A method and system for detecting seismic faults based on an N-type network

Abstract

The application relates to the technical field of seismic exploration data processing, and provides a seismic fault detection method and system based on an N-type network, aiming to solve the technical problems that, in the existing seismic fault detection result, the continuity of some areas is insufficient, and the seismic fault structure is not complete; and the existing prediction process greatly consumes GPU memory. The method comprises the following steps: inputting a seismic fault image into an input layer for down-sampling processing; then inputting a down-sampled fault feature map into an encoder, and utilizing a plurality of convolution layers to perform semantic extraction on the down-sampled fault feature map; inputting the down-sampled fault feature map and a plurality of semantic extraction feature maps into a first decoder and a second decoder respectively, and utilizing the first decoder and the second decoder to decode the input feature maps in a reverse order; then inputting the feature maps output by the encoder and the two decoders into an output layer for splicing processing; and finally performing fault detection to obtain a fault detection result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic exploration data processing, and in particular to a seismic fault detection method and system based on an N-type network. BACKGROUND

[0002] Oil and gas resources are not only important chemical raw materials, but also important energy sources, and are known as "industrial food and blood". Throughout the years, the interpretation of seismic faults is a very important link in oil and gas exploration and development. Seismic faults are not only the boundaries of oil and gas fields, but also the channels for oil and gas migration and accumulation. The distribution and morphology of seismic faults play a key role in the identification and description of oil and gas, and have a great influence on the exploitation and distribution of oil and gas. In underground structures, seismic faults exist in various forms and have different sizes. In the process of oil and gas transportation, seismic faults can act as both channels and barriers. Therefore, determining the location and morphological distribution of seismic faults is an indispensable part of the entire process of oil and gas exploration. With the increasing scale and decreasing cycle of seismic exploration, and the development of exploration theory to a deeper level, the interpretation accuracy of seismic faults, as one of the three major transport systems, is increasingly demanding. Therefore, the research on seismic fault identification has attracted much attention from the industry and academia.

[0003] With the continuous improvement of seismic exploration technology, the volume of seismic data is also increasing. Manual interpretation of seismic faults is not only very tedious and time-consuming, but also difficult to repeat and verify. The complex interpretation process also requires high professional skills of interpreters. Manual interpretation cannot meet the needs of modern production. Traditional methods identify seismic faults by calculating the coherence, curvature, variance, and other attributes of seismic data volumes. However, seismic fault detection methods based on attribute calculation are sensitive to noise and stratigraphic features, and have the disadvantages of being easily disturbed by noise and lacking sufficient accuracy.

[0004] In the prior art, a large amount of seismic fault data is obtained as a training set, and then the training set is preprocessed. The preprocessed seismic data is input into a deep learning model for training. The trained model is applied to new seismic fault data for seismic fault detection.

[0005] The prior art has the following technical problems:

[0006] 1. In the existing deep model generated seismic fault detection results, there are problems of insufficient continuity in some areas and incomplete seismic fault structure.

[0007] 2. The training of the deep learning model is easily limited by the hardware level, and requires the use of high-performance graphics cards. Due to the large volume of three-dimensional seismic data, the prediction process will greatly consume GPU memory. SUMMARY

[0008] The application provides a seismic fault detection method and system based on an N-type network, aiming at solving the technical problems in the prior art that the continuity of some areas is insufficient and the structure of the seismic fault is not complete in the seismic fault detection result, and that the training of the existing deep learning model needs to use a high-performance graphics processing unit (GPU), and the prediction process greatly consumes the GPU memory.

[0009] The technical scheme for solving the above technical problems is as follows: a seismic fault detection method based on an N-type network, the N-type network comprising an input layer, an encoder, a first decoder, a second decoder and an output layer; the seismic fault detection using the N-type network specifically comprises:

[0010] An initial seismic fault image is obtained, the initial seismic fault image is preprocessed to obtain a preprocessed image, the preprocessed image is input into the input layer for down-sampling processing to obtain a down-sampled fault feature map;

[0011] The down-sampled fault feature map is input into the encoder, and a cascaded structure formed by a plurality of first double convolution layers built in the encoder is used to perform semantic extraction on the down-sampled fault feature map to obtain a plurality of semantic extraction feature maps corresponding to outputs of the plurality of first double convolution layers;

[0012] The down-sampled fault feature map and the plurality of semantic extraction feature maps are respectively input into the first decoder and the second decoder, and the first decoder and the second decoder are used to decode the down-sampled fault feature map and the plurality of semantic extraction feature maps in reverse order to respectively obtain a first decoding feature map and a second decoding feature map;

[0013] The first decoding feature map, the second decoding feature map and a semantic extraction feature map output by a last layer of the first double convolution layers in the cascaded structure are input into the output layer for splicing to obtain a fault splicing feature map, the fault splicing feature map is adjusted to an initial resolution, and fault detection is performed on the adjusted fault splicing feature map to obtain a fault detection result; wherein the initial resolution is the resolution of the initial seismic fault image.

[0014] Further, the method further comprises: constructing an initial network, training the initial network to obtain the N-type network, and obtaining the N-type network specifically comprises:

[0015] An original seismic image is obtained, and a corresponding fault label is generated for each original seismic image to obtain an original seismic image set, the original seismic image set is divided into a training set and a validation set;

[0016] inputting each original seismic image in the training set into the initial network for fault prediction in sequence to obtain a plurality of first fault prediction results, calculating a first loss value between the plurality of first fault prediction results and a fault label corresponding to the training set, and updating network parameters of the initial network according to the first loss value;

[0017] repeating the iterative training of the initial network on the training set by using the updated network parameters, and defining that one cycle of the initial network traversing the entire training set as one cycle;

[0018] When each cycle ends, inputting the initial network with updated network parameters into the verification set for fault prediction to obtain a second fault prediction result, and calculating a second loss value between the second fault prediction result and a fault label corresponding to the verification set;

[0019] Setting a patience period, recording a loss difference between the second loss value of the current cycle and the second loss value of the last cycle each time a cycle ends, if the loss difference obtained after a plurality of cycles ends reaches a stop condition, and the sum of the plurality of cycles is greater than the patience period, then stop training the initial network, and select the initial network corresponding to the minimum value of all second loss values as the N-type network.

[0020] Further, the balanced binary cross-entropy loss function is used to calculate the first loss value and the second loss value. The formula of the balanced binary cross-entropy loss function is as follows:

[0021]

[0022] wherein, represents the first loss value or the second loss value, represents a ratio of non-fault pixel points to all pixel points of the original seismic image, N represents all pixel points of the original seismic image, represents a fault label, represents a pixel point predicted as a fault label.

[0023] Further, the input layer includes a first down-sampling module, a first convolutional layer and a second down-sampling module, and the inputting the preprocessed image into the input layer for down-sampling processing specifically includes:

[0024] processing the preprocessed image by using the first down-sampling module, inputting the output of the first down-sampling module into the second down-sampling module after processing by the first convolutional layer, and performing feature extraction processing by the second down-sampling module to output a down-sampling fault feature map.

[0025] Furthermore, the encoder includes a cascaded structure formed by stacking three first double convolutional layers, resulting in multiple semantic extraction feature maps corresponding to the outputs of the multiple first double convolutional layers, specifically including:

[0026] The downsampled tomographic feature map is input into the first double convolutional layer to obtain the first semantic extraction feature map;

[0027] The first semantic extraction feature map is input into the second first double convolutional layer to obtain the second semantic extraction feature map;

[0028] The second semantic feature map is input into the third first double convolutional layer to obtain the third semantic feature map;

[0029] Among them, the first semantic extraction feature map, the second semantic extraction feature map, and the third semantic extraction feature map are multiple semantic extraction feature maps output by multiple first double convolutional layers.

[0030] Furthermore, the encoder is hop-connected to both the first decoder and the second decoder. The first decoder includes three second double convolutional layers, and the first decoded feature map obtained specifically includes:

[0031] The downsampled tomographic feature map and the first semantic extracted feature map are input into the first decoder for feature concatenation to obtain the first concatenated feature map;

[0032] The first concatenated feature map is input into the first second double convolutional layer to obtain the third decoded feature map. The third decoded feature map is then concatenated with the second semantic extraction feature map to obtain the second concatenated feature map.

[0033] The second concatenated feature map is input into the second second double convolutional layer to obtain the fourth decoded feature map. The fourth decoded feature map is then concatenated with the third semantic extraction feature map to obtain the third concatenated feature map.

[0034] The third concatenated feature map is input into the third second double convolutional layer to obtain the first decoded feature map;

[0035] The second decoder includes three third double convolutional layers, resulting in the second decoded feature map, specifically including:

[0036] The third semantic feature map and the second semantic feature map are input into the second decoder for feature concatenation to obtain the fourth concatenated feature map.

[0037] The fourth concatenated feature map is input into the first third double convolutional layer to obtain the fifth decoded feature map. The fifth decoded feature map is then concatenated with the first semantic extraction feature map to obtain the fifth concatenated feature map.

[0038] The fifth concatenated feature map is input into the second third double convolutional layer to obtain the sixth decoded feature map. The sixth decoded feature map is then concatenated with the downsampled tomographic feature map to obtain the sixth concatenated feature map.

[0039] The sixth concatenated feature map is input into the third third double convolutional layer to obtain the second decoded feature map.

[0040] Furthermore, the output layer includes a first upsampling module, a second upsampling module, and a fifth convolutional layer, and the tomography detection results specifically include:

[0041] The third semantic extraction feature map, the first decoding feature map, and the second decoding feature map are input into the output layer and concatenated to obtain a discontinuous concatenated feature map.

[0042] The fault stitching feature map is processed using the first upsampling module and the second upsampling module to restore the fault stitching feature map to its initial resolution, thus obtaining the fault upsampling feature map;

[0043] The fault upsampled feature map is input into the fifth convolutional layer for fault detection, and the fault detection result is obtained.

[0044] Secondly, to solve the above-mentioned technical problems, the present invention also provides an N-type network-based seismic fault detection system, comprising:

[0045] The input layer module is used to acquire an initial seismic fault image, preprocess the initial seismic fault image to obtain a preprocessed image, and downsample the preprocessed image to obtain a downsampled fault feature map.

[0046] The encoder is used to extract semantics from the downsampled tomographic feature map by utilizing the cascaded structure formed by stacking multiple first double convolutional layers built into the encoder, thereby obtaining multiple semantically extracted feature maps corresponding to the outputs of the multiple first double convolutional layers.

[0047] The first decoder is used to receive the downsampled tomographic feature map and multiple semantically extracted feature maps input from the encoder, and decode the downsampled tomographic feature map and multiple semantically extracted feature maps in the input order to obtain the first decoded feature map;

[0048] The second decoder is used to receive the downsampled tomographic feature map and multiple semantic extraction feature maps from the encoder input, and decode the downsampled tomographic feature map and multiple semantic extraction feature maps in the reverse decoding order of the first decoder to obtain the second decoded feature map.

[0049] The output layer module is used to input the first decoded feature map, the second decoded feature map, and the semantic extraction feature map output by the last layer of the first double convolutional layer in the cascaded structure into the output layer for splicing to obtain a fault splicing feature map. The fault splicing feature map is adjusted to an initial resolution, and fault detection is performed on the adjusted fault splicing feature map to obtain a fault detection result; wherein, the initial resolution is the resolution of the initial seismic fault image.

[0050] Thirdly, in order to solve the above-mentioned technical problems, the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the seismic fault detection method based on an N-type network of the present application.

[0051] Fourthly, in order to solve the above-mentioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements an N-type network-based seismic fault detection method of the present application.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] 1. The N-type network designed in this invention, through a skip connection structure between the encoder and decoder, enables the exchange and fusion of low-level and high-level semantic information, resulting in better tomographic detection results. In the N-type network of this invention, the image resolution remains unchanged at different stages of the encoder and decoder, allowing for feature fusion and decoding in different directions, while better preserving image edge information.

[0054] 2. The N-type network proposed in this invention employs a "bidirectional decoding" strategy, which enables it to learn richer and more refined tomographic features. Through this "bidirectional decoding" strategy, low-level semantic information and high-level semantic information can be fully interacted and integrated, resulting in more accurate and continuous tomographic detection results.

[0055] 3. In this invention, seismic data is input into an N-type network and first undergoes two downsampling operations, reducing the image size to one-quarter of its original size, before encoding and decoding. Therefore, the N-type network consumes less computation and is faster. Furthermore, the maximum number of convolutional kernel channels in the N-type network is 16. Thanks to this targeted network structure design, the N-type network can achieve excellent prediction performance with a smaller network size. The compression of the number of channels significantly reduces computational costs.

[0056] 4. The N-type network of this invention uses synthetic seismic data during training and testing. Compared with models that use manually labeled actual seismic data, it can ensure a higher and more sufficient number of training samples. In addition, there are no false negative labels in the training data, which ensures the correctness of model training and improves the model's prediction performance.

[0057] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0058] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A schematic diagram of an N-type network structure for an N-type network-based seismic fault detection method according to an embodiment of the present invention is shown.

[0060] Figure 2(a) shows a schematic diagram of the original seismic images of the training set of a seismic fault detection method based on an N-type network according to an embodiment of the present invention;

[0061] Figure 2(b) shows a schematic diagram of the fault labels corresponding to Figure 2(a);

[0062] Figure 3(a) shows a schematic diagram of the original seismic images of the validation set of a seismic fault detection method based on an N-type network according to an embodiment of the present invention;

[0063] Figure 3(b) shows a schematic diagram of the fault prediction results corresponding to Figure 3(a);

[0064] Figure 3(c) shows a schematic diagram of the fault labels corresponding to Figure 3(a);

[0065] Figure 4(a) shows a schematic diagram of a seismic image to be detected;

[0066] Figure 4(b) shows a schematic diagram of the fault detection results based on Figure 4(a) of the present invention;

[0067] Figure 4(c) shows a schematic diagram of the fault detection results based on Figure 4(a) using the prior art;

[0068] Figure 5(a) shows a three-dimensional schematic diagram of a seismic image to be detected;

[0069] Figure 5(b) shows a schematic diagram of the fault detection results based on Figure 5(a) of the present invention;

[0070] Figure 5(c) shows a schematic diagram of the fault detection results based on Figure 5(a) using the prior art;

[0071] Figure 6(a) shows a schematic diagram of time slices of the seismic image to be detected;

[0072] Figure 6(b) shows a schematic diagram of the fault detection results based on Figure 6(a) of the present invention;

[0073] Figure 6(c) shows a schematic diagram of the fault detection results based on Figure 6(a) using the prior art. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Figure 1 A schematic diagram of an N-type network structure for an N-type network-based seismic fault detection method according to an embodiment of the present invention is shown, as follows: Figure 1 As shown in the figure, an embodiment of the present invention provides a seismic fault detection method based on an N-type network. The N-type network includes an input layer, an encoder, a first decoder, a second decoder, and an output layer. Specifically, the seismic fault detection using the N-type network includes:

[0076] An initial seismic fault image is acquired, and the initial seismic fault image is preprocessed to obtain a preprocessed image. The preprocessed image is then input into the input layer for downsampling processing to obtain a downsampled fault feature map.

[0077] The downsampled tomographic feature map is input into the encoder, and semantic extraction is performed on the downsampled tomographic feature map using the cascaded structure formed by stacking multiple first double convolutional layers built into the encoder, resulting in multiple semantically extracted feature maps output by the multiple first double convolutional layers.

[0078] The downsampled tomographic feature map and multiple semantically extracted feature maps are respectively input into the first decoder and the second decoder. The first decoder and the second decoder are used to decode the downsampled tomographic feature map and multiple semantically extracted feature maps in reverse order to obtain the first decoded feature map and the second decoded feature map respectively.

[0079] The first decoded feature map, the second decoded feature map, and the semantic extraction feature map output from the last layer of the first double convolutional layer in the cascaded structure are input into the output layer for concatenation to obtain a fault concatenation feature map. The fault concatenation feature map is adjusted to the initial resolution, and fault detection is performed on the adjusted fault concatenation feature map to obtain the fault detection result; wherein, the initial resolution is the resolution of the initial seismic fault image.

[0080] This invention treats fault detection as an image segmentation problem, dividing the original 3D seismic data into two parts: fault and non-fault components. Fault pixels occupy a small proportion of the 3D seismic image, presenting a continuous planar structure, requiring the network to accurately learn the detailed information within the image. Based on the characteristics of seismic data and addressing the challenges and problems of fault detection, this invention designs a lightweight and efficient N-type network for fault detection.

[0081] Optionally, it also includes: constructing an initial network, training the initial network to obtain an N-type network, wherein obtaining the N-type network specifically includes:

[0082] Original seismic images are acquired, and a corresponding fault label is generated for each original seismic image to obtain a set of original seismic images. The set of original seismic images is then divided into a training set and a validation set.

[0083] Each original seismic image in the training set is sequentially input into the initial network for fault prediction, resulting in multiple first fault prediction results. A first loss value is calculated between the multiple first fault prediction results and the fault labels corresponding to the training set, and the network parameters of the initial network are updated based on the first loss value.

[0084] The initial network with updated network parameters is used to iteratively train by repeatedly traversing the training set. One cycle is defined as one time the initial network traverses the entire training set.

[0085] At the end of each cycle, the initial network after updating the network parameters is used with the validation set input to perform fault prediction, and a second fault prediction result is obtained. The second loss value between the second fault prediction result and the fault label corresponding to the validation set is calculated.

[0086] Set a patience period. At the end of each cycle, record the loss difference between the second loss value of this cycle and the second loss value of the previous cycle. If the loss difference obtained after multiple consecutive cycles reaches the stopping condition and the total loss of multiple consecutive cycles is greater than the patience period, then stop training the initial network and select the initial network corresponding to the minimum value among all second loss values ​​as the N-type network.

[0087] This invention is based on the principle of seismic image generation and uses mathematical methods to artificially synthesize three-dimensional original seismic images and corresponding fault labels. A total of 200 pairs of training data and 20 pairs of validation data were generated, with a data size of 128×128×128. First, all original seismic images were standardized by subtracting the mean and dividing by the standard deviation, followed by data augmentation. Specifically, each data volume was rotated and flipped around the timeline axis. This operation effectively avoids overfitting during network training, resulting in a more robust model and significantly improving data quality. Finally, the processed data was input into the network for training. Figures 2(a) and 2(b) show the original seismic images and corresponding fault labels of the training set; Figures 3(a), 3(b), and 3(c) show the original seismic images, fault prediction results, and corresponding fault labels of the validation set.

[0088] Optionally, the first loss value and the second loss value are calculated using a balanced binary cross-entropy loss function; the formula for the balanced binary cross-entropy loss function is shown below:

[0089]

[0090] in, Indicates the first loss value or the second loss value. This represents the ratio of non-fault pixels to all pixels in the original seismic image. N represents all pixels in the original seismic image. Indicates a fault label. Represents pixels The probability of being predicted as a fault label.

[0091] Optionally, the input layer includes a first downsampling module, a first convolutional layer, and a second downsampling module. Specifically, inputting the preprocessed image into the input layer for downsampling processing includes:

[0092] The preprocessed image is processed using the first downsampling module. The output of the first downsampling module is processed by the first convolutional layer and then input into the second downsampling module. The second downsampling module performs feature extraction processing and outputs a downsampled tomographic feature map.

[0093] The input layer of this invention includes two convolution-based downsampling stages, each halving the resolution of the preprocessed image. The first downsampling layer uses a convolutional module with a kernel size of 5 and a stride of 2, resulting in a feature map with 8 channels and a size of 64×64×64. The second downsampling layer uses a convolutional kernel size of 3, with other parameters remaining unchanged; the output feature map has 16 channels and a size of 32×32×32. A first convolutional layer with a kernel size of 3×3×3 is included between the two downsampling layers.

[0094] Optionally, the encoder includes a cascaded structure formed by stacking three first double convolutional layers, resulting in multiple semantic extraction feature maps corresponding to the outputs of the multiple first double convolutional layers, specifically including:

[0095] The downsampled tomographic feature map is input into the first double convolutional layer to obtain the first semantic extraction feature map;

[0096] The first semantic extraction feature map is input into the second first double convolutional layer to obtain the second semantic extraction feature map;

[0097] The second semantic feature map is input into the third first double convolutional layer to obtain the third semantic feature map;

[0098] Among them, the first semantic extraction feature map, the second semantic extraction feature map, and the third semantic extraction feature map are multiple semantic extraction feature maps output by multiple first double convolutional layers.

[0099] In this application, the encoder receives the sampled tomographic feature map after processing by the input layer and learns tomographic information from it. The encoder includes three first double convolutional layers, each consisting of two 3×3×3 convolutional layers with two convolutional kernels. The feature map output by each first double convolutional layer has 16 channels and a size of 32×32×32.

[0100] Each stage uses two 3×3×3 convolutional layers to extract features. No pooling layers are used for downsampling between stages, so the encoder's three first double convolutional layers extract features at the same image resolution, which can effectively preserve the image's edge information.

[0101] Optionally, the encoder is skip-connected to both the first decoder and the second decoder. The first decoder includes three second double convolutional layers to obtain the first decoded feature map, specifically including:

[0102] The downsampled tomographic feature map and the first semantic extracted feature map are input into the first decoder for feature concatenation to obtain the first concatenated feature map;

[0103] The first concatenated feature map is input into the first second double convolutional layer to obtain the third decoded feature map. The third decoded feature map is then concatenated with the second semantic extraction feature map to obtain the second concatenated feature map.

[0104] The second concatenated feature map is input into the second second double convolutional layer to obtain the fourth decoded feature map. The fourth decoded feature map is then concatenated with the third semantic extraction feature map to obtain the third concatenated feature map.

[0105] The third concatenated feature map is input into the third second double convolutional layer to obtain the first decoded feature map;

[0106] The second decoder includes three third double convolutional layers, resulting in the second decoded feature map, specifically including:

[0107] The third semantic feature map and the second semantic feature map are input into the second decoder for feature concatenation to obtain the fourth concatenated feature map.

[0108] The fourth concatenated feature map is input into the first third double convolutional layer to obtain the fifth decoded feature map. The fifth decoded feature map is then concatenated with the first semantic extraction feature map to obtain the fifth concatenated feature map.

[0109] The fifth concatenated feature map is input into the second third double convolutional layer to obtain the sixth decoded feature map. The sixth decoded feature map is then concatenated with the downsampled tomographic feature map to obtain the sixth concatenated feature map.

[0110] The sixth concatenated feature map is input into the third third double convolutional layer to obtain the second decoded feature map.

[0111] The decoder receives four feature maps from the encoder: a downsampled tomographic feature map, a first semantic extraction feature map, a second semantic extraction feature map, and a third semantic extraction feature map. The "bidirectional decoding" decoder comprises a first decoder and a second decoder, which decode independently. Each first decoder contains three second double convolutional layers, and each second decoder contains three third double convolutional layers. Both the second and third double convolutional layers are composed of two 3×3×3 convolutional layers. The two decoders decode in opposite directions.

[0112] The first and second decoders, based on a "bidirectional decoding" strategy, can simultaneously decode the encoder's output feature map from both low-level and high-level semantics, extracting richer and more refined tomographic features. Both the second and third dual convolutional layers use two 3×3×3 convolutional layers to extract features, each with 16 channels. No upsampling is performed between the two feature extraction operations, extracting features at the same image resolution. The output feature maps all have 16 channels and a size of 32×32×32.

[0113] Optionally, the output layer includes a first upsampling module, a second upsampling module, and a fifth convolutional layer, and the tomography detection results specifically include:

[0114] The third semantic extraction feature map, the first decoding feature map, and the second decoding feature map are input into the output layer and concatenated to obtain a discontinuous concatenated feature map.

[0115] The fault stitching feature map is processed using the first upsampling module and the second upsampling module to restore the fault stitching feature map to its initial resolution, thus obtaining the fault upsampling feature map;

[0116] The fault upsampled feature map is input into the fifth convolutional layer for fault detection, and the fault detection result is obtained.

[0117] The output layer first concatenates three feature maps from the encoder and decoder branches to obtain a tomographic concatenation feature map. Then, it uses two upsampling operations followed by a 3×3×3 convolution to restore the tomographic concatenation feature map to its original resolution. Finally, a 1×1×1 convolutional module outputs the tomographic detection result. The concatenated feature map has 48 channels and a size of 32×32×32. After the first upsampling, the number of channels is reduced to 16, and the size is restored to 64×64×64. After the second upsampling, the number of channels remains 16, and the size is restored to 128×128×128. Finally, a 1×1×1 convolutional layer reduces the channel dimension to 1, outputting the tomographic detection result.

[0118] The output layer fuses the three feature maps and outputs the tomography detection results after two upsampling operations. The fusion of multiple features allows the network to learn richer tomography information.

[0119] Optionally, upsampling can be performed using nearest neighbor interpolation. This method does not contain learnable parameters, is computationally simple, and can effectively reduce computational costs.

[0120] The following uses two real-world datasets—the F3 earthquake data from the North Sea in the Netherlands and the Kerry 3D data from New Zealand—to validate existing earthquake fault detection methods with the N-type network-based earthquake fault detection method proposed in this application.

[0121] Verification Example 1

[0122] As shown in Figures 4(a), 4(b), and 4(c), Figure 4(a) is a three-dimensional view of the F3 earthquake data volume in the North Sea, Netherlands; Figure 4(b) shows the fault detection results obtained from Figure 4(a) based on the N-type network of this invention; and Figure 4(c) shows the fault detection results obtained from Figure 4(a) based on existing earthquake fault detection methods. As indicated by the black boxes in Figures 4(b) and 4(c), the fault prediction results of this invention are richer. Fault continuity is greatly improved, and the fault structure is more complete.

[0123] Verification Example 2

[0124] As shown in Figures 5(a), 5(b), and 5(c), Figure 5(a) is the Kerry3D data from New Zealand, Figure 5(b) is the fault detection result obtained from Figure 5(a) based on the N-type network of the present invention, and Figure 5(c) is the fault detection result obtained from Figure 5(a) based on the existing seismic fault detection method. As shown by the black arrows in Figures 5(b) and 5(c), it can be seen that the N-type network predicts fault continuity better, while the existing network is more likely to predict broken fault structures.

[0125] Verification Example 3

[0126] As shown in Figures 6(a), 6(b), and 6(c), Figure 6(a) is the Kerry 3D data from New Zealand, Figure 6(b) is the fault detection result obtained from Figure 6(a) based on the N-type network of the present invention, and Figure 6(c) is the fault detection result obtained from Figure 6(a) based on the existing seismic fault detection method. As shown in the black boxes in Figures 6(b) and 6(c), the faults predicted by the existing network have poor continuity and incomplete structures, while the N-type network proposed in this invention provides more accurate prediction results, showing clear and continuous fault structures.

[0127] Compared to existing networks, in terms of GPU memory consumption and time consumption, during the training phase, consuming the same amount of GPU memory, the N-type network can input approximately four times the amount of training data at a time, while the training time is only half that of existing networks. In the prediction phase, using the same amount of GPU memory, the N-type network can predict approximately six times the amount of data at a time. The resolution of the Dutch North Sea F3 earthquake data used in the test was 512*384*128. When tested on a 2080Ti graphics card, the N-type network could input the entire data volume at once, with a prediction time of 0.33 seconds. In contrast, limited by GPU memory, existing networks can only input a 128*128*128 data volume at a time, taking 4.09 seconds to predict the complete Dutch North Sea F3 earthquake data volume.

[0128] Based on the above method, this embodiment of the invention also provides a seismic fault detection system based on an N-type network, corresponding to the above method, comprising:

[0129] The input layer module is used to acquire an initial seismic fault image, preprocess the initial seismic fault image to obtain a preprocessed image, and downsample the preprocessed image to obtain a downsampled fault feature map.

[0130] The encoder is used to extract semantics from the downsampled tomographic feature map by utilizing the cascaded structure formed by stacking multiple first double convolutional layers built into the encoder, thereby obtaining multiple semantically extracted feature maps corresponding to the outputs of the multiple first double convolutional layers.

[0131] The first decoder is used to receive the downsampled tomographic feature map and multiple semantically extracted feature maps input from the encoder, and decode the downsampled tomographic feature map and multiple semantically extracted feature maps in the input order to obtain the first decoded feature map;

[0132] The second decoder is used to receive the downsampled tomographic feature map and multiple semantic extraction feature maps from the encoder input, and decode the downsampled tomographic feature map and multiple semantic extraction feature maps in the reverse decoding order of the first decoder to obtain the second decoded feature map.

[0133] The output layer module is used to input the first decoded feature map, the second decoded feature map, and the semantic extraction feature map output by the last layer of the first double convolutional layer in the cascaded structure into the output layer for splicing to obtain a fault splicing feature map. The fault splicing feature map is adjusted to an initial resolution, and fault detection is performed on the adjusted fault splicing feature map to obtain a fault detection result; wherein, the initial resolution is the resolution of the initial seismic fault image.

[0134] The earthquake fault detection system of this invention can execute the earthquake fault detection method provided in this invention, and their implementation principles are similar. The actions performed by each module and unit in the earthquake fault detection system of each embodiment of this invention correspond to the steps in the earthquake fault detection method of each embodiment of this invention. For detailed functional descriptions of each module of the earthquake fault detection system, please refer to the descriptions in the corresponding earthquake fault detection methods shown above, which will not be repeated here.

[0135] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seismic fault detection method based on N-type networks, characterized in that, The N-type network includes an input layer, an encoder, a first decoder, a second decoder, and an output layer; the use of the N-type network for seismic fault detection specifically includes: An initial seismic fault image is acquired, and the initial seismic fault image is preprocessed to obtain a preprocessed image. The preprocessed image is then input into the input layer for downsampling processing to obtain a downsampled fault feature map. The downsampled tomographic feature map is input into the encoder, and semantic extraction is performed on the downsampled tomographic feature map using the cascaded structure formed by stacking multiple first double convolutional layers built into the encoder, resulting in multiple semantically extracted feature maps output by the multiple first double convolutional layers. The downsampled tomographic feature map and multiple semantically extracted feature maps are respectively input into the first decoder and the second decoder. The first decoder and the second decoder are used to decode the downsampled tomographic feature map and multiple semantically extracted feature maps in reverse order to obtain the first decoded feature map and the second decoded feature map respectively. The first decoded feature map, the second decoded feature map, and the semantic extraction feature map output from the last layer of the first double convolutional layer in the cascaded structure are input into the output layer for concatenation to obtain a fault concatenation feature map. The fault concatenation feature map is adjusted to the initial resolution, and fault detection is performed on the adjusted fault concatenation feature map to obtain the fault detection result; wherein, the initial resolution is the resolution of the initial seismic fault image.

2. The seismic fault detection method based on an N-type network according to claim 1, characterized in that, Also includes: Construct an initial network, train the initial network to obtain an N-type network, and obtain the N-type network specifically includes: Original seismic images are acquired, and a corresponding fault label is generated for each original seismic image to obtain a set of original seismic images. The set of original seismic images is then divided into a training set and a validation set. Each original seismic image in the training set is sequentially input into the initial network for fault prediction, resulting in multiple first fault prediction results. A first loss value is calculated between the multiple first fault prediction results and the fault labels corresponding to the training set, and the network parameters of the initial network are updated based on the first loss value. The initial network with updated network parameters is used to iteratively train by repeatedly traversing the training set. One cycle is defined as one time the initial network traverses the entire training set. At the end of each cycle, the initial network after updating the network parameters is used with the validation set input to perform fault prediction, and a second fault prediction result is obtained. The second loss value between the second fault prediction result and the fault label corresponding to the validation set is calculated. Set a patience period. At the end of each cycle, record the loss difference between the second loss value of this cycle and the second loss value of the previous cycle. If the loss difference obtained after multiple consecutive cycles reaches the stopping condition and the total loss of multiple consecutive cycles is greater than the patience period, then stop training the initial network and select the initial network corresponding to the minimum value among all second loss values ​​as the N-type network.

3. The seismic fault detection method based on an N-type network according to claim 2, characterized in that, The first loss value and the second loss value are calculated using the balanced binary cross-entropy loss function; the formula for the balanced binary cross-entropy loss function is shown below: in, Indicates the first loss value or the second loss value. This represents the ratio of non-fault pixels to all pixels in the original seismic image. N represents all pixels in the original seismic image. Indicates a fault label. Represents pixels The probability of being predicted as a fault label.

4. The seismic fault detection method based on an N-type network according to claim 1, characterized in that, The input layer includes a first downsampling module, a first convolutional layer, and a second downsampling module. Specifically, inputting the preprocessed image into the input layer for downsampling processing includes: The preprocessed image is processed using the first downsampling module. The output of the first downsampling module is processed by the first convolutional layer and then input into the second downsampling module. The second downsampling module performs feature extraction processing and outputs a downsampled tomographic feature map.

5. The seismic fault detection method based on an N-type network according to claim 1, characterized in that, The encoder comprises a cascaded structure formed by stacking three first double convolutional layers, resulting in multiple semantic extraction feature maps corresponding to the outputs of the multiple first double convolutional layers, specifically including: The downsampled tomographic feature map is input into the first double convolutional layer to obtain the first semantic extraction feature map; The first semantic extraction feature map is input into the second first double convolutional layer to obtain the second semantic extraction feature map; The second semantic feature map is input into the third first double convolutional layer to obtain the third semantic feature map; Among them, the first semantic extraction feature map, the second semantic extraction feature map, and the third semantic extraction feature map are multiple semantic extraction feature maps output by multiple first double convolutional layers.

6. The seismic fault detection method based on an N-type network according to claim 5, characterized in that, The encoder is connected to the first decoder and the second decoder in a skip connection. The first decoder includes three second double convolutional layers and the second decoder includes three third double convolutional layers. The process of obtaining the first and second decoded feature maps specifically includes: inputting the downsampled tomographic feature map and the first semantic extraction feature map into the first decoder for feature concatenation to obtain the first concatenated feature map. The first concatenated feature map is input into the first second double convolutional layer to obtain the third decoded feature map. The third decoded feature map is then concatenated with the second semantic extraction feature map to obtain the second concatenated feature map. The second concatenated feature map is input into the second second double convolutional layer to obtain the fourth decoded feature map. The fourth decoded feature map is then concatenated with the third semantic extraction feature map to obtain the third concatenated feature map. The third concatenated feature map is input into the third second double convolutional layer to obtain the first decoded feature map; The third semantic feature map and the second semantic feature map are input into the second decoder for feature concatenation to obtain the fourth concatenated feature map. The fourth concatenated feature map is input into the first third double convolutional layer to obtain the fifth decoded feature map. The fifth decoded feature map is then concatenated with the first semantic extraction feature map to obtain the fifth concatenated feature map. The fifth concatenated feature map is input into the second third double convolutional layer to obtain the sixth decoded feature map. The sixth decoded feature map is then concatenated with the downsampled tomographic feature map to obtain the sixth concatenated feature map. The sixth concatenated feature map is input into the third third double convolutional layer to obtain the second decoded feature map.

7. The seismic fault detection method based on an N-type network according to claim 6, characterized in that, The output layer includes a first upsampling module, a second upsampling module, and a fifth convolutional layer, and the tomography detection results are specifically obtained as follows: The third semantic extraction feature map, the first decoding feature map, and the second decoding feature map are input into the output layer and concatenated to obtain a discontinuous concatenated feature map. The fault stitching feature map is processed using the first upsampling module and the second upsampling module to restore the fault stitching feature map to its initial resolution, thus obtaining the fault upsampling feature map; The fault upsampled feature map is input into the fifth convolutional layer for fault detection, and the fault detection result is obtained.

8. A seismic fault detection system based on an N-type network, characterized in that, include: The input layer module is used to acquire an initial seismic fault image, preprocess the initial seismic fault image to obtain a preprocessed image, and downsample the preprocessed image to obtain a downsampled fault feature map. The encoder is used to extract semantics from the downsampled tomographic feature map by utilizing the cascaded structure formed by stacking multiple first double convolutional layers built into the encoder, thereby obtaining multiple semantically extracted feature maps corresponding to the outputs of the multiple first double convolutional layers. The first decoder is used to receive the downsampled tomographic feature map and multiple semantically extracted feature maps input from the encoder, and decode the downsampled tomographic feature map and multiple semantically extracted feature maps in the input order to obtain the first decoded feature map; The second decoder is used to receive the downsampled tomographic feature map and multiple semantic extraction feature maps from the encoder input, and decode the downsampled tomographic feature map and multiple semantic extraction feature maps in the reverse decoding order of the first decoder to obtain the second decoded feature map. The output layer module is used to input the first decoded feature map, the second decoded feature map, and the semantic extraction feature map output by the last layer of the first double convolutional layer in the cascaded structure into the output layer for splicing to obtain a fault splicing feature map. The fault splicing feature map is adjusted to an initial resolution, and fault detection is performed on the adjusted fault splicing feature map to obtain a fault detection result; wherein, the initial resolution is the resolution of the initial seismic fault image.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the method of any one of claims 1-7.

Images