Three-dimensional seismic fault automatic identification method based on deep learning

Abstract

The invention provides a three-dimensional seismic fault identification method based on deep learning, and the method comprises the following steps: (1), carrying out the automatic modeling, generating three-dimensional seismic data and an accurate fault label, obtaining a three-dimensional seismic image data set, dividing the data set into a training set and a test set, and carrying out the recognition of the three-dimensional seismic fault; (2) training an improved 3D-Unet convolutional neural network of a two-channel structure or an AG-3D-Unet network fused with an attention mechanism and residual learning by taking the three-dimensional synthetic seismic data of the training set and a label of the three-dimensional synthetic seismic data as input, and verifying by using a test set to obtain a trained seismic fault recognition model, (3) inputting real seismic data, and obtaining a seismic fault recognition result. And outputting a fault identification result for the input to-be-predicted image through the identification network. According to the method, the definition, the fault continuity and the accuracy of fault identification are obviously improved, and the method has high practical value in actual seismic data fault interpretation.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and in particular to a three-dimensional seismic fault automatic identification method based on deep learning. Background Art

[0002] With the continuous advancement of industrial modernization, the whole society's demand for energy is also increasing. Although many new energy sources have been introduced, oil and natural gas are still one of the indispensable energy sources at present, so the development and exploration of oil and gas is also very important. Underground faults are fracture aggregation zones caused by stress compression and stretching. These faults are not only the boundaries of oil and gas fields, but also the key channels for oil and gas migration and accumulation. Therefore, accurate interpretation of faults is particularly important for oil exploration. Moreover, the fault morphology is complex and changeable. The conventional fault interpretation method is not only cumbersome and time-consuming, but also difficult to verify repeatedly. Summary of the invention

[0003] In view of the shortcomings of the above-mentioned prior art, the present invention provides a three-dimensional seismic fault automatic identification method based on deep learning, which regards the three-dimensional seismic data volume as a three-dimensional image, and uses deep learning and computer vision methods to interpret faults. This can not only greatly reduce the interference of human factors and the probability of errors, but also significantly shorten the time required for fault interpretation and improve work efficiency.

[0004] The technical solution provided by the present invention is a method for automatic identification of three-dimensional seismic faults based on deep learning, comprising the following steps:

[0005] (1) Automatically model and generate 3D seismic data and accurate fault labels to obtain a 3D seismic image dataset, which is then divided into a training set and a test set;

[0006] (2) Using the 3D synthetic seismic data and its labels of the training set as input, the improved dual-channel structure 3D-Unet convolutional neural network or the AG-3D-Unet network that integrates the attention mechanism and residual learning is trained, and the test set is used for verification to obtain a trained earthquake fault recognition model;

[0007] (3) Input real earthquake data and output fault identification results for the input image to be predicted through the earthquake fault identification model.

[0008] Furthermore, the modeling in step (1) includes constructing a fault structure model and marking the faults. The construction of the fault structure model includes constructing a horizontal reflection coefficient model, then constructing a fold coefficient reflection model, and finally constructing a reflection coefficient model containing both faults and folds.

[0009] Furthermore, the specific method of constructing the horizontal reflection coefficient model is as follows: by using a random number sequence to represent the reflection intensity of different strata to seismic waves, a horizontal reflection coefficient model f(x,y,z) is established in the (x,y,z) coordinate space, where f(x,y,z) represents the reflection coefficient of each point in the space.

[0010] Furthermore, the specific method of constructing the wrinkle coefficient reflection model is as follows: using the vertical shear plane model, the wrinkles are composed of two displacement fields S 1 (x,y,z) and S 2 The combination of (x, y, z) is used to define S 1 (x,y,z) creates a purely tilted structure, S 2 (x,y,z) adds a curved structure by using these two types of displacement fields S 1 (x,y,z) and S 2 (x,y,z), which realizes the transformation of samples from plane space to wrinkle coordinate space.

[0011] Furthermore, the specific method of constructing a reflection coefficient model containing faults and folds is as follows: randomly select a reference point as the starting point of the fault, specify the strike angle and dip angle of the fault to determine the direction of the fault, and construct a local coordinate system, in which the x-axis corresponds to the strike direction, the y-axis corresponds to the dip direction, and the z-axis points to the normal direction. An ideal elliptical function is defined on the fault plane as the displacement field d(x, y, z) = 0 to accurately describe and simulate the displacement characteristics of the fault. Then, the three-dimensional displacement field method is used to create faults in the three-dimensional structural model.

[0012] The wrinkle structure model Perform a coordinate transformation to get (x, y, z), then apply the fault displacement in the local coordinates (x, y, z), and finally perform a coordinate transformation to get (x, y, z) back to A reflection coefficient model containing both faults and folds is obtained.

[0013] Furthermore, the reflection coefficient model containing faults and folds is convolved with the Ricker wavelet, and some random noise is added to the model to simulate the interference in real seismic data, and finally a three-dimensional seismic image is generated.

[0014] Furthermore, the fault labeling is achieved by creating a label array with the same size as the seismic data model.

[0015] Furthermore, the improved dual-channel structure of the 3D-Unet convolutional neural network is derived from the original 3D-Unet architecture. The model contains two independent data streams, which convert and extract the visual features of the input seismic image in different ways. A hollow space pyramid is added to the encoding and decoding connection parts to increase the receptive field and capture multi-scale image information.

[0016] In the improved dual-channel 3D-Unet network, there are two independent branches as downsampling parts, which are responsible for converting the input seismic image into a feature map in the latent space. Each group of layers in the branch consists of a pair of three-dimensional convolutional layers, a convolution kernel of size (2,2,2), and a three-dimensional pooling layer with a stride of 2. A Drop layer is added to the last part of the encoder to prevent the model from overfitting. In the decoding part, the two encoder channel outputs and the output of the previous layer of the decoder are spliced ​​through jump connections, and finally the output result is obtained through a 1×1×1 convolutional layer and a sigmoid function.

[0017] Furthermore, in the convolutional neural network training stage of the improved dual-channel 3D-Unet, a hybrid loss function was used to evaluate the accuracy of fault identification.

[0018] The formula of the mixed loss function is as follows: Mix =α×L 1 +(1-α)×L 2

[0019] Where L 1 To balance the cross entropy loss function, L 2 is the Dice loss function, and α is the weight coefficient.

[0020] Balanced cross entropy loss function:

[0021]

[0022] Dice loss function:

[0023]

[0024] Where:

[0025] N is the total number of pixels in the image, p i is the probability that the model predicts that the i-th pixel is a positive class (i.e., a fault in fault identification), y i is the true label of the ith pixel, and β represents the ratio between non-faulty pixels and total image pixels.

[0026] Furthermore, the AG-3D-Unet network that integrates the attention mechanism and residual learning is based on the structure of the Unet network, reduces the number of convolutional layers and the number of parameters of each convolution layer, and then adds the attention module to the connection layer between the encoder and its corresponding previous layer decoder, passes the fused features to the decoder of the next layer, and adds a residual connection after each convolutional layer of the network.

[0027] The present invention adopts two network improvement methods: one is a dual-channel 3D-Unet based on Unet, which changes the network into a dual-channel structure and adds an ASPP module to enhance the feature extraction capability. At the same time, in view of the imbalance problem between fault samples and non-fault sample data, a hybrid cross entropy function is introduced to improve the accuracy of the model in identifying fault detail information; the other is AG-3D-Unet that integrates attention mechanism and residual learning, which increases the feature expression ability of the model, does not introduce too many network parameters and improves the accuracy of fault identification. By comparing the recognition results of actual seismic data, the present invention has obvious improvements in the clarity, continuity and accuracy of fault identification, and has high practical value in the interpretation of actual seismic data faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a method for fault identification of three-dimensional seismic data according to one of the technical solutions of the present invention;

[0029] Figure 2 A flow chart of automatic modeling and generation of a three-dimensional seismic data set for one of the technical solutions of the present invention;

[0030] Figure 3 It is a schematic diagram of fault marking of the present invention;

[0031] Figure 4 The effect diagram generated by constructing three-dimensional seismic data in the present invention;

[0032] Figure 5 This is the improved dual-channel 3D-Unet model diagram of the present invention;

[0033] Figure 6 It is a schematic diagram of the specific structure of the dual-channel 3D-Unet model of the present invention;

[0034] Figure 7 This is the structural diagram of the AG-ResUnet network model of the present invention;

[0035] Figure 8Comparison diagram of fault identification results of the dual-channel 3D-Unet model of the present invention, (a), (e), (i) are original seismic images; (b), (f), (j) are label images of seismic images; (c) (g) (k) are fault identification images of 3D-Unet before improvement; (d) (h) (l) are fault identification images of improved dual-channel 3D-Unet;

[0036] Fig. 9 Comparison diagram of fault identification results of AG-ResUnet network of the present invention, (a), (e), (i) are original seismic images; (b), (f), (j) are label images of seismic images; (c), (g), (k) are fault identification images of Unet before improvement; (d), (h), (l) are improved AG-ResUnet fault identification images;

[0037] Fig.10 Comparison of the recognition results of the actual seismic data model of the present invention, (a) actual seismic image; (b) Unet fault recognition result; (c) dual-channel Unet recognition result; (d) AG-ResUnet recognition result;

[0038] Fig.11 is a horizontal reflection coefficient model diagram of the present invention;

[0039] Fig.12 is a schematic diagram of an inclined three-dimensional structure constructed by the present invention;

[0040] Fig.13 is a model diagram of the fold reflectance of the present invention;

[0041] Fig.14 This is a reflection coefficient model diagram of the present invention that contains both faults and folds. DETAILED DESCRIPTION

[0042] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] Reference Figure 1 As shown, the present invention proposes a three-dimensional seismic fault identification method based on deep learning, which specifically includes the following steps:

[0045] First, the model was built to generate a three-dimensional seismic data set and accurate fault labels. Second, improvements were made based on the 3D-Unet network to obtain dual-channel 3D-Unet and 3D-AG-Unet. Third, the improved network was used to simulate earthquake fault identification in data and obtain comparative results. Fourth, the improved network was used to identify earthquake faults in actual earthquake data. Finally, the networks before and after the improvement were compared to obtain actual comparative results. Figure 1 As shown, including:

[0046] S1. Modeling generates 3D seismic data sets and accurate fault labels. The overall process is as follows: Figure 2 Shown

[0047] S11. Construction of horizontal reflection coefficient model: By using random number sequences to represent the reflection intensity of different strata to seismic waves, a horizontal reflection coefficient model f(x,y,z) is established in the (x,y,z) coordinate space. Fig.11 Shown: The left picture is a three-dimensional structure, and the right picture is a two-dimensional cross-section.

[0048] Wrinkle structure construction: To simulate the wrinkles of the 3D structure model, we use a vertical shear plane model. The wrinkles are composed of two displacement fields S 1 (x,y,z) and S 2 The first displacement field S 1 (x,y,z) is defined as the following linear function,

[0049] S 1 (X,Y,Z)=aX+bY+c 0

[0050] In this model, where X, Y, Z represent spatial coordinates, parameters a, b and c 0 is randomly selected from a certain preset range. This displacement field S 1 (x, y, z) is used to create a purely tilted structure in the model. The tilt angles in the x and y directions are determined by two parameters, a and b, respectively. We limit the range of these two parameters to [-0.3, 0.3] to avoid combining S 2 An overly tilted structure is then formed.

[0051] The tilted three-dimensional structure and two-dimensional section of the structure are as follows Fig.12 Shown: The left picture shows the inclined three-dimensional structure of the structure, and the right picture shows the two-dimensional section.

[0052] However, the tilted structure alone cannot fully simulate the real folding phenomenon in the model. Therefore, we also introduce a vertical displacement field to add a bending structure to the model. The formula is as follows:

[0053]

[0054] Where: k represents the kth Gaussian function, σ k Indicates the radius width, b k Represents the amplitude. This displacement field is composed of N two-dimensional Gaussian functions and a linear scalar function (1.5 / Z max ). The kth Gaussian function is defined by their center position (c k ,d k ), radius width σ k and amplitude b k In order to prevent the model from producing too sharp curved structures, a smaller σ should be selected. k and b k By combining these N Gaussian functions, we can create a variety of curved structural morphologies. In addition, we assume that the curvature of the structure decreases from the bottom to the top in the vertical direction (z-axis).

[0055] We use these two types of displacement fields S 1 (x,y,z) and S 2 (x,y,z), which realizes the transformation of samples from plane space to wrinkle coordinate space.

[0056] The formula is:

[0057] The three-dimensional model and cross-sectional diagram of the fold reflectivity are shown in Figure 2. Fig.13 shown.

[0058] S12. Fault structure construction: In addition to fold structures, fault structures are another major type of structure in the model. Volume vector fields are used to simulate the fault structure in the structural model. Instead of interpreting specific faults through seismic data, we use the calculated three-dimensional displacement field to construct the fault structure within the model.

[0059] Fault plane simulation: By randomly selecting a reference point (X 0 , Y 0 , Z 0 ) as the starting point of the fault. Next, specify a strike angle and a dip angle θ∈(0°,90°) to indicate the direction of the fault. 0 , Y 0 , Z 0) as the origin to construct a local coordinate system (x, y, z), where the x-axis corresponds to the direction of movement, the y-axis corresponds to the tilt direction, and the z-axis points to the normal direction. According to this coordinate transformation rule, we can transform a point from global coordinates to local coordinates. The specific transformation method is as follows:

[0060]

[0061] Where R is:

[0062]

[0063] Fault displacement:

[0064] We define an ideal elliptic function as the displacement field d(x, y, z) = 0 on the fault plane. This is done to more accurately describe and simulate the displacement characteristics of the fault. 0, y 0 ,z 0 ) is the center, and its longest diameter is l x Along the strike direction, the shortest diameter l y Along the inclination direction.

[0065] Such as the formula:

[0066] Among them, l x , l y and the maximum displacement d at the center point max The expression r(x, y) represents the normalized radial distance from the center point.

[0067] Such as the formula:

[0068] The displacement reaches a maximum at the center point and decreases in all directions. In order to obtain more fault shapes, the elliptical shape can be further perturbed into an irregular shape.

[0069] 3D displacement field:

[0070] To create faults in the 3D structural model, we use a 3D displacement field approach. Specifically, we define a 3D vector field around the fault surface, where each spatial position corresponds to a vector.

[0071] For example, the formula: D(x,y,z)={D x (x,y,z)u x ,D y (x,y,z)u y ,D z (x,y,z)u z}

[0072] Here is the definition of u x ,u y ,u x , which are unit vectors along the x-axis, y-axis, and z-axis respectively. To simplify the implementation, we set the x component D x is zero everywhere, reflecting the absence of strike-slip displacement. This is not a necessary restriction of the model but a simplifying assumption. y There are different expressions for the upper and lower sides. For points in the upper plate block (0≤zf(x,y)≤γ).

[0073] The expression of inclination displacement is: D y (x,y,z) = λgd(x,y,z = 0)gα(x,y,z) where γ is the reverse resistance radius and λ(0 < λ < 1) is the ratio of the displacement of the hanging wall to the footwall. α(x,y,z) is a decreasing nonlinear scalar function that decreases as it moves away from the fault surface z = f(x,y).

[0074]

[0075] Since the fault is a curved surface, the displacement in the z direction must also be non-zero to ensure that the fault block moves seamlessly along the fault surface. z (x,y,z) displacements in the hanging wall and footwall blocks relative to the dip angle D y (x,y,z)

[0076] The formula is: D z (x,y,z)=f(x,y+D y (x,y,z))-f(x,y)

[0077] By defining the volume displacement field D in local coordinates z (x,y,z) and D y (x, y, z), we use the following three steps in the folding structure model (fold structure construction part) Faults occur in:

[0078] First of all, Perform coordinate transformation to get (x, y, z),

[0079] The formula is:

[0080] Second, apply the fault displacement in local coordinates (x,y,z)

[0081] The formula is: Finally, the coordinate transformation is performed to change (x, y, z) back to The formula is:

[0082]

[0083] By using the above steps, we can apply the volume displacement to the fold model and obtain a reflection coefficient model containing both faults and folds, such as Fig.14 shown.

[0084] S13. The labeling of faults is achieved by creating a label array of the same size as the seismic data model. When a fault is added to the seismic data model, the corresponding label array is also updated to indicate the location of the fault.

[0085] S14, further convolve the reflectivity model with the Ricker wavelet. The expression of the Ricker wavelet is:

[0086] Where f is the dominant frequency of the Ricker wavelet. Models with folds and faults help simulate more realistic faults in seismic images after applying convolution, because convolution blurs the sharp faults.

[0087] Then some random noise is added to the model to simulate the interference in real seismic data, making the synthetic seismic image more realistic. The final image is as follows Figure 4 shown.

[0088] S2. Network Improvement

[0089] The improved dual-channel structure of the 3D-Unet convolutional neural network is derived from the original 3D-Unet architecture. The model contains two independent data streams, which convert and extract the visual features of the input seismic image in different ways. The atrous spatial pyramid (ASPP) is added to the encoding and decoding connections to increase the receptive field and capture multi-scale image information.

[0090] The dual-channel 3D-Unet has two independent branches as downsampling parts, which are responsible for converting the input seismic image into a feature map in the latent space. Each group of layers in the branch consists of a pair of three-dimensional convolutional layers, a convolution kernel of size (2,2,2), and a three-dimensional pooling layer with a stride of 2. A Drop layer is added to the last part of the encoder to prevent the model from overfitting. In the decoding part, the two encoder channel outputs and the output of the previous layer of the decoder are spliced ​​through jump connections, and finally the output result is obtained through a 1×1×1 convolutional layer and a sigmoid function. Figure 5 , as shown in 6.

[0091] The improved dual-channel Unet network improves the accuracy of fault recognition by using a dual-channel structure to obtain more useful information and adding a dilated spatial convolutional pooling pyramid to enhance the capture of multi-scale features.

[0092] During the model training phase, a hybrid loss function was used to solve the problem of imbalance between fault data and non-fault sample data.

[0093] The formula of the mixed loss function is as follows: Mix =α×L 1 +(1-α)×L 2

[0094] Where L 1 To balance the cross entropy loss function, L 2 is the Dice loss function, and α is the weight coefficient.

[0095] In order to select the optimal α value, we set the value of α to 0.9, 0.7, 0.5, 0.3 and 0.1 respectively, and the number of iterations was set to 30. α is a parameter obtained through testing, and the optimal value is mainly found through experimental testing. Through experimental testing, the results show that when α is set to 0.5, the model can accurately identify faults while maintaining a low false recognition rate, and the judgment of overlapping areas is also relatively accurate.

[0096] Balanced cross entropy loss function:

[0097]

[0098] Dice loss function:

[0099]

[0100] Where:

[0101] N is the total number of pixels in the image, p i is the probability that the model predicts that the i-th pixel is a positive class (i.e., a fault in fault identification), y i is the true label of the ith pixel, β represents the ratio between non-fault pixels and total image pixels, and 1-β represents the proportion of fault pixels in the 3D seismic image.

[0102] The Dice loss function is particularly suitable for dealing with class imbalance problems because it focuses on the degree of overlap between the predicted area and the target area, rather than simply calculating the degree of matching of the class labels. The cross entropy loss is suitable for classification problems. By combining these two loss functions, the hybrid loss function can simultaneously consider the class imbalance and the accuracy of the prediction, thereby improving the model's recognition accuracy of the fault.

[0103] S3. Improved network earthquake fault identification

[0104] A 3D synthetic seismic image dataset was established, with 200 images in the training set and 20 images in the test set, with the size of (128, 128, 128). Before inputting into the network training, each seismic image of size (128, 128, 128) was normalized using the mean and standard deviation.

[0105] S4. Identification of real earthquake faults by the improved network

[0106] The real seismic data is a subset extracted from the seismic data of the F3 block offshore the Netherlands. The image size is (512,384,128), and the FaultSeg3D method is used as a reference standard for the interpretation of real geological faults.

[0107] Example 2

[0108] On the basis of Example 1, the difference from Example 1 is that the improvement of the network lies in the adoption of a new network structure of attention gate mechanism and residual learning - AG-ResUnet, which reduces the number of convolutional layers and the number of parameters of each convolution layer based on the structure of the Unet network. The attention module is added to the connection layer between the encoder and its corresponding previous decoder, and passes the fused features to the decoder of the next layer. Residual learning is added to the proposed AG-3D-Unet model, which is an improvement based on the original 3D-Unet. In this model, residual learning is used to strengthen network training, which is specifically achieved by adding residual connections (or skip connections) after each convolutional layer of the network. Figure 7 As shown in Figure 2. The attention module can highlight the significant features transmitted through the jump connection and eliminate noise and irrelevant responses to a certain extent. The attention module also filters the activation of neurons during forward propagation and back propagation. During the back propagation process, the gradient from the background area will be reduced in weight, while the fault area will be given a larger weight, so that the update of the shallow network parameters can be mainly based on the spatial area related to the fault.

[0109] Compared with the network before improvement, the dual-channel 3D-Unet and 3D-AG-Unet have improved in all indicators compared with the original 3D-Unet. And the parameter amount of the 3D-AG-Unet network is 1476640, which is the lowest among the three. Overall, although the parameter amount of the dual-channel 3D-Unet has increased, all indicators of fault identification have improved. Except for the accuracy rate value, the other indicators of the 3D-AG-Unet network are the best. The comparison results are as follows Figure 8 ,9 and shown in Table 1. Figure 8 It can be seen that the identified faults are more detailed and closer to the original label map. Fig. 9In the figure, it can be seen that compared with figure (c), the fault recognition effect in figure (d) is better, and the fault edge details can be more finely identified. In the comparison between figure (h) and figure (g), the fault recognition effect is good. In the last area, due to the presence of more faults, figure (l) can more effectively identify the fault area than figure (k), while figure (k) has some fault details lost and a small number of recognition errors. This shows that in the face of more complex faults, the proposed improved network has greater advantages.

[0110] Table 1 Evaluation indicators of the model before and after improvement

[0111]

[0112] Table 1 is a comparison of the data in Table 1 of the evaluation indexes of the model before and after the improvement of the present invention. The results show that the dual-channel 3D-Unet and 3D-AG-Unet are improved in various indicators compared with the original 3D-Unet, indicating that the method proposed in this paper has obvious improvement in fault identification effect.

[0113] The contrast effect is as follows Fig.10 As shown in the figure, by comparing different fault areas, the results show that the two improved networks proposed have significantly improved accuracy in fault identification of actual seismic images compared with the original network. For the dual-channel Unet, the number of network parameters is increased, but more small faults can be identified, and the identified faults are more complete and continuous. Compared with the attention mechanism fusion network, it has fewer network parameters, speeds up the network training speed, and has the best performance in the case of complex earthquake faults.

[0114] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A three-dimensional seismic fault identification method based on deep learning, characterized in that: The following steps are involved: (1) Automatically model and generate 3D seismic data and accurate fault labels to obtain a 3D seismic image dataset, which is then divided into a training set and a test set; (2) Using the 3D synthetic seismic data and its labels of the training set as input, the improved dual-channel structure 3D-Unet convolutional neural network or the AG-3D-Unet network that integrates the attention mechanism and residual learning is trained, and the test set is used for verification to obtain a trained earthquake fault recognition model; (3) Input real earthquake data and output fault identification results for the input image to be predicted through the earthquake fault identification model.

2. The three-dimensional seismic fault identification method based on deep learning according to claim 1, characterized in that: The modeling in step (1) includes constructing a fault structure model and marking the faults. The construction of the fault structure model includes constructing a horizontal reflection coefficient model, then constructing a fold coefficient reflection model, and finally constructing a reflection coefficient model containing both faults and folds.

3. The three-dimensional seismic fault identification method based on deep learning according to claim 2 is characterized in that: The specific method of constructing the horizontal reflection coefficient model is as follows: by using a random number sequence to represent the reflection intensity of different strata to seismic waves, a horizontal reflection coefficient model f(x,y,z) is established in the (x,y,z) coordinate space, where f(x,y,z) represents the reflection coefficient of each point in the space.

4. A three-dimensional seismic fault identification method based on deep learning according to claim 3, characterized in that: The specific method of constructing the wrinkle coefficient reflection model is as follows: using the vertical shear plane model, the wrinkles are defined by the combination of two displacement fields S1(x, y, z) and S2(x, y, z). S1(x, y, z) creates a pure inclined structure, and S2(x, y, z) adds a curved structure. By using these two types of displacement fields S1(x, y, z) and S2(x, y, z), the sample is transformed from the plane space to the wrinkle coordinate space.

5. A three-dimensional seismic fault identification method based on deep learning according to claim 4, characterized in that: The specific method of constructing a reflection coefficient model containing faults and folds is as follows: randomly select a reference point as the starting point of the fault, specify the strike angle and dip angle of the fault to determine the direction of the fault, and construct a local coordinate system, in which the x-axis corresponds to the strike direction, the y-axis corresponds to the dip direction, and the z-axis points to the normal direction. An ideal elliptical function is defined on the fault plane as the displacement field d(x, y, z) = 0 to accurately describe and simulate the displacement characteristics of the fault. Then, the three-dimensional displacement field method is used to create faults in the three-dimensional structural model. The wrinkle structure model Perform a coordinate transformation to get (x, y, z), then apply the fault displacement in the local coordinates (x, y, z), and finally perform a coordinate transformation to get (x, y, z) back to A reflection coefficient model containing both faults and folds is obtained.

6. The three-dimensional seismic fault identification method based on deep learning according to claim 5, characterized in that: The reflection coefficient model containing faults and folds is convolved with the Ricker wavelet, and some random noise is added to the model to simulate the interference in real seismic data, and finally a three-dimensional seismic image is generated.

7. The three-dimensional seismic fault identification method based on deep learning according to claim 2, characterized in that: The faults are labeled by creating a label array of the same size as the seismic data model.

8. The three-dimensional seismic fault identification method based on deep learning according to claim 1, characterized in that: The improved dual-channel structure of the 3D-Unet convolutional neural network is derived from the original 3D-Unet architecture. The model contains two independent data streams, which convert and extract the visual features of the input seismic image in different ways. A hollow space pyramid is added to the encoding and decoding connection parts to increase the receptive field and capture multi-scale image information. In the improved dual-channel 3D-Unet network, there are two independent branches as downsampling parts, which are responsible for converting the input seismic image into a feature map in the latent space. Each group of layers in the branch consists of a pair of three-dimensional convolutional layers, a convolution kernel of size (2,2,2), and a three-dimensional pooling layer with a stride of 2. A Drop layer is added to the last part of the encoder to prevent the model from overfitting. In the decoding part, the two encoder channel outputs and the output of the previous layer of the decoder are spliced ​​through jump connections, and finally the output result is obtained through a 1×1×1 convolutional layer and a sigmoid function.

9. The method for three-dimensional seismic fault identification based on deep learning according to claim 8, characterized in that: In the convolutional neural network training phase of the improved dual-channel 3D-Unet, a mixed loss function is used to evaluate the accuracy of fault recognition. The formula of the mixed loss function is as follows: Mix =α×L1+(1-α)×L2 Where L1 is the balanced cross entropy loss function, L2 is the Dice loss function, and α is the weight coefficient; Balanced cross entropy loss function: Dice loss function: Where: N is the total number of pixels in the image, p i is the probability that the model predicts that the i-th pixel is a positive class, y i is the true label of the ith pixel, and β represents the ratio between non-faulty pixels and total image pixels.

10. The method for three-dimensional seismic fault identification based on deep learning according to claim 8, characterized in that: The AG-3D-Unet network that integrates the attention mechanism and residual learning is based on the structure of the Unet network, reduces the number of convolutional layers and the number of parameters of each convolution layer, then adds the attention module to the connection layer between the encoder and its corresponding previous decoder, passes the fused features to the decoder of the next layer, and adds a residual connection after each convolutional layer of the network.

Images