Three-dimensional seismic fault automatic identification method

Abstract

This invention provides an automatic 3D seismic fault identification method, which includes: Step 1, preprocessing actual data fault interpretation sample data; Step 2, enhancing the actual data fault interpretation sample data; Step 3, automatically identifying 3D seismic faults; Step 4, designing a loss function for automatic 3D seismic fault identification; Step 5, training and validating the 3D seismic fault identification network; and Step 6, conducting practical application tests on the 3D seismic fault identification network method. This automatic 3D seismic fault identification method reconstructs and designs the 3D seismic fault network structure and loss function, ensuring that the training process considers the spatial structural characteristics of actual seismic data and the correspondence between faults, while avoiding interference from incomplete fault interpretation and partial annotation in uninterpreted areas, thus enhancing the reliability, accuracy, and efficiency of automatic fault identification.

Description

Technical Field

[0001] This invention relates to the field of exploration geophysics, and in particular to an automatic three-dimensional seismic fault identification method. Background Technology

[0002] Faults play a crucial role in hydrocarbon migration and accumulation. Obtaining accurate fault identification results requires extensive and tedious 3D seismic interpretation. To effectively improve interpretation efficiency and accuracy, conventional fault identification techniques such as coherence, variance, curvature, and fault likelihood analysis have been continuously developed. The main research approach is to enhance fault characteristics by highlighting discontinuities in phase axes through manually designed algorithms and processes. However, due to the influence of non-geological factors such as observational data and processing noise, the identification results contain considerable interference.

[0003] To address the aforementioned issues, seismic automatic interpretation methods based on deep neural networks have garnered significant attention. These methods first construct sample data through theoretical model forward modeling or practical interpretation. Then, driven by this sample data, deep neural networks learn the mapping relationship between seismic reflection characteristics and faults, achieving end-to-end automatic fault identification. The quantity and quality of the sample data are crucial to the success of these methods. While theoretical model forward modeling offers the advantage of rapidly obtaining large amounts of sample data, it cannot fully reflect the true distribution characteristics of faults and seismic reflection features, potentially leading to low accuracy in fault identification. Practical interpretation, while starting from real seismic reflection characteristics, is often limited by objective factors such as the level of exploration understanding and the workload of interpretation. It typically only performs detailed fault interpretation on key sections and areas of the 3D seismic survey area, resulting in only partial labeling of faults in the overall actual seismic data. This problem severely impacts the judgment of data-driven automatic fault identification network models, posing a significant challenge to the accuracy of deep neural network training and potentially leading to weak generalization ability of the fault identification model. Therefore, further methodological research is needed to improve the utilization rate of existing interpretation results, effectively increase the number and quality of samples, and enhance the reliability and generalization ability of the three-dimensional seismic fault automatic identification network model.

[0004] Chinese patent application CN202011006282.8 discloses a method and system for generating fault detection training data. The method includes: performing the following iterative processing: obtaining a fault deformation seismic data volume based on an initial seismic data volume, and obtaining a fault deformation fault label based on an initial fault label; when the current iteration count reaches a preset iteration count, determining the fault deformation seismic data volume and the fault deformation fault label in the current iteration to generate a fault detection training sample pair; otherwise, updating the initial seismic data volume based on the fault deformation seismic data volume, updating the initial fault label based on the fault deformation fault label, and continuing the iterative processing; generating a fault detection training dataset based on a preset number of fault detection training sample pairs. This invention can provide a low-cost, high-efficiency, and high-precision fault detection training dataset containing a sufficient number of sample pairs, thereby improving the accuracy and generalization ability of deep learning fault detection networks and accelerating the research process. However, this application only addresses the generation of fault sample data and their corresponding seismic forward simulation data using forward modeling, which cannot fully reflect the actual fault development characteristics and seismic reflection characteristics.

[0005] Chinese patent application CN202010909049.4 discloses a method and system for detecting seismic data discontinuities based on a deep learning model. The method includes: classifying seismic data into three categories—simple, moderate, and difficult—and processing them differently according to each category; generating initial data labels for simple seismic data to establish an initial deep learning model; training the initial deep learning model using these initial data labels to obtain a final deep learning model; and inputting the seismic wave signal to be detected into the final deep learning model to detect discontinuities in the signal. This invention detects seismic data discontinuities by simply inputting the actual seismic data into the trained neural network model, achieving a significantly faster calculation speed than traditional methods. Furthermore, this method starts from actual seismic data, eliminating the need for manual labeling and reducing the uncertainty of human factors, thus effectively improving the accuracy of the detection results. However, while this application starts from actual data, the sample data uses time slices with different seismic attributes and does not perform automatic fault identification in three-dimensional space. Additionally, it does not perform any data enhancement processing on the sample data, which can lead to insufficient sample size when actual interpretable data is scarce.

[0006] Chinese patent application CN202010423320.3 discloses a method and apparatus for processing seismic fault data, comprising: acquiring seismic fault data; processing the acquired seismic fault data using a pre-trained fault data processing model to obtain fault data processing results, and conducting seismic fault exploration based on the fault data processing results; wherein the fault data processing model is obtained by training a pre-established convolutional neural network model using acquired fault sample training and validation datasets. This application uses a deep learning method to process the results identified by the fault identification algorithm, and combines image dilation and erosion techniques with skeleton extraction techniques to improve the continuity, purity, and accuracy of faults, thereby providing reliable results for fault interpretation. However, the automatic fault identification in this application uses forward modeling fault data, which does not consider actual fault interpretation data and cannot fully reflect the actual fault development characteristics and seismic reflection characteristics.

[0007] Chinese patent application CN202110946552.1 discloses a seismic fault identification method based on deep learning semantic segmentation, comprising: acquiring three-dimensional seismic data and extracting slices from it to obtain a two-dimensional seismic amplitude image; inputting the obtained two-dimensional seismic amplitude image into a fault segmentation model, and the fault segmentation model inputting a binary fault image; optimizing the binary fault image to obtain a fault identification result; wherein the fault segmentation model is trained using fault training samples and a deep learning semantic segmentation model. This invention, based on deep learning semantic segmentation, employs a pattern recognition approach, leveraging the powerful fitting ability of deep convolutional neural networks to learn patterns for interpreting faults. It utilizes computer vision to pick out faults from seismic amplitude images, achieving fast and accurate fault identification in three-dimensional seismic data. Simultaneously, it reduces human intervention and errors during the identification process and shortens the time required for fault interpretation. Although the application takes into account the seismic reflection characteristics corresponding to the actual faults, the sample data and its enhancement processing are based only on two-dimensional profiles. First, it does not perform automatic fault identification in three-dimensional space, and second, it does not fully consider the partial annotation problems that occur when the fault interpretation of the actual data is incomplete.

[0008] The existing technologies described above are significantly different from the present invention and have failed to solve the technical problem we want to address. Therefore, we have invented a new method for automatic identification of three-dimensional seismic faults. Summary of the Invention

[0009] The purpose of this invention is to provide a three-dimensional seismic fault automatic identification method that enhances the reliability, accuracy, and efficiency of automatic fault identification.

[0010] The objective of this invention can be achieved through the following technical measures: a three-dimensional seismic fault automatic identification method, which includes:

[0011] Step 1: Preprocess the actual data for fault interpretation sample data;

[0012] Step 2: Enhance the sample data for tomographic interpretation of actual data;

[0013] Step 3: Perform automatic identification of three-dimensional seismic faults;

[0014] Step 4: Design the loss function for automatic 3D seismic fault identification;

[0015] Step 5: Train and validate the 3D seismic fault automatic identification network;

[0016] Step 6: Conduct practical application tests on the three-dimensional seismic fault automatic identification network method.

[0017] The objective of this invention can also be achieved through the following technical measures:

[0018] In step 1, the actual seismic profile data and its corresponding fault interpretation results are resampled or interpolated along the vertical time and horizontal seismic trace directions, respectively. The data time sampling interval and trace spacing are unified to obtain the preprocessed seismic profile data and its corresponding fault label data.

[0019] In step 1, for any two-dimensional seismic profile S0, its dimensions in the horizontal X direction and the vertical time T direction are N. x ×N t In this process, the fault control points actually interpreted by experts are segmented and interpolated to obtain continuous fault interpretation results data; the sampling point location closest to the fault line is encoded as 1, indicating that it is an interpreted fault point; other sampling points are encoded as 0, indicating uninterpreted areas, including non-fault points and uninterpreted fault points. After obtaining the encoding of each seismic sampling point location, a two-dimensional fault sample label F0 corresponding to the size of the two-dimensional seismic profile is obtained.

[0020] In step 2, the preprocessed seismic profile data and its corresponding fault label data are augmented in dimension, and then randomly twisted and rotated in azimuth to form a large amount of three-dimensional actual seismic sample feature enhancement data and fault sample label enhancement data.

[0021] In step 2, the two-dimensional sample data is expanded to three-dimensional sample data. The seismic data S0 and fault sample label F0 are copied and expanded along the direction perpendicular to the profile to obtain the expanded three-dimensional seismic data S1 and fault sample label data F1. The dimensions of the two-dimensional sample data S1 in the horizontal Y direction, the horizontal X direction, and the vertical time T direction are N. y ×N x ×N t .

[0022] In step 2, three-dimensional cross-section distortion is performed. To further enrich the style of the fault samples, the three-dimensional seismic data S1 and fault sample label data F1 after dimensional expansion are arbitrarily distorted along the vertical cross-section direction, assuming it is the X direction, to obtain the seismic data S2 and fault label data F2 after cross-section distortion.

[0023] In step 2, a three-dimensional cross-sectional azimuth rotation is performed. The distorted three-dimensional seismic data and cross-sectional sample label data are rotated along different azimuth angles. Then, interpolation is performed according to the new spatial coordinates to obtain the rotated three-dimensional actual seismic sample feature enhancement data volume S and fault sample label enhancement data volume F.

[0024] In step 2, a set of enhanced data volumes for 3D actual earthquake sample features and a set of enhanced data volumes for fault sample labels are generated. Different torsion and azimuth rotation parameters are randomly set, and the above-mentioned fault torsion and azimuth rotation process is repeated to obtain a large number of sets of enhanced data volumes for 3D actual earthquake sample features and sets of enhanced data volumes for fault sample labels with different bending morphologies and azimuth rotation angles, thereby increasing the diversity of actual fault interpretation sample data.

[0025] In step 3, an automatic three-dimensional seismic fault identification network structure is constructed, which is suitable for feature enhancement data of three-dimensional actual seismic samples and fault sample label enhancement data.

[0026] In step 3, the input to the 3D seismic fault automatic identification network is the 3D actual seismic sample feature enhancement data volume, and the output is the 3D actual fault sample label enhancement data volume.

[0027] In step 3, the forward propagation process of the 3D seismic fault automatic identification network includes two steps: encoding and decoding. Encoding involves three feature extraction processes, extracting seismic reflection features from small scale to large scale. Decoding involves three feature recovery processes, recovering fault identification results with features at different scales. The output of each decoding layer is connected to the output of the encoding layer to fuse features from encoders at different scales and upper decoding layers, in order to capture fine-grained and coarse-grained semantics across the entire scale. At the same time, to better analyze the spatial correlation between different fault points on the same fault, attention layers are introduced in each encoding and decoding layer. That is, each encoding layer is a connection of a convolutional layer, an attention layer, an activation function, and a max pooling layer, and each decoding layer is a connection of a transposed convolutional layer, an attention layer, and an activation function, to enhance the network's ability to analyze spatial features of data.

[0028] In step 4, a loss function is designed that is applicable to the feature enhancement data of partially annotated 3D actual seismic samples and the fault sample label enhancement data.

[0029] In step 4, the loss function between the prediction results of the 3D seismic fault automatic identification network model and the known fault labels can be expressed as:

[0030]

[0031] Where F is the label of the actual fault sample. The result is the prediction of the 3D seismic fault automatic identification network. f{} represents the forward propagation process of the 3D seismic fault automatic identification network, and W is the network weight parameter to be optimized.

[0032] In step 4, the fault labels are explicit in the interpreted region, while in the uninterpreted region, both non-fault seismic reflection features and uninterpreted fault reflection features are included. Therefore, the calculation of the training loss function should focus on the interpreted region, and emphasis should be placed on strengthening the correlation analysis between fault labels and seismic reflection features in this region. The quantitative expression focusing on the interpreted fault region is as follows:

[0033] B = g * F

[0034] Where g is the Gaussian convolution kernel, g > 1, * represents the convolution process, and B is the data volume of the focused region, with the same size as the tomographic sample label F; based on this, a mask weight matrix M is constructed, with dimensions N in the horizontal Y direction, horizontal X direction, and vertical time T direction. y ×N x ×N t The specific expression is:

[0035]

[0036] The improved automatic fault identification loss function is:

[0037]

[0038] Among them, M·F and These represent the mask weight matrix M, the fault label F, and the fault prediction result, respectively. The matrix elements are multiplied accordingly. The first term of the loss function focuses on the adjacent region of the explained fault location. The fault sample labels in this region are determined, and the number of fault and non-fault samples is more balanced, which helps to learn the correlation between the seismic discontinuous reflection features and fault labels. The second term focuses on the region of the unexplained fault. This region contains a large number of true non-fault labels and a small number of unlabeled erroneous non-fault labels, focusing on learning the correlation between the seismic continuous reflection features and non-fault labels.

[0039] In step 5, the 3D actual earthquake sample feature enhancement data and fault sample label enhancement data are split into training sample sets and validation sample sets. Key parameters for network training are set. Driven by the training sample set data, the fault automatic identification network is forward propagated, and the loss function between the prediction results and fault labels is calculated. The weight parameters of the fault automatic identification network are iteratively optimized and solved repeatedly. The result that meets the requirements of the validation set is used as the final fault automatic identification network model.

[0040] In step 5, preprocessing is performed to ensure consistency in the size of the 3D actual sample data. The sizes of the enhanced feature data and fault label enhanced data of the 3D actual seismic samples are statistically analyzed, and the minimum size in the horizontal Y direction, horizontal X direction, and vertical time T direction is obtained as N. y_min ×N x_min ×N t_min Based on this, other data in the 3D actual seismic sample feature enhancement data volume set and the fault sample label enhancement data volume set are truncated using a sliding time window method, and the size of all sample data is unified to a minimum size N. y_min ×N x_min ×N t_min .

[0041] In step 5, three-dimensional fault forward modeling sample data is generated according to the minimum size N. y_min ×N x_min ×N t_min Based on the actual geological conditions of the work area, horizontal layered reflection coefficient data volumes are randomly generated, and parameters such as random structural undulations, fault dip angles, and fault displacements are set to obtain reflection coefficient data volumes after tectonic movement. Then, seismic wavelets are set according to the actual seismic data of the work area, and seismic forward modeling is performed to obtain fault sample labels based on forward modeling and their corresponding seismic forward modeling sample feature data, thereby expanding the number of samples used for the automatic fault identification network model.

[0042] In step 5, the training and validation sets of the three-dimensional seismic fault automatic identification sample data are split. The sample data is split into training and validation sets by random selection.

[0043] In step 5, the weight parameters of the 3D seismic fault automatic identification network are optimized. Key parameters such as the number of iterations and the learning rate are set. Driven by the training sample set data, the fault automatic identification network is forward propagated, and the loss function between the prediction results and the fault labels is calculated. The weight parameters of the fault automatic identification network are iteratively optimized and solved repeatedly.

[0044] In step 5, the 3D seismic fault automatic identification network model is output. The trained fault automatic network model is then used to test the validation set. If the prediction results meet the requirements, the final 3D seismic fault automatic identification network model is output. Otherwise, the network parameters are readjusted until the validation set test requirements are met.

[0045] In step 6, the final automatic fault identification network model obtained from the training is used to conduct an application test of three-dimensional automatic fault identification in the actual work area.

[0046] The present invention provides an automatic three-dimensional seismic fault identification method, comprising: preprocessing of actual data fault interpretation sample data; enhancement of actual data fault interpretation sample data; automatic three-dimensional seismic fault identification; design of a loss function for automatic three-dimensional seismic fault identification; training and validation of the automatic three-dimensional seismic fault identification network; and practical application testing of the automatic three-dimensional seismic fault identification network method. This invention expands the dimensionality and quantity of the actual interpretation sample data, effectively increasing the style and scale of the fault sample data while retaining the seismic reflection characteristics of actual faults. It reconstructs and designs the three-dimensional seismic fault network structure and loss function, ensuring that the training process considers the spatial structural characteristics of actual seismic data and the correspondence between faults, while avoiding interference from incomplete fault interpretation and partial annotation in uninterpreted areas, thus enhancing the reliability, accuracy, and efficiency of automatic fault identification.

[0047] Compared with the prior art, the main advantages of the present invention are as follows:

[0048] (1) A three-dimensional data enhancement method for actual fault interpretation samples is proposed to expand the dimension and quantity of actual interpretation sample data. While preserving the seismic reflection characteristics of actual faults, it effectively expands the style and scale of fault sample data and improves the stability of weight parameter solution during the training process of fault automatic identification network.

[0049] (2) A three-dimensional seismic fault network structure and loss function were constructed and designed. This ensures that the training process takes into account the spatial structural characteristics of actual seismic data and the correspondence between faults, and avoids interference caused by incomplete fault interpretation and partial annotation in uninterpreted areas, thereby enhancing the reliability, accuracy and efficiency of automatic fault identification. Attached Figure Description

[0050] Figure 1 This is a flowchart of a specific embodiment of the three-dimensional seismic fault automatic identification method of the present invention;

[0051] Figure 2 This is an overlay image of fault label data and seismic data after preprocessing the actual data fault interpretation results in a specific embodiment of the present invention;

[0052] Figure 3This is an overlay image of the dimensionality expansion result of actual fault sample data and seismic data in a specific embodiment of the present invention;

[0053] Figure 4 This is an overlay image of the distortion and azimuth rotation enhancement results of actual fault sample data and seismic data in a specific embodiment of the present invention;

[0054] Figure 5 This is a diagram of the automatic interpretation network structure of three-dimensional seismic faults in a specific embodiment of the present invention;

[0055] Figure 6 This is an overlay diagram of the earthquake profile to be predicted and some of the labeled fault interpretation sample labels in a specific embodiment of the present invention;

[0056] Figure 7 This is a fault real label map corresponding to the earthquake profile to be predicted in a specific embodiment of the present invention;

[0057] Figure 8 This is a diagram showing the fault identification result of a seismic profile to be predicted using a conventional dice loss function in a specific embodiment of the present invention.

[0058] Figure 9 This is a diagram showing the fault identification result of the seismic profile to be predicted using the loss function proposed in this patent in a specific embodiment of the present invention.

[0059] Figure 10 This is a superimposed image of automatic earthquake identification results and earthquake profiles proposed in this patent in a specific embodiment of the present invention. Detailed Implementation

[0060] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0061] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0062] The automatic 3D seismic fault identification method of this invention first preprocesses and enhances the actual fault interpretation data to obtain a large set of enhanced 3D actual seismic sample feature data and a set of enhanced fault sample label data. Second, it constructs an automatic fault identification network structure suitable for the enhanced 3D sample data of actual fault interpretation. Third, it designs a loss function suitable for the automatic 3D seismic fault identification of partially labeled actual fault interpretation sample data. For example... Figure 1 As shown, Figure 1 This is a flowchart of the automatic three-dimensional seismic fault identification method of the present invention. The automatic three-dimensional seismic fault identification method specifically includes:

[0063] (1) Preprocessing of actual seismic profile data and its corresponding fault interpretation results. The actual seismic profile data and its corresponding fault interpretation results are resampled or interpolated along the vertical time and horizontal seismic trace directions, respectively. The data time sampling interval and trace spacing are unified to obtain the preprocessed seismic profile data and its corresponding fault label data.

[0064] (2) Enhancement of actual data fault interpretation sample data. The preprocessed seismic profile data and its corresponding fault label data are expanded in dimension and randomly twisted and rotated in azimuth to form a large amount of three-dimensional actual seismic sample feature enhancement data and fault sample label enhancement data.

[0065] (3) Construction of a 3D seismic fault automatic identification network structure. A 3D seismic fault automatic identification network structure suitable for feature enhancement data of actual 3D seismic samples and fault sample label enhancement data is constructed.

[0066] (4) Design of loss function for automatic identification of 3D seismic faults. The design is applicable to feature enhancement data of partially labeled 3D actual seismic samples and fault sample label enhancement data.

[0067] (5) Training and validation of the 3D seismic fault automatic identification network. The 3D actual seismic sample feature enhancement data and fault sample label enhancement data were split into training sample sets and validation sample sets. Key parameters for network training were set. Driven by the training sample set data, the fault automatic identification network was forward propagated, and the loss function between the prediction results and the fault labels was calculated. The weight parameters of the fault automatic identification network were iteratively optimized and solved repeatedly. The result that meets the requirements of the validation set was used as the final fault automatic identification network model.

[0068] (6) Practical application test of the three-dimensional seismic fault automatic identification network method. Using the final fault automatic identification network model obtained from the training, the application test of three-dimensional fault automatic identification was carried out in the actual work area.

[0069] The following are several specific embodiments of the application of the present invention.

[0070] Example 1:

[0071] Step 1: Resample or interpolate the actual seismic profile data and its corresponding fault interpretation results along the vertical time and horizontal seismic trace directions, respectively, and unify the data time sampling interval and trace spacing to obtain the preprocessed seismic profile data and its corresponding fault label data.

[0072] For any two-dimensional seismic profile S0, its dimensions in the horizontal X direction and the vertical time T direction are N. x ×N t In this process, segmented interpolation is performed on the fault control points actually interpreted by experts to obtain continuous fault interpretation data. The sampling point location closest to the fault line is encoded as 1, indicating an interpreted fault point; other sampling points are encoded as 0, indicating uninterpreted areas, including both non-fault points and uninterpreted fault points. After obtaining the encoding of each seismic sampling point location, a two-dimensional fault sample label F0 corresponding to the size of the two-dimensional seismic profile is obtained.

[0073] Step 2 involves dimensional expansion of the preprocessed seismic profile data and its corresponding fault label data, followed by random distortion and azimuth rotation to generate a large amount of three-dimensional actual seismic sample feature enhancement data and fault sample label enhancement data.

[0074] Two-dimensional sample data is augmented to three-dimensional sample data. Seismic data S0 and fault sample labels F0 are copied and augmented along a direction perpendicular to the profile to obtain augmented three-dimensional seismic data S1 and fault sample label data F1. The dimensions of S1 in the horizontal Y direction, horizontal X direction, and vertical time T direction are N. y ×N x ×N t .

[0075] 3D cross-section distortion. To further enrich the pattern of the fault samples, the dimensionally expanded 3D seismic data S1 and fault sample label data F1 are arbitrarily distorted along the vertical profile direction (assuming it is the X direction) to obtain the distorted seismic data S2 and fault label data F2.

[0076] 3D cross-sectional azimuth rotation. The distorted 3D seismic data and cross-sectional sample label data are rotated along different azimuth angles. Then, interpolation is performed according to the new spatial coordinates to obtain the rotated 3D actual seismic sample feature enhancement data volume S and fault sample label enhancement data volume F.

[0077] The process involves generating enhanced feature volumes for 3D actual seismic samples and enhanced fault label volumes for fault samples. By randomly setting different torsion and azimuth rotation parameters, the above process of section torsion and azimuth rotation is repeated to obtain a large number of enhanced feature volumes for 3D actual seismic samples and enhanced fault label volumes for fault samples with different bending morphologies and azimuth rotation angles, thereby increasing the diversity of actual fault interpretation sample data.

[0078] Step 3: Construct a three-dimensional earthquake fault automatic identification network structure suitable for feature enhancement data of actual three-dimensional earthquake samples and fault sample label enhancement data.

[0079] The input to the 3D seismic fault automatic identification network is the 3D actual seismic sample feature enhancement data volume, and the output is the 3D actual fault sample label enhancement data volume.

[0080] The forward propagation process of the 3D seismic fault automatic identification network includes two steps: encoding and decoding. Encoding involves three feature extraction processes, extracting seismic reflection features from small to large scales. Decoding involves three feature recovery processes, recovering fault identification results at different scales. The output of each decoding layer is concatenated with the output of the encoding layer to fuse features from encoders at different scales and upper decoding layers, capturing both fine-grained and coarse-grained semantics across the entire scale. Furthermore, to better analyze the spatial correlation between different fault points on the same fault plane, attention layers are introduced in each encoding and decoding layer. Specifically, each encoding layer consists of a convolutional layer, an attention layer, an activation function, and a max-pooling layer; each decoding layer consists of a transposed convolutional layer, an attention layer, and an activation function, enhancing the network's ability to analyze spatial features.

[0081] Step 4: Design a loss function applicable to partially annotated 3D actual seismic sample feature enhancement data and fault sample label enhancement data.

[0082] The loss function between the prediction results of the 3D seismic fault automatic identification network model and the known fault labels can be expressed as:

[0083]

[0084] Where F is the label of the actual fault sample. The result is the prediction of the 3D seismic fault automatic identification network. f{} represents the forward propagation process of the 3D seismic fault automatic identification network, and W is the network weight parameter to be optimized.

[0085] In the interpreted region, the fault labels are explicit, while in the uninterpreted region, both non-fault seismic reflection features and uninterpreted fault reflection features are included. Therefore, the calculation of the training loss function should focus on the interpreted region, and emphasis should be placed on strengthening the correlation analysis between fault labels and seismic reflection features in this region. The quantitative expression focusing on the interpreted fault region is as follows:

[0086] B = g * F

[0087] Where g is the Gaussian convolution kernel (g > 1), * represents the convolution process, and B is the data volume of the focused region, with the same size as the tomographic sample label F. Based on this, a mask weight matrix M is constructed, with dimensions N in the horizontal Y direction, horizontal X direction, and vertical time T direction. y ×N x ×N t The specific expression is:

[0088]

[0089] The improved automatic fault identification loss function is:

[0090]

[0091] Among them, M·F and These represent the mask weight matrix M, the fault label F, and the fault prediction result, respectively. The matrix elements are multiplied accordingly. The first term of the loss function focuses on the adjacent region of the interpreted fault location. Within this region, the fault sample labels are determined, and the number of fault and non-fault samples is more balanced, which helps in learning the correlation between seismic discontinuous reflection features and fault labels. The second term focuses on the region of uninterpreted faults. This region contains a large number of true non-fault labels and a small number of unlabeled erroneous non-fault labels, emphasizing the learning of the correlation between seismic continuous reflection features and non-fault labels.

[0092] Step 5: Split the 3D actual earthquake sample feature enhancement data and fault sample label enhancement data into training and validation sample sets, set key parameters for network training, and perform forward propagation through the fault automatic identification network under the guidance of the training sample set data. Calculate the loss function between the prediction results and the fault labels, iteratively optimize and solve the weight parameters of the fault automatic identification network, and use the results that meet the validation set requirements as the final fault automatic identification network model.

[0093] Preprocessing to ensure consistency in the size of 3D actual sample data. The sizes of the enhanced feature data and fault label enhanced data of 3D actual seismic samples are statistically analyzed, and the minimum size in the horizontal Y-direction, horizontal X-direction, and vertical time T-direction is obtained as N. y_min×N x_min ×N t_min Based on this, other data in the 3D actual seismic sample feature enhancement data volume set and the fault sample label enhancement data volume set are truncated using a sliding time window method, and the size of all sample data is unified to a minimum size N. y_min ×N x_min ×N t_min ;

[0094] Optional, 3D fault forward modeling sample data generation. Based on the minimum size N. y_min ×N x_min ×N t_min Based on the actual geological conditions of the work area, horizontal layered reflection coefficient data volumes are randomly generated, and parameters such as random structural undulations, fault dip angles, and fault displacements are set to obtain reflection coefficient data volumes after tectonic movement. Then, seismic wavelets are set according to the actual seismic data of the work area, and seismic forward modeling is performed to obtain fault sample labels and their corresponding seismic forward modeling sample feature data based on forward modeling, thereby expanding the number of samples used for the automatic fault identification network model.

[0095] Splitting of Training and Validation Sets for Automatic 3D Seismic Fault Identification Sample Data. The sample data is randomly selected to split into training and validation sets.

[0096] Optimization of weight parameters for 3D seismic fault automatic identification network. Key parameters such as iteration count and learning rate are set. Driven by training sample set data, forward propagation is performed through the fault automatic identification network, and the loss function between the prediction results and fault labels is calculated. The weight parameters of the fault automatic identification network are iteratively optimized and solved repeatedly.

[0097] Output of the 3D seismic fault automatic identification network model. Using the trained fault automatic identification network model, the validation set is tested. If the prediction results meet the requirements, the final 3D seismic fault automatic identification network model is output; otherwise, the network parameters are readjusted until the validation set test requirements are met.

[0098] Step 6: Using the final automatic fault identification network model obtained from the training, conduct an application test of automatic three-dimensional fault identification in the actual work area.

[0099] Example 2:

[0100] In a specific embodiment 2 of this invention, 200 pairs of three-dimensional fault models and seismic data, each with a size of 256×256×256 in the horizontal Y-direction, horizontal X-direction, and vertical time T-direction, were constructed using forward modeling. Random sampling of the simulated actual three-dimensional seismic data along different X and Y directions yielded a large number of two-dimensional seismic profiles with a size of 256×256. Local fault interpretation was then performed on these profiles to obtain the actual seismic data and fault interpretation results from the forward modeling. The reliability of the method was verified based on the above data.

[0101] Step 1: Resample or interpolate the actual seismic profile data and its corresponding fault interpretation results along the vertical time and horizontal seismic trace directions, respectively. Unify the data time sampling interval and trace spacing to obtain preprocessed seismic profile data and its corresponding fault label data, such as... Figure 2 As shown.

[0102] Step 2: Enhancement of sample data for actual data fault interpretation.

[0103] like Figure 3 As shown, the preprocessed seismic profile data and its corresponding fault label data are augmented with additional dimensions.

[0104] like Figure 4 As shown, the sample data after dimensional expansion is randomly distorted, and the bending cross-sectional shape is simulated by superimposing multiple sinusoidal signals. The distorted sample data is then rotated by azimuth angle.

[0105] By randomly setting different torsion and azimuth rotation parameters, the above-mentioned cross-section torsion and azimuth rotation process is repeated to obtain a large number of three-dimensional actual seismic sample feature enhancement data sets and fault sample label enhancement data sets with different bending morphologies and azimuth rotation angles, thereby increasing the diversity of actual fault interpretation sample data.

[0106] Step 3, Automatic 3D Seismic Fault Identification. A network structure for automatic 3D seismic fault identification is constructed, suitable for both enhanced feature data and enhanced fault label data from actual 3D seismic samples.

[0107] like Figure 5As shown, the encoding process includes four feature extraction steps, extracting seismic reflection features from small to large scales; the decoding process includes four feature recovery steps, recovering fault identification results at different scales. The output of each decoding layer is concatenated with the output of the encoding layer to fuse features from encoders at different scales and upper decoding layers, capturing both fine-grained and coarse-grained semantics across the entire scale. Furthermore, to better analyze the spatial correlation between different fault points on the same cross section, attention layers are introduced in each encoding and decoding layer. Specifically, each encoding layer consists of a 3×3×3 convolutional layer, an attention layer, a ReLU activation function, and a 2×2×2 max-pooling layer; each decoding layer consists of a 3×3×3 transposed convolutional layer, an attention layer, and a ReLU activation function. This block-based self-attention mechanism not only improves computational efficiency for large-scale sample data but also enhances the network's ability to analyze spatial features.

[0108] Step 4: Design of loss function for automatic 3D seismic fault identification. The designed loss function is applicable to partially annotated 3D actual seismic sample feature enhancement data and fault sample label enhancement data.

[0109] Step 5: Training and validation of the 3D seismic fault automatic identification network.

[0110] The 3D actual earthquake sample feature enhancement data and fault sample label enhancement data are split into training sample sets and validation sample sets. 80% of the incomplete fault interpretation sample data is used as the training set, and the other 20% is used as the validation set.

[0111] Adam was used for parameter optimization, with a learning rate of 0.0001 and 100 iterations. The tomography identification model was obtained through iterative training. Driven by the training sample set data, the automatic tomography identification network was used for forward propagation, and the loss function between the predicted results and the tomography labels was calculated. The weight parameters of the automatic tomography identification network were iteratively optimized, and the result that met the validation set requirements was used as the final automatic tomography identification network model.

[0112] Step 6: Practical Application Test of the 3D Seismic Fault Automatic Identification Network Method. Using the final trained automatic fault identification network model, a 3D fault automatic identification application test is conducted. The seismic profile to be predicted is as follows: Figure 6 As shown, the black dashed lines represent partially annotated fault interpretation sample labels. The true fault labels are as follows: Figure 7 As shown, the tomography results using the conventional dice loss function (Dice) are compared. Figure 8 As shown), the tomographic identification result of the loss function proposed in this patent ( Figure 9 (As shown) has higher reliability.

[0113] Example 3:

[0114] In a specific embodiment 3 of the present invention, the automatic three-dimensional seismic fault identification method of the present invention includes the following steps:

[0115] Step 1: Preprocessing of actual seismic profile fault interpretation sample data. Fault interpretation results from 100 seismic profiles in adjacent work areas of the study region were collected. The actual seismic profile data and their corresponding fault interpretation results were resampled or interpolated along the vertical time and horizontal seismic trace directions. The data time sampling interval and trace spacing were standardized to obtain preprocessed seismic profile data and their corresponding fault label data.

[0116] Step 2: Enhancement of sample data for actual data fault interpretation.

[0117] The preprocessed seismic profile data and its corresponding fault label data are augmented in dimension; the augmented sample data are randomly distorted, and multiple sinusoidal signals are superimposed to simulate the curved cross-sectional shape; the distorted sample data are then rotated by azimuth.

[0118] By randomly setting different torsion and azimuth rotation parameters, the above-mentioned cross-section torsion and azimuth rotation process is repeated to obtain a large number of three-dimensional actual seismic sample feature enhancement data sets and fault sample label enhancement data sets with different bending morphologies and azimuth rotation angles, thereby increasing the diversity of actual fault interpretation sample data.

[0119] Meanwhile, based on forward modeling, 400 pairs of 256×256×256 three-dimensional fault sample data and their seismic forward modeling sample feature data were constructed, further expanding the fault identification sample data.

[0120] Step 3, Automatic 3D Seismic Fault Identification. A network structure for automatic 3D seismic fault identification is constructed, suitable for both enhanced feature data and enhanced fault label data from actual 3D seismic samples.

[0121] The encoding process comprises four feature extraction steps, extracting seismic reflection features from small to large scales. The decoding process comprises four feature recovery steps, recovering fault identification results at different scales. The output of each decoding layer is concatenated with the output of the encoding layer to fuse features from different scale encoders and upper decoding layers, capturing both fine-grained and coarse-grained semantics across the entire scale. Furthermore, to better analyze the spatial correlation between different fault points on the same cross section, attention layers are introduced in each encoding and decoding layer. Specifically, each encoding layer consists of a 3×3×3 convolutional layer, an attention layer, a ReLU activation function, and a 2×2×2 max-pooling layer; each decoding layer consists of a 3×3×3 transposed convolutional layer, an attention layer, and a ReLU activation function. This block-based self-attention mechanism not only improves computational efficiency for large-scale sample data but also enhances the network's ability to analyze spatial features.

[0122] Step 4: Design of loss function for automatic 3D seismic fault identification. The designed loss function is applicable to partially annotated 3D actual seismic sample feature enhancement data and fault sample label enhancement data.

[0123] Step 5: Training and validation of the 3D seismic fault automatic identification network.

[0124] The 3D actual earthquake sample feature enhancement data and fault sample label enhancement data are split into training sample sets and validation sample sets. 80% of the incomplete fault interpretation sample data is used as the training set, and the other 20% is used as the validation set.

[0125] Adam was used for parameter optimization, with a learning rate of 0.0001 and 500 iterations. The tomography identification model was obtained through iterative training. Driven by the training sample set data, the automatic tomography identification network was used for forward propagation, and the loss function between the predicted results and the tomography labels was calculated. The weight parameters of the automatic tomography identification network were iteratively optimized, and the result that met the validation set requirements was used as the final automatic tomography identification network model.

[0126] (6) Practical Application Test of the 3D Seismic Fault Automatic Identification Network Method. Using the final trained automatic fault identification network model, a 3D fault automatic identification application test was conducted in an actual work area. For example... Figure 10 As shown, the fault identification results of this patent are clear and continuous, which can provide effective assistance for rapidly characterizing fracture systems.

[0127] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0128] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.

Claims

1. A method for automatic identification of three-dimensional seismic faults, characterized in that, The automatic three-dimensional seismic fault identification method includes: Step 1: Preprocess the actual data for fault interpretation sample data; Step 2 involves enhancing the actual fault interpretation sample data, including: expanding the dimensions of the preprocessed seismic profile data and its corresponding fault label data, and randomly twisting and rotating them to form a large amount of three-dimensional actual seismic sample feature enhancement data and fault sample label enhancement data. Expanding the dimension of two-dimensional sample data to three-dimensional sample data, and adding seismic data and tomographic sample labels The 3D seismic data is obtained by copying and expanding along the direction perpendicular to the profile. and fault sample label data Its level Direction, Horizontal Direction and vertical time The size of the direction is ; Step 3, automatic identification of 3D seismic faults, including: the forward propagation process of the 3D seismic fault automatic identification network includes two steps: encoding and decoding. Encoding includes... The secondary feature extraction process extracts seismic reflection features from small to large scales; decoding includes... The secondary feature recovery process recovers the tomographic recognition results of features at different scales. The output of each decoding layer is concatenated with the output of the encoding layer to fuse features from encoders at different scales and upper decoding layers, thereby capturing fine-grained and coarse-grained semantics across the entire scale. Simultaneously, to better analyze the spatial correlation between different breakpoints of the same cross section, attention layers are introduced in each encoding and decoding layer. That is, each encoding layer is a concatenation layer, attention layer, activation function, and max pooling layer, and each decoding layer is a concatenation layer, attention layer, and activation function, thereby enhancing the network's ability to analyze spatial features. Step 4: Design a loss function for automatic 3D seismic fault identification. This includes: In interpreted regions, fault labels are clearly defined, while in uninterpreted regions, both non-fault seismic reflection features and uninterpreted fault reflection features are included. Therefore, the calculation of the training loss function should focus on interpreted regions, emphasizing the correlation analysis between fault labels and seismic reflection features in these regions. The quantitative expression focusing on interpreted fault regions is as follows: ; in, The kernel is a Gaussian convolution. , For the convolution process, To focus on the regional data volume, the size and tomographic sample labels enhance the data volume. The same applies; based on this, a mask weight matrix is ​​constructed. Its level Direction, Horizontal Direction and vertical time The size of the direction is The specific expression is: ； The improved automatic fault identification loss function is: ； in, Let be the network weight parameters to be optimized. and These represent the mask weight matrix. Enhanced data volume with tomographic sample labels And prediction results of 3D seismic fault automatic identification network The matrix elements are multiplied accordingly; the first term of the loss function focuses on the adjacent area of ​​the explained fault location, where the fault sample labels are determined and the number of fault and non-fault samples is more balanced, which helps to learn the correlation between the seismic discontinuous reflection features and the fault labels. The second term focuses on the area of ​​the unexplained fault, where there are a large number of true non-fault labels and a small number of unlabeled erroneous non-fault labels, emphasizing the learning of the correlation between the seismic continuous reflection features and the non-fault labels. Step 5: Train and validate the 3D seismic fault automatic identification network; Step 6: Conduct practical application tests on the three-dimensional seismic fault automatic identification network method.

2. The three-dimensional seismic fault auto-identification method according to claim 1, characterized in that, In step 1, the actual seismic profile data and its corresponding fault interpretation results are resampled or interpolated along the vertical time and horizontal seismic trace directions, respectively. The data time sampling interval and trace spacing are unified to obtain the preprocessed seismic profile data and its corresponding fault label data.

3. The automatic three-dimensional seismic fault identification method according to claim 2, characterized in that, In step 1, for any two-dimensional seismic profile Its level Direction and vertical time The dimension of the direction is In this process, segmented interpolation is performed on the fault control points actually interpreted by experts to obtain continuous fault interpretation data. The sampling point location closest to the fault line is encoded as 1, indicating an interpreted fault point; other sampling points are encoded as 0, indicating uninterpreted areas, including both non-fault points and uninterpreted fault points. After obtaining the encoding of each seismic sampling point location, a two-dimensional fault sample label corresponding to the size of the two-dimensional seismic profile is obtained. .

4. The method for automatic identification of three-dimensional seismic faults according to claim 1, characterized in that, In step 2, three-dimensional cross-sectional distortion is performed. To further enrich the pattern of the fault samples, along the vertical cross-sectional direction, it is assumed that... Direction, for 3D seismic data with expanded dimensions and fault sample label data Arbitrary distortion is applied to obtain seismic data with distorted cross-sections. and fault label data .

5. The automatic three-dimensional seismic fault identification method according to claim 4, characterized in that, In step 2, a three-dimensional cross-sectional azimuth rotation is performed. The distorted three-dimensional seismic data and cross-sectional sample label data are rotated along different azimuth angles. Then, interpolation is performed according to the new spatial coordinates to obtain the rotated three-dimensional actual seismic sample feature enhancement data volume. and tomographic sample label augmentation data volume .

6. The method for automatic identification of three-dimensional seismic faults according to claim 5, characterized in that, In step 2, a set of enhanced data volumes for 3D actual earthquake sample features and a set of enhanced data volumes for fault sample labels are generated. Different torsion and azimuth rotation parameters are randomly set, and the above-mentioned fault torsion and azimuth rotation process is repeated to obtain a large number of sets of enhanced data volumes for 3D actual earthquake sample features and sets of enhanced data volumes for fault sample labels with different bending morphologies and azimuth rotation angles, thereby increasing the diversity of actual fault interpretation sample data.

7. The method for automatic identification of three-dimensional seismic faults according to claim 1, characterized in that, In step 3, an automatic three-dimensional seismic fault identification network structure is constructed, which is suitable for feature enhancement data of three-dimensional actual seismic samples and fault sample label enhancement data.

8. The method for automatic identification of three-dimensional seismic faults according to claim 7, characterized in that, In step 3, the input to the 3D seismic fault automatic identification network is the 3D actual seismic sample feature enhancement data volume, and the output is the 3D actual fault sample label enhancement data volume.

9. The method for automatic identification of three-dimensional seismic faults according to claim 1, characterized in that, In step 4, a loss function is designed that is applicable to the feature enhancement data of partially annotated 3D actual seismic samples and the fault sample label enhancement data.

10. The method for automatic identification of three-dimensional seismic faults according to claim 9, characterized in that, In step 4, the loss function between the prediction results of the 3D seismic fault automatic identification network model and the known fault labels is expressed as: ； in, Augment the data volume for tomographic sample labels. The prediction results are from the 3D seismic fault automatic identification network. This represents the forward propagation process of the 3D seismic fault automatic identification network. Let be the network weight parameters to be optimized. Enhanced data volume of rotated 3D actual seismic sample features .

11. The method for automatic identification of three-dimensional seismic faults according to claim 1, characterized in that, In step 5, the 3D actual earthquake sample feature enhancement data and fault sample label enhancement data are split into training sample sets and validation sample sets. Key parameters for network training are set. Driven by the training sample set data, the fault automatic identification network is forward propagated, and the loss function between the prediction results and fault labels is calculated. The weight parameters of the fault automatic identification network are iteratively optimized and solved repeatedly. The result that meets the requirements of the validation set is used as the final fault automatic identification network model.

12. The method for automatic identification of three-dimensional seismic faults according to claim 11, characterized in that, In step 5, preprocessing is performed to ensure consistency in the size of the 3D actual sample data. The size of the 3D actual seismic sample feature enhancement data and fault sample label enhancement data is statistically analyzed to obtain the horizontal data in the statistical results. Direction, Horizontal Direction and vertical time The minimum dimension of the direction is Based on this, other data in the 3D actual seismic sample feature enhancement data volume set and the fault sample label enhancement data volume set are truncated using a sliding time window method, so that the size of all sample data is unified to the minimum size. .

13. The method for automatic identification of three-dimensional seismic faults according to claim 12, characterized in that, In step 5, three-dimensional fault forward modeling sample data is generated, according to the minimum size. Based on the actual geological conditions of the work area, horizontal layered reflection coefficient data volumes are randomly generated, and parameters such as random structural undulations, fault dip angles, and fault displacements are set to obtain reflection coefficient data volumes after tectonic movement. Then, seismic wavelets are set according to the actual seismic data of the work area, and seismic forward modeling is performed to obtain fault sample labels and their corresponding seismic forward modeling sample feature data based on forward modeling, thereby expanding the number of samples used for the automatic fault identification network model.

14. The method for automatic identification of three-dimensional seismic faults according to claim 13, characterized in that, In step 5, the training and validation sets of the three-dimensional seismic fault automatic identification sample data are split. The sample data is split into training and validation sets by random selection.

15. The method for automatic identification of three-dimensional seismic faults according to claim 14, characterized in that, In step 5, the weight parameters of the 3D seismic fault automatic identification network are optimized. Key parameters such as the number of iterations and the learning rate are set. Driven by the training sample set data, the fault automatic identification network is forward propagated, and the loss function between the prediction results and the fault labels is calculated. The weight parameters of the fault automatic identification network are iteratively optimized and solved repeatedly.

16. The method for automatic identification of three-dimensional seismic faults according to claim 15, characterized in that, In step 5, the 3D seismic fault automatic identification network model is output. The trained fault automatic network model is then used to test the validation set. If the prediction results meet the requirements, the final 3D seismic fault automatic identification network model is output. Otherwise, the network parameters are readjusted until the validation set test requirements are met.

17. The method for automatic identification of three-dimensional seismic faults according to claim 1, characterized in that, In step 6, the final automatic fault identification network model obtained from the training is used to conduct an application test of three-dimensional automatic fault identification in the actual work area.

Images