Seismic fault identification method based on adaptive multi-scale fusion and jump attention

Abstract

The invention relates to the technical field of seismic exploration, and particularly provides a seismic fault identification method based on adaptive multi-scale fusion and jump attention. The method comprises the following steps: constructing an earthquake fault identification model; in the encoder, constructing an adaptive multi-scale fusion module, and adaptively adjusting the number of network branches according to input data distribution and task complexity requirements; in the decoder, redesigning jump connection, constructing a jump attention module, and modeling a long-distance dependency relationship between a cross-space dimension and a depth dimension for compensating spatial continuity lost due to two-dimensional segmentation; according to the method, the seismic fault recognition model after the self-adaptive multi-scale fusion module and the jump attention module are constructed is used for recognizing the seismic fault, the spatial continuity of seismic data is improved through the self-adaptive multi-scale fusion module and the jump attention module, and then the quality of seismic image segmentation and the accuracy of seismic exploration are improved.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration technology, and in particular to a seismic fault identification method based on adaptive multi-scale fusion and jump attention. Background Art

[0002] In seismic exploration and development, underground rock strata break and dislocate under the action of tectonic stress to form faults. Fault interpretation is of great significance in geological research and actual production. With the increase in exploration depth and complexity, traditional seismic data interpretation methods face problems such as low efficiency, high cost, and difficult verification of results. Identifying the semantic information of seismic images from the perspective of computer vision and realizing pixel-level classification of seismic images will help improve the accuracy and efficiency of fault interpretation. Considering the extremely simple segmentation of seismic images with semantic information, a well-verified hypothesis is that more refined fault interpretation results can be achieved through local feature enhancement and global feature modeling.

[0003] The core of local feature enhancement lies in multi-scale feature modeling. Image pyramid is one of the effective methods to achieve multi-scale feature expression, which enhances the model's understanding and processing capabilities of fine-grained features and global context features. In addition, in the field of seismic image segmentation, the U-shaped structure has become the de facto standard because of its effective combination of deep and shallow features. However, although traditional convolutional neural networks can expand the receptive field through cascaded convolutions, the effective receptive field is still limited, which limits their ability to explicitly model long-distance dependencies. Summary of the invention

[0004] In view of this, the present invention provides a seismic fault identification method based on adaptive multi-scale fusion and jump attention, so as to improve the spatial continuity of seismic data, thereby improving the quality of seismic image segmentation and the accuracy of seismic exploration.

[0005] In a first aspect, the present invention provides a method for earthquake fault identification based on adaptive multi-scale fusion and jump attention, the method comprising:

[0006] Step 1: Construct an earthquake fault identification model FaultNet, whose structure includes an encoder, a decoder and a skip connection;

[0007] Step 2: In the encoder in step 1, an adaptive multi-scale fusion module AMSA is constructed to adaptively adjust the number of network branches according to the input data distribution and task complexity requirements;

[0008] Step 3: In the decoder in step 1, redesign the skip connection and construct the skip attention module Skip-Att to model the long-distance dependencies across spatial and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation.

[0009] Step 4: Identify earthquake faults using the earthquake fault identification model built with the AMSA module and the Skip-Att module.

[0010] Optionally, the step 2 includes:

[0011] Given an input feature dimension of (H×W,C), where the image resolution is H×W and the depth dimension is C, which represents the slice sequence of the 3D seismic data volume;

[0012] The encoder encodes the input features into high-level semantic features through convolution and downsampling operations. Its expression is:

[0013] F d =Down(Conv(X i ));

[0014] In the formula, X i is the input feature, Down(·) is the downsampling operation, Conv(·) is the convolution operation, and F d is the down-sampled feature output;

[0015] The high-level semantic feature F d The input is sent to the AMSA module, which integrates multi-scale resampled fine-grained features from high-level semantic features through a multi-branch network structure. Its expression is:

[0016]

[0017] In the formula, It means that the number of network branches is adaptively adjusted according to the input data distribution; in addition, the AMSA module is only introduced into the first three stages of the network, that is, let s = 1, 2, 3;

[0018] The AMSA module loss is expressed as the sum of absolute values, and its expression is:

[0019]

[0020] Optionally, step 3 includes:

[0021] The decoder fuses the spatial context semantic information through upsampling and Skip-Att modules, and its expression is:

[0022] F a =SkipAtt(F u ,F s ,F s′ );

[0023] Where SKipAtt(·) is the skip connection Skip-Att module, Fa is a fusion feature with spatial semantic information, F u 、F s and F s′ As the input feature of the Skip-Att module, F u is the upsampled output feature of the current tomographic slice, F s is the downsampled enhanced output feature of the current tomographic slice, F s′ Enhance output features for downsampling of adjacent tomographic slices;

[0024] Specifically, the input feature dimension is Where s = 1, 2, ..., 5;

[0025] F s and F u After concatenation and convolution operations, we get the image feature F that is consistent with the input feature dimension. l , whose expression is:

[0026] F c =Concat(F s ,F u );

[0027] F l =Conv 1×1 (Concat(F s ,F u ));

[0028] In the formula, Conv 1×1 is a 1×1 convolution, and Concat(·) is a concatenation operation;

[0029] F l After 1×1 convolution operation and linear operation, and after Softmax layer, the attention matrix S is obtained, which is expressed as:

[0030]

[0031] In the formula, For feature F l The dimension of the attention matrix S is (C, H×H).

[0032] Attention matrix S and feature F s′ After multiplication, we get the fusion feature F with spatial information. a , whose expression is:

[0033] F a =S×F s′ ;

[0034] In the formula, the fusion feature F aThe dimension of the upsampled output feature F of the current tomographic slice u The dimensions remain consistent;

[0035] Finally, a 3×3 convolutional layer and a Relu activation function layer are used to generate a pixel-level segmentation map F p , whose expression is:

[0036] F p =Relu(Conv 3×3 (F a ));

[0037] In the formula, Conv 3×3 (·) is the convolution operation, Relu(·) is the Relu activation function operation;

[0038] A balanced composite loss function L is constructed by integrating Focal loss and AMSA module loss Total , whose expression is:

[0039] L Total =μL Focal (θ)+(1-μ)L AMSA (θ);

[0040] Where, Focal loss L Focal (θ) is used to alleviate the problem of unbalanced distribution of tomographic pixels, L AMSA (θ) is the AMSA module loss, and the hyperparameter μ is used to control the contribution of Focal loss to the composite loss. The larger the μ, the higher the contribution to the loss.

[0041] In a second aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the earthquake fault identification method based on adaptive multi-scale fusion and jump attention in the first aspect or any possible implementation of the first aspect.

[0042] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the device, enable the device to execute the earthquake fault identification method based on adaptive multi-scale fusion and jump attention in the first aspect or any possible implementation of the first aspect.

[0043] In the technical solution provided by the present invention, the method includes constructing an earthquake fault recognition model FaultNet, whose structure includes an encoder, a decoder and a skip connection; in the encoder, an adaptive multi-scale fusion module AMSA is constructed, and the number of network branches is adaptively adjusted according to the input data distribution and the task complexity requirements; in the decoder, the skip connection is redesigned, and a skip attention module Skip-Att is constructed to model long-distance dependencies across space and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation; earthquake faults are identified by using the earthquake fault recognition model after constructing the AMSA module and the Skip-Att module. The method improves the spatial continuity of seismic data through the adaptive multi-scale fusion module and the skip attention module, thereby improving the quality of seismic image segmentation and the accuracy of seismic exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A flow chart of a method for identifying earthquake faults provided by an embodiment of the present invention;

[0046] Figure 2 A flowchart of another earthquake fault identification method provided by an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] It should be clear that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention are also intended to include plural forms, unless the context clearly indicates other meanings.

[0051] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0052] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0053] The present invention improves an earthquake fault identification method based on adaptive multi-scale fusion and jump attention, such as Figure 1 and Figure 2 As shown, the method includes:

[0054] Step 1: Construct an earthquake fault identification model FaultNet, whose structure includes an encoder, a decoder and a skip connection.

[0055] In the embodiment of the present invention, multiple modules in the earthquake fault identification model cooperate with each other to achieve efficient and accurate fault feature extraction and identification.

[0056] Step 2: In the encoder in step 1, an adaptive multi-scale fusion module (AMSA) is constructed to adaptively adjust the number of network branches according to the input data distribution and task complexity requirements.

[0057] In the embodiment of the present invention, the AMSA module is introduced into the encoder of FaultNet to assist in encoding, and the information of multiple subspaces is processed in parallel to enhance the model's ability to capture multi-scale features. In different application scenarios and tasks, the AMSA module can adaptively adjust the number of network branches according to the input data distribution and task complexity requirements. This dynamic adjustment mechanism enables the module to show good flexibility and adaptability in a variety of different tasks and data sets, which not only enhances the model's sensitivity to fine-grained features, but also improves the ability to characterize multi-scale features and the depth and breadth of multi-scale feature fusion, thereby achieving superior model processing capabilities for high-dimensional and high-complexity feature distributions.

[0058] In the embodiment of the present invention, the AMSA module specifically includes, as shown in Table 1:

[0059] Table 1

[0060]

[0061] In the embodiment of the present invention, step 2 includes:

[0062] Given an input feature dimension of (H×W,C), where the image resolution is H×W and the depth dimension is C, which represents the slice sequence of the 3D seismic data volume;

[0063] The encoder encodes the input features into high-level semantic features through convolution and downsampling operations. Its expression is:

[0064] F d =Down(Conv(X i ));

[0065] In the formula, X i is the input feature, Down(·) is the downsampling operation, Conv(·) is the convolution operation, and F d is the down-sampled feature output;

[0066] The high-level semantic feature F d The input is sent to the AMSA module, which integrates multi-scale resampled fine-grained features from high-level semantic features through a multi-branch network structure. Its expression is:

[0067]

[0068] Where, L = 1, 2, ... m, It means that the number of network branches is adaptively adjusted according to the input data distribution; in addition, the AMSA module is only introduced into the first three stages of the network, that is, let s = 1, 2, 3;

[0069] In the embodiment of the present invention, the AMSA module is only introduced into the first three stages of the network because of the detail loss problem of the output features in the 4th and 5th stages, that is, in the deeper levels of the network, the local detail information gradually decreases, resulting in the feature maps of these stages being insufficient to provide accurate fault location and contour information.

[0070] The AMSA module loss is expressed as the sum of absolute values, and its expression is:

[0071]

[0072] Step 3. In the decoder in step 1, redesign the skip connection and construct the skip attention module Skip-Att to model long-distance dependencies across spatial and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation.

[0073] In the embodiment of the present invention, the skip connection is redesigned and the skip attention module Skip-Att is constructed, which not only makes up for the information loss of the traditional encoding-decoding structure, but also further strengthens the model's sensitivity to pixel-level spatial features, thereby enhancing the model's understanding of the continuity of three-dimensional seismic data.

[0074] In the embodiment of the present invention, the Skip-Att module specifically includes, as shown in Table 2:

[0075] Table 2

[0076]

[0077] In the embodiment of the present invention, step 3 includes:

[0078] The decoder fuses the spatial context semantic information through upsampling and Skip-Att modules, and its expression is:

[0079] F a =SkipAtt(F u ,F s ,F s′ );

[0080] Where SkipAtt(·) is the skip connection Skip-Att module, F a is a fusion feature with spatial semantic information, F u 、F s and F s′ As the input feature of the Skip-Att module, F u is the upsampled output feature of the current tomographic slice, F s is the downsampled enhanced output feature of the current tomographic slice, F s′ Enhance output features for downsampling of adjacent tomographic slices;

[0081] Specifically, the input feature dimension is Where s = 1, 2, ..., 5;

[0082] F s and F u After concatenation and convolution operations, we get the image feature F that is consistent with the input feature dimension. l , whose expression is:

[0083] F c =Concat(F s ,F u );

[0084] F l =Conv 1×1 (Concat(F s ,F u ));

[0085] In the formula, Conv 1×1 is a 1×1 convolution, and Concat(·) is a concatenation operation;

[0086] F l After 1×1 convolution operation and linear operation, and after the Softmax layer, the attention matrix S is obtained, which is expressed as:

[0087]

[0088] In the formula, For feature F l The dimension of the attention matrix S is (C, H×H).

[0089] Attention matrix S and feature F s′ After multiplication, we get the fusion feature F with spatial information. a , whose expression is:

[0090] F a =S×F s′ ;

[0091] In the formula, the fusion feature F a The dimension of the upsampled output feature F of the current tomographic slice u The dimensions remain consistent;

[0092] Finally, a 3×3 convolutional layer and a Relu activation function layer are used to generate a pixel-level segmentation map F p , whose expression is:

[0093] F p =Relu(Conv 3×3 (F a ));

[0094] In the formula, Conv 3×3 (·) is the convolution operation, Relu(·) is the Relu activation function operation;

[0095] A balanced composite loss function L is constructed by integrating Focal loss and AMSA module loss Total , whose expression is:

[0096] L Total =μL Focal (θ)+(1-μ)L AMSA (θ);

[0097] Where, Focal loss L Focal (θ) is used to alleviate the problem of unbalanced distribution of tomographic pixels, L AMSA (θ) is the AMSA module loss, and the hyperparameter μ is used to control the contribution of Focal loss to the composite loss. The larger the μ, the higher the contribution to the loss.

[0098] Step 4: Identify earthquake faults using the earthquake fault identification model built with the AMSA module and the Skip-Att module.

[0099] The Transformer architecture effectively solves the long-distance dependency problem through self-attention mechanism and position encoding, making it perform well in processing long sequence tasks. The present invention aims to solve the problem of lack of spatial continuity caused by the loss of three-dimensional seismic spatial information in the process of two-dimensional seismic image segmentation. In seismic data processing, it is usually necessary to slice three-dimensional seismic data into two-dimensional images for analysis, but this method may ignore the correlation between seismic data in three-dimensional space, thereby affecting the interpretation accuracy of seismic data. Therefore, the purpose of the present invention is to restore and utilize three-dimensional spatial information in two-dimensional seismic image segmentation to improve the spatial continuity of seismic data, thereby improving the quality of seismic image segmentation and the accuracy of seismic exploration.

[0100] The present invention proposes and constructs an adaptive multi-scale fusion module AMSA, which adaptively adjusts the number of network branches according to the input data distribution and task complexity requirements, which not only enhances the model's sensitivity to fine-grained features, but also improves its ability to characterize multi-scale features. The skip connection is redesigned and a skip attention module Skip-Att is constructed, which focuses on modeling long-distance dependencies across spatial and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation, thereby enhancing the model's understanding of the continuity of three-dimensional seismic data.

[0101] In the technical solution provided by the present invention, the method includes constructing an earthquake fault recognition model FaultNet, whose structure includes an encoder, a decoder and a skip connection; in the encoder, an adaptive multi-scale fusion module AMSA is constructed, and the number of network branches is adaptively adjusted according to the input data distribution and the task complexity requirements; in the decoder, the skip connection is redesigned, and a skip attention module Skip-Att is constructed to model long-distance dependencies across space and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation; earthquake faults are identified by using the earthquake fault recognition model after constructing the AMSA module and the Skip-Att module. The method improves the spatial continuity of seismic data through the adaptive multi-scale fusion module and the skip attention module, thereby improving the quality of seismic image segmentation and the accuracy of seismic exploration.

[0102] Each step of the embodiment of the present invention may be performed by an electronic device, which includes but is not limited to a mobile phone, a tablet computer, a portable PC, a desktop computer, etc.

[0103] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the electronic device where the computer-readable storage medium is located is controlled to execute the above-mentioned embodiment of the earthquake fault identification method based on adaptive multi-scale fusion and jump attention.

[0104] Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the electronic device 21 includes: a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, the earthquake fault identification method based on adaptive multi-scale fusion and jump attention in the embodiment is implemented. To avoid repetition, they are not described one by one here.

[0105] The electronic device 21 includes, but is not limited to, a processor 211 and a memory 212. Those skilled in the art will appreciate that Figure 3 It is only an example of the electronic device 21 and does not constitute a limitation of the electronic device 21. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0106] The processor 211 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0107] The memory 212 may be an internal storage unit of the electronic device 21, such as a hard disk or memory of the electronic device 21. The memory 212 may also be an external storage device of the electronic device 21, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 21. Further, the memory 212 may also include both an internal storage unit of the electronic device 21 and an external storage device. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 may also be used to temporarily store data that has been output or is to be output.

[0108] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for earthquake fault identification based on adaptive multi-scale fusion and jump attention, characterized in that: The method comprises: Step 1: Construct an earthquake fault identification model FaultNet, whose structure includes an encoder, a decoder and a skip connection; Step 2: In the encoder in step 1, an adaptive multi-scale fusion module AMSA is constructed to adaptively adjust the number of network branches according to the input data distribution and task complexity requirements; Step 3: In the decoder in step 1, redesign the skip connection and construct the skip attention module Skip-Att to model the long-distance dependencies across spatial and depth dimensions to compensate for the spatial continuity lost due to two-dimensional segmentation. Step 4: Identify earthquake faults using the earthquake fault identification model built with the AMSA module and the Skip-Att module.

2. The method according to claim 1, characterized in that The step 2 comprises: Given an input feature dimension of (H×W,C), where the image resolution is H×W and the depth dimension is C, which represents the slice sequence of the 3D seismic data volume; The encoder encodes the input features into high-level semantic features through convolution and downsampling operations. Its expression is: F d =Down(Conv(X i )); Where, X i is the input feature, Down(·) is the downsampling operation, Conv(·) is the convolution operation, and F d is the down-sampled feature output; The high-level semantic feature F d The input is sent to the AMSA module, which integrates multi-scale resampled fine-grained features from high-level semantic features through a multi-branch network structure. Its expression is: Where, L = 1, 2, ... m, It means that the number of network branches is adaptively adjusted according to the input data distribution; in addition, the AMSA module is only introduced into the first three stages of the network, that is, let s = 1, 2, 3; The AMSA module loss is expressed as the sum of absolute values, and its expression is:

3. The method according to claim 1, characterized in that The step 3 comprises: The decoder fuses the spatial context semantic information through upsampling and Skip-Att modules, and its expression is: F a =SkipAtt(F u ,F s ,F s′ ); Where SkipAtt(·) is the skip connection Skip-Att module, F a is a fusion feature with spatial semantic information, F u 、F s and F s′ As the input feature of the Skip-Att module, F u is the upsampled output feature of the current tomographic slice, F s is the downsampled enhanced output feature of the current tomographic slice, F s′ Enhance output features for downsampling of adjacent tomographic slices; Specifically, the input feature dimension is Where s=1,2,…,5; F s and F u After concatenation and convolution operations, we get the image feature F that is consistent with the input feature dimension. l , whose expression is: F c =Concat(F s ,F u ); F l =Conv 1×1 (Concat(F s ,F u )); In the formula, Conv 1×1 is a 1×1 convolution, and Concat(·) is a concatenation operation; F l After 1×1 convolution operation and linear operation, and after the Softmax layer, the attention matrix S is obtained, which is expressed as: In the formula, For feature F l The dimension of the attention matrix S is (C, H×H). Attention matrix S and feature F s′ After multiplication, we get the fusion feature F with spatial information. a , whose expression is: F a =S×F s′ ; In the formula, the fusion feature F a The dimension of the upsampled output feature F of the current tomographic slice u The dimensions remain consistent; Finally, a 3×3 convolutional layer and a Relu activation function layer are used to generate a pixel-level segmentation map F. p , whose expression is: F p =Release(Conv 3×3 (F a )); In the formula, Conv 3×3 (·) is the convolution operation, Relu(·) is the Relu activation function operation; A balanced composite loss function L is constructed by integrating Focal loss and AMSA module loss Total , whose expression is: L Total =μL Focal (θ)+(1-μ)L AMSA (i); Where, Focal loss L Focal (θ) is used to alleviate the problem of unbalanced distribution of tomographic pixels, L AMSA (θ) is the AMSA module loss, and the hyperparameter μ is used to control the contribution of Focal loss to the composite loss. The larger the μ, the higher the contribution to the loss.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the earthquake fault identification method based on adaptive multi-scale fusion and jump attention as described in any one of claims 1 to 3.

5. An electronic device, characterized in that: include: one or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the device, enable the device to perform the earthquake fault identification method based on adaptive multi-scale fusion and jump attention as described in any one of claims 1 to 3.

Images