Fault identification method and device

Abstract

The invention discloses a fault identification method and device. The method comprises the following steps: acquiring seismic data; cutting the seismic data according to a preset step length; respectively inputting the cut seismic data into a pre-trained fault identification model, and outputting an identification result; the identification result is the probability that the seismic data is the fault data and / or the probability that the seismic data is the non-fault data, the capability of identifying and depicting the minor fault can be improved, the continuity and definition of the identified fault, especially the minor fault, are improved, and the noise immunity and generalization performance of fault identification are improved.

Description

Technical Field

[0001] This invention relates to the field of fracturing technology, and in particular to a fault identification method and apparatus. Background Technology

[0002] This section is intended to provide background or context for embodiments of the present invention. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] Faults, as an important component of structural traps, serve as channels for oil and gas migration and accumulation. Therefore, fault identification is a crucial step in seismic data interpretation. Accurate location of faults, especially small faults, provides a basis and technical support for oil and gas reservoir exploration and development, enhancing the detailed study of structures. Faults with larger displacements control the sedimentary structure and block occurrence of oil-bearing strata, while faults with smaller displacements further divide a complete oil-bearing fault block into numerous smaller blocks, complicating oil-water relationships and controlling the distribution of remaining oil. Because the distribution patterns of oil, gas, and water vary within each small fault block, it presents challenges to oil and gas field development. Improving the identification of small faults is both a hot topic and a difficult challenge for researchers. Therefore, in the process of oil and gas field development, the identification of small faults is key to the detailed depiction of fault block structures. Adopting different and reasonable development measures for each small fault block is essential to maximizing the production potential of the oil and gas field. Accurate identification and detailed characterization of small faults is a crucial research area.

[0004] There are many techniques for small fault detection and identification. In terms of traditional methods, comparing wavelet multi-scale coherence techniques at different frequencies has demonstrated that high-frequency coherence slices are more effective at identifying small faults. By linearly enhancing traditional coherence volume identification techniques, fine characterization of small faults in the Ordovician system of a certain region has been achieved. By jointly interpreting 2D and 3D seismic data, the spatial geometric features of small faults are characterized, thereby identifying small faults in target strata within a basin. Subsequently, various edge-preserving filtering methods have been developed, such as structurally guided constraint filtering and edge-preserving focusing filtering. However, these methods all have certain limitations, such as poor noise resistance in fault identification results and issues with parameter optimization and adaptability in different work areas. Summary of the Invention

[0005] This invention provides a tomographic identification method to improve the ability to identify and characterize small faults, enhance the continuity and clarity of identified faults, especially small faults, and improve the noise resistance and generalization performance of tomographic identification. The method includes:

[0006] Acquire earthquake data;

[0007] The seismic data is cropped according to a preset step size;

[0008] The cropped seismic data are input into a pre-trained fault identification model, and the identification results are output; the identification results are the probability that the seismic data is fault data and / or the probability that it is non-fault data.

[0009] The fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale dilated convolution using historical earthquake data with labels. The labels are used to indicate whether the points corresponding to the historical earthquake data are on faults.

[0010] The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a splicing layer.

[0011] The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers of different porosity. The multi-scale dilated convolution module is used to: perform weighted calculations on the outputs obtained by passing the clipped seismic data through the two convolutional layers of different sizes and the two dilated convolutional layers of different porosity to obtain a first result.

[0012] The splicing layer is used to: splice the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain the second result;

[0013] The multi-scale dilated convolution module is also used to: perform weighted calculations on the outputs of the second result obtained by passing the second result through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the third result;

[0014] The output module is used to: perform weighted calculations on the third results output by multiple improved U-net networks to obtain the recognition result.

[0015] This invention also provides a tomographic identification device to improve the ability to identify and characterize small faults, enhance the continuity and clarity of identified faults, especially small faults, and improve the noise resistance and generalization performance of tomographic identification. The device includes:

[0016] The acquisition module is used to acquire earthquake data;

[0017] The trimming module is used to trim seismic data according to a preset step size;

[0018] The identification module is used to input the cropped seismic data into a pre-trained fault identification model and output the identification results; the identification results are the probability that the seismic data is fault data and / or the probability that it is non-fault data.

[0019] The fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale dilated convolution using historical earthquake data with labels. The labels are used to indicate whether the points corresponding to the historical earthquake data are on faults.

[0020] The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a splicing layer.

[0021] The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers of different porosity. The multi-scale dilated convolution module is used to: perform weighted calculations on the outputs obtained by passing the clipped seismic data through the two convolutional layers of different sizes and the two dilated convolutional layers of different porosity to obtain a first result.

[0022] The splicing layer is used to: splice the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain the second result;

[0023] The multi-scale dilated convolution module is also used to: perform weighted calculations on the outputs of the second result obtained by passing the second result through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the third result;

[0024] The output module is used to: perform weighted calculations on the third results output by multiple improved U-net networks to obtain the recognition result.

[0025] Compared with existing fault identification technologies, this invention acquires seismic data; trims the seismic data according to a preset step size; inputs the trimmed seismic data into a pre-trained fault identification model, and outputs identification results. The identification results represent the probability that the seismic data is fault data and / or the probability that it is not fault data. The fault identification model is trained using labeled historical seismic data on a multi-layer weighted U-net network based on multi-scale dilated convolution. The labels indicate whether points corresponding to historical seismic data are located on faults. The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a stitching layer. The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers with different dilation rates. The multi-scale dilated convolution module is used to: weight the outputs of the cropped seismic data obtained by passing them through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain a first result; the stitching layer is used to: stitch the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain a second result; the multi-scale dilated convolution module is also used to: weight the outputs of the second result obtained by passing them through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain a third result; the output module is used to: weight the third result output by multiple improved U-net networks to obtain the recognition result, which can improve the ability to identify and characterize small faults, improve the continuity and clarity of the identified faults, especially small faults, and improve the noise resistance and generalization performance of fault identification. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0027] Figure 1 This is a flowchart of a tomography identification method provided in an embodiment of the present invention;

[0028] Figure 2 A flowchart illustrating a specific example of a tomography identification method provided in this embodiment of the invention;

[0029] Figure 3 This is a schematic diagram of fault data augmentation provided in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the structure of a multi-scale dilated convolution module provided in an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of the structure of an improved U-net network provided in an embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram of the structure of a multi-layer weighted U-net network based on multi-scale dilated convolution provided in an embodiment of the present invention;

[0033] Figure 7 This is a schematic diagram of seismic data for a certain well area provided in an embodiment of the present invention;

[0034] Figure 8 This is a schematic diagram of the fault identification results obtained by using the method of the present invention based on seismic data of a certain well area, as provided in an embodiment of the present invention.

[0035] Figure 9 This is a schematic diagram of the fault identification results obtained using the coherence attributes of seismic data from a certain well area, provided in an embodiment of the present invention.

[0036] Figure 10 This is a schematic diagram of the fault identification result obtained using the curvature attribute of seismic data from a certain well area, provided in an embodiment of the present invention.

[0037] Figure 11 This is a schematic diagram of a tomography identification device provided in an embodiment of the present invention;

[0038] Figure 12 This is a schematic diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0040] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0041] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0042] In the description of this specification, the terms "comprising," "including," "having," and "containing" are open-ended terms, meaning that they include but are not limited to. The terms "an embodiment," "a specific embodiment," "some embodiments," and "for example," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. The order of steps involved in the various embodiments is used to illustrate the implementation of this application, and the order of steps is not limited and can be adjusted appropriately as needed.

[0043] With the rapid development and widespread application of machine learning technology, machine learning algorithms have been widely used in fault identification of seismic data. Currently, there are semi-automatic fault identification algorithms using artificial neural networks. These algorithms generate a fault identification probability volume by integrating different fault enhancement attributes (similarity, frequency, curvature). Other methods combine convolutional neural networks, traditional machine learning algorithms, and seismic attributes to achieve fault identification. Some researchers have also improved CNN algorithms based on the U-net structure to reliably identify small faults in a certain region. Deep learning algorithms have provided new ideas for small fault identification; however, most network models perform well on test data with high signal-to-noise ratios but perform poorly when applied to real seismic data with complex and diverse fault development. For small fault identification, there are many instances of missed identification or misidentification as phase transitions.

[0044] Accurate location of small faults provides a basis and technical support for oil and gas reservoir exploration and development. Strengthening structural detail research and improving the identification of small faults are both hot topics and challenges for scholars. Existing research also uses image segmentation methods such as U-net for small fault identification. While this network can achieve good fault detection results, it also has certain problems. ① Small faults have a small longitudinal and lateral distribution scale and are highly concealed. Their patterns are disordered over a large area, especially within a small area where even a single small fault often exhibits certain differences in features. ② U-net learns and extracts high-resolution features at a single level. These features can only be used for fault identification within a fixed feature range, which is difficult to meet the needs of practical applications. Moreover, its use of multiple convolutional and pooling layers to increase the receptive field results in the loss of spatial information and low-level features, ultimately increasing the uncertainty in small fault identification. When using U-net for fault identification, the window size and sliding step size of the seismic data can cause prominent boundary effects and poor spatial continuity of the faults.

[0045] To address the issue of unclear boundary recognition in traditional U-net networks when identifying small faults and further improve their recognition performance, we improved the traditional U-net network. We incorporated multi-scale convolutional blocks into U-net, composed of dilated convolutions with varying dilation rates. Different receptive field sizes allow the dilated convolutions to learn features at different scales. We adaptively weighted and summed these features to better integrate global and local fault information from seismic data, thereby improving small fault detection. Furthermore, when using U-net for fault identification, the window size and stride of the seismic data can significantly impact the results, leading to prominent boundary effects and poor fault spatial continuity. Therefore, based on the multi-scale dilated convolutional U-net network, we established a multi-layer U-net network structure. Seismic data clipped at different stride lengths were used as different training sets and input into the branches of the multi-layer U-net network. The prediction results of each branch were weighted and used as the final prediction result to enhance the network's ability to identify and characterize small faults.

[0046] To achieve the above objectives, the present invention provides a method for tomographic identification. Figure 1 This is a flowchart of a tomography identification method provided in an embodiment of the present invention, such as... Figure 1 As shown, the method may include:

[0047] Step 101: Obtain earthquake data;

[0048] Step 102: Trim the seismic data according to a preset step size;

[0049] Step 103: Input the cropped seismic data into the pre-trained fault identification model and output the identification results; the identification results are the probability that the seismic data is fault data and / or the probability that it is non-fault data.

[0050] The fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale dilated convolution using historical earthquake data with labels. The labels are used to indicate whether the points corresponding to the historical earthquake data are on faults.

[0051] The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a splicing layer.

[0052] The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers of different porosity. The multi-scale dilated convolution module is used to: perform weighted calculations on the outputs obtained by passing the clipped seismic data through the two convolutional layers of different sizes and the two dilated convolutional layers of different porosity to obtain a first result.

[0053] The splicing layer is used to: splice the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain the second result;

[0054] The multi-scale dilated convolution module is also used to: perform weighted calculations on the outputs of the second result obtained by passing the second result through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the third result;

[0055] The output module is used to: perform weighted calculations on the third results output by multiple improved U-net networks to obtain the recognition result.

[0056] Starting from seismic data, this invention constructs and trains a multi-layer weighted U-net network based on multi-scale void convolution, and finally performs fault identification in actual work areas, achieving the following beneficial effects: (1) This invention can ensure that the identified small faults conform to the relevant geological laws of faults, and the identification results have strong continuity in the lateral direction, avoiding the splicing effect of the original U-net network results, and the boundary segmentation is clear. (2) This invention can complete fine identification in local areas of actual work areas, and compared with traditional coherence, curvature, machine learning and other methods, the identification effect of small-scale faults is significantly improved.

[0057] In terms of exploration and development potential and technical issues, this invention is applicable to all stages of oil and gas exploration and development. It can effectively improve the accuracy of reflection pattern analysis under small sample conditions, laying a solid foundation for further reservoir prediction and analysis. It can be mainly applied to oil and gas field-related projects.

[0058] Figure 2 A flowchart illustrating a specific example of a tomography identification method provided in this embodiment of the invention is shown below. Figure 2 As shown, the specific process of the fault identification method can be as follows: acquire seismic data, normalize the seismic data, augment the fault data in the seismic data, input the augmented seismic data into a pre-trained fault identification model for fault identification, and obtain the identification results.

[0059] The network proposed in this invention consists of two key parts: (1) adding convolutional blocks based on multi-scale dilated convolution to the network. During the convolution process, the size of the receptive field is changed to extract fault features at different scales in the seismic data, and then the different features are adaptively weighted and summed; (2) constructing a multi-layer U-net weighted network, inputting seismic data of different time lengths into the improved U-net network, and weighting the output results as the final prediction result. This can ensure the overall pattern of faults in a large area, and also fully learn the detailed features of faults in a small area. It can enhance the network's ability to extract spatial information, make the identification of small faults clearer, and obtain small fault identification results with stronger continuity and richer details.

[0060] To better facilitate the training of deep learning models, in one embodiment, the fault identification method may further include: acquiring historical earthquake data; the historical earthquake data including fault data and non-fault data; performing mean normalization on the historical earthquake data; performing a flipping process on the mean-normalized fault data, wherein the flipping process is a horizontal flip, a vertical flip, or a simultaneous horizontal and vertical flip; adding the flipped fault data to the historical earthquake data to obtain updated historical earthquake data; adding labels to the updated historical earthquake data; and training a multi-layer weighted U-net network based on multi-scale dilated convolution using the updated historical earthquake data with labels to generate a fault identification model.

[0061] Specifically, the steps for preprocessing historical earthquake data can be as follows:

[0062] ① Perform data normalization on the training samples (historical earthquake data) to enable the network to converge better;

[0063] ② Use sample augmentation techniques to enrich the sample set by using tomographic data in the training samples, thereby improving the generalization ability of the deep learning model, including horizontal flipping, vertical flipping, and simultaneous horizontal and vertical flipping.

[0064] This step mainly involves preprocessing the earthquake data training set to enable the network model to converge more stably and have higher generalization ability. It mainly includes data normalization and data augmentation.

[0065] Data normalization can eliminate the influence of data dimensions and facilitate network solving and optimization. Normalization methods preserve the characteristics of the data, simply transforming it to different numerical ranges, thus having no impact on the recognition task of this invention.

[0066] Mean normalization is a common data normalization method in machine learning. It can scale the numerical range to the interval [-1, 1]. The normalized data can be obtained using the following formula. Where x max x is the maximum value of the sample data. min x is the minimum value of the sample data. mean This represents the average value of the sample data.

[0067]

[0068] Data augmentation addresses the problem of insufficient training samples. Due to geological structures, the number of fault data points in actual work areas is far less than the number of non-fault data points. During training, the network tends to focus more on the features of non-fault data, over-relying on the limited data samples, leading to overfitting on fault data. When the model is applied to entirely new prediction data, it fails to identify faults effectively. Therefore, before training, fault data must be augmented to include more features. Traditional image processing data augmentation methods include rotation and interpolation. However, faults are geological structures formed by the displacement of seismic data along their phase axes. Therefore, complex data augmentation methods such as interpolation and distortion cannot be used to process seismic data, as this could alter the original seismic amplitude attributes or disrupt the fault location distribution, resulting in unclear distinctions between faults and stratigraphic layers and producing data that does not conform to geological laws. Figure 3 As shown, the earthquake training dataset is expanded by horizontal flipping, vertical flipping, and simultaneous horizontal and vertical flipping. fault(i,j) represents the numerical value of the fault data before transformation, and fault′(i,j) represents the numerical value of the fault data after transformation.

[0069] I. Horizontal Transformation

[0070] For two-dimensional data, perform a symmetrical transformation along the vertical median axis, such as... Figure 2As shown in (b).

[0071] fault′(i,j)=fault(i,hj-1)

[0072] II. Vertical Transformation

[0073] For two-dimensional data, the data is transformed symmetrically along the horizontal midline, such as... Figure 2 As shown in (c).

[0074] fault′(i,j)=fault(wi-1,j)

[0075] III. Horizontal and Vertical Transformation

[0076] For two-dimensional data, the data is first transformed symmetrically along the horizontal central axis, and then along the vertical central axis, such as... Figure 2 As shown in (d).

[0077] fault'(i,j)=fault(wi-1,hj-1)

[0078] The following section introduces the construction and training process of a multi-layer weighted U-net network based on multi-scale dilated convolution.

[0079] This step mainly involves constructing a multi-layer weighted U-net network model based on multi-scale dilated convolution for practical small fault identification. The main steps include the following:

[0080] ① Multi-scale dilated convolution module

[0081] Traditional image segmentation algorithms use pooling layers to increase the receptive field, but this operation reduces the size of the feature map and loses spatial information. Especially in earthquake fault identification, small faults are small targets, and pooling operations can cause discontinuities in the identification process, affecting subsequent identification results. Dilated convolution was proposed to solve the image segmentation problem. Compared with normal convolution, it introduces a dilation rate hyperparameter *r*, which defines the spacing between values ​​processed by the convolution kernel. By changing the dilation rate, dilated convolution can increase the receptive field without changing the size of the output feature map. The actual formula for calculating the dilated convolution kernel is as follows:

[0082] K = k + (k-1)*(r-1)

[0083] Where: k is the original convolution kernel size; r is the dilation rate of the dilated convolution; and K is the size of the dilated convolution.

[0084] Figure 4 This is a schematic diagram of a multi-scale dilated convolution module provided in an embodiment of the present invention. The multi-scale dilated convolution module proposed in this invention is as follows: Figure 4As shown, the module contains four convolutional layers: one regular 1×1 convolutional kernel, one 3×3 convolutional kernel, and two dilated convolutions with dilation rates r of 2 and 3, respectively. Each of the four kernels receives the same output from the previous layer, performs a convolution operation with it, and outputs a result of the same size. The sum of these four outputs is the final output of the layer. The presence of dilated convolutions allows for a larger receptive field, capturing feature information over a wider area. Simultaneously, the different dilation rates avoid the grid effect inherent in single dilated convolutions. The presence of the 3×3 and 1×1 convolutional layers allows the module to simultaneously capture spatial information within a small area, ensuring its sensitivity to detail.

[0085] In one embodiment, the formula for calculating the first result output by the multi-scale dilated convolution module is:

[0086]

[0087] Where, x i The features extracted from the i-th layer in the multi-scale dilated convolution module are described, where the layer is either a convolutional layer or a dilated convolutional layer; r i is the weight corresponding to the i-th layer; X is the first result after weighting the four layers in the multi-scale dilated convolution module.

[0088] We perform a weighted summation of the features extracted from the four convolutions of the dilated convolution to obtain the final feature result. The specific calculation formula is as follows:

[0089]

[0090] Where: x i Features extracted by convolution; r i ...

[0091] ② Multi-layer weighted U-net network (improved U-net network) architecture

[0092] In one embodiment, the multi-scale dilated convolution module may include: a first multi-scale dilated convolution module, a second multi-scale dilated convolution module, a third multi-scale dilated convolution module, a fourth multi-scale dilated convolution module, and a fifth multi-scale dilated convolution module.

[0093] The first multi-scale dilated convolution module is used to perform weighted calculations on the outputs obtained by passing the clipped seismic data through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the first result.

[0094] Each improved U-net network also includes: a first compression layer, a second compression layer, a first deconvolution layer, and a second deconvolution layer;

[0095] The first compression layer is used to extract features and reduce the image size of the first result;

[0096] The second multi-scale dilated convolution module is used to perform weighted calculations on the output of the first compressed layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0097] The second compression layer is used to extract features from the output of the second multi-scale dilated convolution module and reduce the image size.

[0098] The third multi-scale dilated convolution module is used to perform weighted calculations on the output of the second compressed layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0099] The splicing layer includes a first splicing layer and a second splicing layer;

[0100] The first stitching layer is used to stitch together the upsampled low-resolution feature map of the data after deconvolution processing of the output of the third multi-scale dilated convolution module with the corresponding downsampled high-resolution feature map.

[0101] The fourth multi-scale dilated convolution module is used to perform weighted calculations on the output of the first stitched layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0102] The second stitching layer is used to stitch together the upsampled low-resolution feature map of the data after deconvolution processing of the output of the fourth multi-scale dilated convolution module with the corresponding downsampled high-resolution feature map.

[0103] The fifth multi-scale dilated convolution module is used to perform weighted calculations on the output of the second stitching layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0104] Figure 5 This is a schematic diagram of an improved U-net network provided in an embodiment of the present invention. The specific architecture of the improved U-net network model of the present invention is as follows: Figure 5 As shown, U-net is an encoder-decoder structure that consists of two parts: a shrinking path and an expanding path.

[0105] The contraction path consists of multi-scale convolutional blocks and compression layers, serving the purpose of feature extraction. Each step includes a multi-scale dilated convolutional module and a compression layer. After each convolutional operation, a ReLU unit is used for delinearization, enabling subsequent convolutions to effectively extract features. The compression layer is a 3×3 convolutional operation with a stride of 2, achieving image size reduction and field of view expansion while extracting features. In this path, the image size is halved and the number of channels doubles with each step.

[0106] The expansion path consists of a multi-scale dilated convolution module, deconvolution layers, and a stitching layer, serving to recover image information. Each step comprises a 2×2 deconvolution operation, a stitching layer, and a multi-scale dilated convolution module. The deconvolution layer is the inverse operation of max pooling, used to recover the image and achieve feature mapping upsampling. Each 2×2 deconvolution operation doubles the image size, thus layer-by-layer upsampling restores the image to its original size. The stitching layer concatenates the upsampled low-resolution feature maps with the corresponding downsampled high-resolution feature maps to enhance local image features. The multi-scale dilated convolution module, similar to that in the contraction path, performs feature extraction. Finally, a 1×1 convolution and a loss function are applied to perform pixel-level classification, yielding the final segmentation image. In this path, the image size doubles and the number of channels is halved at each step.

[0107] ③ Multi-layer weighted U-net network based on multi-scale dilated convolution

[0108] Figure 6 This is a schematic diagram of a multi-layer weighted U-net network based on multi-scale dilated convolution provided in an embodiment of the present invention. The multi-layer weighted U-net structure proposed in this invention is as follows: Figure 6 As shown. This network contains 4 Figure 5 The U-net network branches shown in the diagram each take seismic data as input, pruned with different step sizes of 8, 16, 32, and 64. The output of each branch is a predicted fault. The final output of the network is a weighted average of the predictions from all branches.

[0109] ④ Training of the U-net network model

[0110] In one embodiment, training a multi-layer weighted U-net network based on multi-scale dilated convolution using updated historical earthquake data with labels to generate a fault identification model may include: using a binary cross-entropy loss function including weight factors, training a multi-layer weighted U-net network based on multi-scale dilated convolution using updated historical earthquake data with labels to generate a fault identification model.

[0111] In one embodiment, the binary cross-entropy loss function, which includes weighting factors, can be:

[0112]

[0113] Where L is the binary cross-entropy loss function including weight factors; m is the number of positive samples, corresponding to the number of tomographic data; n is the number of negative samples, corresponding to the number of non-tomographic data; N is the total number of samples, m+n=N; yi is a one-hot vector, taking the value 0 or 1, taking 1 if the result of identifying the i-th sample is the same as the label of the i-th sample, otherwise taking 0; p represents the probability of the predicted sample belonging to a certain class; α is the weight factor, α∈[0,1], determined according to the ratio of tomographic data and non-tomographic data in the training set.

[0114] After constructing the deep learning model, we need to input the training data into the model for training, so that the deep learning model can be applied to actual fault prediction work. The U-net network uses image segmentation to identify faults. During the training process of the model, we label each point in the seismic data as a fault point (1) or a non-fault point (0) based on whether it is on a fault, and use it as the fault category label y. The final output is the probability that each pixel belongs to a fault point or a non-fault point. The input x and the fault category label y together form the training set.

[0115] Traditional tomography identification networks often employ binary classification with cross-entropy loss, calculated as follows:

[0116] L = -ylog(p) - (1-y)log(1-p)

[0117] Where y is a one-hot vector, taking values ​​of 0 or 1; it is 1 if the class matches the true class of the sample, and 0 otherwise. p represents the probability that the predicted sample belongs to a certain class. For all samples, the loss function is:

[0118]

[0119] Where m is the number of positive samples, corresponding to the number of fault data; n is the number of negative samples, corresponding to the number of non-fault data. N is the total number of samples, m + n = N.

[0120] After data augmentation, the amount of tomographic data increased, but it still differs from the amount of non-tomographic data. During neural network learning, an imbalanced sample distribution, with too many useless negative samples, can cause the model's overall learning direction to deviate, resulting in ineffective learning, poor prediction performance, and weak recognition. Therefore, we added a weighting factor to the loss function to increase the weight of tomographic data and balance the distribution of the loss function, further addressing the problem of imbalanced sample distribution between the two types of data. The modified loss function is as follows:

[0121]

[0122] Where α∈[0,1], we set it to 0.8 based on the ratio of tomographic data to non-tomographic data in the training set, and the corresponding value of 1-α is 0.2.

[0123] The earthquake data we input was 128×128 pixels. The optimization algorithm was Adam, and the learning rate was set to 0.0001. After training 100 times, the network model tended to stabilize.

[0124] ⑤ Fault identification

[0125] This invention first preprocesses the actual earthquake data, and then inputs it into a trained multi-scale dilated convolutional multilayer weighted U-net network to obtain the final fault identification result.

[0126] In this embodiment of the invention, actual data from a certain well area are selected for verification. Figure 7 This is a schematic diagram of seismic data for a certain well area provided in an embodiment of the present invention. Figure 8 This is a schematic diagram of the fault identification results obtained by using the method of the present invention based on seismic data of a certain well area, as provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the fault identification results obtained by utilizing the coherence attributes of seismic data from a certain well area, as provided in an embodiment of the present invention. Figure 10 This is a schematic diagram of the fault identification results obtained by using the curvature properties of seismic data from a certain well area, as provided in an embodiment of the present invention. Figures 7-10 The vertical axis represents time, and the horizontal axis represents the CDP (Common Depth Point) number. From... Figures 7-10It can be seen that traditional geophysical algorithms can identify large-scale faults, but cannot identify faults with small displacements. Compared with traditional coherence, curvature, and machine learning algorithms, the faults identified in this invention have better continuity and clarity, and the identification effect of small-scale faults is significantly improved. It solves the splicing problem in the results of the traditional U-net method, especially in identifying many small-scale faults. Moreover, actual seismic data contains a lot of noise, and the method proposed in this invention has a certain degree of noise resistance and generalization performance.

[0127] Compared with existing fault identification technologies, this invention acquires seismic data; trims the seismic data according to a preset step size; inputs the trimmed seismic data into a pre-trained fault identification model, and outputs identification results. The identification results represent the probability that the seismic data is fault data and / or the probability that it is not fault data. The fault identification model is trained using labeled historical seismic data on a multi-layer weighted U-net network based on multi-scale dilated convolution. The labels indicate whether points corresponding to historical seismic data are located on faults. The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a stitching layer. The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers with different dilation rates. The multi-scale dilated convolution module is used to: weight the outputs of the cropped seismic data obtained by passing them through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain a first result; the stitching layer is used to: stitch the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain a second result; the multi-scale dilated convolution module is also used to: weight the outputs of the second result obtained by passing them through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain a third result; the output module is used to: weight the third result output by multiple improved U-net networks to obtain the recognition result, which can improve the ability to identify and characterize small faults, improve the continuity and clarity of the identified faults, especially small faults, and improve the noise resistance and generalization performance of fault identification.

[0128] Traditional U-net methods exhibit disordered patterns over large areas, especially within smaller areas where even a small fault can reveal varying features. U-net networks extract high-resolution features at a single level, which can only be used for fault identification within a fixed feature range, making it unsuitable for practical applications. Furthermore, its use of multiple convolutional and pooling layers to increase the receptive field leads to the loss of spatial information and low-level features, ultimately increasing the uncertainty in identifying small faults. To address these issues, this invention makes the following improvements:

[0129] This invention starts with seismic data, preprocesses the seismic data in the training set, then constructs and trains a multi-layer weighted U-net network based on multi-scale dilated convolution, and finally identifies faults in actual work areas. Data application in actual work areas demonstrates that this invention can achieve the following beneficial effects:

[0130] (1) The present invention can ensure that the identified small faults conform to the geological laws related to faults, and the identification results have strong continuity in the horizontal direction, avoiding the splicing effect of the original U-net network results, and the boundary segmentation is clear.

[0131] (2) The present invention can complete the fine identification in the local area of ​​the actual work area. Compared with traditional coherence, curvature and machine learning methods, the identification effect of small-scale faults is significantly improved.

[0132] This invention also proposes a tomography identification device, the principle of which is similar to the tomography identification method, and will not be described in detail here.

[0133] Figure 11 This is a schematic diagram of a tomography identification device provided in an embodiment of the present invention, such as... Figure 11 As shown, the tomography identification device may include:

[0134] Module 1101 is used to acquire earthquake data;

[0135] The trimming module 1102 is used to trim seismic data according to a preset step size;

[0136] The identification module 1103 is used to input the cropped seismic data into a pre-trained fault identification model and output the identification results; the identification results are the probability that the seismic data is fault data and / or the probability that it is non-fault data.

[0137] The fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale dilated convolution using historical earthquake data with labels. The labels are used to indicate whether the points corresponding to the historical earthquake data are on faults.

[0138] The multi-layer weighted U-net network based on multi-scale dilated convolution includes multiple improved U-net networks and an output module. Each improved U-net network includes a multi-scale dilated convolution module and a splicing layer.

[0139] The multi-scale dilated convolution module includes two convolutional layers of different sizes and two dilated convolutional layers of different porosity. The multi-scale dilated convolution module is used to: perform weighted calculations on the outputs obtained by passing the clipped seismic data through the two convolutional layers of different sizes and the two dilated convolutional layers of different porosity to obtain a first result.

[0140] The splicing layer is used to: splice the upsampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding downsampled high-resolution feature map to obtain the second result;

[0141] The multi-scale dilated convolution module is also used to: perform weighted calculations on the outputs of the second result obtained by passing the second result through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the third result;

[0142] The output module is used to: perform weighted calculations on the third results output by multiple improved U-net networks to obtain the recognition result.

[0143] In one embodiment, the tomography identification device may further include: a training module, used for:

[0144] Acquire historical earthquake data; the historical earthquake data includes fault data and non-fault data.

[0145] Perform mean normalization on historical earthquake data;

[0146] The fault data after mean normalization is flipped, and the flipping process can be horizontal flipping, vertical flipping, or both horizontal and vertical flipping.

[0147] The inverted fault data is added to the historical earthquake data to obtain updated historical earthquake data.

[0148] Add labels to the updated historical earthquake data;

[0149] A fault identification model is generated by training a multi-layer weighted U-net network based on multi-scale dilated convolution using updated historical earthquake data with labels.

[0150] In one embodiment, the training module is specifically used for:

[0151] A fault identification model is generated by training a multi-layer weighted U-net network based on multi-scale hollow convolution using updated historical earthquake data with labels, employing a binary cross-entropy loss function that includes weighting factors.

[0152] In one embodiment, the binary cross-entropy loss function, including weighting factors, is:

[0153]

[0154] Where L is the binary cross-entropy loss function including weighting factors; m is the number of positive samples, corresponding to the number of tomographic data; n is the number of negative samples, corresponding to the number of non-tomographic data; N is the total number of samples, m + n = N; y i is a one-hot vector with a value of 0 or 1. If the result of the identification of the i-th sample is the same as the label of the i-th sample, it is set to 1; otherwise, it is set to 0. p represents the probability that the predicted sample belongs to a certain class. α is the weight factor, α∈[0,1], which is determined according to the ratio of tomographic data and non-tomographic data in the training set.

[0155] In one embodiment, the formula for calculating the first result output by the multi-scale dilated convolution module is:

[0156]

[0157] Where, x i The features extracted from the i-th layer in the multi-scale dilated convolution module are described, where the layer is either a convolutional layer or a dilated convolutional layer; r i is the weight corresponding to the i-th layer; X is the first result after weighting the four layers in the multi-scale dilated convolution module.

[0158] In one embodiment, the multi-scale dilated convolution module includes: a first multi-scale dilated convolution module, a second multi-scale dilated convolution module, a third multi-scale dilated convolution module, a fourth multi-scale dilated convolution module, and a fifth multi-scale dilated convolution module.

[0159] The first multi-scale dilated convolution module is used to perform weighted calculations on the outputs obtained by passing the clipped seismic data through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates to obtain the first result.

[0160] Each improved U-net network also includes: a first compression layer, a second compression layer, a first deconvolution layer, and a second deconvolution layer;

[0161] The first compression layer is used to extract features and reduce the image size of the first result;

[0162] The second multi-scale dilated convolution module is used to perform weighted calculations on the output of the first compressed layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0163] The second compression layer is used to extract features from the output of the second multi-scale dilated convolution module and reduce the image size.

[0164] The third multi-scale dilated convolution module is used to perform weighted calculations on the output of the second compressed layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0165] The splicing layer includes a first splicing layer and a second splicing layer;

[0166] The first stitching layer is used to stitch together the upsampled low-resolution feature map of the data after deconvolution processing of the output of the third multi-scale dilated convolution module with the corresponding downsampled high-resolution feature map.

[0167] The fourth multi-scale dilated convolution module is used to perform weighted calculations on the output of the first stitched layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0168] The second stitching layer is used to stitch together the upsampled low-resolution feature map of the data after deconvolution processing of the output of the fourth multi-scale dilated convolution module with the corresponding downsampled high-resolution feature map.

[0169] The fifth multi-scale dilated convolution module is used to perform weighted calculations on the output of the second stitching layer, which is obtained by passing the output through two convolutional layers of different sizes and two dilated convolutional layers of different dilation rates.

[0170] The beneficial effects of this invention are as follows:

[0171] 1. This invention improves upon the traditional U-net network. Multi-scale convolutional blocks are added to the U-net, and these blocks consist of dilated convolutions with different dilation rates.

[0172] 2. Based on the multi-scale perforated convolutional U-net network, this invention establishes a multi-layer U-net network structure. Seismic data clipped at different time lengths are used as different training sets and input into the branches of the multi-layer U-net network. The prediction results of each branch are weighted and used as the final prediction result to improve the network's ability to identify and characterize small faults.

[0173] The methods and apparatus described in this invention can be applied to all stages of oil and gas exploration and development. They can ensure that the identified small faults conform to the relevant geological laws of faults, and that the identification results have strong continuity in the lateral direction, avoiding the splicing effect of the original U-net network results, and the boundary segmentation is clear.

[0174] Furthermore, this invention achieves refined identification in local areas of actual work areas, significantly improving the identification effect of small-scale faults compared to traditional methods such as coherence, curvature, and machine learning.

[0175] This invention also provides a computer device. Figure 12 This is a schematic diagram of a computer device in an embodiment of the present invention. The computer device 1200 includes a memory 1210, a processor 1220, and a computer program 1230 stored in the memory 1210 and executable on the processor 1220. When the processor 1220 executes the computer program 1230, it implements the above-mentioned tomography identification method.

[0176] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described tomography identification method.

[0177] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described tomography identification method.

[0178] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0179] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0181] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault identification method characterized by, The method comprises the following steps: acquiring seismic data; cutting the seismic data according to a preset step length; inputting the cut seismic data into a pre-trained fault identification model to output an identification result; the identification result is a probability that the seismic data is fault data and / or a probability that the seismic data is non-fault data; wherein the fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale dilated convolution using historical seismic data carrying labels, and the labels are used to mark whether the points corresponding to the historical seismic data are on a fault; the multi-layer weighted U-net network based on multi-scale dilated convolution comprises a plurality of improved U-net networks and an output module, and each improved U-net network comprises a multi-scale dilated convolution module and a concatenation layer; the multi-scale dilated convolution module comprises two convolution layers of different sizes and two dilated convolution layers of different hole rates; the multi-scale dilated convolution module is configured to perform weighted calculation on the outputs obtained by the cut seismic data through the two convolution layers of different sizes and the two dilated convolution layers of different hole rates, to obtain a first result; the concatenation layer is configured to concatenate the up-sampled low-resolution feature map of the data after convolution or deconvolution processing of the first result with the corresponding down-sampled high-resolution feature map, to obtain a second result; the multi-scale dilated convolution module is further configured to perform weighted calculation on the outputs obtained by the second result through the two convolution layers of different sizes and the two dilated convolution layers of different hole rates, to obtain a third result; the output module is configured to perform weighted calculation on the third results output by the plurality of improved U-net networks, to obtain the identification result.

2. The method of claim 1, wherein, Further comprising: acquiring historical seismic data; the historical seismic data comprises fault data and non-fault data; performing mean value normalization processing on the historical seismic data; performing flip processing on the fault data after the mean value normalization processing; the flip processing is horizontal flip, vertical flip or horizontal and vertical flip at the same time; adding the fault data after the flip processing to the historical seismic data to obtain updated historical seismic data; adding labels to the updated historical seismic data; training the multi-layer weighted U-net network based on multi-scale dilated convolution using the updated historical seismic data carrying labels to generate a fault identification model.

3. The method of claim 2, wherein, Training the multi-layer weighted U-net network based on multi-scale dilated convolution using the updated historical seismic data carrying labels to generate a fault identification model comprises: using a binary classification cross-entropy loss function comprising a weight factor to train the multi-layer weighted U-net network based on multi-scale dilated convolution using the updated historical seismic data carrying labels to generate a fault identification model.

4. The method of claim 3, wherein, The binary classification cross-entropy loss function comprising a weight factor is: wherein, L is a binary cross-entropy loss function including a weight factor; m is the number of positive samples, corresponding to the number of fault data; n is the number of negative samples, corresponding to the number of non-fault data; N is the total number of samples, m+n=N; y i is a one-hot vector, taking values of 0 or 1, taking 1 if the result of identifying the i-th sample is the same as the label of the i-th sample, otherwise taking 0; p represents the probability that the predicted sample belongs to a certain class; α is a weight factor, α∈[0, 1], determined according to the proportion of fault data and non-fault data of the training set, or specified data.

5. The method of claim 1, wherein, The first result output by the multi-scale dilated convolution module is calculated according to the following formula: wherein x i is a feature extracted by an i-th layer in the multi-scale dilated convolution module, the layer being a convolution layer or a dilated convolution layer; r i is a weight corresponding to the i-th layer; and X is a first result after weighting of four layers in the multi-scale dilated convolution module.

6. The method of claim 1, wherein, The multi-scale dilated convolution module comprises a first multi-scale dilated convolution module, a second multi-scale dilated convolution module, a third multi-scale dilated convolution module, a fourth multi-scale dilated convolution module and a fifth multi-scale dilated convolution module. The first multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the cropped seismic data through two convolution layers of different sizes and two hollow convolution layers of different hollow rates, to obtain a first result; Each improved U-net network further comprises a first compression layer, a second compression layer, a first deconvolution layer, and a second deconvolution layer; The first compression layer is configured to perform feature extraction and reduce image size on the first result; The second multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the first compression layer through two convolution layers of different sizes and two hollow convolution layers of different hollow rates; The second compression layer is configured to perform feature extraction and reduce image size on the output of the second multi-scale hollow convolution module; The third multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the second compression layer through two convolution layers of different sizes and two hollow convolution layers of different hollow rates; The concatenation layer comprises a first concatenation layer and a second concatenation layer; The first concatenation layer is configured to concatenate an up-sampled low-resolution feature map of data after deconvolution processing on the output of the third multi-scale hollow convolution module with a corresponding down-sampled high-resolution feature map; The fourth multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the first concatenation layer through two convolution layers of different sizes and two hollow convolution layers of different hollow rates; The second concatenation layer is configured to concatenate an up-sampled low-resolution feature map of data after deconvolution processing on the output of the fourth multi-scale hollow convolution module with a corresponding down-sampled high-resolution feature map; The fifth multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the second concatenation layer through two convolution layers of different sizes and two hollow convolution layers of different hollow rates.

7. A fault identification apparatus characterized by comprising: The method comprises: an acquisition module configured to acquire seismic data; a cropping module configured to crop the seismic data according to a preset step size; an identification module configured to input the cropped seismic data into a pre-trained fault identification model, and output an identification result; the identification result is a probability that the seismic data is fault data and / or a probability that the seismic data is non-fault data; The fault identification model is obtained by training a multi-layer weighted U-net network based on multi-scale hollow convolution using historical seismic data carrying labels, wherein the labels are used to mark whether points corresponding to the historical seismic data are on a fault; The multi-layer weighted U-net network based on multi-scale hollow convolution comprises a plurality of improved U-net networks and an output module, and each improved U-net network comprises a multi-scale hollow convolution module and a concatenation layer; The multi-scale hollow convolution module comprises two convolution layers of different sizes and two hollow convolution layers of different hollow rates; the multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the cropped seismic data through the two convolution layers of different sizes and the two hollow convolution layers of different hollow rates, to obtain a first result; The splicing layer is configured to splice the up-sampling low-resolution feature map of the data after the convolution or deconvolution processing on the first result with the corresponding down-sampling high-resolution feature map to obtain a second result. The multi-scale hollow convolution module is further configured to perform weighted calculation on outputs obtained by the second result respectively passing through two convolution layers of different sizes and two hollow convolution layers of different hollow rates to obtain a third result. The output module is configured to perform weighted calculation on the third results output by the plurality of improved U-net networks to obtain a recognition result.

8. The apparatus of claim 7, wherein, Further comprising: The training module is configured to: obtain historical seismic data; the historical seismic data includes fault data and non-fault data; perform mean value normalization processing on the historical seismic data; perform flipping processing on the mean value normalized fault data, the flipping processing being horizontal flipping, vertical flipping or horizontal and vertical flipping at the same time; add the flipped fault data to the historical seismic data to obtain updated historical seismic data; add labels to the updated historical seismic data; train the multi-layer weighted U-net network based on multi-scale hollow convolution by using the updated historical seismic data carrying the labels to generate a fault recognition model.

9. The apparatus of claim 8, wherein, The training module is specifically configured to: train the multi-layer weighted U-net network based on multi-scale hollow convolution by using the updated historical seismic data carrying the labels to generate a fault recognition model by using a binary classification cross-entropy loss function including a weight factor.

10. The apparatus of claim 9, wherein, The binary classification cross-entropy loss function including the weight factor is: wherein, L is a binary cross-entropy loss function including a weight factor; m is the number of positive samples, corresponding to the number of fault data; n is the number of negative samples, corresponding to the number of non-fault data; N is the total number of samples, m+n=N; y i is a one-hot vector, taking values of 0 or 1, taking 1 if the result of identifying the i-th sample is the same as the label of the i-th sample, otherwise taking 0; p represents the probability that the predicted sample belongs to a certain class; a weight factor, a∈[0,1], determined according to the proportion of fault data and non-fault data of the training set, or specified data.

11. The apparatus of claim 7, wherein, The first result output by the multi-scale hollow convolution module is calculated by the following formula: wherein x i is the feature extracted by the i-th layer in the multi-scale dilated convolution module, the layer being a convolution layer or a dilated convolution layer; r i is the weight corresponding to the i-th layer; and X is the first result after weighting of the four layers in the multi-scale dilated convolution module.

12. The apparatus of claim 7, wherein, The multi-scale hollow convolution module includes a first multi-scale hollow convolution module, a second multi-scale hollow convolution module, a third multi-scale hollow convolution module, a fourth multi-scale hollow convolution module and a fifth multi-scale hollow convolution module. The first multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by the cropped seismic data respectively passing through two convolution layers of different sizes and two hollow convolution layers of different hollow rates to obtain a first result. Each improved U-net network further includes a first compression layer, a second compression layer, a first deconvolution layer and a second deconvolution layer. The first compression layer is configured to perform feature extraction and reduce image size on the first result. The second multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by the output of the first compression layer respectively passing through two convolution layers of different sizes and two hollow convolution layers of different hollow rates. The second compression layer is configured to perform feature extraction and reduce image size on the output of the second multi-scale hollow convolution module. The third multi-scale hollow convolution module is configured to perform weighted calculation on outputs obtained by the output of the second compression layer respectively passing through two convolution layers of different sizes and two hollow convolution layers of different hollow rates. The splicing layer includes a first splicing layer and a second splicing layer. The first splicing layer is configured to splice the up-sampling low-resolution feature map of the data after the deconvolution processing on the output of the third multi-scale hollow convolution module with the corresponding down-sampling high-resolution feature map. The fourth multi-scale dilated convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the first concatenation layer through two convolution layers of different sizes and two dilated convolution layers of different dilated rates; The second concatenation layer is configured to concatenate the up-sampled low-resolution feature map and the corresponding down-sampled high-resolution feature map after performing the inverse convolution processing on the output of the fourth multi-scale dilated convolution module. The fifth multi-scale dilated convolution module is configured to perform weighted calculation on outputs obtained by respectively passing the output of the second concatenation layer through two convolution layers of different sizes and two dilated convolution layers of different dilated rates.

13. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the method in any one of claims 1 to 6 when executing the computer program.

14. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program implements the method in any one of claims 1 to 6 when executed by a processor.

15. A computer program product, characterised in that, The computer program product comprises a computer program, and the computer program implements the method in any one of claims 1 to 6 when executed by a processor.

Images