are suitable for being loaded by the processor and executing the described one A semi-supervised deep learning seismic data inversion method driven by wave equation.

Compared with the prior art, the beneficial effects of the present disclosure are:

The present disclosure deals with the pre-stack seismic data in the seismic exploration method: considering that the single-channel seismic data contains limited information that can reflect the geology, in the original single-channel seismic data, the common shot adjacent feature is extracted by convolution It also adds the global feature of the common shot, the global feature of the common receiver and the location information, which improves the quality of the seismic data, effectively suppresses the noise in the seismic data, and improves the stability of the inversion.

At the same time, the present disclosure also uses full connection to process each enhanced data separately, comprehensively obtains the final result, and makes full use of the geological information contained in all seismic data to accurately describe the inversion results. In this way, the mapping of seismic data to the geological velocity model can be effectively completed, and the inversion effect can be improved.

The present disclosure proposes a semi-supervised inversion architecture based on wave equations to solve the problems that a small amount of seismic data is difficult to obtain a good inversion effect through deep learning methods, lack of physical laws, and poor generalization. When the inversion network calculates the loss function , introducing the wave equation to calculate the seismic data corresponding to the prediction results of the unlabeled data set, and transforming the commonly used wave velocity loss function into a data loss function. The generalization of the inversion method also improves the inversion effect of deep learning networks when there are few labeled data.

Description of drawings

The accompanying drawings that constitute a part of the present disclosure are used to provide further understanding of the present disclosure, and the exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure.

Fig. 1 is the method flow chart of the present embodiment;

2 is a schematic diagram of a convolution-fully connected network of the present embodiment;

3 is a schematic diagram of the loss function calculation and network optimization structure for semi-supervised learning of the present embodiment;

Figure 4 (a)-(c) are respectively the layered geological model in the database established in the present embodiment, a common shot profile of the seismic data, and a common receiver profile of the seismic data;

FIG. 5 is the deep learning inversion result in this embodiment.

Detailed ways:

The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the present disclosure. Unless otherwise defined, 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 disclosure belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

This embodiment discloses a semi-supervised deep learning seismic data inversion method driven by a wave equation, as shown in FIG. 1 , including the following steps:

In step S1, a semi-supervised seismic database is constructed by computer numerical simulation, and the database includes a labeled data group (including the wave velocity model-seismic data pair) and an unlabeled data group (containing only the seismic data);

The method of this example is mainly aimed at the geological information such as the number of layers, undulation shape and shear wave velocity of underground layered geology in a large-scale range. Here, it is represented by wave velocity models with different layers, different undulations and different wave velocities, as shown in Figure 4(a). ) shown;

The size of the model in this embodiment is 1km*1km, the source points are 20, the spacing is 50m, the number of geophones is 32, the spacing is 30m, the model grid size is 10m*10m, and the model wave velocity range is 1500m/s to 4000m/ s, using 20Hz rake wavelet source excitation, the sampling interval of the detector is 1ms, and the sampling time is 1s;

The database of this embodiment contains 5,000 sets of labeled data and 5,000 sets of unlabeled data for training, each of which includes a two-layer media model, a three-layer media model, a four-layer media model, and a five-layer media model, each with 1,250 sets. , which contains 1000 training sets and 1000 testing sets.

Step S2, as shown in Figure 2, constructs a trace convolution-full connection network, and designs the network structure according to the characteristics of the seismic data;

In this embodiment, all seismic data are passed through four groups of different convolutional network layers, as shown in Fig. 2, respectively from the common shot profile shown in Fig. 4(b) (all detectors measured after each source excitation obtained data) and the co-detector profile shown in Fig. 4(c) (the data measured by each detector after all sources are excited), the global features of the common shot are extracted through a multi-layer convolutional network, and the co-detector global feature, common shot point feature, common receiver feature, and at the same time, according to the position code, that is, a set of sequences, the position is marked with 1, the non-position is marked with 0, and the enhanced seismic trace data of each trace data is obtained by combining them; A layered fully connected network is used to process each enhanced seismic data to obtain the corresponding feature map; then through multiple convolution layers, multiple feature maps are processed to obtain the final prediction model, that is, the output result. Since the network performs convolution and fully connected processing for different profile tract data, we name it as a tract convolution-fully connected network.

Specifically, it includes an encoding network structure for feature enhancement of three-dimensional seismic data sections, a feature generation network structure for each single shot and single channel seismic data after encoding, and a decoding network structure for decoding each feature vector.

The feature-enhanced coding network structure includes global features, local information and position encoding. The global feature extraction includes two sets of 6-layer convolutional structures cascaded in sequence, and the adjacent domain information extraction includes two sets of 3-layer convolutional structures cascaded in sequence. , and the location code is a set of vectors that mark the location of the hypocenter and the receiver.

The feature generation network structure includes 5 fully connected layers.

The decoding network structure includes three layers of convolutional structures cascaded in sequence. The outputs of the third layer of convolutional structures are respectively connected to four parallel convolutional structures, and the outputs of the four parallel convolutional structures are superimposed, and finally connected. The input segment of the 1-layer convolution structure, and the final output channel convolution-full connection network result, that is, the predicted wave speed model.

Step S3, as shown in Figure 3, trains a convolution-fully connected network with the supervised data set and unsupervised data set in the database, designs two loss functions, and completes the semi-supervised learning architecture;

In this example, for the prediction results corresponding to the supervised data set data, the wave velocity loss function is calculated, and the calculation formula is:

和M分别为预测波速模型矩阵和原波速模型矩阵，R为以尺度计算两个矩阵的局部相似性(SSIM)，λ_r为不同尺度局部相似性的系数，SSIM^r为不同尺度下两个矩阵的局部相似性，SSIM的计算公式如下：of which, of which

and M are the predicted wave velocity model matrix and the original wave velocity model matrix, respectively, R is the local similarity (SSIM) of the two matrices calculated by scale, λ _r is the coefficient of local similarity at different scales, SSIM ^r is the two matrices at different scales The local similarity of , the calculation formula of SSIM is as follows:

代表在Μ模型中取k点为中心，大小为r×r的窗口，c₁和c₂为稳定分子分母的常数项。Among them, H and W are divided into the height and width of the model,

Represents a window of size r × r taking point k as the center in the M model, and c ₁ and c ₂ are constant terms of the stable numerator and denominator.

In this example, for the unsupervised data group data, the wave equation is further used to calculate the predicted wave velocity model to obtain the predicted seismic data, and the data loss function is calculated. The wave equation calculation formula is:

Among them, v is the predicted velocity model, t is the wave field propagation time, x and y are the spatial position of the model, and p is the pressure, that is, the wave field. The wave field value of the geophone position in the measurement time period is obtained, that is, the predicted seismic data.

The formula for calculating the data loss function is:

和D分别为预测地震数据矩阵和输入地震数据矩阵。in,

and D are the predicted seismic data matrix and the input seismic data matrix, respectively.

Step S4, as shown in Figure 1, trains a convolutional-fully connected network.

The main network parameters and hardware conditions in this embodiment are: the calculation is implemented by a single-chip NVIDIA TITAN Xp. The network is built based on the PyTorch platform. The batch size of the Adam optimizer is 4, the learning rate is 5e-4, and the number of epochs of the learning algorithm in the entire training data set is 200.

Step S5, the trace convolution-full connection network constructs the mapping relationship between the seismic data and the geological velocity model, which can represent the inversion process. Part of the results substituted into the test set is shown in Figure 5. It can be shown that the present invention can obtain the fluctuation and wave velocity of the subsurface layered medium more accurately by inversion under the condition that only part of the data has tags and the other part of the data lacks tags. In the embodiment, it takes about 30 hours to train the network, and about 2 minutes to test 1000 sets of data, and the inversion efficiency can meet the engineering application requirements.

As will be appreciated by one skilled in the art, embodiments of the present disclosure may be provided as a method, system, or computer program product. Accordingly, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present disclosure is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

The above descriptions are only preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they do not limit the protection scope of the present disclosure. Those skilled in the art should understand that on the basis of the technical solutions of the present disclosure, those skilled in the art do not need to pay creative efforts. Various modifications or variations that can be made are still within the protection scope of the present disclosure.

Claims (10)

1. a semi-supervised deep learning seismic data inversion method driven by wave equation, is characterized in that: comprise the following steps: Build a geological model library based on the modeling of the geological velocity model, calculate the corresponding seismic data through the wave equation, classify the data with or without labels, and build a seismic semi-supervised learning database; Construct a deep learning-based trace convolution-full connection network, input the seismic data with labeled data sets and unlabeled data sets into the trace convolution-full connection network, and obtain the corresponding predicted wave velocity model; The prediction model corresponding to the unlabeled data is further subjected to wave field simulation based on the wave equation to obtain the predicted seismic data corresponding to the predicted wave velocity model; Calculate the wave velocity loss function and the data loss function, and integrate the two loss functions to optimize the convolution-fully connected network; The mapping between the seismic data and the geological velocity model is constructed through the trace convolution-full connection network, and the velocity model map is obtained according to the acquired seismic data to realize the inversion of the seismic data. 2. a kind of semi-supervised deep learning seismic data inversion method based on wave equation drive as claimed in claim 1, is characterized in that: constructing seismic semi-supervised learning database comprises: Numerical modeling of subsurface geological layered wave velocity models with various relief structures, wave velocities and/or different layers; For each wave velocity model, the wave field simulation is performed with a fixed source, geophone position and observation time, and the wave field data is recorded at the geophone position to obtain the seismic data corresponding to the wave velocity model; The data is randomly classified in a certain proportion and divided into labeled data groups and unlabeled data groups, and the seismic semi-supervised learning database is obtained. 3. a kind of semi-supervised deep learning seismic data inversion method based on wave equation drive as claimed in claim 1, it is characterized in that: labelled data group is to include wave velocity model-seismic data pair; unlabeled data group only includes earthquake data. 4. a kind of semi-supervised deep learning seismic data inversion method based on wave equation drive as claimed in claim 1, it is characterized in that: construct the trace convolution-full connection network based on deep learning, encode the seismic data of input Preprocessing, including feature enhancement for 3D seismic data sections, feature generation for each single shot and single channel seismic data after encoding, and decoding for each feature vector, respectively, by encoding network structure, feature generation network structure and decoding network Structural realization. 5. The semi-supervised deep learning seismic data inversion method driven by the wave equation as claimed in claim 4, wherein the coding network structure comprises global features, proximity information and position coding, wherein the global feature extraction It includes two sets of 6-layer convolutional structures that are cascaded in sequence, and the adjacent information extraction includes two sets of 3-layered convolutional structures that are cascaded in sequence. Or, the feature generation network structure includes 5 fully connected layers. 6. The semi-supervised deep learning seismic data inversion method driven by the wave equation as claimed in claim 4, wherein the decoding network structure comprises three layers of convolutional structures cascaded in sequence, and the third layer of The output ends of the convolutional structure are respectively connected to the 4-layer parallel convolutional structure, the outputs of the 4-layered parallel convolutional structure are superimposed, and the input segment of the last 1-layer convolutional structure is connected, and the final output is a convolutional-fully connected network. The result is the predicted wave velocity model. 7. a kind of semi-supervised deep learning seismic data inversion method based on wave equation drive as claimed in claim 1, it is characterized in that: carry out acoustic wave field simulation to the prediction model corresponding to unlabeled data, and calculation formula is according to acoustic wave wave equation:

Among them, v is the predicted velocity model, t is the wave field propagation time, x and y are the spatial position of the model, and p is the pressure, that is, the wave field. 8. a kind of semi-supervised deep learning seismic data inversion method driven by wave equation as claimed in claim 1, it is characterized in that: the concrete process of synthesizing two loss function optimization trace convolution-full connection network comprises: The wave velocity loss function is calculated for the prediction model corresponding to the labeled data set and the original wave velocity model. The calculation formula is:

Calculate the data loss function for the predicted seismic data and the input seismic data corresponding to the unlabeled data set, and the calculation formula is:

in

and M are the predicted wave velocity model matrix and the original wave velocity model matrix, respectively,

and D are the predicted seismic data matrix and the original seismic data matrix, respectively, R is the local similarity of the two matrices calculated by scale, λ _r is the coefficient of local similarity at different scales, SSIM ^r is the local similarity of the two matrices at different scales The calculation formula of SSIM ^r is as follows:

Among them, H and W are divided into the height and width of the model,

Represents a window of size r × r taking point k as the center in the M model, and c ₁ and c ₂ are constant terms of the stable numerator and denominator. 9. A computer-readable storage medium, characterized in that: a plurality of instructions are stored therein, and the instructions are adapted to be loaded by a processor of a terminal device and execute the one based on any one of claims 1-8. A wave equation-driven semi-supervised deep learning method for seismic data inversion. 10. A terminal device, characterized in that it comprises a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store a plurality of instructions, and the instructions are suitable for being loaded by the processor and storing the instructions. A semi-supervised deep learning seismic data inversion method driven by a wave equation according to any one of claims 1 to 8 is implemented.

Images