Seismic fault intelligent identification method and device based on semantic segmentation

Abstract

The invention relates to the technical field of seismic data processing, in particular to a seismic fault intelligent identification method and device based on semantic segmentation. The method comprises the following steps: dividing an obtained existing seismic data volume and corresponding fault tag data thereof as well as a synthetic seismic data volume and corresponding fault tag data thereof into a training sample and a prediction sample; constructing a UNet fault recognition model containing a CBAM attention module, inputting the training sample into the fault recognition model for training, testing the precision of the fault recognition model through a prediction sample, and performing optimization; and performing fault identification on the seismic data volume to be identified by using the trained and optimized fault identification model, and outputting an identification result. The problems that fault interpretation in existing seismic data is manually identified, depends on experience of interpretation personnel and is high in subjectivity, so that the accuracy of an identification result cannot be guaranteed, and manual identification is long in consumed time and cannot meet production requirements are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and particularly to an intelligent seismic fault recognition method and device based on semantic segmentation. Background Art

[0002] A fault is a geological structure formed by the displacement of underground rock layers along a fracture surface or fracture zone. It is the main channel for oil and gas migration and accumulation. During the process of oil and gas exploration and development, fault interpretation is the key to seismic data interpretation. Understanding the location and distribution of faults plays a very important role in oil and gas exploration. The result of fault interpretation is directly related to the accuracy of reservoir prediction. Therefore, how to accurately and efficiently identify faults from diverse seismic profile data has become a difficult problem.

[0003] Traditional fault interpretation involves interpreters manually picking the discontinuous points of the seismic phase axis on the seismic profile, connecting these points into fault lines. After interpreting each seismic profile, the fault lines are combined in three-dimensional space to obtain the fault surface. However, the results of manual interpretation are time-consuming and laborious, and highly subjective, relying very much on the experience of the interpreters. To address the deficiencies of traditional fault recognition methods, relevant scholars have proposed methods for identifying faults using seismic attributes such as coherence volume, curvature, and variance. The fault recognition algorithms based on seismic attributes are constantly being optimized and improved, but they rely on the setting of algorithm parameters, and many parameters need to be tested multiple times to obtain ideal results.

[0004] With the increase in the exploration scale, the traditional method of manually interpreting faults can no longer meet the needs of actual production. There is an urgent need for a systematic method that can accurately and quickly identify faults in seismic profiles. Summary of the Invention

[0005] The present invention proposes an intelligent seismic fault recognition method and device based on semantic segmentation to solve the problems that in the existing seismic data, fault interpretation is carried out manually, relying on the experience of interpreters, and is highly subjective, resulting in the inability to guarantee the accuracy of the recognition results, and the time-consuming manual recognition cannot meet the production requirements.

[0006] According to one aspect of the present invention, there is provided an intelligent seismic fault recognition method based on semantic segmentation, including:

[0007] Obtain the existing seismic data volume and corresponding fault label data in the work area, as well as the seismic data volume to be recognized;

[0008] Through the seismic forward model, obtain the synthetic seismic data volume and corresponding fault label data;

[0009] Divide the existing seismic data volume and its corresponding fault label data, as well as the synthetic seismic data volume and its corresponding fault label data, into training samples and prediction samples;

[0010] Construct a UNet fault identification model containing a CBAM attention module, input the training samples into the fault identification model for training, test the accuracy of the fault identification model through the prediction samples, and optimize it;

[0011] Use the trained and optimized fault identification model to identify faults in the seismic data volume to be identified and output the identification results.

[0012] Preferably, the method for obtaining the synthetic seismic data volume through the seismic forward model includes:

[0013] Set the three-dimensional size of the synthetic seismic data, and generate a horizontal layered seismic reflection model through the forward model according to the three-dimensional size;

[0014] Define the number of seismic folds and the Gaussian function parameters, and determine the fold data volume of the horizontal layered seismic reflection model according to the number of folds and the Gaussian function parameters;

[0015] Define the tilt parameter, and obtain the tilted fold data volume according to the fold data volume and the tilt parameter;

[0016] Define the number of faults, and obtain the fault data volume according to the tilted fold data volume and the number of faults;

[0017] Obtain the seismic data volume containing faults according to the fault data volume and the Ricker wavelet;

[0018] Add noise to the seismic data volume containing faults to obtain the synthetic seismic data volume.

[0019] Preferably, the UNet fault identification model containing the CBAM attention module is:

[0020] The encoder part of the UNet fault identification model is 4 consecutive pooling downsamplings of size 2*2, and the decoder part is 4 consecutive upsamplings of size 2*2 convolutions;

[0021] Add the CBAM attention module at the skip connection of the UNet fault identification model, and the CBAM attention module includes channel attention and spatial attention;

[0022] The UNet fault identification model uses a Dropout convolutional block and a Dice loss function;

[0023] The Dropout convolution block includes DropBlock, a batch normalization layer, and a ReLU activation unit.

[0024] Preferably, the calculation formula for the channel attention includes:

[0025] M c (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)));

[0026] In the formula: F is the input feature map, M c (F) is the channel attention module, σ is the sigmoid activation function, MLP is the multi-layer perceptron, AvgPool is the average pooling, and MaxPool is the max pooling;

[0027] The calculation formula for the spatial attention includes:

[0028] M S (F′) = σ(f([AvgPool(F′); MaxPool(F′)]));

[0029] In the formula: Ms(F′) is the spatial attention module, f is the convolution layer operation, σ is the sigmoid activation function, AvgPool is the average pooling, and MaxPool is the max pooling.

[0030] Preferably, the calculation formula for the Dice loss function includes:

[0031]

[0032] Among them,

[0033] In the formula: p is the fault prediction result, g is the fault label, N is the total number of fault images, p i is a certain pixel in p, and g i is a certain pixel in g at the same position.

[0034] Preferably, the method for optimization includes:

[0035] Using the Adam optimization algorithm and a joint loss function composed of the cross-entropy loss function and the Focal Tversky Loss function to optimize the UNet fault recognition model until the error between the output result of the UNet fault recognition model and the label data of the prediction sample is less than a predetermined error.

[0036] Preferably, it further includes:

[0037] Performing post-processing on the output recognition result;

[0038] The post-processing includes: removing noise, extracting the fault line skeleton, and processing the connected fault lines.

[0039] According to one aspect of the present invention, there is provided an intelligent seismic fault recognition device based on semantic segmentation, including:

[0040] An acquisition unit for acquiring the existing seismic data volume and the corresponding fault label data in the work area, as well as the seismic data volume to be recognized;

[0041] A synthetic seismic data generation unit for obtaining a synthetic seismic data volume and the corresponding fault label data through a seismic forward model;

[0042] A sample division unit for dividing the existing seismic data volume and its corresponding fault label data, as well as the synthetic seismic data volume and its corresponding fault label data, into training samples and prediction samples;

[0043] A model training unit for constructing a UNet fault recognition model including a CBAM attention module, inputting the training samples into the fault recognition model for training, testing the accuracy of the fault recognition model through the prediction samples, and optimizing it;

[0044] A fault recognition unit for using the trained and optimized fault recognition model to perform fault recognition on the seismic data volume to be recognized and outputting the recognition result.

[0045] The present invention has at least the following beneficial effects:

[0046] The present invention proposes an intelligent seismic fault recognition method and device based on semantic segmentation. By inputting the training samples segmented from the seismic data volume into a UNet fault recognition model of semantic segmentation including an attention mechanism CBAM for training and optimization, and using the trained and optimized fault recognition model to automatically recognize faults in the seismic data, the fault interpretation efficiency can be effectively improved, the accuracy of the fault recognition result can be improved, and the benefit of fault interpretation in oil exploration and development can be further enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings herein are incorporated into the specification and form a part of this specification. These drawings illustrate embodiments consistent with the present invention and, together with the specification, are used to explain the technical solutions of the present invention.

[0048] Figure 1 A schematic flowchart showing an intelligent seismic fault recognition method based on semantic segmentation according to an embodiment of the present invention;

[0049] Figure 2 Showing according to an embodiment of the present invention Figure 1 The schematic diagram of the CBAM-UNet++ structure;

[0050] Figure 3 Shows the first-layer skip connection diagram of UNet++ according to an embodiment of the present invention;

[0051] Figure 4 Shows the convolutional unit after adding DropBlock according to an embodiment of the present invention;

[0052] Figure 5 Shows the schematic diagram of the process of making synthetic seismic data according to an embodiment of the present invention;

[0053] Figure 6 Shows the schematic diagram of a partial tomographic label constructed according to an embodiment of the present invention;

[0054] Figure 7 Shows the recognition effect diagram of the constructed model on the Dutch F3 dataset according to an embodiment of the present invention;

[0055] Figure 8 Shows the effect diagram after post-processing of the recognition result according to an embodiment of the present invention. Detailed implementation manners

[0056] Various exemplary embodiments, features, and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0057] The special term "exemplary" herein means "serving as an example, embodiment, or illustration". Any embodiment described as "exemplary" herein need not be construed as superior to or better than other embodiments.

[0058] The term "and / or" herein merely describes the associated relationship of associated objects and indicates that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the term "at least one" herein means any one of a plurality or any combination of at least two of a plurality. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set composed of A, B, and C.

[0059] In addition, in order to better illustrate the present invention, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present invention can also be implemented without some specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present invention.

[0060] Figure 1Schematic flowchart of the intelligent seismic fault identification method based on semantic segmentation according to an embodiment of the present invention; Figure 2 Illustrating according to an embodiment of the present invention Figure 1 Schematic diagram of the CBAM-UNet++ structure therein; Figure 3 Illustrating the first layer skip connection diagram of UNet++ according to an embodiment of the present invention; Figure 4 Illustrating the convolutional unit after adding DropBlock according to an embodiment of the present invention; Figure 5 Schematic diagram of the process of making synthetic seismic data according to an embodiment of the present invention; Figure 6 Illustrating the schematic diagram of partial fault labels constructed according to an embodiment of the present invention; Figure 7 Illustrating the recognition effect diagram of the constructed model on the Dutch F3 dataset according to an embodiment of the present invention; Figure 8 Illustrating the effect diagram after post-processing the recognition result. As Figures 1-8 shown, an intelligent seismic fault identification method based on semantic segmentation includes: Step S01: Obtain the existing seismic data volume and corresponding fault label data in the work area, and the seismic data volume to be identified; Step S02: Obtain the synthetic seismic data volume and corresponding fault label data through the seismic forward model; Step S03: Divide the existing seismic data volume and its corresponding fault label data, and the synthetic seismic data volume and its corresponding fault label data into training samples and prediction samples; Step S04: Construct a UNet fault identification model including a CBAM attention module, input the training samples into the fault identification model for training, test the accuracy of the fault identification model through the prediction samples, and optimize it; Step S05: Use the trained and optimized fault identification model to identify the faults in the seismic data volume to be identified, and output the identification result.

[0061] The intelligent seismic fault identification method based on semantic segmentation provided by the embodiment of the present invention specifically includes the following steps:

[0062] Step S01: Obtain the existing seismic data volume and corresponding fault label data in the work area, and the seismic data volume to be identified.

[0063] In the embodiment of the present invention, the format of the seismic data volume is segy. The corresponding seismic amplitude data of the profile is read into the matrix according to the inline number of the seismic data volume, that is, the seismic two-dimensional profile data is read in the profile order, and the corresponding fault label data of each seismic two-dimensional profile is obtained. The two-dimensional profile data and the corresponding fault label data are used as training samples (training dataset) and prediction samples (prediction dataset).

[0064] Step S02: Obtain the synthetic seismic data volume and corresponding fault label data through the seismic forward model.

[0065] In the present invention, the method for obtaining a synthetic seismic data volume through a seismic forward model includes: setting the three-dimensional size of the synthetic seismic data, generating a horizontal layered seismic reflection model through the forward model according to the three-dimensional size; defining the number of seismic folds and Gaussian function parameters, and determining the fold data volume of the horizontal layered seismic reflection model according to the number of folds and the Gaussian function parameters; defining the dip parameter, and obtaining the inclined fold data volume according to the fold data volume and the dip parameter; defining the number of faults, and obtaining the fault data volume according to the inclined fold data volume and the number of faults; obtaining the seismic data volume with faults according to the fault data volume and the Ricker wave; adding noise to the seismic data volume with faults to obtain the synthetic seismic data volume.

[0066] In an embodiment of the present invention, the three-dimensional size xsize, ysize, zsize of the synthetic seismic data is set in the forward model, and a three-dimensional horizontal layered seismic reflection model (layer) is generated according to the input three-dimensional size.

[0067] Define the number N of seismic folds (fold) and the Gaussian function parameters b k 、c k 、d k , and calculate the fold data volume for the three-dimensional horizontal layered reflection model through formula (1).

[0068]

[0069] In the formula, x, y, and z are the X, Y, and Z axis coordinates of a certain point in the three-dimensional data volume, a 0 is the offset of the sum of the Gaussian functions, b k is the height of the peak of the Gaussian curve, c k and d k are the position coordinates of the peak of the Gaussian curve, N is the number of folds, and σ k is the variance.

[0070] Define the dip (planar) parameters a, b, and calculate the fold data volume using formula (2) to obtain the inclined fold data volume.

[0071] s 2 (x, y, z) = a * x + b * y + (-a * x center - b * y center ) (2);

[0072] In the formula, a and b are the inclination degrees of the dip parameters, x center and y center are the central coordinates of the x-axis and y-axis of the data volume respectively.

[0073] Define the number of faults, and calculate the faults for the inclined fold data volume using formula (3) to obtain the fault data volume.

[0074]

[0075] In the formula, X 0 , Y 0 , Z 0 are the coordinates of the reference points randomly selected according to the number of faults, is the strike angle, with a range of [0, 360], and θ is the dip angle size, with a range of [0, 90].

[0076] Calculate all the reflection coefficient data Z(t) of each Z-axis in the fault data volume and the Ricker wavelet using formula (4) to obtain the seismic data volume containing faults.

[0077]

[0078] g(t) = Z(t)) * A(t) (4);

[0079] In the formula, f is the main frequency, t is the time, and g is the seismic record after convolution with the Ricker wavelet.

[0080] Add noise to the seismic data volume containing faults, and finally obtain the synthetic seismic data volume (seismic). The process of making the synthetic seismic data is as Figure 5 shown.

[0081] Extract the two-dimensional seismic profile of the synthetic seismic data volume and its corresponding fault label data, as Figure 6 shown. In Figure 6 , Figures (a), (c), (e), and (g) are the seismic profiles, and Figures (b), (d), (f), and (h) are the corresponding label data.

[0082] Step S03: Divide the existing seismic data volume and its corresponding fault label data, as well as the synthetic seismic data volume and its corresponding fault label data, into training samples and prediction samples.

[0083] In the embodiment of the present invention, the two-dimensional seismic profiles and label data corresponding to the existing seismic data volume, as well as the two-dimensional seismic profiles and label data corresponding to the synthetic seismic data volume, are divided into two parts: training samples and prediction samples. The training samples are used to train the model, and the prediction samples are used to test the model. Through the corresponding relationship between the seismic profiles and the extracted fault label data in the two parts of the existing seismic data volume and the synthetic seismic data volume, a more accurate mapping relationship between the two can be established through the model, thereby further improving the accuracy of the trained model.

[0084] Step S04: Construct a UNet fault recognition model incorporating a CBAM attention module, input the training samples into the fault recognition model for training, test the accuracy of the fault recognition model with prediction samples, and perform optimization.

[0085] In the embodiments of the present invention, CBAM (Convolutional Block Attention Module) represents an attention mechanism module for convolutional modules. It is an attention mechanism module that combines spatial and channel attention. UNet is a semantic segmentation neural network model that combines lightweight and high performance. The constructed UNet fault recognition model incorporating a CBAM attention module is the semantic segmentation model CBAM-UNet++ fault recognition model.

[0086] In the present invention, the UNet fault recognition model incorporating a CBAM attention module is as follows: The encoder part of the UNet fault recognition model is four consecutive pooling downsamplings of size 2*2, and the decoder part is four consecutive upsamplings with a convolutional size of 2*2; the CBAM attention module is added at the skip connection of the UNet fault recognition model, and the CBAM attention module includes channel attention and spatial attention; the UNet fault recognition model uses a Dropout convolutional block and a Dice loss function; the Dropout convolutional block includes DropBlock, a batch normalization layer, and a ReLU activation unit.

[0087] In the embodiments of the present invention, the encoder part of the UNet network model is used to extract the features of the seismic profile image. It is composed of two 3*3 convolutional layers (Conv3*3) plus a 2*2 Maxpooling (maximum pooling) layer to jointly form a downsampling layer. For the corresponding decoder part of the network model, upsampling is performed four times, and the upsampling is a convolution of 2*2. The structure of the CBAM-Unet++ fault recognition model is as Figure 2 shown.

[0088] In the present invention, the calculation formula for the channel attention includes:

[0089] M c (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F))) (5);

[0090] In the formula: F is the input feature map, M c (F) is the channel attention module, σ is the sigmoid activation function, MLP is the multi-layer perceptron, AvgPool is the average pooling, and MaxPool is the maximum pooling;

[0091] The spatial attention calculation formula includes:

[0092] M S (F′) = σ(f([AvgPool(F′); MaxPool(F′)])) (6);

[0093] In the formula: Ms(F′) is the spatial attention module, f is the convolutional layer operation, σ is the sigmoid activation function, AvgPool is average pooling, MaxPool is max pooling, and F′ is the output of the channel attention module.

[0094] In the embodiment of the present invention, a CBAM attention model is added at the UNet++ skip connection, including channel attention and spatial attention. CBAM uses global max pooling and global average pooling to process the input feature map F, that is, the seismic profile, in the channel dimension to obtain two feature maps. Then, these feature maps are input into a multi-layer perceptron MLP for dimensionality reduction and dimensionality increase processing. Then, by adding and processing the two feature maps output by the MLP and passing through the sigmoid activation function, the channel attention M c (F) is obtained.

[0095] CBAM first performs max pooling and average pooling operations on F' and concatenates them. Then, a convolutional network is used to reduce the dimension of the concatenated result to one channel. Finally, after passing through the sigmoid activation function, the spatial attention module is obtained.

[0096] The semantic segmentation model (UNet) uses a Dropout (random dropout regularization) convolutional block (convolutional unit / module), and there is a DropBlock (dropout continuous region regularization), a batch normalization (BN) layer, and a ReLU (rectified linear unit) activation unit behind each convolutional layer. As Figure 4 shown, DropBlock discards continuous regions from the feature map of a layer, making the model more discriminative. DropBlock has two parameters, namely block_size and Υ. block_size is the length and width of DropBlock. When block_size is 1, DropBlock is similar to Dropout. Υ is the probability of the Bernoulli function, which controls the number of features to be deleted. Its calculation formula is as follows:

[0097]

[0098] In the formula: keep_prob is the node retention probability, feat_size is the size of the feature map, and block_size is the length and width of DropBlock.

[0099] In the present invention, the calculation formula of the Dice loss function includes:

[0100]

[0101] Wherein,

[0102] In the formula: p is the fault prediction result, g is the fault label, N is the total number of fault images, p i is a certain pixel in p, and g i is a certain pixel in g at the same position.

[0103] In an embodiment of the present invention, the Dice loss function is used to alleviate the problem of sample imbalance. In image segmentation, the number of pixels in some classes is very small, while the number of pixels in some classes is very large, which leads to the problem of class imbalance. If a traditional loss function is used, the model will tend to recognize classes with a larger number of pixels and ignore classes with a smaller number of pixels. The Dice loss function can effectively solve this problem because it not only considers the correctness of the recognition result but also considers the integrity of the recognition result.

[0104] Input the seismic profile data in the prediction sample into the trained model, run to obtain the corresponding fault recognition result, and calculate the error between the fault recognition result and the label data in the prediction sample through the Dice loss function.

[0105] In the present invention, the method for optimization includes: using the Adam optimization algorithm and a combined loss function composed of a cross-entropy loss function and a Focal Tversky Loss loss function to optimize the UNet fault recognition model until the error between the output result of the UNet fault recognition model and the label data of the prediction sample is less than a predetermined error.

[0106] In an embodiment of the present invention, when training the model, it is necessary to use an optimization algorithm to continuously iterate and update the parameters until the error between the output data and the label data in the prediction sample reaches convergence, that is, less than a predetermined error.

[0107] In the optimization algorithm, the adaptive motion estimation algorithm Adam optimization algorithm is used for gradient descent. This method has high computational efficiency and low memory occupancy, and is very suitable for solving optimization problems with a large amount of data.

[0108] The parameter Batchsize of the UNet fault recognition model (network model) is set to 5. The input size of the image in the model is 128*128*3 pixels, and the output size is 128*128*1 pixel.

[0109] The problem of sample imbalance affects the learning ability of the network. The cross-entropy loss function can avoid the problem of gradient dispersion, thus stably updating the weights of the network model; the Focal Tversky Loss function L FTL can solve the problem of imbalance between positive and negative samples, making the model more sensitive to small target regions and the network having stronger convergence ability.

[0110] Superimpose the binary cross-entropy loss function L BCE (Binary Cross Entropy, BCE) and the Focal Tversky Loss (non-linear focal loss function) to construct a combined loss function L BFT , that is:

[0111] L BFT = L BCE + L FTL (9);

[0112] wherein, L BCE = -(ylog(p)+(1 - y)log(1 - p));

[0113] In the formula: y is the predicted label 0 or 1 of the output, and p is the probability corresponding to the output label y.

[0114] Among them,

[0115] In the formula: c is the number of categories, TP is the number of pixels of the target category correctly predicted by the model as this category, FP is the number of non-target categories wrongly predicted as the target category by the model, FN is the number of pixels of the target category wrongly predicted as non-target categories by the model, ɑ is a hyperparameter controlling the weight of FP, β is a hyperparameter controlling the weight of FN, and γ is a focal parameter, with a range of [1, 3].

[0116] Train and optimize the network model according to the above optimization algorithm to obtain the CBAM-Unet++ network model.

[0117] Step S05: Use the trained and optimized fault identification model to identify faults in the to-be-identified seismic data volume and output the identification result.

[0118] In the embodiment of the present invention, input the seismic profile of the to-be-identified seismic data volume obtained in step S01 into the CBAM-Unet++ network model obtained after training and optimization. After network calculation, obtain the fault identification result corresponding to each seismic profile. The value of each point in the result matrix represents the probability that this point is a fault point. The fault identification results corresponding to all seismic profiles constitute a three-dimensional fault probability data volume.

[0119] In the present invention, it further includes: post-processing the output recognition result; the post-processing includes: removing noise, extracting the skeleton of the fault line, and processing the connection of the fault lines.

[0120] In an embodiment of the present invention, the faults identified by the CBAM-Unet++ network include missed identifications and misidentifications, which exist in the prediction result in a non-connected form and need to be post-processed. The post-processing performs a Hough transform on the fault identification result to connect the disconnected faults. The Hough transform converts the straight-line detection problem in the image space into the parameter space according to the relationship that the collinear points in the image space correspond to the intersecting lines in the parameter space, and accumulates and statistically processes in the parameter space to complete the straight-line detection task. Then, dilation, erosion, and skeleton extraction processing are performed on the fault image processed by the Hough transform. Finally, the fault area is filled, and at the same time, the fault is refined into a line.

[0121] Among them, removing noise is to remove the fault noise area based on the connected area, including: detecting the fault connection in the fault identification result, setting the minimum connection threshold, and when the detected connection threshold is less than the set minimum connection threshold, it indicates that the connected area is an isolated area. Fill the value of the isolated connected area with 0 to achieve noise removal.

[0122] Extracting the skeleton of the fault line is: calculating the distance transformation graph for the matrix after the above noise removal processing. Determine the local center point according to the distance relationship between the points in the current matrix (transformation graph) and its surrounding 8 neighborhood points. Use the minimum covering set algorithm to screen out the pseudo center points, and finally obtain the extracted skeleton of the fault line.

[0123] Processing the connection of the fault lines is: detecting the endpoints of the fault lines in the extracted skeleton of the fault lines. Calculate the angle value between any two fault lines according to the endpoints, and judge whether they pass through the same fault. Judge whether to connect according to the length of the fault line and the angle threshold of the connecting line, that is, if the difference between the angle values of the two lines is within the threshold, then connect the two lines together.

[0124] It can be understood that the above-mentioned various method embodiments mentioned in the present invention can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present invention will not elaborate further.

[0125] The execution entity of the intelligent seismic fault recognition method based on semantic segmentation can be an intelligent seismic fault recognition device based on semantic segmentation. For example, the intelligent seismic fault recognition method based on semantic segmentation can be executed by a terminal device, a server, or other processing devices. Among them, the terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, the intelligent seismic fault recognition method based on semantic segmentation can be implemented by a processor calling computer-readable instructions stored in a memory.

[0126] Those skilled in the art can understand that in the above method of the specific implementation manner, the writing order of each step does not mean a strict execution order and constitutes any limitation on the implementation process. The specific execution order of each step should be determined according to its function and possible internal logic.

[0127] The present invention also provides an intelligent seismic fault recognition device based on semantic segmentation, including: an acquisition unit, configured to acquire the existing seismic data volume and the corresponding fault label data of the work area, and the seismic data volume to be recognized; a synthetic seismic data generation unit, configured to obtain a synthetic seismic data volume and the corresponding fault label data through a seismic forward model; a sample division unit, configured to divide the existing seismic data volume and its corresponding fault label data, and the synthetic seismic data volume and its corresponding fault label data into training samples and prediction samples; a model training unit, configured to construct a UNet fault recognition model including a CBAM attention module, input the training samples into the fault recognition model for training, test the accuracy of the fault recognition model through the prediction samples, and optimize it; a fault recognition unit, configured to use the trained and optimized fault recognition model to perform fault recognition on the seismic data volume to be recognized and output a recognition result.

[0128] In some embodiments, the functions or modules and units included in the device provided by the embodiments of the present invention can be used to execute the methods described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0129] The present invention conducts experiments on the seismic data of a certain work area in F3 of the North Sea in the Netherlands to verify the effectiveness of the method. The fault recognition results are as Figure 7 shown in Figure 7Among them, (a) is the original map of the F3 seismic fault in the North Sea of the Netherlands, and (b) is the fault recognition effect of the F3 in the North Sea of the Netherlands. It can be seen from the figure that the model is better in terms of fault continuity, with a lower misrecognition rate, and can effectively identify the vast majority of fault lines. Further post-process the fault results recognized by the CBAM-UNet++ network model, connect the broken parts of the fault, and the final recognition effect is as Figure 8 shown.

[0130] The present invention extracts a two-dimensional seismic profile matrix from a three-dimensional seismic data segy data volume. Input the two-dimensional seismic profile matrix into a fault recognition model based on semantic segmentation, and calculate and output a fault recognition result binary matrix through the model. Post-process the fault binary matrix to improve the continuity and readability of the fault markings. Among them, the fault recognition model is trained by fault data synthesized by a seismic forward model and actual fault annotation data. The training algorithm is based on an improved Unet++ semantic segmentation model.

[0131] The effect of the present invention is to integrate the attention mechanism CBAM into the Unet++ algorithm, establish a CBAM-UNet++ model, and apply it to seismic fault recognition, suppress the interference of irrelevant features, and at the same time effectively suppress the overfitting problem generated in the network by using DropBlock. By introducing the Dice Loss loss function to alleviate the data imbalance problem in the seismic fault recognition task. Perform Hough transform, extract the skeleton and other post-processing on the fault prediction results, accurately identify the faults with insufficient continuity, improve the accuracy of fault recognition, and make the fault prediction results better applied to geological targets.

[0132] The present invention regards fault interpretation as a semantic segmentation problem, and uses the feature extraction ability of deep learning and the classification ability of neural networks to realize automatic fault recognition and improve the efficiency of fault interpretation. By adding an attention module, the ability to extract fault features is improved, the interference of non-fault signals is suppressed from both the channel and spatial dimensions, the continuity of the fault is improved, and the misrecognition rate is reduced, which has a guiding role in fault interpretation.

[0133] The embodiments of the present invention have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein.

Claims

1. A method for intelligent identification of earthquake faults based on semantic segmentation, characterized in that: include: Obtain the existing seismic data volume and corresponding fault label data in the work area, as well as the seismic data volume to be identified; Through the seismic forward model, synthetic seismic data volume and corresponding fault label data are obtained; Dividing the existing seismic data volume and its corresponding fault label data, as well as the synthetic seismic data volume and its corresponding fault label data into training samples and prediction samples; Constructing a UNet fault recognition model including a CBAM attention module, inputting the training samples into the fault recognition model for training, testing the accuracy of the fault recognition model through prediction samples, and optimizing it; The trained and optimized fault identification model is used to perform fault identification on the seismic data volume to be identified, and the identification result is output.

2. The method for intelligent identification of earthquake faults based on semantic segmentation according to claim 1, characterized in that: The method for obtaining a synthetic seismic data volume through a seismic forward model comprises: Setting a three-dimensional size of the synthetic seismic data, and generating a horizontal layered seismic reflection model through a forward model according to the three-dimensional size; Defining the number of seismic folds and Gaussian function parameters, and determining the fold data volume of the horizontal layered seismic reflection model according to the number of folds and Gaussian function parameters; defining an inclination parameter, and obtaining an inclined fold data volume according to the fold data volume and the inclination parameter; Defining the number of faults, and obtaining a fault data volume according to the inclined fold data volume and the number of faults; According to the fault data volume and the Ricker wavelet, a seismic data volume containing the fault is obtained; Noise is added to the seismic data volume containing the fault to obtain a synthetic seismic data volume.

3. The method for intelligent identification of earthquake faults based on semantic segmentation according to claim 1, characterized in that: The UNet fault recognition model including the CBAM attention module is: The encoder part of the UNet fault recognition model is 4 consecutive pooling downsamplings of size 2*2, and the decoder part is 4 consecutive upsampling convolutions of size 2*2; Add the CBAM attention module at the jump connection of the UNet fault recognition model, where the CBAM attention module includes channel attention and spatial attention; The UNet fault recognition model uses a Dropout convolution block and a Dice loss function; The Dropout convolution block includes a DropBlock, a batch normalization layer, and a ReLU activation unit.

4. The method for intelligent identification of earthquake faults based on semantic segmentation according to claim 3 is characterized in that: The calculation formula of the channel attention is include: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))); Where: F is the input feature map, M c (F) is the channel attention module, σ is the sigmoid activation function, MLP is the multi-layer perceptron, AvgPool is the average pooling, and MaxPool is the maximum pooling; The spatial attention calculation formula includes: M S (F′)=σ(f([AvgPool(F′);MaxPool(F′)])); Where: Ms(F′) is the spatial attention module, f is the convolutional layer operation, σ is the sigmoid activation function, AvgPool is the average pooling, and MaxPool is the maximum pooling.

5. The method for intelligent identification of earthquake faults based on semantic segmentation according to claim 3 is characterized in that: The Dice loss function calculation formula includes: in, Where: p is the fault prediction result, g is the fault label, N is the total number of fault images, p i is a pixel in p, g i is a pixel of g in the same position.

6. The method for intelligent identification of earthquake faults based on semantic segmentation according to claim 1, characterized in that: The optimization method comprises: The Adam optimization algorithm and the joint loss function composed of the cross entropy loss function and the Focal Tversky Loss loss function are used to optimize the UNet fault recognition model until the error between the output result of the UNet fault recognition model and the label data of the predicted sample is less than the predetermined error.

7. The method for intelligent identification of earthquake faults based on semantic segmentation according to any one of claims 1 to 6, characterized in that: Also includes: Post-processing the output recognition result; The post-processing includes: removing noise points, extracting fault line skeletons and connecting fault lines.

8. An intelligent earthquake fault identification device based on semantic segmentation, characterized in that: include: An acquisition unit is used to acquire existing seismic data volumes and corresponding fault label data in the work area, as well as seismic data volumes to be identified; A synthetic seismic data generating unit, used to obtain a synthetic seismic data volume and corresponding fault label data through a seismic forward model; A sample division unit, used for dividing the existing seismic data volume and its corresponding fault label data, as well as the synthetic seismic data volume and its corresponding fault label data, into training samples and prediction samples; A model training unit, used for constructing a UNet fault recognition model including a CBAM attention module, inputting the training samples into the fault recognition model for training, testing the accuracy of the fault recognition model through prediction samples, and optimizing it; The fault identification unit is used to perform fault identification on the seismic data volume to be identified by using the trained and optimized fault identification model, and output the identification result.

Images