Seismic fault recognition model training method, related method and device

Abstract

The invention discloses a seismic fault recognition model training method, a related method and a device. The method comprises the following steps: acquiring a seismic fault data set; obtaining an initial seismic fault recognition model; the initial seismic fault identification model comprises an encoder and a decoder; the encoder comprises a MobileNetV2 module, an ASPP module and an attention mechanism layer; the seismic fault data are input into the MobileNetV2 module, and hidden features are obtained; according to the hidden features, performing calculation to obtain four different-scale feature representations including a shallow feature, a first intermediate feature, a second intermediate feature and a deep feature; inputting the hidden features into an ASPP module and an attention mechanism layer to obtain multi-scale weight features; inputting the four different-scale feature representations and the multi-scale weight features into a decoder to obtain a fault prediction result; updating the initial seismic fault recognition model according to the fault prediction result; and repeating the model training process until a preset condition is reached, and obtaining the seismic fault identification model. According to the method, the robustness to complex conditions and noise interference can be improved.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a method, related methods and apparatus for training a seismic fault identification model. Background Technology

[0002] Traditional fault interpretation relies on interpreters' visual judgment of discontinuities in the phase axes of reflected waves on seismic profiles to identify faults. This method is time-consuming, labor-intensive, and susceptible to the subjective judgment of interpreters. For large amounts of seismic data requiring fault interpretation, and for tiny faults that are difficult to distinguish with the naked eye, manual interpretation is insufficient.

[0003] With the continuous advancement of computer technology, tomographic identification technology has also been promoted, and various tomographic identification techniques have emerged to meet production needs. These include techniques for identifying tomographic enhancement attributes such as coherence volume properties, variance properties, and curvature properties, as well as automatic tomographic tracking and interpretation methods such as ant tracking technology.

[0004] The principle of coherence volume technology is to calculate the similarity between seismic traces, thereby converting data with amplitude attributes into similarity coefficient data and obtaining the coherence attributes of the original amplitude data. By calculating the correlation values ​​at each grid point on a certain time slice, the contours of low correlation values ​​along the fault can be obtained. These low correlation value contours become the fault sections. Similarly, stratigraphic interfaces and discontinuities of special geological bodies can also produce similar low correlation value contours.

[0005] Through continuous progress and development, coherence volume technology has evolved to the third generation of coherence volume algorithms. Recognition techniques enhanced with attributes such as variance and curvature have overcome the challenges of identifying underlying tilt, undulations, minute structures, and faults. Summary of the Invention

[0006] To achieve more accurate earthquake fault identification, this invention provides an earthquake fault identification model training method, related methods, and apparatus.

[0007] In a first aspect, embodiments of the present invention provide a method for training an earthquake fault identification model, which may include:

[0008] Obtain an earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and a label corresponding to each earthquake fault data.

[0009] An initial seismic fault identification model is obtained; the initial seismic fault identification model includes an encoder and a decoder; the encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0010] The seismic fault data is input into the MobileNetV2 module to obtain hidden features;

[0011] Based on the hidden features, shallow features, first intermediate features, second intermediate features, and deep features are calculated;

[0012] The hidden features are input into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features;

[0013] The shallow features, first intermediate features, second intermediate features, deep features, and multi-scale weighted features are input into the decoder to obtain the predicted tomography result.

[0014] Based on the labels corresponding to the earthquake fault identification data and the predicted fault results, the initial earthquake fault identification model is updated based on a preset loss function.

[0015] Repeat the above model training process until the preset conditions are met to obtain the earthquake fault identification model.

[0016] In one or more optional embodiments of this application, the ASPP module includes a convolutional layer, a global pooling layer, and multiple dilated convolutional layers with different dilation rates.

[0017] The process of inputting the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features includes:

[0018] The hidden features are input into the convolutional layer, the global pooling layer, and multiple dilated convolutional layers with different dilation rates to obtain features of different scales.

[0019] The features at multiple different scales are concatenated and input into the attention mechanism layer to obtain multi-scale weighted features.

[0020] In one or more optional embodiments of this application, the step of inputting the shallow features, first intermediate features, second intermediate features, deep features, and multi-scale weighted features into the decoder to obtain the predicted tomography result includes:

[0021] The first intermediate feature is concatenated with the multi-scale weighted feature, and a high-level concatenated feature is obtained through convolution operation.

[0022] The shallow features and the second intermediate features are concatenated, and a low-level concatenated feature is obtained by convolution operation.

[0023] Convolution operations are performed based on the deep features, and then concatenated with the high-level concatenation features to obtain intermediate concatenation features;

[0024] Convolution and upsampling operations are performed based on the intermediate splicing features, and then spliced ​​with the low-level splicing features to obtain multi-scale splicing features;

[0025] The multi-scale spliced ​​features are subjected to convolution and upsampling operations to obtain the predicted tomographic results.

[0026] In one or more optional embodiments of this application, the step of calculating the shallow feature, the first intermediate feature, the second intermediate feature, and the deep feature based on the hidden feature includes:

[0027] Max pooling is performed based on the hidden features to obtain the shallow features;

[0028] Based on the hidden features, downsampling convolution and convolution operations are performed to obtain the first intermediate features;

[0029] Based on the hidden features, a second intermediate feature is obtained through convolutional block operations;

[0030] Multiple downsampling convolution operations are performed based on the hidden features to obtain deep features.

[0031] In one or more optional embodiments of this application, the seismic fault dataset is obtained in the following manner:

[0032] Obtain multiple 3D horizontal reflection models;

[0033] For each three-dimensional horizontal reflection model, a fold structure is added to obtain a folded reflection model;

[0034] Vertical displacement is performed based on the aforementioned folded reflection model to obtain a displacement reflection model;

[0035] Based on the preset fault displacement range, the displacement reflection model is modified to obtain the fault model;

[0036] Based on the fault model, convolution calculations are performed and random noise is added to obtain the earthquake fault model;

[0037] Based on the earthquake fault model, multiple earthquake fault data and the corresponding labels for each stratum fault data were extracted to obtain the earthquake fault dataset.

[0038] In one or more optional embodiments of this application, the step of adding wrinkle structures to each three-dimensional horizontal reflection model to obtain a wrinkled reflection model includes:

[0039] For each three-dimensional horizontal reflection model, the fold construction parameters are calculated based on the following formula:

[0040]

[0041] In the formula, S1(x,y,z) represents the fold construction parameters, x, y, and z represent the horizontal, vertical, and depth dimensions in the three-dimensional horizontal reflection model, respectively, and z... maxIndicates the maximum depth of the strata, N represents the number of strata in the vertical direction, a0, b k e k d k σ k These are all parameters that control the fold structure parameters;

[0042] Based on the three-dimensional horizontal reflection model and the fold construction parameters, the fold reflection model is calculated using the following formula:

[0043] r(x,y,z)=r(x,y,z+S1)

[0044] In the formula, r(x,y,z) represents the folded reflection model, x, y, and z represent the horizontal, vertical, and depth dimensions in the three-dimensional horizontal reflection model and the folded reflection model, respectively, and S1 represents the fold construction parameters.

[0045] In one or more optional embodiments of this application, the step of performing vertical displacement based on the wrinkled reflection model to obtain a displacement reflection model includes:

[0046] Based on the aforementioned folded reflection model, the vertical displacement parameters are calculated using the following formula:

[0047] S2(x,y,z)=e0+fx+gy

[0048] In the formula, S2(x,y,z) represents the vertical displacement parameter, x, y, and z represent the horizontal, vertical, and depth dimensions in the fold reflection model, respectively, and e0, f, and g are all parameters that control the fold construction variables.

[0049] Based on the wrinkle reflection model and the vertical displacement parameters, the displacement reflection model is calculated using the following formula:

[0050] r(x,y,z)=r(x,y,z+S2)

[0051] In the formula, r(x,y,z) represents the displacement reflection model, x, y, and z represent the horizontal, vertical, and depth dimensions in the fold reflection model and the displacement reflection model, respectively, and S2 represents the vertical displacement parameter.

[0052] Secondly, embodiments of the present invention provide a fault identification method based on seismic data. The seismic fault identification model obtained using the above-described seismic fault identification model training method may include:

[0053] Acquire earthquake fault data to be identified;

[0054] The earthquake fault data to be identified is input into the earthquake fault identification model to obtain the predicted fault results.

[0055] Thirdly, embodiments of the present invention provide a seismic fault identification model training device, which may include:

[0056] The first acquisition module is used to acquire an earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and a label corresponding to each earthquake fault data.

[0057] The second acquisition module is used to acquire an initial earthquake fault identification model; the initial earthquake fault identification model includes an encoder and a decoder; the encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0058] The first processing module is used to input the earthquake fault data into the MobileNetV2 module to obtain hidden features;

[0059] The second calculation module is used to calculate the shallow features, the first intermediate features, the second intermediate features, and the deep features based on the hidden features.

[0060] The third computation module is used to input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features;

[0061] The fourth computation module is used to input the shallow features, the first intermediate features, the second intermediate features, the deep features, and the multi-scale weighted features into the decoder to obtain the predicted tomography results;

[0062] The first update module is used to update the initial earthquake fault identification model based on the labels corresponding to the earthquake fault identification data and the predicted fault results, using a preset loss function.

[0063] The first judgment module is used to determine whether the updated initial seismic fault identification model meets the preset conditions: if yes, the seismic fault identification model is obtained; if no, the first calculation module re-executes the process of inputting the seismic fault data into the MobileNetV2 module to obtain the hidden features.

[0064] Fourthly, embodiments of the present invention provide a fault identification device based on seismic data, which may include:

[0065] The first acquisition module is used to acquire an earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and a label corresponding to each earthquake fault data.

[0066] The second acquisition module is used to acquire an initial earthquake fault identification model; the initial earthquake fault identification model includes an encoder and a decoder; the encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0067] The first processing module is used to input the earthquake fault data into the MobileNetV2 module to obtain hidden features;

[0068] The second calculation module is used to calculate the shallow features, the first intermediate features, the second intermediate features, and the deep features based on the hidden features.

[0069] The third computation module is used to input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features;

[0070] The fourth computation module is used to input the shallow features, the first intermediate features, the second intermediate features, the deep features, and the multi-scale weighted features into the decoder to obtain the predicted tomography results;

[0071] The first update module is used to update the initial earthquake fault identification model based on the labels corresponding to the earthquake fault identification data and the predicted fault results, using a preset loss function.

[0072] The first judgment module is used to determine whether the updated initial earthquake fault identification model meets the preset conditions: if yes, the earthquake fault identification model is obtained; if no, the first calculation module re-executes the process of inputting the earthquake fault data into the MobileNetV2 module to obtain the hidden features.

[0073] The third acquisition module is used to acquire earthquake fault data to be identified.

[0074] The first prediction module is used to input the earthquake fault data to be identified into the earthquake fault identification model to obtain the predicted fault results.

[0075] Fifthly, embodiments of the present invention provide a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the earthquake fault identification model training method as described above, and / or, a fault identification method based on earthquake data.

[0076] In a sixth aspect, embodiments of the present invention provide a computer program product, including a computer program / instruction that, when executed by a processor, implements the earthquake fault identification model training method as described above, and / or a fault identification method based on earthquake data.

[0077] In a seventh aspect, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it implements the earthquake fault identification model training method as described above, and / or, a fault identification method based on earthquake data.

[0078] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0079] This invention provides a method for training an earthquake fault identification model. The method constructs an earthquake fault identification model based on a MobileNetV2 module, an ASPP module, and an attention mechanism layer. In the earthquake fault identification model, the hidden features output by the MobileNetV2 module are used to calculate four different scale feature representations: shallow features, first intermediate features, second intermediate features, and deep features. Simultaneously, the ASPP module obtains multi-scale weighted features, which are then fused with the four different scale feature representations in the decoder to obtain the predicted fault result, thus training the earthquake fault identification model. This method, by introducing the MobileNetV2 module and the ASPP module, extracts multi-scale feature representations of earthquake fault data in two different ways and fuses them in the decoder. This effectively improves the generalization ability and robustness to complex situations and noise interference of the earthquake fault identification model, significantly improving its identification accuracy. The model not only depicts the overall fault situation but also provides more detailed depiction of minor faults, offering significant guidance in practical work. It enables automatic and efficient seismic fault identification, which can better serve the efficient interpretation of seismic data and provide effective guidance and assistance for oil and gas exploration.

[0080] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0081] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0082] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0083] Figure 1 This is a schematic diagram illustrating the steps of the earthquake fault identification model training method provided in an embodiment of the present invention;

[0084] Figure 2 A schematic diagram of a three-dimensional horizontal reflection model provided in an embodiment of the present invention;

[0085] Figure 3 A schematic diagram of a wrinkle reflection model provided in an embodiment of the present invention;

[0086] Figure 4A schematic diagram of a displacement reflection model provided in an embodiment of the present invention;

[0087] Figure 5 A schematic diagram of a fault model provided in an embodiment of the present invention;

[0088] Figure 6 This is a schematic diagram of earthquake records provided in an embodiment of the present invention;

[0089] Figure 7 A schematic diagram of an earthquake fault model provided in an embodiment of the present invention;

[0090] Figure 8 This is a schematic diagram of the initial seismic fault identification model provided in an embodiment of the present invention;

[0091] Figure 9 This is a schematic diagram of the training error reduction curve of the earthquake fault identification model provided in an embodiment of the present invention;

[0092] Figure 10 This is a schematic diagram of the labels corresponding to the test earthquake fault identification data provided in an embodiment of the present invention;

[0093] Figure 11 A schematic diagram of the predicted fault results corresponding to the test earthquake fault identification data provided in an embodiment of the present invention;

[0094] Figure 12 This is a schematic diagram of test seismic fault data collected in a real work area, as provided in an embodiment of the present invention.

[0095] Figure 13 A schematic diagram of the predicted fault results corresponding to the test seismic fault data collected in the actual work area provided in this embodiment of the invention;

[0096] Figure 14 This is a schematic diagram illustrating the steps of the fault identification method based on seismic data provided in an embodiment of the present invention;

[0097] Figure 15 This is a schematic diagram of the structure of the earthquake fault identification model training device provided in the embodiments of this application;

[0098] Figure 16 This is a schematic diagram of the structure of a fault identification device based on seismic data provided in an embodiment of this application. Detailed Implementation

[0099] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0100] The inventors discovered that in existing technologies, traditional coherence properties and ant colony algorithms have gradually declined in the field of fault identification due to their respective drawbacks. The mainstream direction of earthquake fault identification has been led by deep learning methods, but they are still difficult to effectively identify faults when faced with complex data or noise interference.

[0101] Based on this, the inventors conducted further research and development and developed this invention, which provides a method for training an earthquake fault identification model, related methods and devices.

[0102] Example 1

[0103] Embodiment 1 of the present invention provides a method for training an earthquake fault identification model, referring to... Figure 1 As shown, the method may include the following steps S101-S108:

[0104] S101: Obtain the earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and the corresponding label for each earthquake fault data.

[0105] S102: Obtain the initial seismic fault identification model; the initial seismic fault identification model includes an encoder and a decoder. The encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0106] S103: Input the seismic fault data into the MobileNetV2 module to obtain the hidden features.

[0107] S104: Based on the hidden features, shallow features, first intermediate features, second intermediate features, and deep features are calculated.

[0108] S105: Input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features.

[0109] S106: Input the shallow features, first intermediate features, second intermediate features, deep features, and multi-scale weighted features into the decoder to obtain the predicted fault results.

[0110] S107: Based on the labels corresponding to the earthquake fault identification data and the predicted fault results, update the initial earthquake fault identification model according to the preset loss function.

[0111] S108: Determine whether the preset conditions have been met: If yes, proceed to step S109; if no, re-execute steps S103-S107.

[0112] S109: Obtain the earthquake fault identification model.

[0113] This invention provides a method for training an earthquake fault identification model. The method constructs an earthquake fault identification model based on a MobileNetV2 module, an ASPP module, and an attention mechanism layer. In the earthquake fault identification model, the hidden features output by the MobileNetV2 module are used to calculate four different scale feature representations: shallow features, first intermediate features, second intermediate features, and deep features. Simultaneously, the ASPP module obtains multi-scale weighted features, which are then fused with the four different scale feature representations in the decoder to obtain the predicted fault result, thus training the earthquake fault identification model. This method introduces the MobileNetV2 module and the ASPP module, extracting multi-scale feature representations of earthquake fault data in two different ways and fusing them in the decoder. This effectively improves the generalization ability and robustness to complex situations and noise interference of the earthquake fault identification model, significantly improving its identification accuracy. This model not only depicts the overall fault situation but also provides more detailed depiction of micro-faults, offering significant guidance in practical work. It enables automatic and efficient seismic fault identification, which can better serve the efficient interpretation of seismic data and provide effective guidance and assistance for oil and gas exploration.

[0114] In step S101 above, the seismic fault dataset is obtained in the following manner, specifically including steps S1011-S1016:

[0115] S1011: Obtain multiple three-dimensional horizontal reflection models.

[0116] Specifically, this can be achieved by randomly generating multiple one-dimensional reflection coefficient sequences within a preset range of reflection coefficient values, and then extending each sequence to three dimensions to obtain multiple three-dimensional horizontal reflection models. This ensures that each horizontal layer in each three-dimensional horizontal reflection model has the same reflection coefficient.

[0117] For example, the preset range of reflection coefficient values ​​is [-1, 1]. A one-dimensional reflection coefficient sequence of 1×128 is randomly generated based on the preset range of reflection coefficient values. An initial three-dimensional matrix of 128×128×128 with an initial value of 0 is set. The one-dimensional reflection coefficient sequence is mapped onto the initial three-dimensional matrix to obtain a three-dimensional horizontal reflection model. All values ​​of each horizontal layer of the three-dimensional horizontal reflection model are equal to the corresponding values ​​in the one-dimensional reflection coefficient sequence.

[0118] For example, the three-dimensional horizontal reflection model obtained in step S1011 is as follows: Figure 2 As shown in the image, different shades of gray represent different values. Figure 2The diagram shows the distribution of values ​​in the three dimensions of x, y, and z. The two vertical planes are the yz plane (constant x) and the xz plane (constant y), and the plane below is the xy plane (constant z). It can be seen that all values ​​in each xy plane are the same in the three-dimensional horizontal reflection model, that is, all values ​​in each horizontal layer are the same.

[0119] S1012: For each three-dimensional horizontal reflection model, add a fold structure to obtain a folded reflection model.

[0120] Specifically, for each three-dimensional horizontal reflection model, the wrinkle construction parameters can be calculated based on the following formula 1:

[0121]

[0122] In the formula, S1(x,y,z) represents the fold construction parameters, x, y, and z represent the horizontal, vertical, and depth dimensions in the three-dimensional horizontal reflection model, respectively, and z... max Indicates the maximum depth of the strata, N represents the number of strata in the vertical direction, a0, b k e k d k σ k These are all parameters that control the fold structure.

[0123] Next, based on the three-dimensional horizontal reflection model and the fold construction parameters, the fold reflection model is calculated using the following formula 2:

[0124] r(x,y,z)=r(x,y,z+S1) Formula 2

[0125] In the formula, r(x,y,z) represents the folded reflection model, x, y, and z represent the horizontal, vertical, and depth dimensions in the three-dimensional horizontal reflection model and the folded reflection model, respectively, and S1 represents the fold construction parameters.

[0126] For example, the three-dimensional horizontal reflection model obtained in step S1011 is used to execute the wrinkled reflection model obtained in step S1012, such as... Figure 3 As shown in the figure, the horizontal curve in the upper part represents the cross section of the reflective layer. It can be seen that compared with the three-dimensional horizontal reflective model, the folded reflective model shows obvious folds in each reflective layer. The three-dimensional horizontal reflective model accurately simulates the fold morphology in the geological structure.

[0127] S1013: Vertical displacement is performed based on the fold reflection model to obtain the displacement reflection model.

[0128] Specifically, the vertical displacement parameters can be calculated based on the folded reflection model and the following formula 3:

[0129] S2(x,y,z)=e0+fx+gy Formula 3

[0130] In the formula, S2(x,y,z) represents the vertical displacement parameter, x, y, and z represent the horizontal, vertical, and depth dimensions in the fold reflection model, respectively, and e0, f, and g are parameters that control the fold construction variables.

[0131] Based on the fold reflection model and vertical displacement parameters, the displacement reflection model is calculated using the following formula 4:

[0132] r(x,y,z)=r(x,y,z+S2) Formula 4

[0133] In the formula, r(x,y,z) represents the displacement reflection model, x, y, and z represent the horizontal, vertical, and depth dimensions in the fold reflection model and the displacement reflection model, respectively, and S2 represents the vertical displacement parameter.

[0134] For example, the wrinkle reflection model obtained in step S1012 is used to execute the displacement reflection model obtained in step S1013, such as... Figure 4 As shown, compared with the folded reflection model, the reflection layer in the displacement reflection model exhibits parallel movement, accurately simulating the vertical displacement of the reflection layer in the folded reflection model, thus forming planar shear.

[0135] S1014: Modify the displacement reflection model based on the preset fault displacement range to obtain the fault model.

[0136] Specifically, it can be done by determining the location of the fault and the points on the fault based on the fault displacement model, randomly selecting the maximum displacement of each point on the fault within a preset fault displacement range, forming the fault in the displacement reflection model, and thus obtaining the fault model.

[0137] For example, the displacement reflection model obtained in step S1013 is used to execute the fault model obtained in step S1014, such as... Figure 5 As shown, a clear fault can be seen.

[0138] S1015: Based on the fault model, perform convolution calculations and add random noise to obtain the earthquake fault model.

[0139] Specifically, it could be that seismic records are obtained based on the fault model and the seismic wavelet convolution, and then random noise is added to the seismic records to obtain the seismic fault model.

[0140] For example, the seismic record obtained by convolution using the fault model obtained in step S1014, such as... Figure 6 As shown, the earthquake fault model obtained by adding random noise is as follows: Figure 7 As shown, the fault model is simulated as a simulated seismic record, and noise is added to increase the diversity and complexity of the seismic fault model.

[0141] S1016: Based on the earthquake fault model, extract multiple earthquake fault data and the corresponding labels for each stratum fault data to obtain the earthquake fault dataset.

[0142] Specifically, this could involve extracting multiple seismic fault data points from a seismic fault model. For example, based on a 128×128×128 seismic fault model, the yz plane could be extracted as seismic fault data every 10 pixels along the x-axis, meaning a total of 13 seismic fault data points could be extracted from this model.

[0143] Meanwhile, since a fault was added to the fold reflection model in steps S1013-S1014, a fault model was obtained. Therefore, based on the maximum displacement of each point on the fault in steps S1013-S1014, the location of the fault in the earthquake fault model can be determined, and thus the location of the fault in the earthquake fault data can be determined. The fault in the earthquake fault data is marked as 1, and the rest of the area is marked as 0, thus obtaining the corresponding label for the earthquake fault data and obtaining the earthquake fault dataset.

[0144] In this embodiment, the seismic fault dataset obtained in step S101, due to its randomness and complexity, contains reflection information of various complex geological structures that can be collected during seismic exploration. It serves as a high-quality training sample and can effectively simulate various fault characteristics in real seismic environments. Using this seismic fault dataset as the training dataset for subsequent seismic fault identification models can effectively improve the model's generalization ability and robustness in complex and noisy environments.

[0145] In step S102 above, the initial earthquake fault identification model is constructed with reference to... Figure 8 As shown, it includes an encoder and a decoder, and the encoder includes a MobileNetV2 module, an ASPP module, and an attention layer.

[0146] In this embodiment of the application, after the initial seismic fault identification model is constructed, data preparation and parameter setting are also required before model training.

[0147] In one specific embodiment, the earthquake fault dataset is divided into a training set and a validation set in a 9:1 ratio for subsequent model training. Then, the maximum number of training iterations (MAX_EPOCH) is set to 500, the batch size (BATCH_SIZE) is set to 64, and the learning rate (LR) is set to 0.0005.

[0148] In step S103 above, seismic fault data is input into the MobileNetV2 module of the initial seismic fault identification model to obtain hidden features. The MobileNetV2 module consists of a MibileVit module, seven Bottleneck modules, a pooling layer, and another MibileVit module arranged sequentially. Those skilled in the art can implement the construction of the MibileVit and Bottleneck modules based on detailed descriptions of existing technologies; further details are omitted here.

[0149] In this embodiment, the initial seismic fault identification model, by introducing inverted residual blocks and depth convolution, can greatly reduce the amount of computation and the number of parameters, thereby improving accuracy while maintaining low computational cost.

[0150] In step S104 above, based on the hidden features obtained in step S103, shallow features, first intermediate features, second intermediate features, and deep features are calculated, specifically including the following steps S1041-S1044:

[0151] S1041: Perform max pooling operation based on hidden features to obtain shallow features.

[0152] S1042: Perform downsampling convolution and convolution operations based on the hidden features to obtain the first intermediate feature.

[0153] S1043: Based on the hidden features, the second intermediate feature is obtained through convolutional block operations.

[0154] The convolutional block consists of a two-dimensional convolutional layer, a batch normalization layer, and the ReLU activation function.

[0155] S1044: Perform multiple downsampling convolution operations based on hidden features to obtain deep features.

[0156] In step S105 above, the hidden features obtained in step S103 are input into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features, specifically including the following steps S1051-S1052:

[0157] S1051: Input the hidden features into a 1×1 convolutional layer, a global pooling layer, and multiple dilated convolutional layers with different dilation rates to obtain features at multiple different scales.

[0158] For example, multiple dilated convolutional layers with different dilation rates can be configured as three dilated convolutional layers with dilation rates of 6, 12, and 18, or three dilated convolutional layers with dilation rates of 1, 2, and 4.

[0159] S1052: Concatenate features of multiple different scales, input them into the attention mechanism layer and a 1×1 convolutional layer to obtain multi-scale weighted features.

[0160] In this embodiment, the initial seismic fault identification module uses dilated convolution through the ASPP module to obtain a larger receptive field without sacrificing image resolution or increasing computational cost, thereby improving computational efficiency while acquiring information at different scales. Simultaneously, an attention mechanism is introduced to weight the feature information, ensuring network accuracy. A 1×1 convolution is used to adjust the number of channels for features, facilitating subsequent calculations.

[0161] In step S106 above, the shallow features, the first intermediate features, the second intermediate features, the deep features, and the multi-scale weighted features are input into the decoder to obtain the predicted tomography result, specifically including the following steps S1061-S1065:

[0162] S1061: Concatenate the first intermediate feature with the multi-scale weighted feature, and obtain the advanced concatenated feature through a 1×1 convolution operation;

[0163] S1062: Concatenate the shallow features and the second intermediate features, and obtain the low-level concatenated features through a 1×1 convolution operation;

[0164] S1063: Perform 1×1 convolution operation based on deep features and concatenate with high-level concatenation features to obtain intermediate concatenation features;

[0165] S1064: Perform 1×1 convolution operation and 4x upsampling operation based on intermediate splicing features, and splice them with low-level splicing features to obtain multi-scale splicing features;

[0166] S1065: Perform 3×3 convolution operation and 4x upsampling operation on the multi-scale spliced ​​features to obtain the predicted fault results.

[0167] In this embodiment, the 4x upsampling operation can be implemented using bilinear interpolation to further enrich the feature information extracted by the network.

[0168] To provide a clearer and more detailed explanation of the process described in S103-S106 above, which involves inputting seismic fault data into an initial seismic fault identification model to obtain predicted fault results, please refer to... Figure 8 As shown, the earthquake data and fault predictions correspond to the earthquake fault data and predicted fault results mentioned above, respectively.

[0169] based on Figure 8 As can be seen, the seismic fault data is input into the MobileNetV2 module in the upper left corner, corresponding to step S103.

[0170] Then, the MobileNetV2 module outputs four features: shallow features, mid-level features 1, mid-level features 2, and high-level features, corresponding to step S104.

[0171] Meanwhile, the MobileNetV2 module outputs to the right into the ASPP module, the attention mechanism layer, and a 1×1 convolutional layer, resulting in... Figure 8 The green cube shown is the multi-scale weighted feature, corresponding to step S105.

[0172] Next, the shallow features, first intermediate features, second intermediate features, deep features, and multi-scale weighted features output by the MobileNetV2 module are input into the decoder to obtain the predicted tomographic results, corresponding to step S106.

[0173] In step S107 above, based on a preset loss function (such as cross-entropy loss, mean square error, etc.), the loss function values ​​of the labels corresponding to the seismic fault identification data and the predicted fault results are calculated.

[0174] Based on the backpropagation algorithm and gradient descent algorithm, the parameters in the initial seismic fault identification model are updated according to the loss function value to obtain the updated seismic fault identification model.

[0175] In step S108 above, based on the updated earthquake fault identification model, steps S103-S107 above are re-executed to iteratively train the earthquake fault identification model until the preset conditions are met, and the earthquake fault identification model is obtained.

[0176] The preset conditions can be set to reach a preset number of iterations, or the accuracy reaches a preset target value, etc.

[0177] In one specific embodiment, the training environment configuration includes: a 12th generation Intel Core i7-12700K processor with a base frequency of 3.60GHz and 12 cores; 64GB of system memory; a 64-bit operating system based on the x64 architecture to fully utilize the processor's performance; an NVIDIA GeForce RTX3090 graphics card with 24GB of video memory, providing powerful graphics computing capabilities; and on the software side, PyCharm is used as the development environment, and model training is performed based on the PyTorch neural network framework.

[0178] Based on the above training environment, the training parameters were set as follows: maximum number of training iterations (MAX_EPOCH) = 500, batch size (BATCH_SIZE) = 64, and learning error (LR) = 0.0005. Model training was performed using these parameters, taking 1 hour and 31 minutes to complete the training of 200 seismic fault data points.

[0179] The training error reduction curve of the earthquake fault identification model is as follows: Figure 9 As shown in the figure, when the number of iterations is less than 200, the curve drops significantly, indicating that the error is relatively large at the beginning of the training process. By continuously optimizing the parameters of the earthquake fault identification model, the error tends to stabilize and is basically maintained at around 0.0008.

[0180] In this embodiment of the application, after training the earthquake fault identification model is completed, the performance of the earthquake fault identification model can be tested. The performance of the earthquake fault identification model is tested using artificially synthesized earthquake fault data obtained in step S101 and earthquake fault data collected from the actual work area, respectively.

[0181] Tests based on artificially synthesized seismic fault data yield the following results: Figure 10 and Figure 11 As shown, Figure 10 This is a schematic diagram showing the labels corresponding to the test earthquake fault identification data. Figure 11 To test the predicted fault results corresponding to the seismic fault identification data, and to compare them... Figure 10 and Figure 11 It is evident that the fault region in the predicted fault results is larger than the actual fault region. This is because the probability of each pixel in the predicted fault result is between [0,1]. When displaying the results, the probability of each pixel in the predicted fault results is processed, and only pixels with a probability greater than 0.5 are displayed. After eliminating this difference, it can be seen that the earthquake fault identification model has very accurately identified the fault region in the test earthquake fault identification data, and the identification result is clearly visible.

[0182] Tests were conducted based on seismic fault data collected from the actual work area, and the test results are as follows: Figure 12 and Figure 13 As shown, Figure 12 This is a schematic diagram of test seismic fault data collected from the actual work area. Figure 13 For based on Figure 12 The obtained predicted fault results, combined with Figure 12 and Figure 13 visible, Figure 12 The faults in the middle are stepped, with visible discontinuities in the same phase axis and low-order faults developing around large faults. Figure 13As shown in the predicted fault results, the earthquake fault identification model has a good identification effect on both typical faults and micro-faults in the face of such complex situations. It has high identification accuracy, not only depicting the overall situation of the fault, but also providing more detailed descriptions of micro-faults. It has great guiding significance in practical work.

[0183] Example 2

[0184] Based on the same inventive concept, this invention also provides a fault identification method based on seismic data, using the seismic fault identification model training method described in Embodiment 1 to obtain the seismic fault identification model, referring to... Figure 14 As shown, the method includes:

[0185] S201: Acquire earthquake fault data to be identified;

[0186] S202: Input the earthquake fault data to be identified into the earthquake fault identification model to obtain the predicted fault results.

[0187] Example 3

[0188] Based on the same inventive concept, embodiments of the present invention also provide a seismic fault identification model training device, referring to... Figure 15 As shown, the device includes:

[0189] The first acquisition module 101 is used to acquire an earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and a label corresponding to each earthquake fault data.

[0190] The second acquisition module 102 is used to acquire an initial earthquake fault identification model; the initial earthquake fault identification model includes an encoder and a decoder; the encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0191] The first calculation module 103 is used to input the earthquake fault data into the MobileNetV2 module to obtain hidden features;

[0192] The second calculation module 104 is used to calculate shallow features, first intermediate features, second intermediate features and deep features based on the hidden features;

[0193] The third operation module 105 is used to input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features;

[0194] The fourth operation module 106 is used to input the shallow features, the first intermediate features, the second intermediate features, the deep features and the multi-scale weighted features into the decoder to obtain the predicted tomography result;

[0195] The first update module 107 is used to update the initial earthquake fault identification model based on a preset loss function, according to the labels corresponding to the earthquake fault identification data and the predicted fault results.

[0196] The first judgment module 108 is used to determine whether the updated initial seismic fault identification model meets the preset conditions: if yes, the seismic fault identification model is obtained; if no, the first calculation module re-executes the process of inputting the seismic fault data into the MobileNetV2 module to obtain the hidden features.

[0197] Example 4

[0198] Based on the same inventive concept, embodiments of the present invention also provide a fault identification device based on seismic data, referring to... Figure 16 As shown, the device includes:

[0199] The first acquisition module 201 is used to acquire an earthquake fault dataset; the earthquake fault dataset includes multiple earthquake fault data and a label corresponding to each earthquake fault data.

[0200] The second acquisition module 202 is used to acquire an initial earthquake fault identification model; the initial earthquake fault identification model includes an encoder and a decoder; the encoder includes a MobileNetV2 module, an ASPP module, and an attention mechanism layer.

[0201] The first calculation module 203 is used to input the seismic fault data into the MobileNetV2 module to obtain hidden features;

[0202] The second calculation module 204 is used to calculate shallow features, first intermediate features, second intermediate features and deep features based on the hidden features;

[0203] The third operation module 205 is used to input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weighted features;

[0204] The fourth operation module 206 is used to input the shallow features, the first intermediate features, the second intermediate features, the deep features and the multi-scale weighted features into the decoder to obtain the predicted tomography result;

[0205] The first update module 207 is used to update the initial earthquake fault identification model based on a preset loss function, according to the labels corresponding to the earthquake fault identification data and the predicted fault results.

[0206] The first judgment module 208 is used to determine whether the updated initial seismic fault identification model meets the preset conditions: if yes, the seismic fault identification model is obtained; if no, the first calculation module re-executes the process of inputting the seismic fault data into the MobileNetV2 module to obtain the hidden features.

[0207] The third acquisition module 209 is used to acquire earthquake fault data to be identified;

[0208] The first prediction module 210 is used to input the earthquake fault data to be identified into the earthquake fault identification model to obtain the predicted fault result.

[0209] Example 5

[0210] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the earthquake fault identification model training method as described in Embodiment 1 above, and / or the fault identification method based on earthquake data as described in Embodiment 2 above.

[0211] Example 6

[0212] Based on the same inventive concept, embodiments of the present invention also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the earthquake fault identification model training method as described in Embodiment 1 above, and / or the fault identification method based on earthquake data as described in Embodiment 2 above.

[0213] Example 7

[0214] Based on the same inventive concept, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it implements the earthquake fault identification model training method as described in Embodiment 1 above, and / or the fault identification method based on earthquake data as described in Embodiment 2 above.

[0215] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0216] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0217] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0218] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0219] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for training a seismic fault identification model, characterized in that, The method comprises the following steps: obtaining a seismic fault data set; the seismic fault data set comprises a plurality of seismic fault data and a corresponding label for each seismic fault data; obtaining an initial seismic fault identification model; the initial seismic fault identification model comprises an encoder and a decoder; the encoder comprises a MobileNetV2 module, an ASPP module and an attention mechanism layer; inputting the seismic fault data into the MobileNetV2 module to obtain hidden features; according to the hidden features, calculating shallow features, first intermediate features, second intermediate features and deep features; inputting the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weight features; inputting the shallow features, the first intermediate features, the second intermediate features, the deep features and the multi-scale weight features into the decoder to obtain a predicted fault result; updating the initial seismic fault identification model based on a preset loss function according to the corresponding label of the seismic fault identification data and the predicted fault result; repeating the above model training process until a preset condition is reached to obtain a seismic fault identification model.

2. The method of claim 1, wherein, The ASPP module comprises a convolution layer, a global pooling layer and a plurality of dilated convolution layers with different dilated rates; the step of inputting the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weight features comprises: inputting the hidden features into the convolution layer, the global pooling layer and the plurality of dilated convolution layers with different dilated rates respectively to obtain a plurality of features with different scales; concatenating the plurality of features with different scales and inputting them into the attention mechanism layer to obtain multi-scale weight features.

3. The method of claim 1, wherein, the step of inputting the shallow features, the first intermediate features, the second intermediate features, the deep features and the multi-scale weight features into the decoder to obtain a predicted fault result comprises: concatenating the first intermediate features and the multi-scale weight features and performing convolution operation to obtain high-level concatenated features; concatenating the shallow features and the second intermediate features and performing convolution operation to obtain low-level concatenated features; performing convolution operation based on the deep features and concatenating the high-level concatenated features to obtain intermediate concatenated features; performing convolution operation and up-sampling operation based on the intermediate concatenated features and concatenating the low-level concatenated features to obtain multi-scale concatenated features; performing convolution operation and up-sampling operation on the multi-scale concatenated features to obtain the predicted fault result.

4. The method of claim 1, wherein, the step of calculating shallow features, first intermediate features, second intermediate features and deep features based on the hidden features comprises: performing maximum pooling operation based on the hidden features to obtain the shallow features; performing down-sampling convolution operation and convolution operation based on the hidden features to obtain the first intermediate features; obtaining the second intermediate features by convolution block operation based on the hidden features; performing a plurality of down-sampling convolution operations based on the hidden features to obtain the deep features.

5. The method of claim 1, wherein, The seismic fault data set is obtained by the following method: obtaining a plurality of three-dimensional horizontal reflection models; for each three-dimensional horizontal reflection model, adding a fold structure to obtain a fold reflection model; performing vertical displacement based on the fold reflection model to obtain a displacement reflection model; modifying the displacement reflection model based on a preset fault displacement range to obtain a fault model; performing convolution calculation based on the fault model and adding random noise to obtain a seismic fault model; extracting a plurality of seismic fault data and a label corresponding to each stratum fault data based on the seismic fault model to obtain a seismic fault data set.

6. The method of claim 5, wherein, The method comprises the following steps: For each three-dimensional horizontal reflection model, a fold structure is added to obtain a fold reflection model, comprising: In the formula, S1(x, y, z) represents a fold structure parameter, x, y, and z respectively represent three dimensions of lateral, longitudinal, and depth in a three-dimensional horizontal reflection model, z max represents the maximum depth of a stratum, N represents the number of layers of a longitudinal stratum, a0, b k , e k , d k , σ k are all parameters for controlling the fold structure parameter; For each three-dimensional horizontal reflection model, a fold structure parameter is calculated based on the following formula: According to the three-dimensional horizontal reflection model and the fold structure parameter, the fold reflection model is calculated based on the following formula: r(x,y,z) = r(x,y,z + S1) 7. The method of claim 5, wherein, In the formula, r(x,y,z) represents the fold reflection model, x, y, and z represent the three dimensions of the three-dimensional horizontal reflection model and the fold reflection model, respectively, and S1 represents the fold structure parameter. The method comprises the following steps: According to the fold reflection model, a vertical displacement parameter is calculated based on the following formula: S2(x,y,z) = e0 + fx + gy In the formula, S2(x,y,z) represents the vertical displacement parameter, x, y, and z represent the three dimensions of the fold reflection model, and e0, f, and g are parameters for controlling the fold structure variables; According to the fold reflection model and the vertical displacement parameter, the displacement reflection model is calculated based on the following formula: r(x,y,z) = r(x,y,z + S2) 8. A method of fault identification based on seismic data, characterized in that, In the formula, r(x,y,z) represents the displacement reflection model, x, y, and z represent the three dimensions of the fold reflection model and the displacement reflection model, respectively, and S2 represents the vertical displacement parameter. The method comprises the following steps: Obtaining seismic fault data to be identified; 9.A device for training a seismic fault identification model, characterized in that, Inputting the seismic fault data to be identified into the seismic fault identification model to obtain a predicted fault result. The method comprises the following steps: A first obtaining module is configured to obtain a seismic fault data set; The seismic fault data set comprises a plurality of seismic fault data and a label corresponding to each seismic fault data; A second obtaining module is configured to obtain an initial seismic fault identification model; the initial seismic fault identification model comprises an encoder and a decoder; the encoder comprises a MobileNetV2 module, an ASPP module, and an attention mechanism layer; A first operation module is configured to input the seismic fault data into the MobileNetV2 module to obtain hidden features; A second operation module is configured to calculate shallow features, first intermediate features, second intermediate features, and deep features based on the hidden features; A third operation module is configured to input the hidden features into the ASPP module and the attention mechanism layer to obtain multi-scale weight features; A fourth operation module is configured to input the shallow features, the first intermediate features, the second intermediate features, the deep features, and the multi-scale weight features into the decoder to obtain a predicted fault result. The first updating module is configured to update the initial seismic fault identification model based on a preset loss function according to the label corresponding to the seismic fault identification data and the predicted fault result. The first determining module is configured to determine whether the updated initial seismic fault identification model meets a preset condition, and if yes, obtain the seismic fault identification model, and if no, re-execute the process of inputting the seismic fault data into the MobileNetV2 module to obtain the hidden feature by the first operation module.

10. A fault identification apparatus based on seismic data, characterized by, The method comprises: The first obtaining module is configured to obtain a seismic fault data set. The seismic fault data set comprises a plurality of seismic fault data and a label corresponding to each seismic fault data. The second obtaining module is configured to obtain an initial seismic fault identification model. The initial seismic fault identification model comprises an encoder and a decoder. The encoder comprises a MobileNetV2 module, an ASPP module, and an attention mechanism layer. The first operation module is configured to input the seismic fault data into the MobileNetV2 module to obtain a hidden feature. The second operation module is configured to calculate a shallow feature, a first intermediate feature, a second intermediate feature, and a deep feature according to the hidden feature. The third operation module is configured to input the hidden feature into the ASPP module and the attention mechanism layer to obtain a multi-scale weight feature. The fourth operation module is configured to input the shallow feature, the first intermediate feature, the second intermediate feature, the deep feature, and the multi-scale weight feature into the decoder to obtain a predicted fault result. The first updating module is configured to update the initial seismic fault identification model based on a preset loss function according to the label corresponding to the seismic fault identification data and the predicted fault result. The first determining module is configured to determine whether the updated initial seismic fault identification model meets a preset condition, and if yes, obtain the seismic fault identification model, and if no, re-execute the process of inputting the seismic fault data into the MobileNetV2 module to obtain the hidden feature by the first operation module.

11. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The third obtaining module is configured to obtain to-be-identified seismic fault data.

12. A computer program product comprising computer programs / instructions, characterized in that, The first prediction module is configured to input the to-be-identified seismic fault data into the seismic fault identification model to obtain a predicted fault result.

13. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-12. The computer program / instruction is executed by the processor to implement the seismic fault identification model training method of any one of claims 1-7 and / or the fault identification method based on seismic data of claim 8. The computer program / instruction is executed by the processor to implement the seismic fault identification model training method of any one of claims 1-7 and / or the fault identification method based on seismic data of claim 8. The processor executes the computer program to implement the seismic fault identification model training method of any one of claims 1-7 and / or the fault identification method based on seismic data of claim 8.

Images