A Deep Learning-Based Automatic Method for 3D Seismic Fault Identification

Abstract

This invention provides a deep learning-based method for three-dimensional seismic fault identification, comprising the following steps: (1) automatically modeling and generating three-dimensional seismic data and accurate fault labels to obtain a three-dimensional seismic image dataset, which is then divided into a training set and a test set; (2) using the three-dimensional synthetic seismic data and its labels from the training set as input to train an improved dual-channel structure 3D-Unet convolutional neural network or an AG-3D-Unet network incorporating attention mechanisms and residual learning, and validating the model using a test set to obtain a trained seismic fault identification model; (3) inputting real seismic data and outputting fault identification results from the input image to be predicted through the identification network. This invention significantly improves the clarity, continuity, and accuracy of fault identification, and has high practical value in the interpretation of actual seismic data faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic data processing, and particularly relates to a three-dimensional seismic fault automatic identification method based on deep learning. BACKGROUND

[0002] With the continuous advancement of industrial modernization, the demand for energy in the whole society is also increasing. Although many new energy sources have emerged, oil and natural gas are still one of the indispensable energy sources at present, so the development and exploration of oil and gas are particularly important. Underground faults are fracture accumulation 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 variable, and the use of conventional fault interpretation methods not only has a tedious process and takes a long time, but also the results are difficult to repeat and verify. SUMMARY

[0003] In view of the above problems of the prior art, the present application provides a three-dimensional seismic fault automatic identification method based on deep learning, which regards the three-dimensional seismic data body as a three-dimensional image, and uses deep learning and computer vision methods for fault interpretation. 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 scheme provided by the present application is a three-dimensional seismic fault automatic identification method based on deep learning, comprising the following steps:

[0005] (1) automatically modeling to generate three-dimensional seismic data and accurate fault labels, obtaining a three-dimensional seismic image data set, and dividing the data set into a training set and a test set;

[0006] (2) taking the three-dimensional synthetic seismic data of the training set and its labels as input, training the convolutional neural network of the improved double-channel structure 3D-Unet or the AG-3D-Unet network which fuses attention mechanism and residual learning, verifying by using the test set, and obtaining a trained seismic fault identification model;

[0007] (3) inputting real seismic data, and outputting a fault identification result by the seismic fault identification model for the inputted image to be predicted.

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

[0009] Further, the specific method for constructing the horizontal reflection coefficient model is as follows: a random number sequence is used to represent the reflection intensity of different strata on seismic waves, and a horizontal reflection coefficient model f(x, y, z) is established in the (x, y, z) coordinate space, f(x, y, z) representing the reflection coefficient of each point in the space.

[0010] Further, the specific method for constructing the fold coefficient reflection model is as follows: a vertical shear plane model is used, and the fold is defined by the combination of two displacement fields S1(x, y, z) and S2(x, y, z), S1(x, y, z) creating a pure tilt structure and S2(x, y, z) adding a curved structure, and by using the two types of displacement fields S1(x, y, z) and S2(x, y, z), the sample is converted from a plane space to a fold coordinate space.

[0011] Further, the specific method for constructing the reflection coefficient model containing faults and folds is as follows: a reference point is randomly selected as the starting point of the fault, the trend angle and the dip angle of the fault are specified to determine the direction of the fault, and a local coordinate system is constructed, in which the x-axis corresponds to the trend direction, the y-axis corresponds to the tilt direction, and the z-axis points to the normal direction. An ideal elliptic function is defined on the fault surface as a displacement field d(x, y, z) = 0, which accurately describes and simulates the displacement characteristics of the fault, and a three-dimensional displacement field method is used to create the fault in the three-dimensional structure model.

[0012] For the fold structure model (x, y, z) is obtained, secondly, the fault displacement is applied in the local coordinate (x, y, z), and finally the coordinate transformation is performed to change (x, y, z) back to to obtain a reflection coefficient model containing faults and folds.

[0013] Further, 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 the real seismic data, and finally a three-dimensional seismic image is generated.

[0014] Further, the label of the fault is realized by creating a label array with the same size as the seismic data model.

[0015] Further, the convolutional neural network of the improved dual-channel structure 3D-Unet is derived based on the original architecture of 3D-Unet, and the model contains two independent data streams, which respectively convert and extract the visual features of the input seismic image in different ways. In the coding and decoding connection part, a hollow space pyramid is added to increase the receptive field to capture multi-scale image information.

[0016] In the improved dual-channel 3D-Unet network, there are two independent branches as downsampling parts. These branches are responsible for converting the input seismic image into feature maps in the latent space. Each set of layers in the branches consists of a pair of three-dimensional convolutional layers with a convolutional kernel of size (2,2,2) and a three-dimensional pooling layer with a stride of 2. A drop layer is added at the end of the encoder to prevent the model from overfitting. In the decoding part, the output of the two encoder channels is concatenated with the output of the previous layer of the decoder through skip connections. Finally, the output is obtained by passing through a 1×1×1 convolutional layer and a sigmoid function.

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

[0018] The formula for the hybrid loss function is as follows: L Mix = α×L1+(1-α)×L2

[0019] Where L1 is the balanced cross-entropy loss function, L2 is the Dice loss function, and α is the weight coefficient.

[0020] Balanced cross-entropy loss function:

[0021]

[0022] Dice loss function:

[0023]

[0024] In the formula:

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

[0026] Furthermore, the AG-3D-Unet network, which integrates attention mechanisms and residual learning, is based on the structure of the Unet network. It reduces the number of convolutional layers and the number of parameters in each convolutional layer. Then, the attention module is added to the connection layer between the encoder and its corresponding previous layer decoder, and the fused features are passed to the next layer decoder. Residual connections are added after each convolutional layer of the network.

[0027] This invention employs two network improvement methods: one is a dual-channel 3D-Unet based on Unet, which modifies the network to a dual-channel structure and adds an ASPP module to enhance feature extraction capabilities. Simultaneously, to address the imbalance between fault and non-fault sample data, a hybrid cross-entropy function is introduced, thereby improving the model's accuracy in identifying fault details. The other method is AG-3D-Unet, which integrates attention mechanisms and residual learning, increasing the model's feature representation capabilities without introducing excessive network parameters and improving fault identification accuracy. By comparing the identification results with actual seismic data, this invention significantly improves fault identification clarity, fault continuity, and accuracy, demonstrating high practical value in the interpretation of faults in actual seismic data. Attached Figure Description

[0028] Figure 1 This is a flowchart of a three-dimensional seismic data fault identification method according to one of the technical solutions of the present invention;

[0029] Figure 2 A flowchart illustrating the automatic modeling and generation of a 3D seismic dataset for one of the technical solutions of this invention;

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

[0031] Figure 4 This is a rendering of the effect generated by constructing three-dimensional seismic data according to the present invention;

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

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

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

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

[0036] Figure 9The images show a comparison of the fault identification results of the AG-ResUnet network in this invention. (a), (e), and (i) are the original seismic images; (b), (f), and (j) are the labeled images of the seismic images; (c), (g), and (k) are the fault identification images of Unet before the improvement; and (d), (h), and (l) are the improved AG-ResUnet fault identification images.

[0037] Figure 10 For comparison of the actual earthquake data model identification results of this invention, (a) actual earthquake image; (b) Unet fault identification results; (c) dual-channel Unet identification results; (d) AG-ResUnet identification results;

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

[0039] Figure 12 This is a schematic diagram of the inclined three-dimensional structure constructed in this invention;

[0040] Figure 13 This is a model diagram of the wrinkle reflection coefficient of the present invention;

[0041] Figure 14 This is a reflection coefficient model diagram of the present invention that simultaneously contains faults and folds. Detailed Implementation

[0042] To make the objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

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

[0045] First, a 3D seismic dataset and accurate fault labels are generated through modeling. Second, the 3D-Unet network is improved to obtain dual-channel 3D-Unet and 3D-AG-Unet. Third, the improved network is used to identify seismic faults in simulated data, and comparative results are obtained. Fourth, the improved network is used to identify seismic faults in actual seismic data. Finally, the networks before and after improvement are compared to obtain actual comparative results. Figure 1 As shown, it includes:

[0046] S1. Modeling and generating a 3D seismic dataset and accurate fault labels, the overall process is as follows: Figure 2 As shown

[0047] S11. Construction of the horizontal reflection coefficient model: By using a random number sequence to represent the reflection intensity of seismic waves from different strata, a horizontal reflection coefficient model f(x,y,z) is established in the (x,y,z) coordinate space. For example... Figure 11 As shown: the left image is a three-dimensional structure, and the right image is a two-dimensional cross-sectional view.

[0048] Fold Structure Construction: To simulate the folds in the 3D structural model, we adopted a vertical shear plane model. The folds are defined by a combination of two displacement fields, S1(x,y,z) and S2(x,y,z). The first displacement field, S1(x,y,z), is defined as the following linear function:

[0049] S1(X,Y,Z)=aX+bY+c0

[0050] In this model, X, Y, and Z represent spatial coordinates, and parameters a, b, and c0 are randomly selected from a certain preset range. The displacement field S1(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 an excessively tilted structure after combining with S2.

[0051] The constructed inclined three-dimensional structure and two-dimensional cross-section are as follows Figure 12 As shown: the left figure is the constructed inclined three-dimensional structure, and the right figure is the two-dimensional cross-section.

[0052] However, the tilted structure alone cannot adequately simulate the true 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] In the formula: k represents the k-th Gaussian function, σ k Indicates the radius and width, b k The magnitude is represented by N two-dimensional Gaussian functions and one linear scalar function (1.5 / Z). max The k-th Gaussian function is defined by its center position (c). k ,d k ), radius and width σ k and amplitude b k The decision was made jointly, and these parameters were all randomly selected. To prevent overly sharp bends in the model, a smaller σ should be chosen. k and b kBy combining these N Gaussian functions, we can create a variety of curved structural forms. Furthermore, we assume that the degree of curvature of the structure gradually decreases from bottom to top in the vertical direction (z-axis).

[0055] We achieved the transformation of samples from planar space to folded coordinate space by using two types of displacement fields S1(x,y,z) and S2(x,y,z).

[0056] The formula is:

[0057] The 3D model and cross-sectional view of the wrinkle reflection coefficient are as follows: Figure 13 As shown.

[0058] S12. Fault Structure Construction: Besides folded structures, fault structures are another major structural type in the model. Volume vector fields are used to simulate fault structures within the structural model. Instead of interpreting specific faults using seismic data, we utilize calculated three-dimensional displacement fields to construct fault structures within the model.

[0059] Fault plane simulation: A reference point (X0, Y0, Z0) is randomly selected as the starting point of the fault. Then, a strike angle is specified. And a dip angle θ∈(0°,90°) is used to represent the direction of the fault. Then, we construct a local coordinate system (x,y,z) with the reference point (X0, Y0, Z0) as the origin, where 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. 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. This elliptic function is defined with respect to the origin (x,y,z)=0. 0, Centered on (y0, z0), its longest diameter is l x Along the direction of travel, the shortest diameter l y Along the direction of the inclination angle.

[0065] As in the formula:

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

[0067] As in the formula:

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

[0069] Three-dimensional displacement field:

[0070] To create faults in a three-dimensional structural model, we employed a three-dimensional displacement field approach. Specifically, we defined a three-dimensional vector field around the fault surface, where each spatial location 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] u is defined here. x u y u x These are unit vectors along the x-axis, y-axis, and z-axis, respectively. To simplify the implementation, we define the x-component D... x The value is zero everywhere, reflecting the absence of strike-slip displacement. This is not a necessary constraint on the model, but rather a simplifying assumption. The y-component D y There are different expressions on the upper and lower sides. For points in the upper block (0≤zf(x,y)≤γ).

[0073] The expression for the dip displacement is: D y (x,y,z)=λgd(x,y,z=0)gα(x,y,z) where γ is the radius of the reverse resistance, and λ (0<λ<1) is the ratio of the displacement of the hanging wall to the displacement of 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 The relative tilt angle displacement D of (x,y,z) in the upper and lower blocks.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 the local coordinate system z (x,y,z) and D y (x, y, z), we use the following three steps in the folded structure model (the folded structure construction part). Faults are generated in the middle:

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

[0079] The formula is:

[0080] Secondly, fault displacement is applied in the local coordinates (x, y, z).

[0081] The formula is: Finally, a coordinate transformation is performed to convert (x, y, z) back to its original form. The formula is:

[0082]

[0083] By using the steps above, we can apply volumetric displacement to the folding model and obtain a reflection coefficient model that simultaneously contains faults and folds, such as... Figure 14 As shown.

[0084] S13. Fault labeling 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 updated to indicate the location of the fault.

[0085] S14. Further convolve the reflectivity model with the Ricker wavelet. The expression for the Ricker wavelet is:

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

[0087] Then, some random noise is added to the model to simulate interference in real-world earthquake data, making the synthesized earthquake image more realistic. The final generated image is as follows. Figure 4 As shown.

[0088] S2, Network Improvements

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

[0090] The dual-channel 3D-Unet has two independent branches as downsampling parts, responsible for converting the input seismic image into feature maps in the latent space. Each layer in the branch consists of a pair of 3D convolutional layers with kernels of size (2,2,2) and a 3D pooling layer with a stride of 2. A drop layer is added at the end of the encoder to prevent overfitting. In the decoding part, the outputs of the two encoder channels are concatenated with the output of the previous layer of the decoder through skip connections, and finally passed through a 1×1×1 convolutional layer and a sigmoid function to obtain the output result. Figure 5 As shown in Figure 6.

[0091] The improved dual-channel Unet network enhances the accuracy of tomography identification 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 address the imbalance between tomographic and non-tomographic sample data in the tomographic data.

[0093] The formula for the hybrid loss function is as follows: L Mix = α×L1+(1-α)×L2

[0094] Where L1 is the balanced cross-entropy loss function, L2 is the Dice loss function, and α is the weight coefficient.

[0095] To select the optimal value of α, we set 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 was mainly found through experimental testing. Experimental results show that when α = 0.5, the model can accurately identify faults while maintaining a low false recognition rate, and the identification of overlapping regions is also relatively accurate.

[0096] Balanced cross-entropy loss function:

[0097]

[0098] Dice loss function:

[0099]

[0100] In the formula:

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

[0102] The Dice loss function is particularly suitable for handling class imbalance problems because it focuses on the degree of overlap between the predicted and target regions, rather than simply calculating the matching degree of class labels. Cross-entropy loss, on the other hand, is suitable for classification problems. By combining these two loss functions, the hybrid loss function can simultaneously consider class imbalance and prediction accuracy, thereby improving the model's accuracy in identifying tortuosities.

[0103] S3, Improved Network Seismic 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, each with a size of (128, 128, 128). Before inputting the data into the network for training, each seismic image of size (128, 128, 128) was normalized using the mean and standard deviation.

[0105] S4. Improved network for identifying real earthquake faults

[0106] The actual seismic data used was a subset extracted from the seismic data of the F3 block off the coast of the Netherlands. The image size was (512,384,128), and the FaultSeg3D method was used as a reference standard for interpreting the actual geological faults.

[0107] Example 2

[0108] Building upon Example 1, the improvement lies in the adoption of an attention gate mechanism and a novel network structure—AG-ResUnet—using residual learning. Based on the Unet network structure, the number of convolutional layers and the number of parameters per convolutional layer are reduced. An attention module is added to the connection layer between the encoder and its corresponding previous decoder, passing the fused features to the next layer's decoder. Residual learning is incorporated into the proposed AG-3D-Unet model, an improvement upon the original 3D-Unet. In this model, residual learning is used to enhance network training, specifically by adding residual connections (or skip connections) after each convolutional layer. Figure 7As shown, the attention module can highlight salient features passed through skip connections and, to some extent, eliminate noise and irrelevant responses. The attention module also filters neuron activations during forward and backward propagation. During backward propagation, gradients from background regions are weighted less, while fault regions receive greater weight. This allows the updates of shallow network parameters to be primarily based on spatial regions related to the faults.

[0109] Compared to the original 3D-Unet, both the dual-channel 3D-Unet and 3D-AG-Unet show improvements in all metrics. Furthermore, the 3D-AG-Unet network has the lowest number of parameters (1,476,640) among the three. Overall, although the dual-channel 3D-Unet has a higher parameter count, all metrics for tomography identification are improved. The 3D-AG-Unet network is the best in all metrics except accuracy. The comparison results are as follows: Figure 8 As shown in ,9 and Table 1. Figure 8 As can be seen, the identified faults are more refined and closer to the original label map. From Figure 9 As can be seen from Figure (d), compared to Figure (c), the fault identification performance in Figure (d) is better, and it can identify fault edge details more precisely. In the comparison between Figures (h) and (g), the fault identification performance is good in both. In the last region, due to the presence of more faults, Figure (l) can identify the fault region more effectively than Figure (k), while Figure (k) suffers from some loss of fault details and a small number of identification errors. This indicates that the proposed improved network has a greater advantage when dealing with more complex fault conditions.

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

[0111]

[0112] Table 1 shows the comparison of the data in Table 1 with the evaluation index 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 all indicators compared with the original 3D-Unet, indicating that the method proposed in this paper has a significant improvement in the tomographic identification effect.

[0113] Comparison effect as follows Figure 10 As shown, by comparing different fault regions, the results demonstrate that both proposed improved networks significantly improve accuracy in fault identification of actual seismic images compared to the original network. For the dual-channel Unet, the number of network parameters is increased, but it can identify more minute faults, and the identified faults are more complete and continuous. Compared to the Unet, the attention mechanism fusion network has fewer network parameters, accelerates network training, and exhibits the best performance in complex seismic fault scenarios.

[0114] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based three-dimensional seismic fault identification method, characterized in that, The method comprises the following steps: (1) automatically modeling to generate three-dimensional seismic data and accurate fault labels, to obtain three-dimensional seismic image data sets, and dividing the data sets into a training set and a test set; (2) taking the three-dimensional synthetic seismic data of the training set and the labels thereof as inputs, training a convolutional neural network of an improved 3D-Unet with a double-channel structure, verifying by using the test set, and obtaining a trained seismic fault identification model; (3) inputting real seismic data, and outputting a fault identification result by the seismic fault identification model for the inputted image to be predicted; The convolutional neural network of the improved 3D-Unet with a double-channel structure is derived based on the original architecture of the 3D-Unet, the model comprises two independent data streams, respectively converts and extracts visual features of inputted seismic images in different ways, and a hollow spatial pyramid is added in a coding and decoding connection part to increase a receptive field to capture multi-scale image information. In the improved dual-channel structure of 3D-Unet network, there are two independent branches as the down-sampling part, which are responsible for converting the input seismic image into feature maps in the latent space. Each set of layers in the branch is composed of a pair of three-dimensional convolutional layers, a convolution kernel with a size of (2, 2, 2), and a three-dimensional pooling layer with a step size of 2. A Drop layer is added at the last part of the encoder to prevent model overfitting. In the decoding part, the outputs of the two encoder channels are connected through a jump connection to splice the features with the output of the previous layer of the decoder. Finally, a convolutional layer and a sigmoid function are used to obtain the output result. In the training stage of the convolutional neural network of the improved 3D-Unet with a double-channel structure, a hybrid loss function is used to evaluate the accuracy of fault identification. The mixed loss function is formulated as follows , Wherein L1 is a balanced cross-entropy loss function, L2 is a Dice loss function, and a is a weight coefficient. 2.The method of claim 1, wherein, The modeling in the step (1) comprises constructing a fault structure model and marking a fault, the construction of the fault structure model comprises constructing a horizontal reflection coefficient model, then constructing a fold coefficient reflection model, and finally constructing a reflection coefficient model containing both a fault and a fold. 3.The method of claim 2, wherein, The specific method of constructing the horizontal reflection coefficient model is as follows: by using random number sequences to represent the reflection intensity of different strata to seismic waves, in a coordinate space, a horizontal reflection coefficient model is established , which represents the reflection coefficient of each point in the space. 4.The method of claim 3, wherein, The specific method of constructing the fold coefficient reflection model is as follows: adopting a vertical shear plane model, the fold is defined by the combination of two displacement fields and , creating a pure tilt structure, adding a bending structure, by using the two types of displacement fields and , the sample is converted from the plane space to the fold coordinate space. 5.The method of claim 4, wherein, The specific method for constructing the reflection coefficient model containing faults and folds is as follows: a reference point is randomly selected as the starting point of the fault, the trend angle and the dip angle of the fault are specified to determine the direction of the fault, and a local coordinate system is constructed, wherein the x axis corresponds to the trend direction, the y axis corresponds to the tilt direction, and the z axis points to the normal direction, and an ideal elliptic function is defined on the fault surface as a displacement field , which accurately describes and simulates the displacement characteristics of the fault, and the three-dimensional displacement field method is used to create the fault in the three-dimensional structure model, Model of fold structure The coordinate transformation is carried out to obtain Secondly, the fault displacement is applied in the local coordinate Finally, the coordinate transformation is carried out to change back to , and a reflection coefficient model containing both faults and folds is obtained.

6. The method of claim 5, wherein: The reflection coefficient model containing both the fault and the fold is convolved with a Ricker wavelet, some random noise is added in the model to simulate interference in real seismic data, and finally three-dimensional seismic images are generated.

7. The method of claim 2, wherein: The marking of the fault is realized by creating a label array with 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: Balanced cross-entropy loss function: ， Dice loss function: ， In the formula: N is the total number of pixels in the image, p i is the model's prediction of the probability that the ith pixel is positive, y i is the true label of the ith pixel, and β represents the ratio between non-defective pixels and total image pixels.

Images