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

Abstract

The present application relates to the technical field of seismic image processing, and specifically provides a method, apparatus, and device for fault detection based on an earthquake fault segmentation model. The method comprises: inputting an earthquake fault profile image into an encoder of a network model, and gradually extracting multi-scale semantic features from the earthquake fault profile image through a multi-level downsampling module of the encoder; fusing each level of semantic features output by the encoder layer by layer through a decoder of the network model, as well as each level of upsampling modules and jump connections, gradually restoring the spatial resolution of the multi-scale semantic features and refining the fault boundaries to obtain an earthquake fault prediction map; calculating 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; optimizing the network model based on the loss value to determine the earthquake fault segmentation model. The embodiments of the present application can achieve accurate prediction of earthquake faults.

Description

Technical Field

[0001] The present application relates to the technical field of seismic image processing, and in particular to a method, apparatus and device for fault detection based on a seismic fault segmentation model. Background Art

[0002] Faults are structures where crustal rock fractures due to stress reaching a certain intensity, with significant relative movement along the fracture surface. It is generally believed that earthquakes are often caused by fault activity, and earthquakes can also create new faults. Therefore, earthquake fault analysis plays an important role.

[0003] The most commonly used method for seismic fault segmentation currently relies on manual visual interpretation, which relies on the manual annotation of geologists. This method involves preprocessing the raw seismic data, including noise suppression and calculation of coherence or curvature volume attributes, to highlight fault features. Geologists then manually delineate fault lines based on areas of reflection event anomalies in the seismic data. This method relies on expert experience and subjective judgment, is time-consuming, and is susceptible to personal bias.

[0004] Therefore, how to provide a technical solution for an efficient and accurate method for automatic detection of earthquake fault areas has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The purpose of some embodiments of the present application is to provide a method, device and equipment for fault detection based on an earthquake fault segmentation model. Through the technical solutions of the embodiments of the present application, accurate and efficient detection of earthquake fault areas can be achieved.

[0006] In a first aspect, some embodiments of the present application provide a method for training an earthquake fault segmentation model, the method comprising: inputting an earthquake fault profile image into an encoder of a network model, and gradually extracting 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; through the decoder of the network model, and each level of upsampling module and jump connection, each level of semantic features output by the encoder is fused layer by layer, and the spatial resolution of the multi-scale semantic features is gradually restored and the fault boundaries are refined to obtain an earthquake fault prediction map; using a hybrid boundary loss function to calculate the loss value between the earthquake fault prediction map and the earthquake fault real map; wherein, the hybrid boundary loss function includes pixel-level classification loss, fault segmentation loss and fault boundary perception loss; based on the loss value, the network model is optimized to determine the earthquake fault segmentation model.

[0007] Some embodiments of the present application implement dynamic image sampling by inputting earthquake fault profile images into the encoder of a network model. A decoder is then used to process the multi-scale semantic features of the image output by the encoder to obtain an earthquake fault prediction map. Finally, a hybrid boundary loss function is used to calculate the loss between the earthquake fault prediction map and the actual earthquake fault map to optimize the network model and obtain an earthquake fault segmentation model. Some embodiments of the present application can accurately train the designed network model using earthquake fault profile images, thereby improving the detection effect and robustness of the model.

[0008] In some embodiments, the multi-level downsampling module has N levels, where 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 is dynamically adjustable; the sampling position changes with the value of the offset variable; the decoder includes M levels of upsampling modules, where 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;

[0009] In which, the boundary deformable convolution module extracts features from 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, which is related to the input image features and the boundary offset parameters; the boundary offset parameters are related to the vertical and horizontal gradients of the input image features.

[0010] Some embodiments of the present application provide support for model training by setting different boundary deformable convolution modules in multi-level downsampling modules and upsampling modules to facilitate accurate processing of earthquake fault profile images.

[0011] In some embodiments, the earthquake fault profile image is input into the encoder of the network model, and the multi-scale semantic features are gradually extracted from the earthquake fault profile image through the multi-stage downsampling module of the encoder, including: inputting the i-th stage feature map into the i+1-th stage downsampling module, and outputting the i+1-th stage feature map, i∈[0,N-1]; wherein, when i=0, the 0th stage feature map is the earthquake fault profile image; when i=N-1, the Nth stage feature map is the multi-scale semantic features.

[0012] Some embodiments of the present application process feature maps at different stages through different downsampling modules to obtain multi-scale semantic features of the image, thereby achieving efficient image processing.

[0013] In some embodiments, the decoder of the network model, and the semantic features of each level output by the encoder are fused layer by layer through the upsampling module of each level and the jump connection, so as to gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundaries to obtain an earthquake fault prediction map, including: inputting the multi-scale semantic features and the feature map of the N-1th stage into the Mth level upsampling module, and outputting the Mth sampling feature; inputting the feature map of the Njth stage and the Mth sampling feature into the M-j+1th level upsampling module, and outputting the M-j+1th sampling feature, j∈[2,N-1]; when j=N-1, the first sampling feature is the earthquake fault prediction map.

[0014] Some embodiments of the present application jointly process the sampling features of different stages and the feature maps of different stages of downsampling output through a multi-stage upsampling module to output an earthquake fault prediction map, thereby achieving effective prediction of earthquake faults.

[0015] 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 of the pixel-level classification loss, the fault segmentation loss, and the fault boundary perception loss with the weight values to obtain the hybrid boundary loss function.

[0016] Some embodiments of the present application construct a loss function through multiple losses to achieve accurate optimization of the earthquake fault segmentation model during the training process.

[0017] In some embodiments, the fault boundary perception loss is obtained by the following method: obtaining the gradient amplitude prediction values of the earthquake fault prediction map in the horizontal and vertical directions; obtaining the gradient amplitude true values of the earthquake fault real map in the horizontal and vertical directions; and determining the fault boundary perception loss based on the gradient amplitude prediction values and the gradient amplitude true values.

[0018] Some embodiments of the present application determine the fault boundary perception loss through the predicted values and true values of the gradient amplitude of the earthquake fault prediction map in different directions, which can subsequently improve the accuracy of model training.

[0019] In the second aspect, some embodiments of the present application provide a method for detecting an earthquake fault area, comprising: obtaining an earthquake fault image to be detected; inputting 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 any method embodiment in the first aspect.

[0020] Some embodiments of the present application can implement earthquake fault detection on images to be detected through a trained earthquake fault segmentation model, with high detection efficiency and accuracy.

[0021] In a third aspect, some embodiments of the present application provide a device for training an earthquake fault segmentation model, an encoding module for inputting an earthquake fault profile image into an encoder of a network model, and gradually extracting 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 for gradually restoring the spatial resolution of the multi-scale semantic features and refining the fault boundaries through a decoder of the network model, as well as each level of upsampling module and a jump connection, to obtain an earthquake fault prediction map; a loss calculation module for calculating the loss value between the earthquake fault prediction map and the earthquake fault real map corresponding to the earthquake fault profile image 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; a model optimization module for optimizing the network model based on the loss value to determine the earthquake fault segmentation model.

[0022] In a fourth aspect, a device for detecting an earthquake fault area includes: an acquisition module for acquiring an earthquake fault image to be detected; a detection module for inputting 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 any method embodiment in the first aspect.

[0023] In a fifth aspect, some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any embodiment of the first aspect.

[0024] In a sixth aspect, some embodiments of the present application provide an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor can implement a method as described in any embodiment of the first aspect when executing the program.

[0025] In a seventh aspect, some embodiments of the present application provide a computer program product, comprising a computer program, wherein the computer program, when executed by a processor, can implement the method described in any embodiment of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of some embodiments of the present application, the following is a brief introduction to the drawings required for use in some embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 Schematic diagram of the BDConv process provided for some embodiments of the present application;

[0028] Figure 2 A schematic diagram of the calculation process of BDConv provided for some embodiments of the present application;

[0029] Figure 3 A flow chart of a method for training an earthquake fault segmentation model provided in some embodiments of the present application;

[0030] Figure 4 A comparison diagram of image resolution distribution before and after cropping provided in some embodiments of the present application;

[0031] Figure 5 A diagram of a network model structure provided for some embodiments of the present application;

[0032] Figure 6 Schematic diagram of gradient extraction using the Sobel operator provided in some embodiments of the present application;

[0033] Figure 7 One of the flow charts of the method for detecting earthquake fault regions provided in some embodiments of the present application;

[0034] Figure 8 Flowchart 2 of the method for detecting earthquake fault regions provided in some embodiments of the present application;

[0035] Figure 9 A block diagram of the apparatus for training an earthquake fault segmentation model provided in some embodiments of the present application;

[0036] Figure 10 A block diagram of an apparatus for detecting earthquake fault regions provided in some embodiments of the present application;

[0037] Figure 11 A schematic diagram of an electronic device is provided for some embodiments of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in some embodiments of the present application will be described below in conjunction with the drawings in some embodiments of the present application.

[0039] 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 or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0040] In related technologies, machine learning-based methods manually extract seismic attributes and combine them with statistical learning models for fault identification. Common methods include support vector machines (SVMs) and multi-layer perceptrons (MLPs). Specifically, they extract multiple attributes from seismic data, such as coherence coefficients and azimuthal gradients. Using machine learning models like SVMs or MLPs, weight parameters are optimized through backpropagation to identify faults. Seismic data is then input into the model, and a fault probability map is output. This approach requires extensive feature engineering and parameter tuning and is sensitive to noise. In recent years, deep learning models, particularly convolutional neural networks, have made significant progress in seismic fault segmentation. Among them, the U-Net model uses an encoder-decoder architecture with skip connections to preserve low-level spatial details and incorporate high-level semantic information for fault segmentation. However, it still has limitations in capturing long-range dependencies and maintaining boundary continuity. Standard convolution operations are limited to a local receptive field, making it difficult to establish cross-scale geological structural correlations, resulting in insufficient ability to capture long-range dependencies. In addition, the existing earthquake fault segmentation model does not pay attention to the continuity requirements of the fault boundary. The model is not sensitive enough to weak continuity boundaries, resulting in discontinuous faults in the segmentation results.

[0041] In view of this, some embodiments of the present application adopt a method for training an earthquake fault segmentation model. This 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 loss types to obtain an earthquake fault segmentation model. Some embodiments of the present application can obtain a highly accurate earthquake fault segmentation model, provide model support for earthquake fault detection, and also improve the efficiency of earthquake fault detection.

[0042] To improve the accuracy of fault identification, this application proposes a boundary deformable convolution (BDConv) module that can dynamically adjust the sampling position. The BDConv module in the network model provided in some embodiments of this application is significantly different from that used in traditional technologies. Therefore, the following first illustrates the data processing process of the BDConv module in this application.

[0043] Traditional convolutional layers use a fixed grid receptive field, which limits their adaptability when detecting irregular geological structures. In seismic images, faults often exhibit varying strikes, displacements, and lateral continuity. Standard convolution operations, constrained by their fixed grid sampling pattern, struggle to adaptively capture complex fault morphologies. Furthermore, due to the local receptive field nature of convolutional operations, the model is unable to effectively establish long-range feature dependencies across scales, leading to discontinuous fault segmentation results.

[0044] 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 variable; the offset variable is determined by an offset learning function, which is related to the input image features and the boundary offset parameters; the boundary offset parameters are related to the vertical gradient and the horizontal gradient of the input image features.

[0045] To overcome these problems, this application proposes BDConv, which uses a learnable offset and boundary-aware attention mechanism to dynamically adjust the convolution sampling position. This enables the network to focus on fault boundaries more effectively while expanding its receptive field in an adaptive manner. Figure 1 As shown in Figure 2, 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 fixed-space kernels, deformable convolution learns to move the sampling points to the most relevant areas, improving its ability to capture uneven features.

[0046] Faults in seismic data typically appear as long, narrow lines, with their most prominent features often aligned in the vertical direction. Therefore, a vertical Sobel kernel is used in BDConv to maintain continuity in fault boundary identification. The vertical Sobel kernel effectively emphasizes gradients in the vertical direction, which is consistent with the structure of most faults, ensuring that the network focuses on fault edges while maintaining continuity across long fault segments. By dynamically adjusting the receptive field and focusing on the vertical direction, the model is able to better capture long, continuous fault lines and maintain continuity during segmentation.

[0047] The calculation process of BDConv is as follows Figure 2 As shown in Figure 2. The standard convolution operation samples features at fixed positions within the K×K convolution kernel. The sampling formula is:

[0048]

[0049] in, p 0 represents the current position, w k is the weight of the kth position in the convolution kernel, x k is in positionp 0+ p k The input eigenvalue at p k Indicates a fixed sampling offset.

[0050] In BDConv, each sampling position is offset by a learnable p k This allows the receptive field to be modified dynamically. The formula for the deformable convolution is:

[0051]

[0052] in, p k is a learned offset (as a specific example of an offset variable) that dynamically moves the sampling position, enabling the network model to adaptively focus on more relevant regions in the feature map.

[0053] To improve the extraction of key fault boundaries during convolution, this embodiment of the application also introduces a boundary-aware directional attention mechanism. This mechanism refines the network model's focus on fault edges, guides the offset learning process, and ensures that the deformation of the receptive field is aligned with the fault structure. Furthermore, this mechanism enhances the ability to extract features along fault discontinuities, especially for small faults and slight displacements, while increasing directional sensitivity. This allows the model to effectively distinguish fault edges from surrounding noise, ultimately improving segmentation performance.

[0054] 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 is x (as a specific example of input image features) and boundary map A (as a specific example of boundary offset parameters) to adjust the offset. The attention map A is calculated using a directional Sobel filter to emphasize high gradient areas and ensure that the offset points to the fault boundary:

[0055]

[0056] 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.

[0057] 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 exemplified, wherein the terminal device may be a server device.

[0058] Please see the attached Figure 3 , Figure 3 A flow chart of a method for training an earthquake fault segmentation model is provided for some embodiments of the present application. The method for training an earthquake fault segmentation model may include:

[0059] S310, inputting the earthquake fault profile image into the encoder of the network model, and gradually extracting 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.

[0060] For example, in some embodiments of the present application, a training dataset is first obtained for training a network model. This training dataset includes seismic fault profile images and their corresponding ground-truth values. This seismic fault profile image is then input into the designed network model's encoder for feature extraction. Each downsampling module in the encoder employs a dynamic adjustment mechanism that adaptively optimizes sampling positions based on the local characteristics of the fault structure, effectively preserving key geological information and thereby obtaining multi-scale semantic features of the seismic fault profile image.

[0061] The following is an example of a method for obtaining an earthquake fault profile image.

[0062] In this application, the Thebe dataset, a 3D seismic data volume originally acquired for geological fault interpretation, was used. The raw data in the dataset was originally stored in a 3D voxel format. This application used a preprocessing pipeline to process the data to train a deep learning model (i.e., the network model of this application).

[0063] First, we used the Python NumPy library to load the 3D seismic data volume and slice it equally spaced along the depth dimension, dividing the 3D seismic data volume and its corresponding 3D fault annotations into 2D slices. Each seismic data slice was saved as a PNG seismic image file and normalized to the [0, 1] range using the "seismic" color map. The fault annotation slices were directly converted to binary images.

[0064] Initial visualization revealed a significant amount of redundant data in the binary image, typically manifested as large white areas. To improve data quality and relevance, an adaptive thresholding method was employed to detect white areas. Specifically, the image was converted to grayscale, and a mask was generated using an empirical threshold. The bounding box of the non-white areas was then determined by calculating the effective range of the mask in the row and column directions. In particular, to ensure consistency between the seismic data and its annotations, both were cropped simultaneously using identical bounding box parameters.

[0065] 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, Figure 4 The left image in the figure shows the scale before cropping, and the right image shows the scale after cropping. After these preprocessing steps, the resulting dataset contains high-quality 2D seismic images (i.e., earthquake fault profile images) that can be directly used to train network models for fault segmentation tasks.

[0066] 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 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.

[0067] For example, in some embodiments of the present application, Figure 5 The network model structure is shown in Figure 1. This network model includes an encoder and decoder. The encoder includes five (N=5) downsampling modules, namely BDC_1, BDC_2, and BDC_3, from bottom to top: the first downsampling module, the second downsampling module, and the fifth downsampling module. Each downsampling module includes a BDConv module (a specific example of a boundary deformable convolution module) and a Maxpooling layer (a specific example of a maximum pooling layer). BDC_1 also includes a convolutional layer (Conv). Specifically, the BDConv module dynamically adjusts the convolution sampling position using a learnable offset and boundary-aware attention mechanism, enabling the network model to adaptively focus on fault boundary regions. Specifically, the first downsampling module, BDC_1, processes the original (640, 640, 3) fault profile image to extract low-level features, such as fault edge texture. As the network depth increases, the 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, ultimately forming a highly abstract feature representation. Figure 5 The numbers on the left side of the figure represent the changes in spatial resolution and number of channels during the downsampling process.

[0068] Figure 5 The decoder of the network model is composed of 4 BDC_2 on the right. The decoder part gradually restores the spatial resolution by upsampling. Each level of the decoder not only receives the feature input of the previous level, but also receives the feature input of the previous level through the jump connection (i.e. Figure 5 The Skip Connect in

[15] fuses the feature maps from the corresponding encoder stages. This design effectively addresses the vanishing gradient problem in deep networks while preserving the spatial details of the fault structure. Furthermore, all convolutions in the encoder are BDConv, ensuring that the fused features have consistent boundary-aware properties. It should be noted that the number of downsampling and upsampling modules in the encoder and decoder of the network model can be adjusted based on the actual application scenario and is not specifically limited in this embodiment of the present application.

[0069] In some embodiments of the present application, S310 may include: inputting the i-th stage feature map into the i+1-th level downsampling module, and outputting the i+1-th stage feature map, i∈[0,N-1]; wherein, when i=0, the 0-th stage feature map is the seismic fault profile image; when i=N-1, the N-th stage feature map is the multi-scale semantic feature.

[0070] For example, in some embodiments of the present application, the input fault profile image is processed step by step by using the feature map output by the previous downsampling module as the input of the next downsampling module. During each stage of processing, the sampling position of the downsampling module is not fixed and can be dynamically adjusted to focus on the fault boundary area.

[0071] S320: The decoder of the network model, along with each level of upsampling modules and skip connections, layer by layer fuses the semantic features of each level output by the encoder, gradually restoring the spatial resolution of the multi-scale semantic features and refining the fault boundaries to obtain an earthquake fault prediction map. The decoder includes M levels of upsampling modules, where M is a positive integer and NM = 1; each of the M levels of upsampling modules includes the boundary deformable convolution module.

[0072] For example, in some embodiments of the present application, the multi-scale semantic features output by the encoder are fused layer by layer through a decoder, upsampling, and skip connections to gradually restore the spatial resolution and refine the fault boundaries to obtain a high-precision earthquake fault prediction map.

[0073] In some embodiments of the present application, S320 may include: inputting the multi-scale semantic features and the N-1th stage feature map into the Mth level upsampling module, and outputting the Mth sampling feature; inputting the Njth stage feature map and the Mth sampling feature into the M-j+1th level upsampling module, and outputting the M-j+1th sampling feature, j∈[2,N-1]; when j=N-1, the first sampling feature is the earthquake fault prediction map.

[0074] For example, in some embodiments of the present application, Figure 5 Taking the network model as an example, the BDC_2 on the right is divided from bottom to top into the first upsampling module, the second upsampling module, and the fourth upsampling module, that is, M=4. The input data of the fourth 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 upsampling module is the fourth sampling features output by the fourth upsampling module and the third-stage feature map output by BDConv in the third downsampling module. And so on, until the earthquake fault prediction map is output.

[0075] Traditional deep learning models for fault segmentation typically 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 segmentation balance but does not explicitly emphasize fault boundaries. 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 a loss function) is introduced, which integrates BCE loss, Dice loss, and a novel fault boundary-aware loss. This loss function enhances the model's recognition of the continuity of fault boundaries.

[0076] Specifically, in some embodiments of the present application, it is necessary to construct a mixed boundary loss function before executing S330. Specifically, the mixed 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 of the pixel-level classification loss, the fault segmentation loss, and the fault boundary perception loss with the weight value to obtain a mixed boundary loss function.

[0077] 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 perception loss.

[0078] The BCE loss function is used to measure the error of pixel-level classification and is defined as:

[0079]

[0080] in, p i,c is the predicted probability of category 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 contribution of different classes in the segmentation task, especially when dealing with class imbalance problems.

[0081] The Dice loss function is widely used to evaluate segmentation tasks and is defined as:

[0082]

[0083] in, y true and y pred Represent the true and predicted binary masks respectively; This is used to prevent the denominator from being zero, ensuring numerical stability, especially when there are no faults in the image.

[0084] To refine fault boundaries, a novel boundary-aware loss is proposed, which helps maintain the continuity of fault boundaries. The boundary-aware loss calculates the difference between the gradient of the predicted segmentation and the gradient of the true segmentation. Vertical and horizontal Sobel kernels are utilized in the boundary-aware loss function. By combining vertical and horizontal Sobel kernels in the calculation of the boundary-aware loss, boundary information is captured from different directions, improving the model's ability to detect subtle faults and weak discontinuities.

[0085] Finally, the total loss function used for training L total are BCE loss, Dice loss and boundary perception loss L BoundaryAware The weighted sum of:

[0086]

[0087] in, W 1. W 2 and W 3 is the weight of each loss, which is used to balance the contributions of pixel-level classification, overall segmentation, and boundary preservation to ensure the performance of the model under different task requirements. The weights can be set flexibly and are not specifically limited in this embodiment of the application.

[0088] In some embodiments of the present application, the fault boundary perception loss is obtained by the following method: obtaining the gradient amplitude prediction value of the earthquake fault prediction map in the horizontal and vertical directions; obtaining the gradient amplitude true value of the earthquake fault real map in the horizontal and vertical directions; determining the fault boundary perception loss based on the gradient amplitude prediction value and the gradient amplitude true value.

[0089] For example, in some embodiments of the present application, the Sobel operator is a discrete differential operator that is 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. Figure 6 As shown, the Sobel kernel is applied to the horizontal (sobel x ) and vertical sobel y direction. These kernels are applied to the predicted and true segmentation maps by convolving the image. The resulting convolution gives gradient values in the horizontal and vertical directions, representing the edges in the predicted and true images. The gradient magnitude is then calculated as follows:

[0090]

[0091] in, pred x and pred y are the gradients of the predicted segmentation along the horizontal and vertical axes, respectively (i.e., the predicted values of the gradient amplitude of the earthquake fault prediction map in the horizontal and vertical directions); target x and target y is the gradient of the true segmentation (i.e., the true value of the gradient amplitude of the earthquake fault true map in the horizontal and vertical directions). Through the amplitude of these gradients, the boundary between the predicted and true segmentation can be identified. Then the boundary perception loss is calculated L BoundaryAware :

[0092]

[0093] in, Grad pred and Grad ture are the gradient amplitudes of the predicted segmentation and the true segmentation, respectively, calculated by the Sobel operator; Used to prevent the denominator from being zero. By minimizing this loss, the model refines its boundary predictions, focusing on areas where the predicted boundary overlaps with the true boundary.

[0094] S330, using a hybrid boundary loss function to calculate the loss value between the earthquake fault prediction image and the earthquake fault real image corresponding to the earthquake fault profile image; wherein the hybrid boundary loss function includes pixel-level classification loss, fault segmentation loss and fault boundary perception loss.

[0095] For example, in some embodiments of the present application, the hybrid boundary loss function constructed above is used L total The earthquake fault prediction map and the earthquake fault real map are calculated to obtain the loss value.

[0096] S340: Optimize the network model based on the loss value to determine an earthquake fault segmentation model.

[0097] For example, in some embodiments of the present application, the parameters in the network model are optimized by the loss value obtained by the above calculation until the loss value of the network model is lower than the loss threshold or the model reaches the number of iterations, and the training is stopped to obtain the earthquake fault segmentation model.

[0098] The following is combined with Figure 7 The implementation process of earthquake fault area detection performed by a terminal device provided in some embodiments of the present application is exemplified.

[0099] Please see the attached Figure 7 , Figure 7 A flow chart of a method for detecting an earthquake fault region is provided for some embodiments of the present application. The method for detecting an earthquake fault region may include:

[0100] S710: Acquire an image of the earthquake fault to be detected.

[0101] For example, in some embodiments of the present application, an initial image may be first acquired, and the same processing measures may be taken on the initial image through the process of processing the 3D seismic data volume in the above-mentioned training model process to obtain the seismic fault image to be detected.

[0102] S720: Input the earthquake fault image to be detected into an earthquake fault segmentation model to obtain an earthquake fault prediction map.

[0103] For example, in some embodiments of the present application, the earthquake fault image to be detected is input into the earthquake fault segmentation model trained by the above method embodiment to obtain an earthquake fault prediction map.

[0104] Furthermore, since the BDCNet proposed above is a creative invention of this application, four evaluation metrics commonly used in semantic segmentation were used to evaluate BDCNet's performance in the earthquake fault segmentation task: intersection over union (IoU), Dice coefficient, recall, and precision. These metrics were selected to assess the accuracy and robustness of segmentation, particularly for the complex and challenging fault structures in seismic images. Earthquake fault identification was formulated as a binary semantic segmentation problem, where fault pixels were considered the positive class and non-fault pixels were considered the negative class. The confusion matrix for this classification task is shown in Table 1.

[0105] Table 1

[0106]

[0107] The IoU metric evaluates the overlap between the predicted segmentation and the true segmentation by calculating the ratio of the intersection over union of the predicted and true pixels. A higher IoU value indicates better segmentation performance in identifying fault boundaries.

[0108]

[0109] The Dice coefficient is another widely used metric for evaluating the overlap between predicted and true segmentations. It is similar to IoU but focuses more on balancing precision and recall. A higher Dice coefficient indicates better segmentation performance.

[0110]

[0111] Precision is a measure of how many of the pixels predicted as faults are actually true fault pixels. It is calculated as the ratio of true positive pixels to the total number of pixels predicted as faults.

[0112]

[0113] 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 true fault pixels.

[0114]

[0115] These metrics provide a comprehensive assessment that considers the accuracy of the fault segmentation model and the balance between correctly and incorrectly detected faults.

[0116] Ablation experiments analyzed the impact of the following key aspects: Deformable Convolution (DConv): Standard deformable convolution without boundary guidance; Boundary Deformable Convolution (BDConv), which dynamically adjusts the receptive field; and Mixed Boundary Loss (MBLoss): This application's loss function enhances fault boundary sensitivity. Comparative experiments showed that this application's BDCNet outperformed existing segmentation models in earthquake fault detection. By utilizing BDCv and MBLoss, BDCNet enhanced fault boundary delineation and improved prediction accuracy, making it an effective method for earthquake fault interpretation.

[0117] The following is combined with Figure 8 The specific process of earthquake fault area detection provided by some embodiments of the present application is exemplified.

[0118] Please see the attached Figure 8 , Figure 8 A flow chart of a method for detecting earthquake fault regions is provided for some embodiments of the present application.

[0119] The above process is explained below as an example.

[0120] S810 , acquiring a 3D seismic data volume, and obtaining a seismic fault profile image by processing the 3D seismic data volume.

[0121] S820: Input the earthquake fault profile image into the encoder of the network model to obtain multi-scale semantic features of the image.

[0122] S830, through the decoder of the network model, based on the multi-scale semantic features of the image and the stage feature map output by each level of downsampling module, an earthquake fault prediction map is obtained.

[0123] S840: Calculate the loss value between the earthquake fault prediction image and the earthquake fault real image corresponding to the earthquake fault profile image using a mixed boundary loss function.

[0124] S850: Optimize the network model based on the loss value to determine the earthquake fault segmentation model.

[0125] S860: Acquire an image of the earthquake fault to be detected.

[0126] S870: Input the earthquake fault image to be detected into the earthquake fault segmentation model to obtain an earthquake fault prediction map.

[0127] It should be noted that the specific implementation process of S810 to S870 can refer to the method embodiment provided above. To avoid repetition, detailed description is appropriately omitted here.

[0128] Please refer to Figure 9 , Figure 9A block diagram of the components of an apparatus for training an earthquake fault segmentation model provided in some embodiments of the present application is shown. It should be understood that the apparatus for training an earthquake fault segmentation model corresponds to the aforementioned method embodiment and is capable of executing each step involved in the aforementioned method embodiment. The specific functions of the apparatus for training an earthquake fault segmentation model can be found in the description above, and a detailed description is omitted here to avoid repetition.

[0129] Figure 9 The 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 skip 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; the 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; the model optimization module 940 is used to optimize the network model based on the loss value to determine the earthquake fault segmentation model.

[0130] Please refer to Figure 10 , Figure 10 A block diagram of the components of an apparatus for detecting earthquake fault regions provided by some embodiments of the present application is shown. It should be understood that the apparatus for detecting earthquake fault regions corresponds to the aforementioned method embodiments and is capable of performing each of the steps involved in the aforementioned method embodiments. The specific functions of the apparatus for detecting earthquake fault regions can be found in the description above, and a detailed description is omitted here to avoid repetition.

[0131] Figure 10 The device for detecting earthquake fault areas includes at least one software functional 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.

[0132] Those skilled in the art will 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.

[0133] 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 corresponding to any of the above methods provided in the above embodiments.

[0134] Some embodiments of the present application further provide a computer program product, which includes a computer program, wherein when the computer program is executed by a processor, it can implement the operations corresponding to any of the above methods provided in the above embodiments.

[0135] like Figure 11 As 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 in the memory 1110 and executable on the processor 1120, wherein the processor 1120 can implement a method as described in any of the above embodiments when reading the program from the memory 1110 through the bus 1130 and executing the program.

[0136] Processor 1120 can process digital signals and can include various computing architectures, such as 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, processor 1120 can be a microprocessor.

[0137] The memory 1110 can be used to store instructions executed by the processor 1120 or data related to the execution of instructions. These instructions and / or data may include code for implementing some or all functions of one or more modules described in the embodiments of this application. The processor 1120 of the embodiment of the present disclosure can be used to execute the instructions in the memory 1110 to implement the method shown above. The memory 1110 includes dynamic random access memory, static random access memory, flash memory, optical storage, or other memory known to those skilled in the art.

[0138] The foregoing is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures.

[0139] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0140] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A method for training an earthquake fault segmentation model, characterized in that: include: An earthquake fault profile image is input into an encoder of a network model, and multi-scale semantic features are gradually extracted from the earthquake fault profile image through a multi-level downsampling module of the encoder. The semantic features of each level output by the encoder are layer-by-layer fused through a decoder of the network model, as well as upsampling modules at each level and skip connections, to gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundaries, thereby obtaining an earthquake fault prediction map. Each upsampling module and each downsampling module includes: a boundary deformable convolution module; the boundary deformable convolution module uses a learnable offset and a boundary-aware attention mechanism to dynamically adjust the convolution sampling position in the earthquake fault profile image, allowing the sampling position to deviate from the fixed sampling grid structure and moving the sampling point to the most relevant fault edge area; the boundary deformable convolution module extracts features from the sampling position by introducing the dynamically moving offset variable; the offset variable is determined by an offset learning function, the offset learning function is related to the input image feature and a boundary offset parameter; the boundary offset parameter is related to the vertical gradient and the horizontal gradient of the input image feature; Calculating a loss value between the earthquake fault prediction map and the earthquake fault real map corresponding to the earthquake fault profile image 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; the fault boundary perception loss is calculated based on vertical and horizontal Sobel kernels of the earthquake fault prediction map; The network model is optimized based on the loss value to determine an earthquake fault segmentation model.

2. The method according to claim 1, wherein The multi-level downsampling module has N levels, where N is a positive integer; each level of the downsampling module includes: a boundary deformable convolution module and a maximum pooling layer; the sampling position changes with the value of the offset variable; the decoder includes M levels of upsampling modules, where 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.

3. The method according to claim 2, wherein The method includes inputting the earthquake fault section image into the encoder of the network model, and gradually extracting multi-scale semantic features from the earthquake fault section image through the multi-stage downsampling module of the encoder, including: 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, wherein: The decoder of the network model, as well as the upsampling module and the skip connection of each level, fuse the semantic features of each level output by the encoder layer by layer, gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundary 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, wherein The hybrid 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 hybrid boundary loss function.

6. The method according to any one of claims 1 to 3, wherein The fault boundary perception loss is obtained by the following method: Obtaining predicted values of gradient amplitudes of the earthquake fault prediction map in the horizontal and vertical directions; Obtaining 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: The apparatus is configured to execute the method according to claim 1, comprising: An encoding module is configured to input the earthquake fault profile image into an encoder of the 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 the downsampling module in the encoder adopts a dynamic adjustment mechanism to adaptively optimize the sampling position; A decoding module is used to gradually restore the spatial resolution of the multi-scale semantic features and refine the fault boundaries by fusing the semantic features of each level output by the encoder layer by layer through the decoder of the network model, as well as the upsampling module and skip connection of each level, to obtain an earthquake fault prediction map; a loss calculation module, configured to calculate a loss value between the earthquake fault prediction image and the earthquake fault real image corresponding to the earthquake fault profile image 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; 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 areas, characterized in that: include: An acquisition module, used for acquiring an image of an earthquake fault to be detected; 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 the processor runs the computer program.

Images