A method and apparatus for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas.

Abstract

This application provides a method and apparatus for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas. The method includes: preprocessing the original post-stack depth-domain 3D seismic volumes of the target area sequentially to obtain preprocessed seismic volumes; inputting the preprocessed seismic volumes into a 3D convolutional neural network model to obtain an initial fault probability volume; overlaying the initial fault probability volume with the original seismic profile to form an initial overlay image, constructing a learning sample set and a validation sample set, and determining the optimal number of iteration weights; constructing a final optimized neural network model based on the optimal number of iteration weights to recalculate the fault probability volume; performing parameterized post-processing on the recalculated fault probability volume to obtain a final fault probability volume; and using the final fault probability volume and the final overlay image as the final fracture identification result. This approach achieves the technical effect of improving the identification accuracy of complex fractures, weak fractures, and small-scale fractures.

Description

Technical Field

[0001] This application belongs to the field of oil and gas exploration technology, and in particular relates to a method and device for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas. Background Technology

[0002] Some oil and gas reservoirs belong to important zones for exploration and development of ultra-deep carbonate oil and gas. These reservoirs have complex fault systems, especially strike-slip faults and their derivatives. Faults play a crucial role in reservoir stimulation, hydrocarbon migration and enrichment, trap formation, and well placement. However, due to problems such as low signal-to-noise ratio, weak fault response, complex fault combinations, and difficulty in identifying weak and small-scale faults in the back-stack depth-domain seismic data of these zones, detailed fault interpretation is challenging. Currently, for general oil and gas reservoirs, coherence attributes, tectonic entropy attributes, structural tensor attributes, maximum curvature attributes, maximum likelihood attributes, and likelihood ant body attributes are commonly used for fault identification.

[0003] However, conventional attribute identification results for ultra-deep carbonate buried mountain areas often have the following problems: 1) limited ability to identify weak and small-scale fractures; 2) unclear characterization of complex cross fractures and strike-slip fracture zone combinations; 3) poor continuity of fracture distribution; 4) insufficient stability of single attribute interpretation results and high interpretation uncertainty; 5) difficulty in meeting the engineering requirements for fine fracture characterization in well location deployment and trap evaluation.

[0004] To improve the accuracy of fault identification, artificial intelligence methods are used for seismic interpretation to learn fault characteristics from seismic data and output fault probability volumes. However, for fault identification of low signal-to-noise ratio post-stack depth-domain seismic volumes in ultra-deep carbonate buried mountain areas, existing general-purpose intelligent identification methods still lack a learning sample construction method that matches the local fault pattern, a transfer learning control method that determines the optimal weight based on the peak value of the validation sample, and a parameterized probability volume post-processing workflow that matches the fault scale of the target work area. This leads to technical problems such as insufficient characterization of the continuity of complex, weak, and small-scale faults.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] The purpose of this application is to provide a method and device for identifying fractures in ultra-deep carbonate reservoirs in buried mountainous areas, which can improve the identification accuracy of complex fractures, weak fractures and small-scale fractures under low signal-to-noise ratio data.

[0007] This application provides a method and apparatus for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas, which is implemented as follows: A method for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas, the method comprising: Obtain the original post-stack depth domain 3D seismic volume of the target work area; The original post-stack depth domain 3D seismic volume is sequentially subjected to noise reduction and resolution enhancement processing to obtain a preprocessed seismic volume. The preprocessed seismic volume is input into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; The initial fault probability volume is overlaid with the original seismic profile to form an initial overlay map. Fault interpretation lines that conform to the seismic geological characteristics are selected as annotation samples on the initial overlay map, and a learning sample set and a validation sample set are constructed according to the profile blocks. Using the three-dimensional convolutional neural network model as the benchmark model, transfer learning is performed based on the learning sample set, and iterative training is performed in combination with the validation sample set. The evaluation curves of the learning sample set and the validation sample set are recorded. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, the number of iterations corresponding to the peak of the validation sample set evaluation curve is determined as the optimal number of iteration weights. The final optimized neural network model is constructed based on the optimal number of iteration weights, and the fault probability volume is recalculated using the final optimized neural network model. The recalculated fault probability volume is parameterized and post-processed to obtain the final fault probability volume. Based on the final fault probability volume, a probability volume profile is formed. The original seismic profile is overlaid with the probabilistic volume profile to form a final overlay image. The final fault probabilistic volume and the final overlay image are used as the final fault identification result.

[0008] In one implementation, fault interpretation lines conforming to seismic geological characteristics are selected as annotation samples on the initial overlay map, and a learning sample set and a validation sample set are constructed according to profile blocks, including: On the initial superposition diagram, an interpretation line that meets at least one of the following conditions is selected as the fault interpretation line: consistent with the fault relationship of the reflection layer, having a clear linear extension characteristic, matching the local structural pattern, and maintaining a reasonable consistency with the distribution direction of the adjacent known faults. The selected fault interpretation lines are converted into label data that can be used for model training. The label data is a narrow band label formed by extending the fault interpretation lines on both sides with a preset width. The selected fault interpretation lines are rasterized using the label data to form a label matrix corresponding to the profile block. The profile block is a sample block of fixed size or uniform specification formed by extracting a local window from the original seismic profile, the corresponding probability response profile, and, if necessary, the adjacent trace information, with the fault interpretation line as the center. Based on the label matrix, a learning sample set is formed by combining multiple labeled profile blocks, and a validation sample set is constructed from different profile blocks that have never participated in training.

[0009] In one implementation, the three-dimensional convolutional neural network model is trained according to the following loss function:

[0010] in, The weighting coefficient for the first loss term. The weighting coefficient for the second loss term. The weighting coefficients for the continuity constraint loss term. As the first loss item, This is the second loss item. The loss term represents the continuity constraint, where N represents the total number of sample points. This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. Indicates the foreground sample weights. Represents the background sample weights, where, > , This is a smoothing term used to avoid the denominator being zero. This represents the fault neighborhood mask value corresponding to the i-th sample point. This represents the predicted probability value of the i-th sample point among its adjacent sample points in the tilt direction. This represents the predicted probability value of the i-th sample point in the direction of its adjacent sample points. The continuity weights represent the direction of tendency. The continuity weight represents the direction of movement.

[0011] In one embodiment, the original post-stack depth domain 3D seismic volume is sequentially subjected to noise reduction and resolution enhancement processing to obtain a preprocessed seismic volume, including: The original post-stack depth-domain 3D seismic volume in the spatial domain is converted into a wavenumber-domain spectrum using a 3D Fourier transform:

[0012] in, This is the original post-stack depth domain 3D seismic volume. For seismic data plane coordinates, The coordinates are in the depth direction. The complex amplitude spectrum of the earthquake body in the wavenumber domain. It is a three-dimensional discrete Fourier transform operator. The wave number in the x-direction. The wave number in the y-direction. The wave number in the z-direction; Constructing a fracture-guided frequency domain mask using a sigmoid function:

[0013] in, As a fracture-guided frequency domain mask, For three-dimensional total wavenumber mode, This is the mask steepness coefficient. The center wavenumber of the fracture-sensitive frequency band, It is a natural constant; The denoised seismic body is obtained by performing inverse Fourier transform on the fracture-guided frequency domain mask and wavenumber domain spectrum.

[0014] in, The noise-reduced seismic body after adaptive frequency domain noise reduction. This is a three-dimensional inverse Fourier transform, converting the wavenumber domain result back to the spatial domain; An adaptive deconvolution operator is constructed using the working zone layer velocity and the formation dip angle as constraints. The adaptive deconvolution operator is convolved with the denoised seismic volume in 3D to obtain a high-resolution preprocessed seismic volume:

[0015] in, For high-resolution preprocessed seismic volumes, For adaptive deconvolution operators, This is a three-dimensional convolution operation.

[0016] In one implementation, forming a probability volume profile based on the final fault probability volume includes: The original seismic profile is extracted from the longitudinal or transverse profile of the original post-stack depth domain three-dimensional seismic body under preset coordinates. At a location, orientation, and sampling interval that are completely consistent with the original seismic profile, the corresponding profile is extracted from the final fault probability volume and used as the probability volume profile. Accordingly, the original seismic profile and the probabilistic volume profile are overlaid and displayed to form a final overlay image, including: Using unified spatial coordinates, sampling points, and display range, the original seismic profile is used as the base map, and the probabilistic volume profile is overlaid on the base map with color and transparency to form the final overlay map.

[0017] In one embodiment, the parameterized post-processing includes: probability threshold filtering, seed point spacing control, yaw direction window constraint, traverse direction window constraint, normal direction window constraint, yaw direction smoothing, traverse direction smoothing, yaw direction deformation constraint, and traverse direction deformation constraint.

[0018] A fracture identification device for ultra-deep carbonate reservoirs in buried mountainous areas includes: The acquisition module is used to acquire the original post-stack depth domain 3D seismic volume of the target work area; The preprocessing module is used to sequentially perform noise reduction and resolution improvement processing on the original post-stack depth domain 3D seismic volume to obtain a preprocessed seismic volume; The input module is used to input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; The overlay module is used to overlay and display the initial fault probability volume with the original seismic profile to form an initial overlay map. On the initial overlay map, fault interpretation lines that conform to the seismic geological characteristics are selected as annotation samples, and a learning sample set and a validation sample set are constructed according to the profile blocks. The learning module is used to perform transfer learning based on the learning sample set, using the three-dimensional convolutional neural network model as the benchmark model, and to perform iterative training in combination with the validation sample set. It records the evaluation curve of the learning sample set and the evaluation curve of the validation sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, the number of iterations corresponding to the peak of the evaluation curve of the validation sample set is determined as the optimal number of iteration weights. The calculation module is used to construct the final optimized neural network model based on the optimal number of iteration weights, and to recalculate the fault probability volume using the final optimized neural network model. The parameterization module is used to perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and to form a probability volume profile based on the final fault probability volume. The identification module is used to overlay and display the original seismic profile with the probability volume profile to form a final overlay image, and to use the final fault probability volume and the final overlay image as the final fault identification result.

[0019] An electronic device includes a processor and a memory for storing processor-executable instructions, wherein the processor, when executing the instructions, implements the steps of the method described above.

[0020] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0021] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0022] The method and apparatus for fault identification in ultra-deep carbonate reservoirs in buried mountainous areas provided in this application first preprocesses the original post-stack depth domain 3D seismic volume to improve data quality, and then inputs it into a 3D convolutional neural network to obtain an initial fault probability volume. Based on the initial overlay of the probability volume and the seismic profile, fault interpretation lines that conform to geological characteristics are selected to construct a learning and validation sample set. Transfer learning is carried out using the 3D convolutional neural network as the benchmark model, and the optimal iteration weights are determined based on the peak value of the validation set evaluation curve to obtain an optimized model and recalculate the fault probability volume. After parameterization post-processing, the final fault probability volume is obtained, the probability volume profile is extracted and accurately overlaid with the original seismic profile for display, and the final fault identification result is output. The above scheme can solve the technical problems of low fault identification accuracy, difficulty in characterizing weak and small-scale faults, poor continuity of fault distribution, and easy overfitting of intelligent models in existing ultra-deep low signal-to-noise ratio seismic data. It achieves the technical effect of improving the identification accuracy and continuity of complex, weak, and small-scale faults, and reducing the risk of model overfitting and interpretation uncertainty. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of one embodiment of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas provided in this application; Figure 2 This is a flowchart of the method for line noise reduction and resolution improvement provided in this application; Figure 3 This is a flowchart of one embodiment of the migration learning fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas provided in this application; Figure 4 This is a comparative diagram of the original seismic body and the noise-reduced seismic body provided in this application; Figure 5 This is a schematic diagram comparing the original seismic volume and the seismic volume after noise reduction and resolution improvement provided in this application; Figure 6 This is a schematic diagram of the initial tomographic probability volume generated by the convolutional neural network provided in this application; Figure 7 This is a schematic diagram of the optimized result of the initial fault probability volume provided in this application; Figure 8 This is a schematic diagram of the fault probability volume after transfer learning provided in this application; Figure 9 This is a schematic diagram of the probability volume of a tomography in a conventional convolutional neural network provided in this application; Figure 10 This is a schematic diagram of the superposition effect of the earthquake body provided in this application and the superposition effect of the fault probability body and the earthquake body after transfer learning; Figure 11 This is a schematic diagram of the fracture results identified after transfer learning provided in this application; Figure 12 This is a hardware structure block diagram of an electronic device for a fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas, as provided in this application. Figure 13 This is a schematic diagram of the module structure of one embodiment of the fracture identification device for ultra-deep carbonate rock reservoirs in buried mountainous areas provided in this application. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0026] It should also be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0027] To address the technical challenges of insufficient identification of weak and small-scale faults, unclear characterization of complex faults, poor continuity of fault distribution, weak generalization ability of intelligent models, high risk of overfitting, and difficulty in directly supporting engineering applications in ultra-deep carbonate rock low signal-to-noise ratio post-stack depth-domain seismic data, this study considers improving data quality through seismic preprocessing, extracting initial fault features using a 3D convolutional neural network, constructing a localized dataset using profile block samples that fit the geological characteristics of the work area, selecting optimal model weights through transfer learning and validation set peak determination, and completing probabilistic volume optimization and morphological constraints through parameterized post-processing. This approach improves the identification accuracy and continuity of distribution of complex strike-slip faults, weak faults, and small-scale faults under low signal-to-noise ratio data, reduces interpretation uncertainty and model overfitting risk, and generates stable, reproducible, and geologically consistent fault identification results that can directly support structural interpretation, trap evaluation, and well location deployment.

[0028] Figure 1 This is a flowchart illustrating one embodiment of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas provided in this application. While this application provides method operation steps or apparatus structures as shown in the following embodiments or figures, more or fewer operation steps or module units may be included in the method or apparatus based on conventional or non-inventive effort. In steps or structures where there is no logically necessary causal relationship, the execution order of these steps or the module structure of the apparatus is not limited to the execution order or module structure described in the embodiments and figures of this application. When the method or module structure is applied in actual devices or end products, it can be executed sequentially or in parallel according to the method or module structure shown in the embodiments or figures (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed processing environment).

[0029] Specifically, such as Figure 1 As shown, the above-mentioned fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas may include the following steps: Step 101: Obtain the original post-stack depth domain 3D seismic volume of the target work area; Step 102: Perform noise reduction and resolution enhancement processing on the original post-stack depth domain 3D seismic volume in sequence to obtain a preprocessed seismic volume; Step 103: Input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; Step 104: Overlay the initial fault probability volume with the original seismic profile to form an initial overlay map. Select fault interpretation lines that conform to the seismic geological characteristics as annotation samples on the initial overlay map, and construct a learning sample set and a validation sample set according to the profile blocks. Step 105: Using the three-dimensional convolutional neural network model as the baseline model, perform transfer learning based on the learning sample set, and perform iterative training in combination with the verification sample set. Record the evaluation curve of the learning sample set and the evaluation curve of the verification sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the verification sample set rises first and then falls, determine the number of iterations corresponding to the peak of the evaluation curve of the verification sample set as the optimal number of iteration weights. Step 106: Construct the final optimized neural network model based on the optimal number of iteration weights, and recalculate the fault probability volume using the final optimized neural network model; Step 107: Perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and form a probability volume profile based on the final fault probability volume; The parameterized post-processing can include: probability threshold screening, seed point spacing control, dip direction window constraint, strike direction window constraint, normal direction window constraint, dip direction smoothing, strike direction smoothing, dip direction deformation constraint, and strike direction deformation constraint. The parameterized post-processing uses a set of fixed, quantifiable, and reproducible engineering parameters to perform threshold screening, seed point control, spatial window constraint, smoothing, and deformation constraint on the fault probability volume output by the model, so that the fault identification results are more continuous, cleaner, and more in line with geological laws, and finally form an engineering result that can be directly used for structural interpretation, trap evaluation, and well location deployment.

[0030] Specific parameterized post-processing may include: filtering the fault probability volume using probability thresholds; extracting initial seed points according to the set seed point spacing; setting neighborhood windows in the dip direction, strike direction, and normal direction respectively; performing smoothing processing in the dip direction and strike direction respectively; applying deformation constraints in the dip direction and strike direction respectively; extracting the continuity-enhanced fracture volume based on the above processing and outputting the final fracture identification result.

[0031] Step 108: Overlay the original seismic profile with the probability volume profile to form a final overlay image, and use the final fault probability volume and the final overlay image as the final fault identification result.

[0032] In one embodiment, step 105 above, selecting fault interpretation lines that conform to seismic geological characteristics as annotation samples on the initial overlay map and constructing a learning sample set and a validation sample set according to profile blocks, may include: selecting interpretation lines on the initial overlay map that meet at least one of the following conditions as fault interpretation lines: consistent with the faulting relationship of the reflection layer, having a clear linear extension characteristic, matching the local structural pattern, and maintaining a reasonable consistency with the direction of the distribution of nearby known faults; converting the selected fault interpretation lines into label data that can be used for model training, wherein the label data is a narrow band label formed by expanding the fault interpretation lines on both sides of a preset width; performing rasterization processing on the selected fault interpretation lines using the label data to form a label matrix corresponding to the profile blocks, wherein the profile block is a sample block of fixed size or uniform specification formed by truncating a local window from the original seismic profile, the corresponding probability response profile, and, if necessary, the adjacent trace information, with the location of the fault interpretation line as the center; and constructing a validation sample set in different profile blocks that have never participated in the training, based on the label matrix.

[0033] Specifically, by strictly selecting fault interpretation lines that conform to the geological characteristics of the earthquake on the initial overlay map, the geological rationality and label accuracy of the labeled samples can be guaranteed, avoiding interference from invalid or erroneous samples to the model training. Converting the fault interpretation lines into narrow-band labels with preset widths on both sides and performing rasterization can form a label matrix that matches the size of the seismic profile block, making the label data more consistent with the spatial distribution of the faults and adapting to the input format requirements of the three-dimensional convolutional neural network. Extracting profile block samples of uniform specifications centered on the fault location can highlight local fault features and reduce global data redundancy, improving model training efficiency. Strictly partitioning the sample set according to training and validation profile blocks can ensure the independence of model training and generalization ability evaluation, providing a high-quality and reliable data foundation for subsequent transfer learning iteration optimization and optimal weight determination, thereby significantly improving the model's ability to learn and identify fault features in the target work area.

[0034] To improve the continuity representation of weak faults, small-scale faults, and complex strike-slip fault combinations in low signal-to-noise ratio post-stack depth-domain seismic data, a fault continuity constraint loss term is introduced into the loss function. This term suppresses isolated noise responses and unreasonable interruptions in fault prediction results and enhances the continuous distribution of faults along the dip and strike directions. Based on this, the three-dimensional convolutional neural network model can be trained according to the following loss function:

[0035] in, The weighting coefficient for the first loss term. The weighting coefficient for the second loss term. The weighting coefficients for the continuity constraint loss term. As the first loss item, This is the second loss item. The loss term represents the continuity constraint, where N represents the total number of sample points. This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. Indicates the foreground sample weights. Represents the background sample weights, where, > , This is a smoothing term used to avoid the denominator being zero. This represents the fault neighborhood mask value corresponding to the i-th sample point. This represents the predicted probability value of the i-th sample point among its adjacent sample points in the tilt direction. This represents the predicted probability value of the i-th sample point in the direction of its adjacent sample points. The continuity weights represent the direction of tendency. The continuity weight represents the direction of movement.

[0036] In this example, the original post-stack depth-domain 3D seismic volume is sequentially subjected to denoising and resolution enhancement to obtain a preprocessed seismic volume. This preprocessed seismic volume is obtained by first converting the spatial domain seismic volume into a wavenumber domain spectrum using a 3D Fourier transform; then, a fault-guided frequency domain mask is constructed using a S-shaped function to accurately preserve fault characteristic wavenumbers while suppressing noise wavenumbers. The denoised seismic volume is then obtained through an inverse Fourier transform. Finally, an adaptive deconvolution operator is constructed using the layer velocity and dip angle of the work area as constraints, and this operator is convolved in 3D with the denoised seismic volume after adaptive frequency domain denoising to obtain a high-resolution preprocessed seismic volume. This approach enhances the response to fault boundaries and discontinuities while avoiding noise amplification.

[0037] Based on this, the original post-stack depth domain 3D seismic volume is sequentially subjected to noise reduction and resolution enhancement processing to obtain a preprocessed seismic volume, which can be as follows: Figure 2 As shown, it includes the following steps: S201: Converting the original post-stack depth-domain 3D seismic volume in the spatial domain into a wavenumber-domain spectrum using a 3D Fourier transform:

[0038] in, This is the original post-stack depth domain 3D seismic volume. For seismic data plane coordinates, The coordinates are in the depth direction. The complex amplitude spectrum of the earthquake body in the wavenumber domain. It is a three-dimensional discrete Fourier transform operator. The wave number in the x-direction. The wave number in the y-direction. The wave number in the z-direction; S202: Constructing a fracture-guided frequency domain mask using a sigmoid function:

[0039] in, As a fracture-guided frequency domain mask, For three-dimensional total wavenumber mode, This is the mask steepness coefficient. The center wavenumber of the fracture-sensitive frequency band, It is a natural constant; S203: Obtain the denoised seismic body by performing inverse Fourier transform on the fracture-guided frequency domain mask and wavenumber domain spectrum:

[0040] in, The noise-reduced seismic body after adaptive frequency domain noise reduction. This is a three-dimensional inverse Fourier transform, converting the wavenumber domain result back to the spatial domain; S204: Construct an adaptive deconvolution operator with the working zone layer velocity and the formation dip angle as constraints; S205: Perform a 3D convolution between the adaptive deconvolution operator and the denoised seismic volume to obtain a high-resolution preprocessed seismic volume.

[0041] in, For high-resolution preprocessed seismic volumes, For adaptive deconvolution operators, This is a three-dimensional convolution operation.

[0042] In practical implementation, forming a probability volume profile based on the final fault probability volume can include: extracting the original seismic profile from the longitudinal or transverse survey profile of the original post-stack depth domain 3D seismic body under preset coordinates; extracting the corresponding profile from the final fault probability volume at a position, orientation, and sampling interval completely consistent with the original seismic profile, as the probability volume profile; correspondingly, overlaying the original seismic profile and the probability volume profile to form a final overlay image can include: using unified spatial coordinates, sampling points, and display range, using the original seismic profile as the base map, and overlaying the probability volume profile on the base map with color and transparency to form the final overlay image.

[0043] In other words, by extracting the original seismic profile and probabilistic volume profile under unified spatial coordinates, the same profile location, consistent profile direction, and sampling interval, it is possible to ensure precise spatial alignment between the two, avoid misalignment and distortion, and guarantee a strict correspondence between fault identification results and seismic geological features. By using unified sampling points and display ranges, the probabilistic volume profile is overlaid on the original seismic profile using color mapping and semi-transparent overlay. This allows for a clear and intuitive presentation of the matching degree between the predicted fault location and the seismic reflection faulting relationship, significantly improving the visualization and verification efficiency of fault results, enhancing the reliability and persuasiveness of interpretation results, and providing intuitive, accurate, and rapidly verifiable visual support conditions for detailed structural interpretation, trap evaluation, and well location deployment.

[0044] The above method will be described below with reference to a specific embodiment. However, it should be noted that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.

[0045] To address the issue of unsatisfactory conventional attribute fracture identification performance in existing low signal-to-noise ratio post-stack depth-domain seismic data for buried mountain reservoirs, this paper presents a migration learning fracture identification method for ultra-deep carbonate reservoirs in buried mountain areas. This method aims to improve the identification capability of complex fractures, weak fractures, and small-scale fractures in low signal-to-noise ratio post-stack depth-domain seismic data, and to enhance the continuity of fracture distribution and the stability of interpretation.

[0046] Specifically, a migration learning-based fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas is provided, which can, for example... Figure 3 As shown, it includes the following steps: Step 301: Obtain the original post-stack depth domain 3D seismic volume of the target work area; Step 302: Perform noise reduction and resolution enhancement processing on the original post-stack depth domain 3D seismic volume in sequence to obtain a preprocessed seismic volume; Noise reduction processing can suppress random noise and local interference, improve the clarity of the response in the fracture neighborhood, and enhance fracture boundaries and discontinuity features by increasing resolution processing.

[0047] Step 303: Input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; Step 304: Overlay the initial fault probability volume with the original seismic profile, select fault interpretation lines that conform to the seismic geological characteristics as annotation samples on the overlay map, and construct a learning sample set and a validation sample set according to the profile blocks; The labeled sample uses fault line labels, and the fault lines should meet at least one of the following conditions: consistent with the fault relationship of the reflective layer, have clear linear extension characteristics, match the local structural pattern, and have reasonable consistency with the distribution direction of nearby known faults.

[0048] Selecting fault interpretation lines that conform to seismic geological characteristics as annotation samples on the overlay map can include: overlaying the initial fault probability volume and the original seismic profile in the same spatial coordinate system, and having interpreters select fault interpretation lines that conform to seismic geological characteristics on the overlay map; preferably, the fault interpretation lines meet at least one of the following conditions: consistent with the fault relationship of the reflection layer, having a clear linear extension characteristic, matching the local structural pattern, and maintaining a reasonable consistency with the direction of the distribution of nearby known faults. Then, the manually selected fault interpretation lines are converted into label data that can be used for model training. The label data can be line labels or narrow band labels formed by extending the fault interpretation lines to a preset width on both sides, in order to solve the problems of excessively thin fault lines, extreme imbalance of positive and negative samples, and unstable network training. In this example, the annotation samples use fault line labels, and the fault interpretation lines are rasterized to form a label matrix corresponding to the profile blocks.

[0049] The aforementioned profile blocks are not directly input from the entire long profile. Instead, they are local windows extracted from the original seismic profile, the corresponding probabilistic response profile, and, if necessary, neighboring trace information, centered on the location of the fault interpretation line, forming sample blocks of fixed size or uniform specifications. By decomposing the long profile into local samples suitable for training a 3D convolutional network, the computational load is reduced. Furthermore, the samples can focus on amplitude abrupt changes, discontinuities, linear extensions, and local texture differences in the fault neighborhood, and the balance of sample distribution can be improved, preventing the network from learning only large-scale background information. The training sample set can consist of multiple labeled profile blocks, while the validation sample set can be preferably selected from different profile blocks that have not participated in the training, to reduce the risk of sample leakage and more realistically evaluate the model's generalization ability.

[0050] The aforementioned profile block can also be a local three-dimensional sub-body extracted from the target profile along the survey line direction, the connecting survey line direction, and the depth direction. The label is obtained by mapping or extending the fault interpretation line on the central profile to the adjacent sampling direction, thus matching the three-dimensional convolutional neural network. It is not a simple two-dimensional profile recognition, but takes into account the continuity of the fracture in space.

[0051] Step 305: Using the three-dimensional convolutional neural network model as the baseline model, perform transfer learning based on the learning sample set, and perform iterative training in combination with the verification sample set; Step 306: Record the evaluation curves of the learning sample set and the validation sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, the number of iterations corresponding to the peak of the validation sample set evaluation curve is determined as the optimal number of iteration weights. The learning samples are input in the form of profile blocks, the number of training iterations is 1000 to 1500, the batch size is 50 to 100, and the optimal number of iteration weights is in the range of 500 to 650.

[0052] Step 307: Construct the final optimized neural network model based on the optimal number of iteration weights, and recalculate the fault probability volume using the final optimized neural network model; Step 308: Perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume. The parameterization post-processing includes: probability threshold filtering, seed point spacing control, dip direction window constraint, strike direction window constraint, normal direction window constraint, dip direction smoothing, strike direction smoothing, dip direction deformation constraint, and strike direction deformation constraint. Among them, parameterized post-processing uses a set of fixed, quantifiable, and reproducible engineering parameters to perform threshold screening, seed point control, spatial window constraints, smoothing, and deformation constraints on the fault probability volume output by the model, so that the fault identification results are more continuous, cleaner, and more in line with geological laws, and finally form engineering results that can be directly used for structural interpretation, trap evaluation, and well location deployment.

[0053] Step 309: Based on the final fault probability volume, form a fracture planar diagram; Step 310: Generate an overlay image of the original seismic profile and the probability volume profile, and use the final fault probability volume and the overlay image as the final fault identification result.

[0054] That is, the above-mentioned fault probability volume can be divided into three layers: the first layer is the initial fault probability volume, which is the probability volume generated after the preprocessed seismic volume is input into the reference three-dimensional convolutional neural network; the second layer is the recalculated fault probability volume, which is the probability volume recalculated by the final optimized neural network model after the optimal number of iteration weights is determined by transfer learning; the third layer is the final fault probability volume after parameterization and processing, which is the final usable probability volume formed after threshold screening, seed point control, window constraint, smoothing and deformation constraint are performed on the recalculated probability volume.

[0055] The fault probability volume obtained in step 308 above is the third layer, which is the final fault probability volume after processing.

[0056] The aforementioned fracture plane map is not a separately generated original data volume, but a planarized interpretation result based on the final fault probability volume. That is, based on the final fault probability volume, high-probability fault responses are extracted along the target layer, depth slice, time slice, or interpretation layer surface, and then projected onto a plane to form a fracture plane map.

[0057] The aforementioned seismic profile can be extracted from the original post-stack depth domain 3D seismic body in the same coordinate system, and can be a longitudinal profile, a transverse profile, or an arbitrary line profile. The aforementioned probabilistic volume profile is the corresponding profile extracted from the corresponding fault probabilistic volume at the same position, orientation, and sampling interval as the seismic profile. During overlay, a unified spatial coordinate, sampling point location, and display range are used. The seismic profile is used as the base map, and the fault probabilistic response is covered on it with color and transparency, thereby forming an overlay map of the probabilistic volume and the seismic profile, thus obtaining the spatial correspondence between the original seismic profile, the probabilistic volume profile, and the fault interpretation line.

[0058] In one embodiment, the aforementioned three-dimensional convolutional neural network model can be a three-dimensional encoder-decoder structure network model, which may include: an input layer, an encoder, a bottleneck layer, a decoder, and an output layer. The input layer receives the preprocessed three-dimensional seismic volume; the encoder extracts features related to fault discontinuity, linear distribution, and local texture differences; the bottleneck layer characterizes the deep spatial semantic features of the fault zone; the decoder progressively restores the spatial resolution and fuses shallow boundary features with deep semantic features through skip connections; and the output layer outputs the fault probability volume corresponding to the input three-dimensional seismic volume.

[0059] The training of this 3D convolutional neural network model can employ the following first loss function to alleviate the imbalance between foreground and background samples caused by sparse tomographic line labels:

[0060] Where N represents the total number of sample points, This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. Indicates the foreground sample weights. Represents the background sample weights, where, > .

[0061] To further improve the consistency of overlap between the predicted fault region and the labeled fault region, the training of this 3D convolutional neural network model can employ the following second loss function:

[0062] Where N represents the total number of sample points, This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. This is a smoothing term used to avoid the denominator being zero.

[0063] The model training employs a joint loss function comprised of a weighted binary cross-entropy loss function and a Dice loss function, which can be expressed as:

[0064] To improve the continuity representation of weak faults, small-scale faults, and complex strike-slip fault combinations in low signal-to-noise ratio post-stack depth-domain seismic data, a fault continuity constraint loss term is introduced based on the loss function. This term suppresses isolated noise responses and unreasonable interruptions in fault prediction results and enhances the continuous distribution of faults along the dip and strike directions. Fault continuity constraint loss term It can be represented as:

[0065] in, This represents the fault neighborhood mask value corresponding to the i-th sample point, used to limit the continuity constraint to mainly act on the fault label neighborhood; This represents the predicted probability value of the i-th sample point in the adjacent sample points in the tilt direction; This represents the predicted probability value of the i-th sample point in the direction of its adjacent sample points; The continuity weights represent the direction of tendency. The continuity weight represents the direction of movement.

[0066] Accordingly, the total loss function can be expressed as:

[0067] in, and For loss weighting coefficients, These are the weighting coefficients for the continuity constraint loss term.

[0068] That is, by constraining the difference in predicted probabilities of dip and strike directions only within the fault neighborhood, the network training results can better match the directional continuous distribution characteristics of the fault in space. The fault continuity constraint loss term works in conjunction with the dip direction window constraint, strike direction window constraint, dip direction smoothing, strike direction smoothing and deformation constraint in the subsequent parameterized post-processing, thereby improving the fault continuity representation ability in both the training and post-processing stages and reducing discrete false anomaly responses under low signal-to-noise ratio conditions.

[0069] Furthermore, in implementation, fracture-sensitive attributes can be extracted based on the preprocessed seismic body. These attributes may include: coherence attributes, tectonic entropy attributes, structural tensor attributes, maximum curvature attributes, maximum likelihood attributes, and likelihood ant body attributes. Fracture-sensitive attributes can be used for conventional fracture identification and performance comparison analysis to demonstrate that in deep, low-signal-to-noise-ratio, and complex strike-slip fault development zones like buried mountain areas, relying solely on conventional attributes is insufficient for consistently identifying weak fractures, small-scale fractures, and complex combinations of fractures. Based on this, a three-dimensional convolutional neural network and transfer learning process are introduced to obtain more continuous and stable fracture identification results. In other words, fracture-sensitive attributes primarily serve as a means of characterizing fracture sensitivity features and a basis for comparative evaluation. Fracture sensitivity attributes can assist interpreters in more accurately identifying fault interpretation lines consistent with seismic geological characteristics on probabilistic volumes and original seismic profile overlays, thereby improving the quality of labeled samples. They can also serve as a benchmark for comparing results before and after transfer learning with those of conventional methods, determining whether the model truly improves the identification of weak and complex faults. Furthermore, they can be used to verify the rationality of post-processing parameters, such as whether the continuity of fault strike, fault combination relationships, and local anomaly responses are consistent with attribute response trends. Therefore, fracture sensitivity attributes can improve the reliability of sample labeling, the credibility of performance evaluation, and the stability of result interpretation. By introducing six types of attributes—coherence, tectonic entropy, structural tensor, maximum curvature, maximum likelihood, and likelihood ant volume—fracture sensitivity characteristics of the target work area can be characterized from different perspectives, providing a basis for comparing the performance of conventional and intelligent methods.

[0070] In the example above, by sequentially denoising and resolving the original post-stack depth domain seismic volume, the fracture response can be enhanced and the identification basis under low signal-to-noise ratio conditions can be improved. The fault probability volume output by a three-dimensional convolutional neural network can more fully extract fracture feature information from the seismic volume than conventional attribute methods. By constructing a local sample set by selecting fault interpretation lines that conform to seismic geological characteristics on the overlay map of the probability volume and the original seismic profile, the samples can better fit the fracture geometry of the target work area. By introducing a transfer learning approach that combines learning and validation sample sets, and determining the optimal number of iteration weights based on the peak value of the validation sample set evaluation curve, the risk of overfitting can be reduced, and the model's adaptability to local fracture features can be improved. By setting parameterized post-processing constraints such as thresholds, seed points, windows, smoothing, and deformation on the fault probability volume, the continuity of fracture distribution can be improved, discrete noise response reduced, and engineering-usable fracture identification results formed. The final output fault probability volume, fracture plan view, and overlay profile can directly serve well location deployment, trap evaluation, and structural interpretation in the Halahatang buried mountain area.

[0071] The original post-stack depth-domain 3D seismic volume of the target layer in the buried mountainous area of ​​the Halahatang Oilfield in the Tarim Basin was used as input data. First, the original seismic volume underwent noise reduction processing to suppress random noise and local interference. Then, based on the noise reduction, further resolution enhancement processing was performed to improve the response of fault boundaries and discontinuities. For example... Figure 4 The image shows a comparison diagram of the original seismic body and the denoised seismic body. Figure 5 The image shows a comparison diagram of the original seismic volume and the seismic volume after noise reduction and resolution improvement.

[0072] After obtaining the preprocessed seismic body, six types of fracture-sensitive attributes were extracted, including coherence attributes, tectonic entropy attributes, structural tensor attributes, maximum curvature attributes, maximum likelihood attributes, and likelihood ant body attributes, which were used for conventional fracture identification and effect comparison analysis. Different attributes have different sensitivities to fracture response in the target work area, but overall there are still problems with insufficient characterization of weak fractures, small-scale fractures, and complex combinations.

[0073] Based on this, the preprocessed seismic volume is input into a 3D convolutional neural network model to generate an initial fault probability volume. The 3D convolutional neural network is a 3D encoder-decoder structure. The encoder extracts fault boundaries, discontinuities, and local texture differences through multi-layer 3D convolution. The decoder restores spatial resolution through progressive upsampling and incorporates skip connections to enhance the retention of shallow boundary information. Finally, it outputs the fault probability volume corresponding to the input 3D seismic volume. Figure 6 The diagram shows the initial fault probability volume generated by the convolutional neural network. Figure 7 The diagram shown is a schematic of the optimized result of the initial fault probability volume.

[0074] During model training, training samples are input as profile blocks, and fault line labels are used. Specifically, the fault probability volume generated by the convolutional neural network is overlaid with the original seismic profile. Fault interpretation lines that conform to seismic geological characteristics are manually selected as annotation samples on the overlay image. The selected fault interpretation lines must meet at least one of the following conditions: consistent with seismic reflection faulting, have clear linear extension characteristics, match the local tectonic pattern, and maintain reasonable consistency with the distribution direction of adjacent faults. Fault line samples are selected as many as possible from multiple profiles to form a training sample set; at the same time, fault line samples are selected from profiles that have not participated in the training to form a validation sample set.

[0075] During model training, a weighted binary cross-entropy loss function can be used, and the Adam optimizer can be used as the optimizer. In some implementations, in order to further improve the overlap consistency between the fault foreground and the fault line label, a joint loss function can be used. The number of training iterations is preferably set to 1000 to 1500 times, and the batch size is preferably set to 50 to 100.

[0076] The above loss function interrupts the neighborhood mask value. It can be determined by one or more of the following: fault label, fault label extension region, and high-confidence fault candidate regions in the initial fault probability volume, with a bias direction continuity weight. and direction of continuity weight The settings can be the same, or different settings can be made according to the local fracture propagation direction, local discontinuity response intensity, or initial fault probability value.

[0077] By constraining the difference in predicted probabilities for dip and strike directions only within the fault neighborhood, the network training results can better reflect the directional continuity of faults in space. Furthermore, the fault continuity constraint loss term, in conjunction with subsequent parameterized post-processing constraints such as dip direction window constraints, strike direction window constraints, dip direction smoothing, strike direction smoothing, and deformation constraints, collectively improves the fault continuity representation capability during both the training and post-processing stages, reducing discrete false anomaly responses under low signal-to-noise ratio conditions.

[0078] To improve the model's ability to identify local fault patterns in the Qianshan area, transfer learning can be introduced on top of the baseline 3D convolutional neural network model. Specifically, the baseline model can be fine-tuned locally using multiple cross-sectional fault line samples from the learning sample set, while the model's generalization ability is evaluated using the validation sample set. During the transfer learning process, the evaluation curves of the learning sample set and the validation sample set are recorded. When the evaluation result of the learning sample set continuously increases, while the evaluation result of the validation sample set first increases and then decreases, the number of iterations corresponding to the peak value of the validation sample set evaluation curve is determined as the optimal number of iterations for the weight. For example, the optimal number of iterations for the weight can be preferably located in the range of 500 to 650.

[0079] The final optimized neural network model is reconstructed using the optimal number of iterations for the weights, and the fault probability volume is recalculated based on this model. For example... Figure 8 The diagram shown is a schematic of the fault probability volume after transfer learning. Figure 9 The diagram shown is a schematic of the probability volume of a tomography fault in a typical convolutional neural network. Figure 10 The image shows the superposition effect of earthquake bodies and the superposition effect of fault probability volumes and earthquake bodies after transfer learning. Figure 11 The diagram shows the fracture recognition results after transfer learning. It is evident that the results after transfer learning are significantly superior to those of ordinary convolutional neural networks in terms of fracture continuity and fracture distribution representation, and also superior to the results of various conventional attribute recognition methods.

[0080] After obtaining the fault probability volume output by transfer learning, a parameterized post-processing procedure can be performed on it, which may include: S1: Probability threshold screening: Set probability threshold = 0.6 to suppress low-confidence responses below this threshold; S2: Seed point spacing control: Set distance between seedpoints = 4 stps to control the spatial distribution density of the initial seed points and avoid excessively dense seed points that may lead to repeated tracking or noise propagation; S3: Window constraints: Set window on dip = 150 ms, window on strike = 500 m, and window on normal = 250 m to limit the neighborhood scale of fracture extraction in the dip direction, strike direction, and normal direction, respectively. S4: Smoothing: Set smooth on dip = 2.00 and smooth on strike = 2.00 to improve the continuity of the fracture extraction results in the dip and strike directions; S5: Deformation constraint: Set strain on dip = 0.25 and strain on strike = 0.25 to limit excessive bending or abnormal displacement in local areas during fracture extraction.

[0081] After the above parameterization post-processing is completed, the output fault probability volume, fault plane map, and the overlay map of the probability volume and seismic profile are used as the final fault identification results for well location deployment, trap evaluation and structural interpretation in the target work area.

[0082] In the example above, a migration learning-based fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas is presented. This method can improve the identification accuracy and distribution continuity of complex fractures, weak fractures, and small-scale fractures in low signal-to-noise ratio post-stack depth domain seismic data, providing a basis for well location deployment, trap evaluation, and structural interpretation.

[0083] The methods and embodiments provided in the above-described embodiments of this application can be executed in a mobile terminal, computer terminal, or similar computing device. Taking operation on an electronic device as an example... Figure 12 This is a hardware structure block diagram of an electronic device for a fracture identification method in ultra-deep carbonate reservoirs in buried mountainous areas, as provided in this application. Figure 12 As shown, the electronic device 10 may include one or more (only one is shown in the figure) processors 02 (processors 02 may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), a memory 04 for storing data, and a transmission module 06 for communication functions. Those skilled in the art will understand that... Figure 12The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, electronic device 10 may also include... Figure 12 The more or fewer components shown, or having the same Figure 12 The different configurations shown.

[0084] The memory 04 can be used to store software programs and modules for application software, such as the program instructions / modules corresponding to the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas in this embodiment of the application. The processor 02 executes various functional applications and data processing by running the software programs and modules stored in the memory 04, thereby realizing the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas described above. The memory 04 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 04 may further include memory remotely located relative to the processor 02, and these remote memories can be connected to the electronic device 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0085] The transmission module 06 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the electronic device 10. In one example, the transmission module 06 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 06 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0086] At the software level, the aforementioned fracture identification device for ultra-deep carbonate reservoirs in buried mountainous areas can, for example, Figure 13 As shown, it includes: Module 1301 is used to acquire the original post-stack depth domain three-dimensional seismic volume of the target work area; The preprocessing module 1302 is used to sequentially perform noise reduction and resolution improvement processing on the original post-stack depth domain three-dimensional seismic volume to obtain a preprocessed seismic volume. Input module 1303 is used to input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain an initial fault probability volume; The overlay module 1304 is used to overlay and display the initial fault probability volume and the original seismic profile to form an initial overlay map. On the initial overlay map, fault interpretation lines that conform to the seismic geological characteristics are selected as annotation samples, and a learning sample set and a verification sample set are constructed according to the profile blocks. The learning module 1305 is used to use the three-dimensional convolutional neural network model as a benchmark model, perform transfer learning based on the learning sample set, and perform iterative training in combination with the verification sample set. It records the evaluation curve of the learning sample set and the evaluation curve of the verification sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the verification sample set rises first and then falls, the number of iterations corresponding to the peak of the evaluation curve of the verification sample set is determined as the optimal number of iteration weights. The calculation module 1306 is used to construct the final optimized neural network model based on the optimal number of iteration weights, and to recalculate the fault probability volume using the final optimized neural network model. The parameterization module 1307 is used to perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and to form a probability volume profile based on the final fault probability volume. The identification module 1308 is used to overlay the original seismic profile with the probability volume profile to form a final overlay image, and to use the final fault probability volume and the final overlay image as the final fault identification result.

[0087] In one embodiment, the overlay module 1304 can be specifically used to select interpretation lines on the initial overlay map that meet at least one of the following conditions as fault interpretation lines: consistent with the fault relationship of the reflection layer, having a clear linear extension characteristic, matching the local structural pattern, and maintaining a reasonable consistency with the direction of the distribution of nearby known faults; converting the selected fault interpretation lines into label data that can be used for model training, wherein the label data is a narrow band label formed by expanding the fault interpretation lines on both sides of a preset width; performing rasterization processing on the selected fault interpretation lines through the label data to form a label matrix corresponding to the profile blocks, wherein the profile blocks are sample blocks of fixed size or uniform specifications formed by truncating local windows from the original seismic profile, the corresponding probability response profile, and, if necessary, the adjacent trace information, with the fault interpretation line as the center; and forming a learning sample set by combining multiple labeled profile blocks according to the label matrix, and constructing a validation sample set in different profile blocks that have never participated in the training.

[0088] In one implementation, the three-dimensional convolutional neural network model is trained according to the following loss function:

[0089] in, The weighting coefficient for the first loss term. The weighting coefficient for the second loss term. The weighting coefficients for the continuity constraint loss term. As the first loss item, This is the second loss item. The loss term represents the continuity constraint, where N represents the total number of sample points. This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. Indicates the foreground sample weights. Represents the background sample weights, where, > , This is a smoothing term used to avoid the denominator being zero. This represents the fault neighborhood mask value corresponding to the i-th sample point. This represents the predicted probability value of the i-th sample point among its adjacent sample points in the tilt direction. This represents the predicted probability value of the i-th sample point in the direction of its adjacent sample points. The continuity weights represent the direction of tendency. The continuity weight represents the direction of movement.

[0090] In one implementation, the preprocessing module 1302 can specifically be used to convert the original post-stack depth-domain three-dimensional seismic volume in the spatial domain into a wavenumber-domain spectrum through a three-dimensional Fourier transform:

[0091] in, This is the original post-stack depth domain 3D seismic volume. For seismic data plane coordinates, The coordinates are in the depth direction. The complex amplitude spectrum of the earthquake body in the wavenumber domain. It is a three-dimensional discrete Fourier transform operator. The wave number in the x-direction. The wave number in the y-direction. The wave number in the z-direction; Constructing a fracture-guided frequency domain mask using a sigmoid function:

[0092] in, As a fracture-guided frequency domain mask, For three-dimensional total wavenumber mode, This is the mask steepness coefficient. The center wavenumber of the fracture-sensitive frequency band, It is a natural constant; The denoised seismic body is obtained by performing inverse Fourier transform on the fracture-guided frequency domain mask and wavenumber domain spectrum.

[0093] in, The noise-reduced seismic body after adaptive frequency domain noise reduction. This is a three-dimensional inverse Fourier transform, converting the wavenumber domain result back to the spatial domain; An adaptive deconvolution operator is constructed using the working zone layer velocity and the formation dip angle as constraints. The adaptive deconvolution operator is convolved with the denoised seismic volume in 3D to obtain a high-resolution preprocessed seismic volume:

[0094] in, For high-resolution preprocessed seismic volumes, For adaptive deconvolution operators, This is a three-dimensional convolution operation.

[0095] In one embodiment, forming a probability volume profile based on the final fault probability volume may include: extracting an original seismic profile from the longitudinal or transverse survey profile of the original post-stack depth domain 3D seismic body under preset coordinates; extracting a corresponding profile from the final fault probability volume at a position, orientation, and sampling interval completely consistent with the original seismic profile, as the probability volume profile; correspondingly, overlaying the original seismic profile and the probability volume profile to form a final overlay image may include: using unified spatial coordinates, sampling points, and display range, using the original seismic profile as a base map, and overlaying the probability volume profile on the base map with color and transparency to form the final overlay image.

[0096] In one implementation, the parameterization post-processing may include: probability threshold filtering, seed point spacing control, yaw direction window constraint, homing direction window constraint, normal direction window constraint, yaw direction smoothing, homing direction smoothing, yaw direction deformation constraint, and homing direction deformation constraint.

[0097] This application also provides a specific implementation of an electronic device capable of implementing all steps of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas as described in the above embodiments. The electronic device specifically includes: a processor, a memory, a communication interface, and a bus; wherein the processor, memory, and communication interface communicate with each other via the bus; the processor is used to call a computer program in the memory, and when the processor executes the computer program, it implements all steps of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas as described in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step 1: Obtain the original post-stack depth domain 3D seismic volume of the target work area; Step 2: Perform noise reduction and resolution enhancement processing on the original post-stack depth domain 3D seismic volume in sequence to obtain the preprocessed seismic volume; Step 3: Input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; Step 4: Overlay the initial fault probability volume with the original seismic profile to form an initial overlay map. Select fault interpretation lines that conform to the seismic geological characteristics as annotation samples on the initial overlay map, and construct a learning sample set and a validation sample set according to the profile blocks. Step 5: Using the three-dimensional convolutional neural network model as the baseline model, perform transfer learning based on the learning sample set, and combine it with the validation sample set for iterative training. Record the evaluation curve of the learning sample set and the evaluation curve of the validation sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, determine the number of iterations corresponding to the peak of the validation sample set evaluation curve as the optimal number of iteration weights. Step 6: Construct the final optimized neural network model based on the optimal number of iteration weights, and recalculate the fault probability volume using the final optimized neural network model; Step 7: Perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and form a probability volume profile based on the final fault probability volume; Step 8: Overlay the original seismic profile with the probability volume profile to form a final overlay image, and use the final fault probability volume and the final overlay image as the final fault identification result.

[0098] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas as described in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the fracture identification method for ultra-deep carbonate reservoirs in buried mountainous areas as described in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step 1: Obtain the original post-stack depth domain 3D seismic volume of the target work area; Step 2: Perform noise reduction and resolution enhancement processing on the original post-stack depth domain 3D seismic volume in sequence to obtain the preprocessed seismic volume; Step 3: Input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; Step 4: Overlay the initial fault probability volume with the original seismic profile to form an initial overlay map. Select fault interpretation lines that conform to the seismic geological characteristics as annotation samples on the initial overlay map, and construct a learning sample set and a validation sample set according to the profile blocks. Step 5: Using the three-dimensional convolutional neural network model as the baseline model, perform transfer learning based on the learning sample set, and combine it with the validation sample set for iterative training. Record the evaluation curve of the learning sample set and the evaluation curve of the validation sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, determine the number of iterations corresponding to the peak of the validation sample set evaluation curve as the optimal number of iteration weights. Step 6: Construct the final optimized neural network model based on the optimal number of iteration weights, and recalculate the fault probability volume using the final optimized neural network model; Step 7: Perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and form a probability volume profile based on the final fault probability volume; Step 8: Overlay the original seismic profile with the probability volume profile to form a final overlay image, and use the final fault probability volume and the final overlay image as the final fault identification result.

[0099] As described above, the embodiments of this application first preprocess the original post-stack depth domain 3D seismic volume to improve data quality, and then input it into a 3D convolutional neural network to obtain an initial fault probability volume. Based on the initial overlay of the probability volume and the seismic profile, fault interpretation lines that conform to geological characteristics are selected to construct a learning and validation sample set. Transfer learning is carried out using the 3D convolutional neural network as the benchmark model, and the optimal iteration weights are determined based on the peak value of the validation set evaluation curve to obtain an optimized model and recalculate the fault probability volume. After parameterization postprocessing, the final fault probability volume is obtained, the probability volume profile is extracted and accurately overlaid with the original seismic profile for display, and the final fault identification result is output. The above scheme can solve the technical problems of low fault identification accuracy, difficulty in characterizing weak and small-scale faults, poor continuity of fault distribution, and easy overfitting of intelligent models in existing ultra-deep low signal-to-noise ratio seismic data. It achieves the technical effect of improving the identification accuracy and continuity of complex, weak, and small-scale faults, and reducing the risk of model overfitting and interpretation uncertainty.

[0100] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0101] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0102] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in the embodiments or drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0103] While this specification provides method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or end product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in the process, method, product, or apparatus that includes said elements is not excluded.

[0104] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing the embodiments of this specification, the functions of each module can be implemented in one or more software and / or hardware components, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0105] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0106] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of computer program products implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] The embodiments described in this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. The embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0109] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, system embodiments are basically similar to method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. In the description of this specification, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments in this specification. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0110] The above description is merely an embodiment of the present specification and is not intended to limit the embodiments of the present specification. For those skilled in the art, various modifications and variations can be made to the embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present specification should be included within the scope of the claims of the embodiments of the present specification.

Claims

1. A method for fracture identification in ultra-deep carbonate reservoirs in buried mountainous areas, characterized in that, The method includes: Obtain the original post-stack depth domain 3D seismic volume of the target work area; The original post-stack depth domain 3D seismic volume is sequentially subjected to noise reduction and resolution enhancement processing to obtain a preprocessed seismic volume. The preprocessed seismic volume is input into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; The initial fault probability volume is overlaid with the original seismic profile to form an initial overlay map. Fault interpretation lines that conform to the seismic geological characteristics are selected as annotation samples on the initial overlay map, and a learning sample set and a validation sample set are constructed according to the profile blocks. Using the three-dimensional convolutional neural network model as the benchmark model, transfer learning is performed based on the learning sample set, and iterative training is performed in combination with the validation sample set. The evaluation curves of the learning sample set and the validation sample set are recorded. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, the number of iterations corresponding to the peak of the validation sample set evaluation curve is determined as the optimal number of iteration weights. The final optimized neural network model is constructed based on the optimal number of iteration weights, and the fault probability volume is recalculated using the final optimized neural network model. The recalculated fault probability volume is parameterized and post-processed to obtain the final fault probability volume. Based on the final fault probability volume, a probability volume profile is formed. The original seismic profile is overlaid with the probabilistic volume profile to form a final overlay image. The final fault probabilistic volume and the final overlay image are used as the final fault identification result.

2. The method according to claim 1, characterized in that, On the initial overlay map, fault interpretation lines that conform to seismic geological characteristics are selected as annotation samples, and a learning sample set and a validation sample set are constructed according to profile blocks, including: On the initial superposition diagram, an interpretation line that meets at least one of the following conditions is selected as the fault interpretation line: consistent with the fault relationship of the reflection layer, having a clear linear extension characteristic, matching the local structural pattern, and maintaining a reasonable consistency with the distribution direction of the adjacent known faults. The selected fault interpretation lines are converted into label data that can be used for model training. The label data is a narrow band label formed by extending the fault interpretation lines on both sides with a preset width. The selected fault interpretation lines are rasterized using the label data to form a label matrix corresponding to the profile block. The profile block is a sample block of fixed size or uniform specification formed by extracting a local window from the original seismic profile, the corresponding probability response profile, and the information of adjacent traces, with the fault interpretation line as the center. Based on the label matrix, a learning sample set is formed by combining multiple labeled profile blocks, and a validation sample set is constructed from different profile blocks that have never participated in training.

3. The method according to claim 1, characterized in that, The three-dimensional convolutional neural network model is trained using the following loss function: in, The weighting coefficient for the first loss term. The weighting coefficient for the second loss term. These are the weighting coefficients for the continuity constraint loss term. As the first loss item, This is the second loss item. The loss term represents the continuity constraint, where N represents the total number of sample points. This represents the label value of the i-th sample point. This represents the probability value that the model predicts the i-th sample point to be a fault. Indicates the foreground sample weights. Represents the background sample weights, where, > , This is a smoothing term used to avoid the denominator being zero. This represents the fault neighborhood mask value corresponding to the i-th sample point. This represents the predicted probability value of the i-th sample point among its adjacent sample points in the tilt direction. This represents the predicted probability value of the i-th sample point in the direction of its adjacent sample points. The continuity weights represent the direction of tendency. The continuity weight represents the direction of movement.

4. The method according to claim 1, characterized in that, The original post-stack depth domain 3D seismic volume is subjected to noise reduction and resolution enhancement processes sequentially to obtain a preprocessed seismic volume, including: The original post-stack depth-domain 3D seismic volume in the spatial domain is converted into a wavenumber-domain spectrum using a 3D Fourier transform: in, This is the original post-stack depth domain 3D seismic volume. For seismic data plane coordinates, The coordinates are in the depth direction. The complex amplitude spectrum of the earthquake body in the wavenumber domain. It is a three-dimensional discrete Fourier transform operator. The wave number in the x-direction. The wave number in the y-direction. The wave number in the z-direction; Constructing a fracture-guided frequency domain mask using a sigmoid function: in, As a fracture-guided frequency domain mask, For three-dimensional total wavenumber mode, This is the mask steepness coefficient. The center wavenumber of the fracture-sensitive frequency band, It is a natural constant; The denoised seismic body is obtained by performing inverse Fourier transform on the fracture-guided frequency domain mask and wavenumber domain spectrum. in, The noise-reduced seismic body after adaptive frequency domain noise reduction. This is a three-dimensional inverse Fourier transform, converting the wavenumber domain result back to the spatial domain; An adaptive deconvolution operator is constructed using the working zone layer velocity and the formation dip angle as constraints. The adaptive deconvolution operator is convolved with the denoised seismic volume in 3D to obtain a high-resolution preprocessed seismic volume: in, For high-resolution preprocessed seismic volumes, For adaptive deconvolution operators, This is a three-dimensional convolution operation.

5. The method according to claim 1, characterized in that, Based on the final fault probability volume, a probability volume profile is formed, including: The original seismic profile is extracted from the longitudinal or transverse profile of the original post-stack depth domain three-dimensional seismic body under preset coordinates. At a location, orientation, and sampling interval that are completely consistent with the original seismic profile, the corresponding profile is extracted from the final fault probability volume and used as the probability volume profile. Accordingly, the original seismic profile and the probabilistic volume profile are overlaid and displayed to form a final overlay image, including: Using unified spatial coordinates, sampling points, and display range, the original seismic profile is used as the base map, and the probabilistic volume profile is overlaid on the base map with color and transparency to form the final overlay map.

6. The method according to any one of claims 1 to 5, characterized in that, The parameterized post-processing includes: probability threshold filtering, seed point spacing control, directional window constraint, directional window constraint, normal window constraint, directional smoothing, directional smoothing, directional deformation constraint, and directional deformation constraint.

7. A fracture identification device for ultra-deep carbonate reservoirs in buried mountainous areas, characterized in that, include: The acquisition module is used to acquire the original post-stack depth domain 3D seismic volume of the target work area; The preprocessing module is used to sequentially perform noise reduction and resolution improvement processing on the original post-stack depth domain 3D seismic volume to obtain a preprocessed seismic volume; The input module is used to input the preprocessed seismic volume into a three-dimensional convolutional neural network model to obtain the initial fault probability volume; The overlay module is used to overlay and display the initial fault probability volume with the original seismic profile to form an initial overlay map. On the initial overlay map, fault interpretation lines that conform to the seismic geological characteristics are selected as annotation samples, and a learning sample set and a validation sample set are constructed according to the profile blocks. The learning module is used to perform transfer learning based on the learning sample set, using the three-dimensional convolutional neural network model as the benchmark model, and to perform iterative training in combination with the validation sample set. It records the evaluation curve of the learning sample set and the evaluation curve of the validation sample set. When the evaluation result of the learning sample set continues to rise and the evaluation result of the validation sample set rises first and then falls, the number of iterations corresponding to the peak of the evaluation curve of the validation sample set is determined as the optimal number of iteration weights. The calculation module is used to construct the final optimized neural network model based on the optimal number of iteration weights, and to recalculate the fault probability volume using the final optimized neural network model. The parameterization module is used to perform parameterization post-processing on the recalculated fault probability volume to obtain the final fault probability volume, and to form a probability volume profile based on the final fault probability volume. The identification module is used to overlay and display the original seismic profile with the probability volume profile to form a final overlay image, and to use the final fault probability volume and the final overlay image as the final fault identification result.

8. An electronic device comprising a processor and a memory for storing processor-executable instructions, characterized in that, When the processor executes the instructions, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 6.

Images