An automatic seismic facies recognition method based on the combination of self-attention mechanism and U-shaped structure

Abstract

The present invention discloses an automatic seismic facies recognition method based on self-attention mechanism and U-shaped structure. First, an encoder structure based on Segformer is proposed. Then, a block dilation module is introduced to construct a decoder structure, and Hypercolumns technology is introduced to fuse features for seismic facies recognition. Image seismic facies segmentation is performed through a reasonable training and inference process. Finally, a seismic facies body is formed based on the seismic facies recognition result to interpret different underground sedimentary environments. Through experiments on public data and actual data, compared with the seismic facies recognition model of traditional convolutional neural networks, the automatic seismic facies recognition method proposed by the present invention has fewer noise points, more obvious boundary details, stronger lateral continuity, and can more accurately interpret different underground sedimentary environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic exploration, and relates to a seismic facies automatic recognition technology combining self-attention mechanism and U-shaped structure. In particular, it is a seismic image semantic segmentation method that combines Segformer self-attention segmentation network, UNet network structure and Hypercolumn semantic segmentation technology, and applies this seismic image semantic segmentation method to the automatic recognition and classification of seismic facies of seismic data. Background Art

[0002] With the increasing demand for oil and gas and the rapid development of artificial intelligence technology, modern oil and gas exploration technology is gradually developing towards intelligence and automation. At present, oil and gas exploration mainly adopts seismic exploration methods, that is, post-stack seismic data is obtained through artificial seismic wave reflection, and the underground structure, lithology, oil and gas bearing properties, etc. contained therein are mined and analyzed through multidisciplinary knowledge, so as to locate the distribution of underground oil and gas reservoirs. The traditional seismic facies division scheme is that interpreters manually interpret or use some mathematical methods to semi-automatically extract features and segment seismic facies. However, these traditional methods have great subjectivity, the semi-automatic methods have insufficient accuracy and poor timeliness, and they cannot accurately locate oil and gas reservoirs under complex underground structures and sedimentary conditions. How to use computer resources to achieve an efficient automatic seismic facies recognition method has become a problem that needs to be solved currently.

[0003] In order to improve the seismic facies automatic recognition method, many researchers have proposed seismic facies classification methods based on deep learning. This type of method is to learn a non-linear mapping from seismic data to seismic facies labels in an end-to-end manner based on existing labeled seismic data, and apply it to unseen seismic data for seismic facies classification. The advantage of this method is that it can directly perform end-to-end seismic facies classification, saving a large amount of labor costs, and improving the interpretation effect to a certain extent. Jesper (2018) migrated the pre-trained VGG16 network on ImageNet to manually labeled seismic data, and used a sliding window method to identify the seismic facies at the center of the sliding window, so as to achieve automatic seismic facies classification. Zhao (2019) used a CNN network with an encoder-decoder structure to classify seismic facies. Di (2019) labeled 4 Inline profiles and used a network similar to UNet to achieve automatic seismic facies recognition. Although the above-mentioned seismic facies recognition methods based on deep learning can improve the accuracy of seismic facies recognition, these methods have the following disadvantages, resulting in unable to obtain more efficient and accurate seismic facies interpretation results.

[0004] The above technologies have the following disadvantages:

[0005] (1) Treat the seismic facies recognition task as a single-region classification problem instead of a pixel-level semantic segmentation problem. Therefore, the computational load is large and there is a lot of redundant calculation.

[0006] (2) The above methods are all implemented based on windows. Use the pixels within the window to form blocks as input to predict the seismic facies type of a central point, only using local features and not using global information. Summary of the Invention

[0007] The object of the present invention is to overcome the above-mentioned shortcomings of the prior art and propose a seismic facies automatic recognition method based on the combination of self-attention mechanism and U-shaped structure. Based on the semantic segmentation network of the self-attention mechanism, an encoding-decoding U-shaped structure is introduced. The semantic segmentation network is used as the encoder module, and a block expansion module based on fully connected layer upsampling is introduced in the decoder, and supercolumn technology is used to fuse features, so as to obtain a global attention seismic facies segmentation method with low computational load and capable of extracting multi-scale features.

[0008] The present invention provides a seismic facies automatic recognition method based on the combination of attention mechanism and U-shaped structure. The method includes the following steps:

[0009] Obtain post-stack seismic data and preprocess the post-stack seismic data volume to construct a sample training and validation data set;

[0010] Construct an encoder using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation;

[0011] Construct a decoder using a block expansion module with linear upsampling function, the efficient self-attention transformer module and a skip connection module capable of fusing high and low level features;

[0012] Construct a seismic facies recognition model using the encoder, the decoder and a supercolumn module. The seismic facies recognition model includes a supercolumn unitary segmenter;

[0013] Construct a hybrid loss function, and use the training and validation set in the sample training and validation data set to iteratively train the seismic facies recognition model; input test data to obtain the seismic facies of the test seismic data.

[0014] In addition, according to the seismic facies automatic recognition method based on the combination of self-attention mechanism and U-shaped structure of the present invention, it may also have the following technical features:

[0015] The step of obtaining post-stack seismic data and preprocessing the post-stack seismic data volume to construct a sample training and validation data set specifically includes the following steps:

[0016] Collect the original seismic data, perform preprocessing to obtain a post-stack seismic data volume, and normalize the amplitude of the post-stack seismic data to [0, 1];

[0017] Divide the post-stack seismic data volume into N cross-sectional blocks along the connecting survey line direction, where each cross-sectional block consists of a first sub-block and a second sub-block. Take the first sub-block as the training set and the second sub-block as the validation set; where N is a positive integer greater than 2, and the number of first sub-blocks is not less than the number of second sub-blocks;

[0018] Use the method of linear interpolation to adjust the size of the seismic profile image to a multiple of 16;

[0019] Perform data augmentation on the seismic profile image through left-right flipping and Gaussian noise transformation.

[0020] Constructing the model encoder using the overlapping block merging module with downsampling function and the efficient self-attention transformer module capable of global modeling representation specifically includes the following steps:

[0021] For a seismic image with input height and width of H and W respectively Construct an encoder composite function such that and for the feature map there is a first sub-function where C i is the number of channels of the i-th encoder output feature map, and the first sub-function consists of 1 overlapping block merging module and 2 efficient self-attention transformer modules. The first sub-function includes 4. The 4 first sub-functions form 4 consecutive stages of the encoder. Among them, the overlapping block merging module is implemented through a convolutional layer with a stride smaller than the kernel size. The efficient self-attention transformer module contains a self-attention sub-module and a feed-forward neural network sub-module. The calculation formulas of the self-attention sub-module and the feed-forward neural network sub-module are:

[0022] sAtt(x) = MHSA(LN(x)) + x, (1)

[0023] FFN(x) = L2(cv(L1(LN(x)))) + x (2)

[0024] where LN is the layer normalization function, MHSA is the multi-head self-attention calculation function, L1 and L2 are two fully connected functions, and cv is the convolutional layer function.

[0025] Constructing the model decoder using the block expansion module with linear upsampling function, the efficient self-attention transformer module, and the skip connection module capable of fusing high and low level features specifically includes the following steps:

[0026] Construct a decoder function f d , such that the decoder function f d includes 4 second sub-functions, and the 4 second sub-functions constitute 4 stages of the decoder. For the encoded feature map and the decoded feature map there is where i = {2, 3, 4}, d (4) = x (4) , concat(·, ·) is the tensor concatenation operation along the channel dimension. The feature map of the last stage

[0027] the second sub-function is composed of the block expansion module and 2 of the efficient Transformer modules; when the input feature map is x, the calculation formula of the block expansion module is:

[0028] x = Linear[C, 2C](x), (3)

[0029]

[0030] where, Linear is the fully connected layer, and Reshape is the dimension reshaping operation;

[0031] Connect through the skip connection module in the corresponding stages of the encoder and the decoder, so that the decoder receives the features from the encoder of the same stage for fusion.

[0032] Constructing the seismic facies identification model using the encoder, the decoder, and the hypercolumn module specifically includes the following steps:

[0033] Introduce a hypercolumn structure to fuse the output feature maps of each stage of the decoder and perform pixel-level seismic facies classification on the fused feature map. Its calculation formula is:

[0034]

[0035]

[0036]

[0037] M = Linear[C, N C (df ) (8)

[0038] Among them, Upsample[2 i ×] represents a bilinear interpolation operation for upsampling by a factor of 2 i times, Concat is an operation for concatenating along the channel dimension, and Linear[C, N C is a linear mapping from dimension C to dimension N C , and N C is the number of seismic facies categories; the encoder and the decoder form a U-shaped structure. The left side of the U-shaped structure is the encoder, and the right side of the U-shaped structure is the decoder. The outputs of each stage of the decoder are used for seismic facies classification by fusing features with the supercolumn structure.

[0039] The steps of constructing the hybrid loss function and using the training and validation set in the sample training and validation dataset to iteratively train the seismic facies recognition model; inputting test data to obtain the specific seismic facies of the test seismic data specifically include the following steps:

[0040] Train the seismic facies recognition model using a hybrid loss of the cross-entropy loss function and the Dice Loss. The calculation formula of its loss function is:

[0041] Loss = 0.7 * CE + 0.3 * Dice (9)

[0042] Among them, y is the true seismic facies label of the seismic image, p is the predicted mask of the seismic image, and y i，j represents the one-hot encoded label corresponding to the pixel at i, j in the seismic image, represents the class probability that the pixel at i, j in the seismic image is predicted as the k-th class of seismic facies.

[0043] The method further includes the following steps:

[0044] When training the seismic facies recognition model, use the batch stochastic gradient descent algorithm to iteratively update and learn the parameters of the seismic facies recognition model;

[0045] The steps of using the batch stochastic gradient descent algorithm to iteratively update and learn the parameters of the model specifically include the following steps:

[0046] By calculating the gradient of the loss function, update the parameters of the seismic facies recognition model along the negative direction of the gradient to achieve continuous decrease of the loss function.

[0047] The method of the present invention has the following beneficial effects:

[0048] The present invention combines the self-attention mechanism with the U-shaped structure, and uses the idea of supercolumn technology segmentation for automatic seismic facies segmentation and recognition. At the same time, a hybrid loss function is introduced, which can make the model pay more attention to the continuity of segmentation. Compared with the existing UNet and Segformer models, the seismic facies recognition network model proposed in the present invention not only enables the computer to have a lower computational load, but also can obtain a higher seismic facies recognition accuracy. Further, the position and structure of the underground sedimentary environment are more effectively predicted by using the seismic facies recognition results with high accuracy, providing favorable technical support for oil and gas exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the encoder structure diagram of an embodiment of the present invention;

[0050] Figure 2 is the block expansion module structure diagram of an embodiment of the present invention;

[0051] Figure 3 is the HUSeg network structure diagram of an embodiment of the present invention;

[0052] Figure 4 is the comparison diagram of different semantic segmentation methods for seismic facies recognition in an embodiment of the present invention, where Figure 4 (a) is the seismic profile image; Figure 4 (b) is the seismic facies label map corresponding to the seismic profile; Figure 4 (c) is the seismic facies recognition result map of SegNet; Figure 4 (d) is the seismic facies recognition result map of SegNet; Figure 4 (e) is the seismic facies recognition result map of SegNet; Figure 4 (f) is the seismic facies recognition result map proposed by the present invention;

[0053] Figure 5 is the post-stack three-dimensional seismic data volume map of the Bohai Bay Basin in an embodiment of the present invention;

[0054] Figure 6 is the seismic data horizon slice and recognition result map of an embodiment of the present invention, where Figure 6 (a) is the seismic data horizon slice map; Figure 6 (b) is the horizon slice map of the seismic facies recognition proposed by the present invention. DETAILED IMPLEMENTATION METHOD

[0055] The following further describes in detail the automatic seismic facies recognition method based on the self-attention mechanism and the U-shaped structure provided by the present invention in conjunction with the attached Figures 1-6 drawings:

[0056] The seismic phase interpretation of post-stack seismic data can be regarded as the semantic segmentation of seismic images, which divides the seismic images into different areas, thereby reflecting the different sedimentary environments underground, and can be used to locate oil and gas reservoirs. However, it is difficult to obtain good results by directly using image segmentation methods to divide seismic phases. These methods do not model the correlation between pixels in seismic images, and the effect of depicting the boundary details of some seismic phase categories is poor. Therefore, with the help of an attention mechanism deep neural network model, it can be used to learn the regional correlation and continuity of seismic images, so as to more accurately depict the boundary details of seismic phases. The present invention proposes a seismic phase automatic identification method based on a self-attention mechanism combined with a U-shaped structure, which is used to predict the corresponding seismic phases of seismic data, thereby performing seismic interpretation.

[0057] The present invention provides a method for automatic identification of earthquake phases based on a self-attention mechanism combined with a U-shaped structure, the method comprising the following steps:

[0058] Obtaining a post-stack seismic data volume, and preprocessing the post-stack seismic data volume to construct a sample training and verification data set;

[0059] The encoder is constructed using an overlapping block merging module with downsampling capabilities and an efficient self-attention transformer module capable of globally modeling representations.

[0060] Constructing a decoder using a block expansion module with a linear upsampling function, the efficient self-attention transformer module, and a skip connection module capable of fusing high- and low-level features;

[0061] Constructing a seismic phase identification model using the encoder, the decoder and the supercolumn module, wherein the seismic phase identification model includes a supercolumn unitary partitioner;

[0062] A mixed loss function is constructed, and the seismic phase identification model is iteratively trained using the training validation set in the sample training validation data set; and test data is input to obtain the seismic phase of the test seismic data.

[0063] In a specific implementation, an acquisition device can be used to collect seismic waves to obtain shot-gather seismic data, which can be stacked to form a post-stack seismic data volume. In seismic exploration, in order to facilitate observation and analysis, it is usually necessary to process the post-stack seismic data volume to improve its resolution. Here, shot gather is a special term for seismic data processing, which is a type of seismic data received by a single source excitation.

[0064] In a specific implementation, the efficient Transformer module is a semantic segmentation network.

[0065] In a specific implementation, a post-stack seismic data volume is obtained, and the post-stack seismic data volume is preprocessed to construct a sample training and validation data set, which specifically includes the following steps:

[0066] Collect the original seismic data, and preprocess the original seismic data to obtain post-stack seismic data, denoted as where I, C, and D are the number of main survey lines (Inline), the number of connecting survey lines (Crossline), and the number of time sampling points respectively, and the amplitude of the entire volume is normalized to [0, 1]. The entire seismic volume training data is evenly divided into 10 cross-sectional blocks along the Crossline direction; the first 70% of the sub-blocks in the Crossline direction within each cross-sectional block are taken as the training set, and the last 30% of the sub-blocks are taken as the validation set. The seismic images formed by the cross-sections within each sub-block are resized, and the height and width (H, W) of the seismic images are transformed into the nearest multiple of 16 using linear interpolation, that is, the resolution is transformed into Finally, the cross-sectional seismic images obtained after left-right flipping and Gaussian noise transformation are used as the input sample data set of the model.

[0067] In a specific implementation, constructing a model encoder using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation specifically includes the following steps:

[0068] For a seismic image with input height and width of H and W respectively Construct an encoder composite function such that and for the feature map there is a first sub-function where C i is the number of channels of the output feature map of the i-th sub-function of the encoder. As Figure 1 shown, each first sub-function of the encoder consists of 1 overlapping block merging module and 2 efficient Transformer modules, which is called a sub-encoding block. Each sub-encoding block can halve the spatial dimension of the input feature map. Thus, 4 sub-encoding blocks form 4 consecutive stages of the encoder, and 4 feature maps with different scales can be obtained to provide features at different abstraction levels.

[0069] In each sub-encoding block, the overlapping block merging module performs matrix multiplication on the input data through an overlapping sliding window method, thereby realizing linear embedding of the feature map and being able to downsample the spatial resolution dimension of the input data by 2 times. The efficient Transformer module includes a self-attention sub-module and a feed-forward neural network sub-module, which are used to learn the global features and fusion features between different positions in the cross-sectional image. For the input feature map x of the i-th stage (i), the sub - coding block calculation can be expressed as:

[0070]

[0071] Among them, is a composite function operator, Conv is an overlapping block merging module implemented using convolutional operations, sAtt is a self - attention sub - module calculation function, and FFN is an efficient feed - forward neural network sub - module calculation function. The calculation formulas of the two are as follows:

[0072] sAtt(x) = MHSA(LN(x)) + x, (1)

[0073] FFN(x) = L2(cv(L1(LN(x)))) + x (2)

[0074] Among them, LN is layer normalization, MHSA is a multi - head self - attention calculation function, L1 and L2 are two fully - connected functions, and cv is a convolutional layer function for position encoding.

[0075] In specific implementation, constructing a model decoder using a block dilation module with linear up - sampling function, the efficient self - attention transformer module, and a skip - connection module capable of fusing high - and low - level features specifically includes the following steps:

[0076] For the multi - stage encoded features encoded by the encoder Construct a decoder function f d , such that The decoding process adopts a symmetric manner with the encoding process. The decoder function f d includes 4 second - order sub - functions, and the 4 second - order sub - functions constitute 4 stages of the decoder. For the encoded feature map and the decoded feature map There is where i = {2, 3, 4}, d (4) = x (4) , [·, ·] is a tensor concatenation operation along the channel dimension. The feature map of the last stage

[0077] Among them, the second - order sub - function is composed of 1 block dilation module based on fully - connected layer up - sampling and 2 efficient Transformer modules, and is called a sub - decoding block. As Figure 2Shown is a block expansion module. The idea is to exchange the channel dimension for the spatial dimension. After doubling the channel dimension of the data through a fully connected layer, a dimension reshaping operation is performed. The specific operation is to split its channels into four parts, and every two parts are interleaved and concatenated along the spatial dimension, thereby doubling the spatial dimension while halving the channel dimension. Assuming the input feature map is x, the described block expansion module can be described as:

[0078] x = Linear[C, 2C](x), (3)

[0079]

[0080] Among them, Linear is the fully connected layer, and Reshape is the dimension reshaping operation.

[0081] At the same time, in the corresponding stages of the encoder and the decoder, they are connected through a "skip connection" structure, and the features of the encoder and the decoder are concatenated, so that the decoder can receive the features of the encoder in the same stage for fusion, in order to fuse different semantic features of shallow coarse-grained and high-level fine-grained.

[0082] In specific implementation, constructing a seismic facies recognition model using the described encoder, decoder, and supercolumn module specifically includes the following steps:

[0083] Introduce a supercolumn structure to fuse the output feature maps of each stage of the decoder and then perform pixel-level seismic facies classification on the fused feature maps. Classifying using features at multiple levels in this way can improve the effect of dense prediction tasks. The calculation process is as follows:

[0084]

[0085]

[0086]

[0087] M = Linear[C, N C (d f )(8)

[0088] Among them, Upsample[2 i ×] represents the bilinear interpolation operation of upsampling by 2 i times, Concat is the operation of concatenating along the channel dimension, Linear[C, N C is the linear mapping from dimension C to dimension N C , N C is the number of seismic facies categories. The constructed seismic facies recognition model is as Figure 3As shown, the encoder and the decoder form a U-shaped structure. The left side of the U is the encoder, and the right side is the decoder. The outputs of each stage of the decoder are fused by the hypercolumn structure to classify seismic facies. The present invention names this network the Hypercolumns-U-Segformer (HUSeg) network.

[0089] In a specific implementation, a hybrid loss function is constructed, and the seismic facies recognition model is iteratively trained using the training and validation set in the sample training and validation dataset; input test data to obtain the seismic facies of the test seismic data, which specifically includes the following steps:

[0090] Select pixel-level cross-entropy loss (CE) as the main optimization objective for model training. At the same time, in order to enable the model to consider regional correlation, Dice loss is used for assistance. Therefore, the specific loss function is:

[0091] Loss = 0.7 * CE + 0.3 * Dice (9)

[0092] Where, y is the true seismic facies label of the seismic image, p is the predicted mask of the seismic image, and y i，j represents the one-hot label corresponding to the pixel at i, j in the seismic image, represents the class probability that the pixel at i, j in the seismic image is predicted as the k-th type of seismic facies. When training the model, an Adam optimizer with a batch size of 8, an initial learning rate of 1e-3, and a weight decay of 1e-4 is used to train and learn the model.

[0093] In a specific implementation, the method further includes the following steps:

[0094] When training the seismic facies recognition model, the batch stochastic gradient descent algorithm is used to iteratively update and learn the parameters of the seismic facies recognition model;

[0095] After the seismic facies recognition model is trained, the size of the seismic profile image is adjusted to a multiple of 16 to input the seismic facies recognition model to predict the seismic facies category.

[0096] In a specific implementation, using the batch stochastic gradient descent algorithm to iteratively update and learn the parameters of the model specifically includes the following steps:

[0097] By calculating the gradient of the loss function, update the parameters of the seismic facies recognition model along the negative direction of the gradient to achieve continuous decrease of the loss function.

[0098] After the model is trained, for any seismic profile, the size of the seismic profile image is adjusted to a multiple of 16 using linear interpolation and then input into the model for prediction. The size of the output tensor of the model is [H, W, N C , and the index of the maximum probability in the third dimension is taken as the predicted seismic facies category. Finally, a seismic facies matrix of size [H, W] is output, where the value at each position represents the seismic facies category corresponding to the position in the input seismic image.

[0099] Numerical simulation results

[0100] First, to verify the effectiveness of the present invention, a publicly available seismic dataset from the F3 block in the Dutch North Sea, annotated by Alaudah et al. (2019), is used for verification. The part of Inline 300 - 700 and Crossline 300 - 1000 is selected as the training set, and the rest is used as the test set. Preprocessing is carried out according to step (2) of the specific implementation method in the specification, and then training and testing are carried out according to step (6) to obtain a trained model, which is then evaluated.

[0101] As Figure 4 shown, the seismic facies segmentation results of the seismic facies recognition model proposed by the present invention on the inline400 seismic profile of the test set are presented. In this experiment, the proposed HUSeg seismic facies automatic recognition method of the present invention is compared with SegNet, UNet, and Segformer respectively. Figure 4 (a) and (b) are respectively the seismic profile image of the inline400 of the test set and its corresponding seismic facies label, Figure 4(c)-(f) are the seismic facies identification results of SegNet, UNet, Segformer, and the proposed HUSeg model of the present invention, respectively. The accuracy rate of the SegNet test set is 0.851, the accuracy rate of the UNet test set is 0.861, the accuracy rate of the Segformer test set is 0.903, and the accuracy rate of the proposed HUeg test set of the present invention is 0.931. To further illustrate the advantages of the network model proposed in the present invention, the computational amounts of the UNet, Segformer, and the proposed HUSeg model are compared. Since the accuracy rate of the SegNet model is relatively low, the computational amount of the SegNet model is not compared here. When the input sample is an image of 128X128, the computational amounts of the UNet, Segformer, and the proposed HUSeg model are 13.65 (billion), 4.41 (billion), and 4.37 (billion), respectively. That is to say, the proposed HUSeg network model of the present invention not only enables the computer to have a relatively low computational amount, but also can obtain a higher seismic facies identification accuracy rate. SegNet introduces partial shallow seismic facies noise in the deep seismic facies region of the seismic image, and introduces a large amount of deep seismic facies noise in the middle seismic facies region. UNet introduces a small amount of adjacent seismic facies noise in the shallow seismic facies region, and introduces various seismic facies noises in the middle seismic facies region. Segformer only introduces a large amount of shallow seismic facies in the middle-layer seismic facies region. However, the effect of the proposed HUSeg model of the present invention is better than the results of all the above models. It not only removes the noise in the middle-layer seismic facies, but also reduces a large amount of other seismic facies noise in the deep seismic facies region. The boundary details of its seismic facies are smoother and the lateral continuity is better.

[0102] Finally, the effectiveness of the present invention is further verified on a post-stack seismic data in the Bohai Bay Basin. The main frequency of the post-stack seismic data is 29 Hz and the frequency band is 6 - 52 Hz. The proposed HUSeg model of the present invention is used to identify the seismic facies of all inline profiles of the seismic volume, and the results are stitched into a seismic facies volume, and the layer slices are taken to interpret the sedimentary facies. Figure 5 The 3D seismic data is shown, where the Inline number is 7596 and the Xline number is 3180. As Figure 6 shown in the layer slice and the seismic facies identification result, the layer slice along the middle sub-member of the Es3 formation can be interpreted as sedimentary microfacies such as delta plain, swamp, distributary channel, natural levee, interdistributary bay, and semi-deep lake.

Claims

1. An automatic seismic facies identification method based on the combination of self-attention mechanism and U-shaped structure, characterized in that, Including the following steps: Obtain a post-stack seismic data volume, and preprocess the post-stack seismic data volume to construct a sample training and validation data set; Construct an encoder using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation; Construct a decoder using a block expansion module with linear upsampling function, the efficient self-attention transformer module, and a skip connection module capable of fusing high and low level features; Construct a seismic facies identification model using the encoder, the decoder, and a hypercolumn module, the seismic facies identification model including a hypercolumn unitary segmenter; Construct a hybrid loss function, and iteratively train the seismic facies identification model using the training and validation set in the sample training and validation data set; input test data to obtain the seismic facies of the test seismic data; The specific steps of constructing the model encoder using the overlapping block merging module with downsampling function and the efficient self-attention transformer module capable of global modeling representation include the following steps: For an earthquake image with an input height and width of H and W respectively Construct an encoder composite function such that and for the feature map there is a first sub-function wherein is a composite function operator, C i is the number of channels of the output feature map of the i-th encoder, and the first sub-function is composed of 1 overlapping block merging module and 2 efficient self-attention transformer modules. The first sub-function includes 4. Four such first sub-functions form 4 consecutive stages of the encoder. Among them, the overlapping block merging module is implemented through a convolutional layer with a stride smaller than the kernel size, and the efficient self-attention transformer module includes a self-attention sub-module and a feed-forward neural network sub-module. The calculation formulas of the self-attention sub-module and the feed-forward neural network sub-module are as follows: sAtt(x) = MHSA(LN(x)) + x, (1) FFN(x) = L2(cv(L1(LN(x)))) + x (2) Where, LN is a layer normalization function, MHSA is a multi-head self-attention calculation function, L1, L2 are two fully connected functions, and cv is a convolutional layer function; The specific steps of constructing the model decoder using the block expansion module with linear upsampling function, the efficient self-attention transformer module, and the skip connection module capable of fusing high and low level features include the following steps: Construct a decoder function f d , such that the decoder function f d includes 4 second sub-functions, and the 4 second sub-functions constitute 4 stages of the decoder. For the encoded feature map and the decoded feature map there is where i = {2, 3, 4}, d (4) = x (4) , concat(·, ·) is the tensor concatenation operation along the channel dimension, and the feature map of the last stage The second sub-function is composed of the block dilation module and two of the efficient Transformer modules; when the input feature map is x, the calculation formula of the block dilation module is: x = Linear[C, 2C](x), (3) Where, Linear is a fully connected layer, Reshape is a dimension reshaping operation, C is the number of crosslines, Connect through the skip connection module in the corresponding stages of the encoder and the decoder, so that the decoder receives the features from the encoder in the same stage for fusion; The specific steps of constructing the seismic facies identification model using the encoder, the decoder, and the hypercolumn module include the following steps: Introduce a super-column structure to fuse the output feature maps of each stage of the decoder and perform pixel-level seismic facies classification on the fused feature maps, and its calculation formula is as follows: M = Linear[C, N C (d f ) (8) Among them, Upsample[2 i ×] represents a bilinear interpolation operation with an upsampling factor of 2 i times. Concat is an operation for concatenating along the channel dimension. Linear[C, N C is a linear mapping from dimension C to dimension N C , and N C is the number of seismic facies categories. The supercolumn unitary splitter is a U-shaped structure composed of the encoder and the decoder. The left side of the U-shaped structure is the encoder, and the right side of the U-shaped structure is the decoder. The outputs of each stage of the decoder are used for seismic facies classification by fusing features with the supercolumn structure.

2. The automatic seismic facies identification method based on the combination of self-attention mechanism and U-shaped structure according to claim 1, characterized in that, The specific steps of obtaining the post-stack seismic data volume and preprocessing the post-stack seismic data volume to construct a sample training and validation data set include the following steps: Collect the original seismic data, obtain the post-stack seismic data volume after preprocessing, and normalize the amplitude of the post-stack seismic data to [0, 1]; Divide the post-stack seismic data volume into N cross-sectional blocks along the crossline direction, where each cross-sectional block is composed of a first sub-block and a second sub-block, and use the first sub-block as the training set and the second sub-block as the validation set; where, N is a positive integer greater than 2, and the number of the first sub-blocks is not less than the number of the second sub-blocks; Use the method of linear interpolation to adjust the size of the seismic profile image to a multiple of 16; Perform data augmentation on the seismic profile image through left-right flipping and Gaussian noise transformation.

3. The automatic seismic facies identification method based on the combination of self-attention mechanism and U-shaped structure according to claim 1, characterized in that, The specific steps of constructing the hybrid loss function, and iteratively training the seismic facies identification model using the training and validation set in the sample training and validation data set; input test data to obtain the seismic facies of the test seismic data include the following steps: The seismic facies recognition model is trained using a hybrid loss function that combines cross-entropy loss and Dice Loss. The calculation formula for the loss function is as follows: Loss = 0.7 * CE + 0.3 * Dice (9) Among them, y is the true seismic phase label of the seismic image, p is the predicted mask of the seismic image, and y i，j represents the one-hot encoded label corresponding to the pixel at (i, j) in the seismic image, represents the class probability that the pixel at (i, j) in the seismic image is predicted to be the k-th seismic phase.

4. The automatic seismic facies identification method based on the combination of self-attention mechanism and U-shaped structure according to claim 3, characterized in that, The method further includes the following steps: When training the seismic facies recognition model, the batch stochastic gradient descent algorithm is used to iteratively update and learn the parameters of the seismic facies recognition model; After the seismic facies recognition model is trained, the size of the seismic profile image is adjusted to a multiple of 16 to input the seismic facies recognition model to predict the seismic facies category.

5. The automatic seismic facies identification method based on the combination of self-attention mechanism and U-shaped structure according to claim 4, characterized in that, The use of the batch stochastic gradient descent algorithm to iteratively update and learn the parameters of the model specifically includes the following steps: By calculating the gradient of the loss function, the parameters of the seismic facies recognition model are updated along the negative direction of the gradient to achieve continuous decrease of the loss function.

Images