Transform enhancement-based U-NET fracture identification method and device

Abstract

The invention relates to the technical field of AI seismic attribute detection, and particularly discloses a Transform enhancement-based U-NET fracture recognition method and device, and the method comprises the steps: carrying out the artificial fault interpretation of original three-dimensional seismic data, obtaining a fault pattern recognition sample, inputting the fault pattern recognition sample into a network of a U-NET architecture, and carrying out the recognition of the fault pattern; performing feature extraction, feature fusion, up-sampling, feature processing, loss calculation and parameter optimization to obtain a training model; evaluating the training model to obtain an optimal training model; and applying the optimal training model to actual work, and carrying out fault identification on new seismic data. According to the invention, after the network is fully trained, the fault position in the seismic data can be accurately identified, and especially for an obvious fault, the identification accuracy can reach a high level; for seismic data with multiple staggered faults, irregular fault forms and noise interference, the characteristics of different faults can be distinguished, so that each fault can be accurately identified.

Description

Technical Field

[0001] This invention relates to the field of AI seismic attribute detection technology, specifically to a U-NET fracture identification method and apparatus based on Transformer enhancement. Background Technology

[0002] Fault prediction is a core task in seismic exploration interpretation, directly impacting the efficiency and effectiveness of oil and gas field exploration and development. The main method for fault prediction utilizes pre-stack and post-stack seismic data to extract and analyze spatial attributes sensitive to fault characteristics. In the early days, fault identification based on seismic data relied primarily on manual work by geological experts. This method heavily depended on the experts' professional knowledge and practical experience, and required significant time and manpower. With the development of machine learning technology, some methods based on traditional machine learning algorithms have been applied to fault identification based on seismic data. The first step is feature engineering, which involves manually extracting various features from seismic data, such as the amplitude, frequency, and phase of seismic waves, as well as features obtained based on texture analysis. Commonly used models include Support Vector Machines (SVM), decision trees, and random forests. Typically, a relatively large amount of representative and high-quality data is needed to train the model to achieve good results. Early deep learning methods, such as simple convolutional neural networks (CNNs), can progressively extract features from low to high levels through multiple convolutional and pooling operations. However, it is relatively simple to integrate features at different levels, mainly relying on pooling and fully connected layers to integrate features, which has limitations when processing data with complex structures such as seismic data.

[0003] Currently, machine learning methods have been applied to some extent in the field of seismic exploration. The two main types of methods that are most suitable are deep learning and pattern recognition. However, the samples they are based on are all single-point in form, and they do not actually use the fracture pattern as a sample for training. This makes the predicted fracture detection results scattered on both the profile and the plane, and even contrary to the actual fracture direction.

[0004] Based on this technical background, this invention studies a U-NET fracture identification method and device based on Transformer enhancement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a Transformer-enhanced U-NET fault identification method and apparatus. After sufficient training, the network can accurately identify fault locations in seismic data, especially for obvious faults, achieving a high level of accuracy. It can clearly delineate the strike and extent of faults, reducing misjudgments and omissions. For seismic data with multiple intersecting faults, irregular fault morphologies, or noise interference, the network's multi-level feature extraction and fusion mechanism can distinguish the characteristics of different faults, thereby accurately identifying each fault.

[0006] To achieve the above objectives, a first aspect of the present invention provides a Transformer-enhanced U-NET fracture identification method, comprising:

[0007] Artificial fault interpretation was performed on the raw 3D seismic data to obtain fault pattern recognition samples;

[0008] The tomographic pattern recognition samples are input into a U-NET architecture network for feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization, to obtain a training model.

[0009] The training model is evaluated, and the training model is adjusted and optimized based on the evaluation results to obtain the optimal training model;

[0010] The optimal training model is then applied to actual seismic data fault identification work to identify faults in new seismic data.

[0011] A second aspect of the present invention provides a Transformer-enhanced U-NET fracture detection device, comprising:

[0012] The fault interpretation module is used to perform artificial fault interpretation on the raw 3D seismic data to obtain fault pattern recognition samples.

[0013] The model training module is used to input the tomographic pattern recognition samples into the U-NET architecture network, perform feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization, to obtain the trained model.

[0014] The model evaluation module is used to evaluate the training model and adjust and optimize the training model based on the evaluation results to obtain the optimal training model.

[0015] The model application module is used to apply the optimal training model to actual seismic data fault identification work to identify faults in new seismic data.

[0016] A third aspect of the present invention provides an electronic device, the electronic device comprising:

[0017] Memory, which stores executable instructions;

[0018] A processor that executes the executable instructions in the memory to implement the Transformer-enhanced U-NET fracture identification method described in the first aspect.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the Transformer-enhanced U-NET fracture identification method described in the first aspect.

[0020] The beneficial effects of this invention include:

[0021] (1) The U-NET fault identification method based on Transformer enhancement proposed in this invention can accurately identify the fault location in seismic data after the network has been fully trained. In particular, for more obvious faults, its identification accuracy can reach a high level, and it can clearly delineate the direction and extent of the fault, reducing the situation of misjudgment and omission. For seismic data with multiple faults intersecting, irregular fault morphology and noise interference, the features of different faults can be distinguished through the multi-level feature extraction and fusion mechanism of the network, thereby accurately identifying each fault.

[0022] (2) The U-NET fault identification method based on Transformer enhancement proposed in this invention realizes the automation of fault identification in seismic data. Once the network training is completed, it can quickly process a large amount of seismic data, which greatly improves work efficiency. Compared with the traditional manual fault identification method, it can give results in a short time when processing large-scale seismic data, saving manpower and time costs.

[0023] (3) The U-NET fault identification method based on Transformer enhancement proposed in this invention has the following advantages over traditional manual identification methods: Traditional manual fault identification requires geological experts to analyze and label seismic data one by one, which is an extremely time-consuming process. However, this method can process a large amount of data in a short time, and the processing speed is far greater than that of manual identification. Stability advantage: Manual identification is easily affected by subjective factors, such as the experience and fatigue of experts. Different experts may have different judgments on faults in the same seismic data. However, once the network of this method is trained, its judgment criteria are based on the learned data features, which has high stability and consistency.

[0024] (4) The U-NET fault identification method based on Transformer enhancement proposed in this invention has the following advantages over other traditional machine learning methods: Traditional machine learning methods often require manually designed features for fault identification, and these manually designed features may not fully reflect the fault information in seismic data; while the network of this method can automatically extract features from the original seismic data layer by layer, including features at different scales, without manual intervention, and can mine richer and more representative feature information; It also has the following advantages: When facing seismic data of different regions and types, traditional machine learning methods may need to readjust the feature extraction method and model parameters; while the network of this method, after being trained with a large amount of data, has a stronger generalization ability and can better adapt to different seismic data, reducing the complexity of adjusting the model for different data.

[0025] (5) The U-NET fault identification method based on Transformer enhancement proposed in this invention has the advantages of U-shaped structure and skip-layer connection compared with other deep learning architectures: Early convolutional neural networks may have shortcomings when dealing with problems such as seismic data that require simultaneous consideration of global and local features. However, the U-shaped structure and skip-layer connection in the U-NET architecture of this method can effectively fuse shallow local detail features and deep global semantic features. In seismic fault identification, shallow features can help determine the precise boundary of the fault, while deep features can help grasp the overall direction and distribution of the fault. This fusion mechanism makes fault identification more accurate. It also has the ability to process multi-scale features: Early networks may not be flexible enough when processing multi-scale features. The U-NET architecture of this method, through the downsampling process of the encoder and the upsampling process of the decoder, combined with the feature fusion of different levels, can better process fault features of different scales in seismic data. Small-scale fault details can be captured in the shallow layer, and large-scale fault direction can be extracted in the deep layer and well integrated and processed in the network.

[0026] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0027] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating the U-NET fracture identification method based on Transformer enhancement proposed in this invention.

[0029] Figure 2This is a schematic diagram of the network structure of the U-NET architecture in a specific implementation of the Transformer-enhanced U-NET fracture identification method proposed in this invention.

[0030] Figure 3 This is a schematic diagram of a tomographic pattern recognition sample in a specific implementation of the Transformer-enhanced U-NET fracture recognition method proposed in this invention.

[0031] Figure 4 This is a schematic diagram of AI training execution curve analysis in a specific implementation of the Transformer-enhanced U-NET fracture recognition method proposed in this invention.

[0032] Figure 5 This is a schematic diagram of the AI ​​training network model parameters in a specific implementation of the Transformer-enhanced U-NET fracture recognition method proposed in this invention.

[0033] Figure 6 This is a schematic diagram of a fault prediction profile in a specific implementation of the Transformer-enhanced U-NET fault identification method proposed in this invention.

[0034] Figure 7 This is a schematic diagram of the layer prediction plane in a specific implementation of the Transformer-enhanced U-NET fracture identification method proposed in this invention. Detailed Implementation

[0035] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0036] This invention provides a U-NET fracture identification method based on Transformer enhancement, such as... Figure 1 As shown, it includes:

[0037] Artificial fault interpretation was performed on the raw 3D seismic data to obtain fault pattern recognition samples;

[0038] The tomographic pattern recognition samples are input into the U-NET architecture network for feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization to obtain the training model.

[0039] The training model is evaluated, and the model is adjusted and optimized based on the evaluation results to obtain the optimal training model.

[0040] The optimal training model is applied to actual seismic data fault identification work to identify faults in new seismic data.

[0041] In this invention, after the network is fully trained, it can accurately identify the location of faults in seismic data. In particular, for obvious faults, the identification accuracy can reach a high level, clearly outlining the direction and extent of the faults and reducing misjudgments and omissions. For seismic data with multiple intersecting faults, irregular fault morphology, and noise interference, the network's multi-level feature extraction and fusion mechanism can distinguish the characteristics of different faults, thereby accurately identifying each fault.

[0042] According to the present invention, the original three-dimensional seismic data is subjected to artificial fault interpretation to obtain fault pattern recognition samples, including:

[0043] The original 3D seismic data is converted into 2D slices, namely isochronous / depth slices and seismic profiles;

[0044] Artificial fault interpretation is performed on two-dimensional slices, and the results of seismic profiles and artificial fault interpretation are converted into image files;

[0045] Image files are used as samples for tomographic pattern recognition.

[0046] According to the present invention, feature extraction includes:

[0047] In the encoder of the network, features are extracted from the input data through a series of convolution operations. With each convolution operation, the network extracts more abstract and representative features.

[0048] Max pooling is used to reduce data dimensionality while retaining key feature information, helping the network gradually focus on more critical features.

[0049] According to the present invention, feature fusion, i.e., cross-layer connectivity, includes:

[0050] During the downsampling process of the network, features at different levels are fused to enable the network to comprehensively utilize feature information at different levels of abstraction, thus avoiding excessive information loss during dimensionality reduction.

[0051] This invention automates the identification of faults in seismic data. Once the network is trained, it can quickly process large amounts of seismic data, greatly improving work efficiency. Compared with traditional manual fault identification methods, it can provide results in a short time when processing large-scale seismic data, saving manpower and time costs.

[0052] According to the present invention, upsampling and feature processing include:

[0053] In the network's decoder, the dimensions of the data are gradually restored through upsampling operations to make the data size closer to the original input data.

[0054] After upsampling, the features are further adjusted and processed by convolution operations to make the feature information more accurate and targeted.

[0055] The upsampling method used is either deconvolution or interpolation.

[0056] Compared to traditional manual identification methods, this invention offers several advantages: Traditional manual fault identification requires geological experts to analyze and label seismic data one by one, an extremely time-consuming process. This method, however, can process large amounts of data in a short time, far exceeding the processing speed of manual methods. Furthermore, it offers stability advantages: manual identification is easily affected by subjective factors, such as the expert's experience and fatigue level. Different experts may have different judgments about faults in the same seismic data. Once the network of this method is trained, its judgment criteria are based on the learned data features, exhibiting high stability and consistency.

[0057] Preferably, loss calculation and parameter optimization include:

[0058] The network output is compared with manually labeled labels, and the cross-entropy loss function is used to calculate the difference between the prediction results and the true labels to obtain the loss value;

[0059] Based on the loss value, stochastic gradient descent and its variants are used to backpropagate and update the network parameters, continuously adjusting the network weights and biases to make the network's prediction results increasingly closer to the actual fault situation.

[0060] Compared to other traditional machine learning methods, this invention has the following advantages: It possesses superior feature extraction capabilities. Traditional machine learning methods often require manually designed features for fault identification, which may not fully reflect the fault information in seismic data. In contrast, the network in this invention can automatically extract features layer by layer from raw seismic data, including features at different scales, without human intervention, thus uncovering richer and more representative feature information. Furthermore, it exhibits data adaptability. When faced with seismic data from different regions and of different types, traditional machine learning methods may require readjusting feature extraction methods and model parameters. However, the network in this invention, trained on a large amount of data, has stronger generalization capabilities and can adapt well to different seismic data, reducing the complexity of model adjustments for different data.

[0061] According to the present invention, evaluating the training model and adjusting and optimizing the training model based on the evaluation results to obtain the optimal training model includes:

[0062] The trained model was evaluated using an independent test dataset, and the network’s performance in identifying faults in seismic data was measured by calculating accuracy, recall, and F1-score.

[0063] Based on the evaluation results, the network structure and hyperparameters are adjusted and optimized to obtain the optimal training model.

[0064] Compared to other deep learning architectures, this invention offers advantages in its U-shaped structure and skip connections: Early convolutional neural networks may have limitations when dealing with problems like seismic data that require simultaneous consideration of global and local features. However, the U-shaped structure and skip connections in this method's U-NET architecture effectively fuse shallow local detail features with deep global semantic features. In seismic fault identification, shallow features help determine the precise boundaries of faults, while deep features help grasp the overall strike and distribution of faults. This fusion mechanism makes fault identification more accurate. Furthermore, it possesses multi-scale feature processing capabilities: Early networks may not be flexible enough in handling multi-scale features. This method's U-NET architecture, through encoder downsampling and decoder upsampling processes combined with feature fusion at different levels, can better handle fault features at different scales in seismic data. Small-scale fault details can be captured in the shallow layer, while large-scale fault strikes are extracted in the deep layer, and these features are well integrated and processed within the network.

[0065] The present invention will be described in more detail below through embodiments.

[0066] Example 1:

[0067] like Figure 1 As shown, this embodiment proposes a U-NET fracture identification method based on Transformer enhancement, which is aimed at accurate and efficient identification of fractures under complex structural conditions. The identification results are closer to the manual fracture interpretation style, that is, the fracture ductility on the cross section and plane is more consistent with the actual fracture direction. The results can be directly input into the automatic fault identification system to automatically generate a fault interpretation file that can be used directly.

[0068] The specific steps of this method are as follows:

[0069] I. Data Preparation Stage:

[0070] (1) Data collection:

[0071] Collecting raw 3D seismic data volumes allows for pre-processing of seismic data interpretation; to facilitate network processing, these volumes need to be converted into 2D slices (isochronous / depth slices and seismic profiles).

[0072] (2) Label creation:

[0073] By having professional geologists interpret artificial faults, the cross-sections and the results of the artificial fault interpretations are converted into image files, which serve as fault pattern recognition samples, thus forming the labeled data required for supervised learning. These labels need to clearly indicate which areas in the slice contain faults.

[0074] II. Model Training Phase:

[0075] (1) Input data:

[0076] Two-dimensional seismic data slices are input into a network based on the U-NET architecture;

[0077] (2) Feature extraction (encoder part):

[0078] The network's encoder begins to work, extracting features from the input data through a series of convolutional (3*3) operations; with each convolutional operation, the network extracts more abstract and representative features.

[0079] Then, max pooling is performed. The pooling window is usually 2*2. While reducing the data dimensionality, it retains the main feature information and helps the network gradually focus on more critical features.

[0080] (3) Feature fusion (cross-layer connection):

[0081] During the downsampling process of the network, features from different levels are fused through cross-layer connections. This feature fusion mechanism enables the network to comprehensively utilize feature information from different levels of abstraction, avoiding excessive information loss during dimensionality reduction.

[0082] (4) Upsampling and feature processing (decoder part):

[0083] In the decoder stage, the dimensions of the data are gradually restored through upsampling operations. Upsampling typically employs methods such as deconvolution or interpolation to make the data size close to the original input data.

[0084] After upsampling, the features are further adjusted and processed by convolution (1x1) to make the feature information more accurate and targeted.

[0085] (5) Loss calculation and parameter optimization:

[0086] The network output is compared with manually labeled tags. Typically, loss functions suitable for classification problems, such as cross-entropy loss function, are used to calculate the difference between the predicted results and the true labels.

[0087] Based on the loss value, optimization algorithms (such as stochastic gradient descent and its variants) are used to backpropagate and update the network parameters, continuously adjusting the network's weights and biases so that the network's prediction results become closer and closer to the actual fault situation.

[0088] III. Model Evaluation and Application Phase:

[0089] (1) Model evaluation:

[0090] The trained network (model) was evaluated using an independent test dataset, and its performance in identifying faults in seismic data was measured by calculating metrics such as accuracy, recall, and F1-score.

[0091] The network is adjusted and optimized based on the evaluation results, such as adjusting the network structure and hyperparameters, until satisfactory performance is achieved.

[0092] (2) Practical application:

[0093] The evaluated and optimized network (model) is applied to actual seismic data fault identification work to quickly and accurately identify faults in new seismic data.

[0094] In this embodiment, Figure 2 This is a schematic diagram of the network structure of the U-NET architecture; Figure 3 Samples for tomographic pattern recognition; Figure 4 Analyze the training execution curve for AI; Figure 5 Train network model parameters for AI; such as Figure 6 Fault prediction profile and Figure 7 As can be seen from the fault prediction plan view, the identification results of the method in this embodiment are closer to the manual fracture interpretation style, that is, the fracture extension on the cross section and plane is more consistent with the actual fracture direction. The fracture detection results can be directly input into the automatic fault identification system to automatically generate a fault interpretation file that can be used directly.

[0095] Example 2:

[0096] This embodiment provides a Transformer-enhanced U-NET fracture identification method, such as... Figure 1 As shown, it includes:

[0097] Artificial fault interpretation was performed on the raw 3D seismic data to obtain fault pattern recognition samples;

[0098] The tomographic pattern recognition samples are input into the U-NET architecture network for feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization to obtain the training model.

[0099] The training model is evaluated, and the model is adjusted and optimized based on the evaluation results to obtain the optimal training model.

[0100] The optimal training model is applied to actual seismic data fault identification work to identify faults in new seismic data.

[0101] In this embodiment, the original 3D seismic data is subjected to artificial fault interpretation to obtain fault pattern recognition samples, including:

[0102] The original 3D seismic data is converted into 2D slices, namely isochronous / depth slices and seismic profiles;

[0103] Artificial fault interpretation is performed on two-dimensional slices, and the results of seismic profiles and artificial fault interpretation are converted into image files;

[0104] Use image files as samples for tomographic pattern recognition;

[0105] In this embodiment, feature extraction includes:

[0106] In the encoder of the network, features are extracted from the input data through a series of convolution operations. With each convolution operation, the network extracts more abstract and representative features.

[0107] Max pooling is used to reduce data dimensionality while retaining key feature information, helping the network to gradually focus on more critical features.

[0108] In this embodiment, feature fusion, i.e., cross-layer connectivity, includes:

[0109] During the downsampling process of the network, features at different levels are fused to enable the network to comprehensively utilize feature information at different levels of abstraction, thus avoiding excessive information loss during dimensionality reduction.

[0110] In this embodiment, upsampling and feature processing include:

[0111] In the network's decoder, the dimensions of the data are gradually restored through upsampling operations to make the data size closer to the original input data.

[0112] After upsampling, the features are further adjusted and processed by convolution operations to make the feature information more accurate and targeted.

[0113] The upsampling method used is either deconvolution or interpolation.

[0114] In this embodiment, loss calculation and parameter optimization include:

[0115] The network output is compared with manually labeled labels, and the cross-entropy loss function is used to calculate the difference between the prediction results and the true labels to obtain the loss value;

[0116] Based on the loss value, stochastic gradient descent and its variants are used to backpropagate and update the network parameters, continuously adjusting the network weights and biases to make the network's prediction results increasingly closer to the actual fault situation.

[0117] In this embodiment, evaluating the training model and adjusting and optimizing it based on the evaluation results to obtain the optimal training model includes:

[0118] The trained model was evaluated using an independent test dataset, and the network’s performance in identifying faults in seismic data was measured by calculating accuracy, recall, and F1-score.

[0119] Based on the evaluation results, the network structure and hyperparameters are adjusted and optimized to obtain the optimal training model.

[0120] Example 3:

[0121] This embodiment provides a Transformer-enhanced U-NET fracture detection device, including:

[0122] The fault interpretation module is used to perform artificial fault interpretation on the raw 3D seismic data to obtain fault pattern recognition samples.

[0123] The model training module is used to input tomographic pattern recognition samples into the U-NET architecture network for feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization to obtain the trained model.

[0124] The model evaluation module is used to evaluate the trained model and adjust and optimize it based on the evaluation results to obtain the optimal trained model.

[0125] The model application module is used to apply the optimal training model to actual seismic data fault identification work, and to identify faults in new seismic data.

[0126] In this embodiment, the original 3D seismic data is subjected to artificial fault interpretation to obtain fault pattern recognition samples, including:

[0127] The original 3D seismic data is converted into 2D slices, namely isochronous / depth slices and seismic profiles;

[0128] Artificial fault interpretation is performed on two-dimensional slices, and the results of seismic profiles and artificial fault interpretation are converted into image files;

[0129] Use image files as samples for tomographic pattern recognition;

[0130] In this embodiment, feature extraction includes:

[0131] In the encoder of the network, features are extracted from the input data through a series of convolution operations. With each convolution operation, the network extracts more abstract and representative features.

[0132] Max pooling is used to reduce data dimensionality while retaining key feature information, helping the network to gradually focus on more critical features.

[0133] In this embodiment, feature fusion, i.e., cross-layer connectivity, includes:

[0134] During the downsampling process of the network, features at different levels are fused to enable the network to comprehensively utilize feature information at different levels of abstraction, thus avoiding excessive information loss during dimensionality reduction.

[0135] In this embodiment, upsampling and feature processing include:

[0136] In the network's decoder, the dimensions of the data are gradually restored through upsampling operations to make the data size closer to the original input data.

[0137] After upsampling, the features are further adjusted and processed by convolution operations to make the feature information more accurate and targeted.

[0138] The upsampling method used is either deconvolution or interpolation.

[0139] In this embodiment, loss calculation and parameter optimization include:

[0140] The network output is compared with manually labeled labels, and the cross-entropy loss function is used to calculate the difference between the prediction results and the true labels to obtain the loss value;

[0141] Based on the loss value, stochastic gradient descent and its variants are used to backpropagate and update the network parameters, continuously adjusting the network weights and biases to make the network's prediction results increasingly closer to the actual fault situation.

[0142] In this embodiment, evaluating the training model and adjusting and optimizing it based on the evaluation results to obtain the optimal training model includes:

[0143] The trained model was evaluated using an independent test dataset, and the network’s performance in identifying faults in seismic data was measured by calculating accuracy, recall, and F1-score.

[0144] Based on the evaluation results, the network structure and hyperparameters are adjusted and optimized to obtain the optimal training model.

[0145] Example 4:

[0146] This invention provides an electronic device including a memory and a processor, comprising:

[0147] Memory, which stores executable instructions;

[0148] The processor executes executable instructions in memory to implement a Transformer-enhanced U-NET fracture detection method.

[0149] This memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0150] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of the invention, the processor is used to execute computer-readable instructions stored in the memory.

[0151] Those skilled in the art should understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this invention.

[0152] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0153] Example 5:

[0154] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a Transformer-enhanced U-NET fracture identification method.

[0155] A computer-readable storage medium according to embodiments of the present invention stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the methods described in the foregoing embodiments of the present invention are performed.

[0156] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0157] The U-NET fault identification method based on Transformer enhancement proposed in the embodiments of the present invention can accurately identify fault locations in seismic data after the network has been fully trained. In particular, for relatively obvious faults, its identification accuracy can reach a high level, clearly delineating the strike and extent of the fault and reducing misjudgments and omissions. For seismic data with multiple intersecting faults, irregular fault morphology, and noise interference, the characteristics of different faults can be distinguished through the network's multi-level feature extraction and fusion mechanism, thereby accurately identifying each fault.

[0158] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A U-NET fracture identification method based on Transformer enhancement, characterized in that, include: Artificial fault interpretation was performed on the raw 3D seismic data to obtain fault pattern recognition samples; The tomographic pattern recognition samples are input into a U-NET architecture network for feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization, to obtain a training model. The training model is evaluated, and the training model is adjusted and optimized based on the evaluation results to obtain the optimal training model; The optimal training model is then applied to actual seismic data fault identification work to identify faults in new seismic data.

2. The method according to claim 1, characterized in that, Artificial fault interpretation was performed on the raw 3D seismic data, resulting in fault pattern recognition samples, including: The original 3D seismic data is converted into 2D slices, namely isochronous / depth slices and seismic profiles; Artificial fault interpretation is performed on the two-dimensional slices, and the results of the seismic profile and artificial fault interpretation are converted into image files; The image file was used as a sample for tomographic pattern recognition.

3. The method according to claim 1, characterized in that, The feature extraction includes: In the encoder of the network, features are extracted from the input data through a series of convolution operations, and with each convolution operation, the network extracts more abstract and representative features. Max pooling is used to reduce data dimensionality while retaining key feature information, helping the network gradually focus on more critical features.

4. The method according to claim 1, characterized in that, The feature fusion, i.e., cross-layer connectivity, includes: During the downsampling process of the network, features at different levels are fused to enable the network to comprehensively utilize feature information at different levels of abstraction, thus avoiding excessive information loss during dimensionality reduction.

5. The method according to claim 1, characterized in that, The upsampling and feature processing include: In the decoder of the network, the dimensions of the data are gradually restored through upsampling operations to make the size of the data close to that of the original input data. After the upsampling, the features are further adjusted and processed by convolution operations to make the feature information more accurate and targeted. The upsampling method used is either deconvolution or interpolation.

6. The method according to claim 1, characterized in that, The loss calculation and parameter optimization include: The output of the network is compared with the manually labeled labels, and the difference between the prediction result and the true label is calculated using the cross-entropy loss function to obtain the loss value; Based on the loss value, stochastic gradient descent and its variants are used to backpropagate and update the network parameters, continuously adjusting the network's weights and biases to make the network's prediction results increasingly closer to the actual fault situation.

7. The method according to claim 1, characterized in that, Evaluating the training model and adjusting and optimizing it based on the evaluation results to obtain the optimal training model includes: The trained model was evaluated using an independent test dataset, and the network's performance in identifying faults in seismic data was measured by calculating accuracy, recall, and F1-score. Based on the evaluation results, the structure and hyperparameters of the network are adjusted and optimized to obtain the optimal training model.

8. A U-NET fracture detection device based on Transformer enhancement, characterized in that, include: The fault interpretation module is used to perform artificial fault interpretation on the raw 3D seismic data to obtain fault pattern recognition samples. The model training module is used to input the tomographic pattern recognition samples into the U-NET architecture network, perform feature extraction, feature fusion, upsampling and feature processing, as well as loss calculation and parameter optimization, to obtain the trained model. The model evaluation module is used to evaluate the training model and adjust and optimize the training model based on the evaluation results to obtain the optimal training model. The model application module is used to apply the optimal training model to actual seismic data fault identification work to identify faults in new seismic data.

9. An electronic device, characterized in that, The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the Transformer-enhanced U-NET fracture identification method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the Transformer-enhanced U-NET fracture identification method according to any one of claims 1-7.

Images