Fault detection method, device and equipment based on seismic fault segmentation model

Abstract

The invention relates to the technical field of seismic image processing, and particularly provides a fault detection method, device and equipment based on a seismic fault segmentation model, and the method comprises the steps: inputting a seismic fault profile image into an encoder of a network model, extracting the seismic fault profile image step by step through a multi-stage down-sampling module of an encoder to obtain multi-scale semantic features; fusing each level of semantic features output by an encoder layer by layer through a decoder of the network model, each level of up-sampling module and jump connection, gradually recovering the spatial resolution of the multi-scale semantic features and refining the fault boundary to obtain a seismic fault prediction map; calculating a loss value between the seismic fault prediction map and a seismic fault real map corresponding to the seismic fault profile image by using a mixed boundary loss function; and optimizing the network model based on the loss value, and determining a seismic fault segmentation model. According to the embodiment of the invention, accurate prediction of the seismic fault can be realized.

Description

Technical Field

[0001] This application relates to the technical field of seismic image processing. Specifically, it relates to a method, device, and equipment for fault detection based on a seismic fault segmentation model. Background Art

[0002] When the crustal rock strata are stressed to a certain intensity, they rupture, and a structure with obvious relative movement along the rupture surface is called a fault. Generally, it is believed that earthquakes are often caused by fault activities, and earthquakes may also cause new faults. Therefore, the analysis of seismic faults plays an important role.

[0003] Currently, the commonly used seismic fault segmentation method is manual visual interpretation, which relies on the manual annotation of geological experts. This method specifically includes: preprocessing the original seismic data, including noise suppression and the calculation of coherent body or curvature body attributes to highlight fault features. Then, geological experts manually draw fault lines based on the abnormal area of the reflection event axis of the seismic data. This method relies on the experience and subjective judgment of experts, is time-consuming, and is easily affected by personal biases.

[0004] Therefore, how to provide a technical solution for an efficient and accurate method for automatically detecting seismic fault regions has become a technical problem that urgently needs to be solved. Summary of the Invention

[0005] Some embodiments of this application aim to provide a method, device, and equipment for fault detection based on a seismic fault segmentation model. Through the technical solutions of the embodiments of this application, accurate and efficient detection of seismic fault regions can be achieved.

[0006] In a first aspect, some embodiments of this application provide a method for training a seismic fault segmentation model. The method includes: inputting a seismic fault profile image into the encoder of a network model, and gradually extracting multi-scale semantic features from the seismic fault profile image through the multi-level downsampling module of the encoder; wherein, each level of the downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position; through the decoder of the network model, and each level of upsampling module and skip connection, gradually fuse the semantic features of each level output by the encoder, gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundary to obtain a seismic fault prediction map; calculate the loss value between the seismic fault prediction map and the seismic fault ground truth map using a hybrid boundary loss function; wherein, the hybrid boundary loss function includes pixel-level classification loss, fault segmentation loss, and fault boundary perception loss; optimize the network model based on the loss value to determine the seismic fault segmentation model.

[0007] Some embodiments of the present application achieve dynamic sampling of an image by inputting a seismic fault profile image into the encoder of a network model. Subsequently, the decoder processes the multi-scale semantic features of the image output by the encoder to obtain a seismic fault prediction map. Finally, the loss value between the seismic fault prediction map and the seismic fault ground truth map is calculated through a hybrid boundary loss function to optimize the network model and obtain a seismic fault segmentation model. Some embodiments of the present application can accurately train the designed network model through seismic fault profile images, improving the detection effect and robustness of the model.

[0008] In some embodiments, the multi-level downsampling module is N levels, where N is a positive integer. Each level of the downsampling module includes: a boundary deformable convolution module and a max pooling layer. The offset variable in the boundary deformable convolution module can be dynamically adjusted. The sampling position changes with the value of the offset variable. The encoder includes M levels of upsampling modules, where M is a positive integer and N - M = 1. Each level of the M-level upsampling module includes the boundary deformable convolution module. Among them, the boundary deformable convolution module extracts features of the corresponding sampling position in the seismic fault profile image by introducing the dynamically moving offset variable. The offset variable is determined by an offset learning function, and the offset learning function is related to the input image features and the boundary offset parameters. The boundary offset parameters are related to the gradients of the input image features in the vertical and horizontal directions.

[0009] Some embodiments of the present application set different boundary deformable convolution modules in the multi-level downsampling module and the upsampling module to facilitate accurate processing of the seismic fault profile image and provide support for model training.

[0010] In some embodiments, the step of inputting the seismic fault profile image into the encoder of the network model and gradually extracting multi-scale semantic features of the seismic fault profile image through the multi-level downsampling module of the encoder includes: inputting the feature map of the i-th stage into the (i + 1)-th level downsampling module to output the feature map of the (i + 1)-th stage, where i ∈ [0, N - 1]. Among them, when i = 0, the feature map of the 0-th stage is the seismic fault profile image. When i = N - 1, the feature map of the N-th stage is the multi-scale semantic feature.

[0011] Some embodiments of the present application process feature maps of different stages through different downsampling modules to obtain multi-scale semantic features of the image, which can achieve efficient processing of the image.

[0012] In some embodiments, through the decoder of the network model, and the upsampling module and skip connection at each level, the semantic features at each level output by the encoder are fused layer by layer, gradually restoring the spatial resolution of the multi-scale semantic features and refining the fault boundary to obtain a seismic fault prediction map, including: inputting the multi-scale semantic features and the feature map of the (N-1)th stage into the upsampling module at the Mth level to output the Mth sampled feature; inputting the feature map of the (N-j)th stage and the Mth sampled feature into the upsampling module at the (M-j+1)th level to output the (M-j+1)th sampled feature, where j ∈ [2, N-1]; when j = N-1, the first sampled feature is the seismic fault prediction map.

[0013] In some embodiments of the present application, by jointly processing the sampled features at different stages and the feature maps at different stages output by downsampling through a multi-level upsampling module, a seismic fault prediction map is output, which can achieve effective prediction of seismic faults.

[0014] In some embodiments, the hybrid boundary loss function is constructed in the following manner: obtaining the weight values corresponding to the pixel-level classification loss, the fault segmentation loss, and the fault boundary perception loss respectively; performing weighted summation on the pixel-level classification loss, the fault segmentation loss, the fault boundary perception loss, and the weight values to obtain the hybrid boundary loss function.

[0015] In some embodiments of the present application, a loss function is composed of multiple losses to achieve precise optimization of the seismic fault segmentation model during the training process.

[0016] In some embodiments, the fault boundary perception loss is obtained by the following method: obtaining the predicted gradient magnitude values of the seismic fault prediction map in the horizontal and vertical directions; obtaining the true gradient magnitude values of the seismic fault true map in the horizontal and vertical directions; determining the fault boundary perception loss based on the predicted gradient magnitude values and the true gradient magnitude values.

[0017] In some embodiments of the present application, the fault boundary perception loss is determined by the predicted gradient magnitude values and the true values of the seismic fault prediction map in different directions, which can improve the accuracy of model training subsequently.

[0018] In a second aspect, some embodiments of the present application provide a method for detecting a seismic fault area, including: obtaining an image of a seismic fault to be detected; inputting the image of the seismic fault to be detected into a seismic fault segmentation model to obtain a seismic fault prediction map; wherein, the seismic fault segmentation model is obtained by any method embodiment in the first aspect.

[0019] In some embodiments of the present application, the trained seismic fault segmentation model can be used to detect seismic faults in the image to be detected, with high detection efficiency and accuracy.

[0020] In a third aspect, some embodiments of the present application provide an apparatus for training a seismic fault segmentation model. The encoding module is configured to input a seismic fault profile image into the encoder of the network model, and gradually extract multi-scale semantic features from the seismic fault profile image through the multi-level downsampling modules of the encoder; wherein, each level of downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position. The decoding module is configured to gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundary through the decoder of the network model, and each level of upsampling module and skip connection layer by layer to fuse each level of semantic features output by the encoder, so as to obtain a seismic fault prediction map. The loss calculation module is configured to calculate the loss value between the seismic fault prediction map and the true seismic fault map corresponding to the seismic fault profile image by using a hybrid boundary loss function; wherein, the hybrid boundary loss function includes a pixel-level classification loss, a fault segmentation loss, and a fault boundary perception loss. The model optimization module is configured to optimize the network model based on the loss value to determine a seismic fault segmentation model.

[0021] In a fourth aspect, an apparatus for detecting a seismic fault area includes: an acquisition module configured to acquire a seismic fault image to be detected; a detection module configured to input the seismic fault image to be detected into a seismic fault segmentation model to obtain a seismic fault prediction map; wherein, the seismic fault segmentation model is obtained by any method embodiment in the first aspect.

[0022] In a fifth aspect, some embodiments of the present application provide a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the method described in any embodiment of the first aspect can be implemented.

[0023] In a sixth aspect, some embodiments of the present application provide an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, when the processor executes the program, the method described in any embodiment of the first aspect can be implemented.

[0024] In a seventh aspect, some embodiments of the present application provide a computer program product, the computer program product includes a computer program, wherein, when the computer program is executed by a processor, the method described in any embodiment of the first aspect can be implemented. Description of the Drawings

[0025] To more clearly illustrate the technical solutions of some embodiments of the present application, the following will briefly introduce the drawings required to be used in some embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0026] Figure 1 Schematic diagram of the BDConv process provided for some embodiments of the present application; Figure 2 Schematic diagram of the calculation process of BDConv provided for some embodiments of the present application; Figure 3 Method flow chart for training a seismic fault segmentation model provided for some embodiments of the present application; Figure 4 Comparison chart of image resolution distributions before and after cropping provided for some embodiments of the present application; Figure 5 Network model structure diagram provided for some embodiments of the present application; Figure 6 Schematic diagram of extracting gradients using the Sobel operator provided for some embodiments of the present application; Figure 7 One of the method flow charts for seismic fault area detection provided for some embodiments of the present application; Figure 8 Another method flow chart for seismic fault area detection provided for some embodiments of the present application; Figure 9 Block diagram of the device for training a seismic fault segmentation model provided for some embodiments of the present application; Figure 10 Block diagram of the device for seismic fault area detection provided for some embodiments of the present application; Figure 11 Schematic diagram of an electronic device provided for some embodiments of the present application. Detailed implementation manners

[0027] The following will describe the technical solutions in some embodiments of the present application in combination with the drawings in some embodiments of the present application.

[0028] It should be noted that: similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present application, terms such as "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0029] In related technologies, machine learning-based methods perform fault discrimination by manually extracting seismic attributes and combining statistical learning models. Common methods include support vector machines (SVMs) and multi-layer perceptrons (MLPs). Specifically, multiple attributes are extracted from seismic data, such as coherence coefficients, azimuth gradients, etc. Machine learning models such as SVMs or MLPs are used to optimize weight parameters through backpropagation for fault discrimination. Then, seismic data is input into the model, and a fault probability map is output. This method requires a large amount of feature engineering and parameter tuning and is sensitive to noise. In recent years, deep learning models, especially convolutional neural networks, have made significant progress in seismic fault segmentation tasks. Among them, the U-Net model uses an encoder-decoder structure, retains low-level spatial detail information through skip connections, and combines high-level semantic information for fault segmentation. However, it still has deficiencies in the ability to capture long-range dependencies and maintain boundary continuity. Standard convolutional operations are limited to local receptive fields and are difficult to establish cross-scale geological structure associations, resulting in insufficient ability to capture long-range dependencies. In addition, existing seismic fault segmentation models do not pay attention to the continuity requirements of fault boundaries, and the models are not sensitive enough to weakly continuous boundaries, resulting in discontinuous faults in the segmentation results.

[0030] In view of this, some embodiments of the present application disclose a method for training a seismic fault segmentation model. The method trains a network model composed of an encoder and a decoder containing multi-level downsampling modules and upsampling modules, and optimizes it using a loss function containing multiple types of losses to obtain a seismic fault segmentation model. Some embodiments of the present application can obtain a seismic fault segmentation model with higher accuracy, provide model support for seismic fault detection, and improve the efficiency of seismic fault detection at the same time.

[0031] To improve the accuracy of fault identification, a boundary deformable convolution (BDConv) module that can dynamically adjust the sampling position is proposed in the present application. The BDConv module in the network model provided in some embodiments of the present application is a module with significant differences from traditional technologies. Therefore, the data processing process of the BDConv module in the present application is first exemplarily generated below.

[0032] Traditional convolutional layers use a fixed grid receptive field, which limits their adaptability in detecting irregular geological structures. In seismic images, faults usually exhibit different orientations, displacement amounts, and lateral continuities, and standard convolutional operations are limited by their fixed grid sampling patterns and are difficult to adaptively capture complex fault morphologies. In addition, due to the local receptive field characteristics of convolutional operations, the model cannot effectively establish cross-scale long-range feature dependencies, resulting in discontinuities in the fault segmentation results.

[0033] In some embodiments of the present application, the boundary deformable convolution module extracts features from the corresponding sampling positions in the seismic fault profile image by introducing the dynamically moving offset variables; the offset variables are determined by an offset learning function, and the offset learning function is related to the input image features and the boundary offset parameters; the boundary offset parameters are related to the gradients of the input image features in the vertical direction and the horizontal direction.

[0034] To overcome these problems, the present application proposes BDConv, which uses learnable offsets and a boundary-aware attention mechanism to dynamically adjust the convolution sampling positions. This enables the network to more effectively focus on the fault boundaries while expanding its receptive field in an adaptive manner. As Figure 1 shown, BDConv introduces learnable offsets into the standard convolution operation, allowing the sampling positions to deviate from the fixed grid structure. Instead of extracting features from a kernel with fixed intervals, deformable convolution learns to move the sampling points to the most relevant regions, improving its ability to capture non-uniform features.

[0035] Faults in seismic data usually appear as long and narrow lines, and their most prominent features are usually arranged along the vertical direction. Therefore, a vertical Sobel kernel is used in BDConv to maintain the continuity of fault boundary recognition. The vertical Sobel effectively emphasizes the gradients in the vertical direction, which is consistent with the structure of most faults, ensuring that the network focuses on the fault edges while maintaining the continuity of long fault segments. By dynamically adjusting the receptive field and focusing on the vertical direction, the model is allowed to better capture long and continuous fault lines and maintain continuity during the segmentation process.

[0036] The calculation process of BDConv is as Figure 2 shown. The standard convolution operation samples features at fixed positions within a K×K convolution kernel, and the sampling formula is:

[0037] where, p 0 represents the current position, w k is the weight at the kth position in the convolution kernel, x k is the input feature value at the position p 0 + p k and p k represents the fixed sampling offset.

[0038] In BDConv, each sampling position is adjusted by a learnable offset p k which allows the receptive field to be dynamically modified. The formula for deformable convolution is:

[0039] in, p k is a learned offset (as a specific example of an offset variable) that dynamically moves the sampling positions, allowing the network model to adaptively focus on more relevant areas in the feature map.

[0040] In order to improve the extraction of key fault boundaries during the convolution process, the embodiment of the present application also introduces a boundary-aware directional attention mechanism, which refines the network model's attention to the fault edge, guides the offset learning process, and ensures that the deformation of the receiving domain is aligned with the fault structure. In addition, the mechanism enhances the ability to extract features along fault discontinuities, especially for small faults and small displacements, while increasing directional sensitivity. This allows the model to effectively distinguish between fault edges and surrounding noise, ultimately improving segmentation performance.

[0041] The boundary-aware attention map A(p) is defined as a function that regulates the offset learning process, namely: p k = f ( x ,A). Among them, f is a learnable transformation function (as a specific example of an offset learning function) that x (as a specific example of input image features) and boundary map A (as a specific example of boundary offset parameters) adjust the offset. The attention map A is calculated using a directional Sobel filter, emphasizing high gradient areas and ensuring that the offset points to the fault boundary:

[0042] in, Represents the input feature map x The gradient in the vertical direction (i.e., the gradient of the input image feature in the vertical direction) is used to capture the vertical changes of the fault boundary; Represents the gradient of the input feature map in the horizontal direction, which is used to capture the horizontal changes of the fault boundary.

[0043] The following is combined with Figure 3 The implementation process of training an earthquake fault segmentation model performed by a terminal device provided in some embodiments of the present application is exemplarily described, wherein the terminal device may be a server device.

[0044] Please see attached Figure 3 , Figure 3 A flow chart of a method for training an earthquake fault segmentation model provided for some embodiments of the present application, the method for training an earthquake fault segmentation model may include: S310, inputting the earthquake fault profile image into the encoder of the network model, and gradually extracting the multi-scale semantic features from the earthquake fault profile image through the multi-level downsampling module of the encoder; wherein each level of the downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position.

[0045] For example, in some embodiments of the present application, a training data set for training a network model is first obtained, and the training data set includes an earthquake fault profile image and its corresponding earthquake fault true value. After that, the earthquake fault profile image is input into the encoder of the designed network model for feature extraction. Each level of downsampling module in the encoder adopts a dynamic adjustment mechanism, which can adaptively optimize the sampling position according to the local characteristics of the fault structure, thereby effectively retaining key geological information, thereby obtaining multi-scale semantic features of the earthquake fault profile image.

[0046] The following is an example of a method for acquiring an earthquake fault profile image.

[0047] In the embodiment of the present application, the Thebe dataset is used, which is a 3D seismic data volume originally collected for geological fault interpretation. The raw data in the dataset is originally stored in a 3D voxel format, and the present application processes the data through a preprocessing pipeline to train a deep learning model (i.e., the network model of the present application).

[0048] First, use Python's NumPy library to load the 3D seismic data volume and perform equally spaced slicing along the depth dimension to divide the 3D seismic data volume and its corresponding 3D fault annotations into 2D slices. Each seismic data slice is saved as a seismic image file in PNG format and normalized to the [0,1] range using the "seismic" color map. The fault annotation slices are directly converted into binary images.

[0049] After preliminary visualization, it is observed that the binary image contains a large number of redundant data areas, which are usually manifested as large white areas. In order to improve the quality and relevance of the data, the adaptive threshold method is used to detect blank areas. Specifically, the image is first converted into a grayscale image, and an empirical threshold is set to generate a mask; then the bounding box of the non-blank area is determined by calculating the effective range of the mask in the row and column directions. In particular, to ensure the consistency of the seismic data and its annotation, the two are synchronously cropped using exactly the same bounding box parameters.

[0050] Since the original seismic image has a high resolution, directly inputting the original high-resolution image into the network model will consume too much computing resources, while directly reducing the resolution of the original seismic image will cause feature loss. Therefore, the 2D seismic image and its corresponding fault annotation are cropped to half of the original width. The image resolution before and after cropping is compared as follows:Figure 4 as shown, where Figure 4 the left image in shows the ratio before cropping, and the right image shows the ratio after cropping. After these preprocessing steps, the obtained dataset contains high-quality 2D seismic images (i.e., seismic fault profile images), which can be directly used to train a network model for the fault segmentation task.

[0051] In some embodiments of the present application, the multi-level downsampling module is N-level, where N is a positive integer; each level of the downsampling module includes: a boundary deformable convolution module and a max pooling layer; the offset variable in the boundary deformable convolution module can be dynamically adjusted; the sampling position changes with the value of the offset variable.

[0052] For example, in some embodiments of the present application, as Figure 5 shown in the schematic diagram of the network model structure. The network model includes an encoder and a decoder. Among them, the encoder includes 5 (i.e., N = 5) levels of downsampling modules such as BDC_1, BDC_2, and BDC_3, which are the first downsampling module, the second downsampling module... the fifth downsampling module from bottom to top. Each level of the downsampling module includes: a BDConv module (as a specific example of the boundary deformable convolution module) and a Maxpooling (as a specific example of the max pooling layer). Among them, BDC_1 also includes a convolutional layer Conv. Specifically, the BDConv module dynamically adjusts the convolutional sampling position through a learnable offset and a boundary-aware attention mechanism, enabling the network model to adaptively focus on the fault boundary region. Specifically, the first-level downsampling module BDC_1 processes the original (640, 640, 3) fault profile diagram, extracts low-level features such as fault edge textures; as the network depth increases, the subsequent receptive field is gradually expanded to capture higher-level semantic features. During the encoding process, after each level of processing, the spatial resolution of the feature map is halved and the number of channels is doubled, finally forming a highly abstract feature representation. Figure 5 The numbers on the left in represent the changes in the spatial resolution and the number of channels during the downsampling process.

[0053] Figure 5 On the right, 4 BDC_2s form the decoder of the network model. This part of the decoder gradually restores the spatial resolution through upsampling. Each level of the decoder not only receives the feature input from the previous level but also through a skip connection (i.e., Figure 5The Skip Connect in it) fuses the feature maps corresponding to the encoder stage. This design effectively solves the problem of vanishing gradients in deep networks while retaining the spatial details of the fault structure. In addition, the convolutions in the encoder part are all BDConv, ensuring that the fused features have consistent boundary awareness characteristics. It should be noted that the number of downsampling modules and upsampling modules in the encoder and decoder of the network model can be adjusted according to the actual application scenario, and no specific limitation is made in this embodiment of the present application.

[0054] In some embodiments of the present application, S310 may include: inputting the feature map of the i-th stage into the (i + 1)-th level downsampling module to output the feature map of the (i + 1)-th stage, where i ∈ [0, N - 1]; among them, when i = 0, the feature map of the 0-th stage is the seismic fault profile image; when i = N - 1, the feature map of the N-th stage is the multi-scale semantic feature.

[0055] For example, in some embodiments of the present application, by using the feature map output by the previous-level downsampling module as the input of the next-level downsampling module, the input fault profile diagram is processed level by level. When processing at each stage, the sampling positions of the downsampling module are not fixed and can be dynamically adjusted to focus on the fault boundary area.

[0056] S320, through the decoder of the network model, and each level of upsampling module and skip connection, layer by layer fuses the semantic features of each level output by the encoder, gradually restores the spatial resolution of the multi-scale semantic feature and refines the fault boundary to obtain a seismic fault prediction map. The encoder includes M levels of upsampling modules, M is a positive integer and N - M = 1; each level of upsampling module in the M levels of upsampling modules includes the boundary deformable convolution module.

[0057] For example, in some embodiments of the present application, through the decoder, upsampling and skip connection layer by layer fuse the multi-scale semantic features output by the encoder, gradually restore the spatial resolution and refine the fault boundary to obtain a high-precision seismic fault prediction map.

[0058] In some embodiments of the present application, S320 may include: inputting the multi-scale semantic feature and the feature map of the (N - 1)-th stage into the M-th level upsampling module to output the M-th sampling feature; inputting the feature map of the (N - j)-th stage and the M-th sampling feature into the (M - j + 1)-th level upsampling module to output the (M - j + 1)-th sampling feature, where j ∈ [2, N - 1]; when j = N - 1, the first sampling feature is the seismic fault prediction map.

[0059] For example, in some embodiments of the present application, Figure 5Taking the network model as an example, the BDC_2 on the right is divided into the first upsampling module, the second upsampling module... the fourth upsampling module from bottom to top, that is, M = 4. Among them, the input data of the fourth-level upsampling module is the multi-scale semantic features of the image and the fourth-stage feature map output by BDConv in the fourth downsampling module; the input data of the third-level upsampling module is the fourth sampling feature output by the fourth-level upsampling module and the third-stage feature map output by BDConv in the third downsampling module, and so on, until the seismic fault prediction map is output.

[0060] Traditional deep learning models for fault segmentation usually rely on standard loss functions such as BCE loss and Dice loss. BCE loss treats each pixel independently, resulting in over-segmentation or under-segmentation in noisy seismic images. Dice loss improves the segmentation balance but does not explicitly emphasize the fault boundary. These loss functions do not enforce spatial continuity, resulting in discontinuities in fault identification. Therefore, in the embodiments of the present application, a new MBLoss function (as a specific example of the loss function) is introduced, which integrates BCE loss, Dice loss, and a novel fault boundary-aware loss. This loss function enhances the model's continuous identification of the fault boundary.

[0061] Specifically, in some embodiments of the present application, a hybrid boundary loss function needs to be constructed before executing S330. Specifically, the hybrid boundary loss function is constructed in the following manner: obtain the weight values corresponding to the pixel-level classification loss, the fault segmentation loss, and the fault boundary-aware loss respectively; perform weighted summation on the pixel-level classification loss, the fault segmentation loss, the fault boundary-aware loss, and the weight values to obtain the hybrid boundary loss function.

[0062] For example, in some embodiments of the present application, the hybrid boundary loss function includes pixel-level classification loss (BCE loss), fault segmentation loss (Dice loss), and fault boundary-aware loss.

[0063] The BCE loss function is used to measure the error of pixel-level classification and is defined as:

[0064] Where, p i,c is the predicted probability of class c at pixel i; y i is the true label of pixel i; w c is the class weight, which is used to balance the contributions of different classes in the segmentation task, especially when dealing with class imbalance problems.

[0065] The Dice loss function is widely used to evaluate the segmentation task and is defined as:

[0066] Among them, y true and y pred represent the true and predicted binary masks respectively; is used to prevent the denominator from being zero and ensure numerical stability, especially when there are no faults in the image.

[0067] To refine the fault boundary, a novel boundary-aware loss is proposed, which helps to maintain the continuity of the fault boundary. The boundary-aware loss calculates the difference between the gradients of the predicted segmentation and the true segmentation. In the boundary-aware loss function, vertical and horizontal Sobel kernels are utilized. By combining vertical and horizontal Sobel kernels when calculating the boundary-aware loss, it is ensured that boundary information is captured from different directions, improving the model's ability to identify subtle faults and weak discontinuities.

[0068] Finally, the total loss function for training L total is the weighted sum of the BCE loss, Dice loss, and boundary-aware loss L BoundaryAware :

[0069] Among them, W 1, W 2, and W 3 are the weights of various losses, used to balance the contributions of pixel-level classification, overall segmentation, and boundary preservation, ensuring the performance of the model under different task requirements. The weights can be set flexibly, and the embodiments of this application do not make specific limitations here.

[0070] In some embodiments of this application, the fault boundary-aware loss is obtained by the following method: obtaining the predicted values of the gradient magnitudes of the seismic fault prediction map in the horizontal and vertical directions; obtaining the true values of the gradient magnitudes of the seismic fault true map in the horizontal and vertical directions; determining the fault boundary-aware loss based on the predicted values of the gradient magnitudes and the true values of the gradient magnitudes.

[0071] For example, in some embodiments of this application, the Sobel operator is a discrete differential operator used to calculate an approximation of the image gradient, highlighting regions of high spatial frequency, corresponding to edges or boundaries in the image. In the context of seismic image segmentation, these edges correspond to fault boundaries. As Figure 6 shown, the Sobel kernel is applied to the horizontal (sobel x ) and vertical sobel yDirection. These kernels are applied to the predicted and ground truth segmentation maps by performing a convolution Conv on the images. The resulting convolution gives the gradient values in the horizontal and vertical directions, representing the edges in the predicted and ground truth images. Then, the gradient magnitude is calculated as follows:

[0072] where pred x and pred y are the gradients of the predicted segmentation along the horizontal and vertical axes respectively (i.e., the predicted gradient magnitude values of the seismic fault prediction map in the horizontal and vertical directions); target x and target y are the gradients of the ground truth segmentation (i.e., the true gradient magnitude values of the seismic fault ground truth map in the horizontal and vertical directions). Through the magnitudes of these gradients, the boundaries of the predicted and ground truth segmentations can be identified. Then, the boundary-aware loss L BoundaryAware is calculated as follows:

[0073] where Grad pred and Grad ture are the gradient magnitudes of the predicted segmentation and the ground truth segmentation respectively, calculated using the Sobel operator; is used to prevent the denominator from being zero. By minimizing this loss, the model refines its boundary predictions and focuses on the regions where the predicted boundaries overlap with the ground truth boundaries.

[0074] S330, calculating the loss value between the seismic fault prediction map and the seismic fault ground truth map corresponding to the seismic fault profile image using the hybrid boundary loss function; wherein, the hybrid boundary loss function includes pixel-level classification loss, fault segmentation loss, and fault boundary-aware loss.

[0075] For example, in some embodiments of the present application, the above-constructed hybrid boundary loss function L total is used to calculate the loss value for the seismic fault prediction map and the seismic fault ground truth map.

[0076] S340, optimizing the network model based on the loss value to determine the seismic fault segmentation model.

[0077] For example, in some embodiments of the present application, the parameters in the network model are optimized by the loss value obtained above until the loss value of the network model is lower than the loss threshold or the model reaches the number of iterations, and then the training is stopped to obtain the seismic fault segmentation model.

[0078] The following will exemplarily elaborate on the implementation process of seismic fault area detection performed by a terminal device provided in some embodiments of the present application in conjunction with the accompanying Figure 7 drawings.

[0079] Please refer to the accompanying Figure 7 drawings. Figure 7 FIG. is a flowchart of a method for seismic fault area detection provided in some embodiments of the present application. The method for seismic fault area detection may include: S710, obtaining a seismic fault image to be detected.

[0080] For example, in some embodiments of the present application, an initial image may be obtained first, and the same processing measures may be taken for the initial image through the process of processing the 3D seismic data volume in the above training model process to obtain the seismic fault image to be detected.

[0081] S720, inputting the seismic fault image to be detected into a seismic fault segmentation model to obtain a seismic fault prediction map.

[0082] For example, in some embodiments of the present application, the seismic fault image to be detected is input into the seismic fault segmentation model trained through the above method embodiments to obtain a seismic fault prediction map.

[0083] In addition, since BDCNet in the above text is creatively proposed in the present application, in order to evaluate the performance of BDCNet in the seismic fault segmentation task, four evaluation metrics commonly used in semantic segmentation are adopted: intersection over union (IoU), Dice coefficient, recall, and precision. These metrics are selected to evaluate the accuracy and robustness of the segmentation, especially for the complex and challenging fault structures in seismic images. The seismic fault recognition is regarded as a binary semantic segmentation problem, where the fault pixels are regarded as the positive class and the non-fault pixels are regarded as the negative class. The confusion matrix of this classification task is shown in Table 1.

[0084] Table 1

[0085] The IoU metric evaluates the overlap between the predicted segmentation and the ground truth segmentation by calculating the ratio of the intersection to the union of the predicted pixels and the ground truth pixels. A higher IoU value indicates better segmentation performance in identifying the fault boundary.

[0086]

[0087] The Dice coefficient is another widely used metric for evaluating the overlap between the predicted segmentation and the ground truth segmentation. It is similar to IoU but places more emphasis on balancing precision and recall. A higher Dice coefficient indicates better segmentation performance.

[0088]

[0089] Precision is used to measure how many of the pixels predicted to be faults are actually true fault pixels. It is calculated as the ratio of true positive pixels to the total number of pixels predicted to be faults.

[0090]

[0091] Recall is used to measure the ability of the model to correctly identify pixels belonging to the fault class. It is defined as the ratio of true positive pixels to the total number of ground truth fault pixels.

[0092]

[0093] These metrics provide a comprehensive evaluation, considering the accuracy of the fault segmentation model and the balance between correctly detecting faults and misdetecting faults.

[0094] Through ablation experiments, the impacts are mainly analyzed from the following aspects: deformable convolution (DConv): standard deformable convolution without boundary guidance, boundary deformable convolution (BDConv) for dynamically adjusting the receptive field, and mixed boundary loss (MBLoss): the loss function of this application for enhancing fault boundary sensitivity. Through comparative experiments, it is found that the BDCNet of this application is superior to existing segmentation models in seismic fault detection. By utilizing BDConv and MBLoss, BDCNet enhances the depiction of fault boundaries and improves the prediction accuracy, making it an effective method for seismic fault interpretation.

[0095] The following combines the attached Figure 8 Exemplarily elaborates the specific process of seismic fault area detection provided by some embodiments of this application.

[0096] Please refer to the attached Figure 8 , Figure 8 which is a flowchart of a method for seismic fault area detection provided by some embodiments of this application.

[0097] The following exemplarily elaborates the above process.

[0098] S810, obtain a 3D seismic data volume, and through processing the 3D seismic data volume, obtain a seismic fault profile image.

[0099] S820. Input the seismic fault profile image into the encoder of the network model to obtain the multi-scale semantic features of the image.

[0100] S830. Through the decoder of the network model, according to the multi-scale semantic features of the image and the stage feature maps output by each level of downsampling module, obtain the seismic fault prediction map.

[0101] S840. Use the hybrid boundary loss function to calculate the loss value between the seismic fault prediction map and the true seismic fault map corresponding to the seismic fault profile image.

[0102] S850. Optimize the network model based on the loss value to determine the seismic fault segmentation model.

[0103] S860. Obtain the seismic fault image to be detected.

[0104] S870. Input the seismic fault image to be detected into the seismic fault segmentation model to obtain the seismic fault prediction map.

[0105] It should be noted that the specific implementation processes of S810 to S870 can refer to the method embodiments provided above. To avoid repetition, the detailed descriptions are appropriately omitted here.

[0106] Please refer to Figure 9 , Figure 9 The block diagram showing the composition of the device for training the seismic fault segmentation model provided by some embodiments of the present application is shown. It should be understood that this device for training the seismic fault segmentation model corresponds to the above method embodiments and can execute each step involved in the above method embodiments. The specific functions of this device for training the seismic fault segmentation model can be seen in the descriptions above. To avoid repetition, the detailed descriptions are appropriately omitted here.

[0107] Figure 9The device for training an earthquake fault segmentation model includes at least one software function module that can be stored in a memory in the form of software or firmware or solidified in the device for training an earthquake fault segmentation model. The device for training an earthquake fault segmentation model includes: an encoding module 910, which is used to input an earthquake fault profile image into an encoder of a network model, and gradually extract multi-scale semantic features from the earthquake fault profile image through a multi-level downsampling module of the encoder; wherein each level of downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position; a decoding module 920, which is used to decode the network model, And each level of upsampling module and jump connection fuses each level of semantic features output by the encoder layer by layer, gradually restores the spatial resolution of the multi-scale semantic features and refines the fault boundaries to obtain an earthquake fault prediction map; a loss calculation module 930 is used to calculate the loss value between the earthquake fault prediction map and the earthquake fault real map corresponding to the earthquake fault profile image using a mixed boundary loss function; wherein the mixed boundary loss function includes pixel-level classification loss, fault segmentation loss and fault boundary perception loss; a model optimization module 940 is used to optimize the network model based on the loss value to determine the earthquake fault segmentation model.

[0108] Please refer to Figure 10 , Figure 10 The block diagram of the composition of the device for detecting earthquake fault regions provided by some embodiments of the present application is shown. It should be understood that the device for detecting earthquake fault regions corresponds to the above method embodiment and can execute each step involved in the above method embodiment. The specific functions of the device for detecting earthquake fault regions can be found in the description above. To avoid repetition, the detailed description is appropriately omitted here.

[0109] Figure 10 The device for detecting earthquake fault areas includes at least one software function module that can be stored in a memory in the form of software or firmware or solidified in the device for detecting earthquake fault areas. The device for detecting earthquake fault areas includes: an acquisition module 1010, used to acquire an earthquake fault image to be detected; a detection module 1020, used to input the earthquake fault image to be detected into an earthquake fault segmentation model to obtain an earthquake fault prediction map.

[0110] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method, and will not be described in detail here.

[0111] Some embodiments of the present application further provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the operations of the method corresponding to any of the above methods provided in the above embodiments.

[0112] Some embodiments of the present application also provide a computer program product, which includes a computer program. When the computer program is executed by a processor, it can implement the operations corresponding to any of the methods provided in the above embodiments of the above methods.

[0113] As Figure 11 shown, some embodiments of the present application provide an electronic device 1100, which includes: a memory 1110, a processor 1120, and a computer program stored on the memory 1110 and executable on the processor 1120. When the processor 1120 reads the program from the memory 1110 through the bus 1130 and executes the program, it can implement the methods of any of the above embodiments.

[0114] The processor 1120 can process digital signals and can include various computing architectures. For example, a complex instruction set computer architecture, a reduced instruction set computer architecture, or an architecture that implements a combination of multiple instruction sets. In some examples, the processor 1120 can be a microprocessor.

[0115] The memory 1110 can be used to store instructions executed by the processor 1120 or data related to the execution of the instructions. These instructions and / or data can include code for implementing some or all of the functions of one or more modules described in the embodiments of the present application. The processor 1120 of the present disclosure embodiment can be used to execute the instructions in the memory 1110 to implement the methods shown above. The memory 1110 includes a dynamic random access memory, a static random access memory, a flash memory, an optical memory, or other memories well known to those skilled in the art.

[0116] The above description is only for the embodiments of the present application and does not limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0117] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0118] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

Claims

1. A method for training an earthquake fault segmentation model, characterized in that: include: Inputting the earthquake fault profile image into the encoder of the network model, and gradually extracting the earthquake fault profile image through the multi-level downsampling module of the encoder to obtain multi-scale semantic features; wherein each level of downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position; Through the decoder of the network model, and the upsampling modules and skip connections of each level, the semantic features of each level output by the encoder are fused layer by layer, the spatial resolution of the multi-scale semantic features is gradually restored and the fault boundary is refined, so as to obtain an earthquake fault prediction map; The loss value between the earthquake fault prediction image and the earthquake fault real image corresponding to the earthquake fault profile image is calculated using a mixed boundary loss function; wherein the mixed boundary loss function includes pixel-level classification loss, fault segmentation loss and fault boundary perception loss; The network model is optimized based on the loss value to determine an earthquake fault segmentation model.

2. The method according to claim 1, characterized in that The multi-level downsampling module has N levels, N is a positive integer; each level of the downsampling module includes: a boundary deformable convolution module and a maximum pooling layer; the offset variable in the boundary deformable convolution module can be dynamically adjusted; the sampling position changes with the value of the offset variable; the encoder includes M levels of upsampling modules, M is a positive integer and NM=1; each level of the upsampling module in the M levels of upsampling modules includes the boundary deformable convolution module; Among them, the boundary deformable convolution module extracts features of the corresponding sampling positions in the seismic fault profile image by introducing the dynamically moving offset variable; the offset variable is determined by an offset learning function, and the offset learning function is related to the input image features and boundary offset parameters; the boundary offset parameters are related to the gradient of the input image features in the vertical direction and the gradient in the horizontal direction.

3. The method according to claim 2, characterized in that The step of inputting the earthquake fault section image into the encoder of the network model and gradually extracting the earthquake fault section image through the multi-level downsampling module of the encoder to obtain multi-scale semantic features includes: The i-th stage feature map is input into the i+1-th stage downsampling module, and the i+1-th stage feature map is output, i∈[0,N-1]; wherein, when i=0, the 0-th stage feature map is the earthquake fault profile image; when i=N-1, the N-th stage feature map is the multi-scale semantic feature.

4. The method according to claim 2 or 3, characterized in that The decoder of the network model, as well as each level of upsampling modules and skip connections, layer by layer fuse the semantic features of each level output by the encoder, gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundaries to obtain an earthquake fault prediction map, including: Input the multi-scale semantic features and the N-1th stage feature map into the Mth level upsampling module, and output the Mth sampling feature; The Nj-th stage feature map and the M-th sampling feature are input into the M-j+1-th level upsampling module, and the M-j+1-th sampling feature is output, j∈[2,N-1]; when j=N-1, the first sampling feature is the earthquake fault prediction map.

5. The method according to any one of claims 1 to 3, characterized in that The mixed boundary loss function is constructed as follows: Obtaining weight values ​​corresponding to the pixel-level classification loss, the fault segmentation loss, and the fault boundary perception loss respectively; The pixel-level classification loss, the fault segmentation loss, the fault boundary perception loss and the weight value are weightedly summed to obtain the mixed boundary loss function.

6. The method according to any one of claims 1 to 3, characterized in that The fault boundary perception loss is obtained by the following method: Obtaining predicted values ​​of the gradient amplitude of the earthquake fault prediction map in the horizontal and vertical directions; Obtaining the true values ​​of the gradient amplitudes of the earthquake fault true map in the horizontal and vertical directions; The fault boundary perception loss is determined based on the gradient amplitude prediction value and the gradient amplitude true value.

7. A method for detecting earthquake fault regions, characterized in that: include: Acquire an image of the earthquake fault to be detected; The earthquake fault image to be detected is input into an earthquake fault segmentation model to obtain an earthquake fault prediction map; wherein the earthquake fault segmentation model is obtained by the method described in any one of claims 1-6.

8. A device for training an earthquake fault segmentation model, characterized in that: include: An encoding module, used for inputting the earthquake fault profile image into the encoder of the network model, and gradually extracting the earthquake fault profile image through the multi-level downsampling module of the encoder to obtain multi-scale semantic features; wherein each level of downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position; A decoding module, used to gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundaries through the decoder of the network model, as well as the upsampling modules and skip connections of each level to fuse the semantic features of each level output by the encoder layer by layer, so as to obtain an earthquake fault prediction map; A loss calculation module, used for calculating the loss value between the earthquake fault prediction image and the earthquake fault real image corresponding to the earthquake fault profile image by using a mixed boundary loss function; wherein the mixed boundary loss function includes pixel-level classification loss, fault segmentation loss and fault boundary perception loss; A model optimization module is used to optimize the network model based on the loss value to determine the earthquake fault segmentation model.

9. A device for detecting earthquake fault regions, characterized in that: include: An acquisition module, used for acquiring an image of a to-be-detected earthquake fault; A detection module is used to input the earthquake fault image to be detected into an earthquake fault segmentation model to obtain an earthquake fault prediction map; wherein the earthquake fault segmentation model is obtained by the method described in any one of claims 1-6.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program executes the method according to any one of claims 1 to 7 when being run by the processor.

Images