CN116027404A - A Small-Sample Prestack Seismic Reflection Pattern Analysis Method Based on Large Kernel Attention - Google Patents

Abstract

The invention discloses a small sample pre-stack seismic reflection mode analysis method based on large nuclear attention, which is applied to the field of seismic data processing and aims at the problem of insufficient identification capability of complex geological structures in the prior art; the method comprises the steps of constructing an image classification seismic reflection mode analysis model based on an improved ConvNext framework; the receptive field of the model is increased by introducing a large nuclear attention (LKA) mechanism, the receptive capacity of the model to global information is enhanced, and the space transverse characteristics and reflection information of the stratum are fully utilized and learned; training the model of the step S1 by adopting the expanded labeled pre-stack seismic data; and finally, predicting unlabeled pre-stack seismic pictures according to the model trained in the step S2, and inserting the result into the seismic section according to xline and inline coordinates of the pictures to finally obtain the whole section prediction result.

Description

A Small-Sample Prestack Seismic Reflection Pattern Analysis Method Based on Large Kernel Attention

technical field

The invention belongs to the field of seismic data processing, in particular to a pre-stack seismic reflection mode analysis technology.

Background technique

Seismic exploration is a commonly used method in the field of geophysics, which provides the necessary basis for the identification of oil and gas reservoirs. In seismic exploration, different lithological parameters and different structural features (such as cracks and caves) in the ground will cause differences in seismic reflection signals. In order to emphasize the similarity of the same type of seismic reflection signals and the difference of different types of seismic reflection signals , we define the unique characteristics of a class of seismic reflection signals that are different from other seismic reflection signals as the mode of this type of seismic reflection signal. By analyzing the seismic reflection mode, we can mine the common characteristics of the same type of seismic reflection signals, predict different underground geological structures, and provide a theoretical basis for subsequent geological exploration.

In the early days, seismic reflection pattern analysis was usually performed manually, which relied heavily on expert knowledge and rich experience, and the results were highly subjective. Moreover, some subtle reflection information in seismic data is difficult to observe and describe artificially. In recent years, seismic reflection pattern analysis technology has begun to develop toward automation and intelligence. Artificial intelligence methods such as k-means, SVN, and CLS have also been gradually applied to seismic pattern analysis, which can more effectively describe formation lithology and related formation structures. Changes are characterized by rapidity, quantification and objectivity.

In the past, limited by the signal-to-noise ratio of seismic signals, it is usually chosen to process and superimpose the collected raw data with multiple coverages to form post-stack data, and analyze the reflection mode of seismic signals based on a single post-stack seismic signal. However, due to the loss of multi-angle information, the post-stack seismic signal cannot represent the change of the seismic reflection signal with azimuth or offset, and the single-channel reflection mode analysis cannot make full use of the spatial and lateral characteristics of the formation. Therefore, based on the post-stack seismic data Reflection pattern analysis is often difficult to accurately predict subsurface reservoirs. Nowadays, with the development of seismic signal processing technology, the quality of pre-stack multi-channel seismic signals has been significantly improved. The intelligent analysis method based on pre-stack signals can make full use of the rich reflection information between strata and effectively reduce the The multi-solution nature of the analysis makes geological exploration continue to develop towards refinement.

According to whether well logging labels are used in the classification process, seismic signal reflection pattern analysis methods can be divided into supervised and unsupervised categories. At present, traditional seismic reflection pattern analysis based on prestack data is basically an unsupervised method. The unsupervised method is mainly based on the clustering algorithm and the feature mapping method: the clustering algorithm starts from the data as a whole, and automatically classifies the seismic reflection signals into different clusters according to different difference criteria, and each type of signal belongs to the same seismic reflection model. The feature mapping method transforms the seismic attributes in order to display the underlying structural characteristics or reservoir characteristics. The supervised method establishes constraints based on existing logging information, and obtains an optimal classification model through manual or computer training, and then realizes seismic reflection pattern recognition. Among them, the supervised algorithm can make full use of the accurate information of well logging data and has high reliability, so it has gradually received widespread attention.

In actual projects, due to the high cost of obtaining well logging data, the quantity is often very scarce, which cannot meet the training needs of supervised models. Under such small sample conditions, model training is prone to overfitting, and it is difficult to apply to the reflection mode analysis of large batches of unlabeled prestack seismic data. At the same time, the current traditional supervised models are based on the idea of traditional image processing, with a small receptive field, and cannot capture the connection between different angle gathers in the pre-stack seismic data.

Supervised learning refers to the machine learning task of learning intrinsic features from labeled data and inferring target values. Supervised seismic reflection pattern analysis mainly uses the convolutional neural network as the core architecture, establishes learning tasks by constructing a network model, and then trains model parameters with the help of labels (usually logging data), and finally enables the model to realize the prediction of seismic phases. Among them, supervised seismic pattern analysis can be roughly divided into two methods, image classification and semantic segmentation, according to the task category of building the model.

1. Seismic pattern analysis based on image classification

The image classification method refers to calculating the category of the image as a whole through the algorithm model. In 2012, Alex Krizhevsky proposed the AlexNet network, and won the first place in the ISLVRC Challenge through a more reasonable network construction, which also greatly increased the attention in the field of image classification. Subsequently, the image classification model developed rapidly. In 2014, Simonyan designed the VGG network architecture to achieve a larger receptive field by stacking multiple small convolution kernels; in 2015, He et al. proposed a residual network architecture and designed a ResNet model. The module realizes the fusion between different levels of features; in 2017, howard et al. proposed a lightweight network mobileNet, which greatly reduces the model parameters while ensuring the training effect, and is often used in devices with weaker performance such as mobile terminals ; In 2021, the Vision transformer proposed by Alexey et al. introduced the attention mechanism in natural language processing into image processing, and divided the image into small patches to calculate the self-attention and replace the convolutional network, which significantly improved the classification accuracy.

In the field of seismic exploration, image classification methods have also been widely used: in 2016, Alaudah et al. proposed a weakly supervised label mapping algorithm, which can generate a large amount of training data by inputting a small number of pre-stack labels, and use it for seismic pattern analysis; 2018 In 2010, Chevitarese et al. proposed that by dividing the pre-stack seismic profile image into multiple small image patches along the length and width, and assuming that each patch belongs to the same category, these patches are input into the classification network for training and prediction, and then combined. The complete classification results of seismic sections based on pre-stack data can be obtained; Dramsch et al. proposed that the sliding window mechanism can be used for pre-stack seismic sections to greatly increase the number of cut patches and improve the effect of model training. The basic framework of prestack seismic model analysis based on image classification is shown in Fig. 1.

2. Analysis of Seismic Patterns Based on Semantic Segmentation

In addition to image classification, semantic segmentation can also be used to classify images, that is, to distinguish different geological targets on the same seismic section. Unlike classification, semantic segmentation does not output a single category for each image, but outputs the category for each pixel on the image. In 2015, Long et al. proposed the FCN model (Full Convolutional Neural Network), which built the first semantic segmentation model by migrating the feature extraction part of the classification task to the segmentation task and adding upsampling to restore the image size. Subsequently, the field of semantic segmentation developed rapidly, and achieved good results in medical images, road real scenes and other aspects. In the same year, Ronneberger et al. proposed the Unet network, which solved the problem of insufficient labeling of traditional medical images to a certain extent; Chen et al. proposed the Deeplab network, which expanded the receptive field by introducing hole convolution, obtained global information of the image, and reduced the details of the image during the downsampling process. Loss; In 2016, Zhao et al. proposed the PSPNet network. By introducing a pyramid pooling module, multi-scale pooling of deep features was realized, which effectively improved the classification effect.

In the field of geological exploration, the idea of image segmentation has also begun to be widely used: in 2018, Dramsch implemented a patch-based CNN seismic pattern analysis, and the classification results can be obtained by inputting pre-stack seismic data. In 2019, Alaudah et al. proposed a deconvolution network based on small-scale images and seismic sections. The model can not only realize pre-stack seismic model analysis, but also complete data expansion and increase the number of samples. In 2020, Zhang proposed that the Deeplabv3+ network model could be used to extract multi-scale features of pre-stack seismic data through the ASPP pyramid structure, and good results were also obtained. The basic framework of prestack seismic pattern analysis based on semantic segmentation is shown in Fig. 2.

In the reflection mode analysis method of pre-stack multi-channel seismic data, traditional supervised methods often rely on abundant logging sample data:

(1) Seismic pattern analysis based on semantic segmentation requires a large number of complete seismic sections with labels as training data. However, it is very expensive and difficult to obtain well logging label data in actual projects. The fitting phenomenon cannot be extended to large-scale unlabeled data;

(2) Although the method based on image classification expands the number of samples to a certain extent by dividing the labeled seismic image into patches, the repetitiveness of the labels is high, which still cannot meet the needs of model training, and it is necessary to assume that each patch belongs to Same seismic pattern, less accurate.

In addition, the existing supervised pre-stack models are all directly transplanted based on traditional image processing models. When processing pre-stack seismic data as ordinary images, it is impossible to learn the unique characteristics of pre-stack seismic data. Rich features. As shown in Fig. 3, the multi-angle gathers obtained by stacking are rich in formation reflection information and anisotropy features, but the receptive field of the traditional supervised model is usually small, and it is difficult to capture different angle gathers with long distances between global deep features.

Contents of the invention

In order to solve the above technical problems, the present invention proposes a small-sample pre-stack seismic reflection pattern analysis method based on large-core attention. The large-core attention mechanism is introduced into the ConvNext model to learn the deep global features of the data, and through single-point prediction and Stitching completes pre-stack seismic reflection pattern analysis.

The technical solution adopted in the present invention is: a small sample pre-stack seismic reflection mode analysis method based on large core attention, including:

S1. Construct an image classification seismic reflection pattern analysis model based on the improved ConvNext framework;

S2. Using the expanded labeled pre-stack seismic data to train the model in step S1;

S3. Predict the unmarked pre-stack seismic pictures according to the model trained in step S2, and then splice the single-point prediction results at all positions into a complete seismic profile prediction result.

The model described in step S1 is specifically: using four cascaded improved ConvNext modules to expand the number of channels and extract deep features of pre-stack seismic images, and add a downsampling module between two adjacent improved ConvNext modules for Change the size of the pre-stack seismic image, and finally convert the length and width of the input pre-stack seismic image to 1×1 through the global pooling layer, and then map the number of channels to the number of categories to be classified through the fully connected layer to obtain the input pre-stack seismic image Seismic pictures are respectively judged as the probability of each category, and the maximum probability is the final classification result corresponding to the pre-stack seismic picture.

The improved ConvNext module is specifically based on the anti-bottleneck structure, adding LayerNorm regularization, GELU activation function, Layer Scale parameter scaling and Drop Path layer; the anti-bottleneck structure is a level of an Attention module and two Linear fully connected layers The connection structure; the Attention module replaces the 7×7 convolutional layer in ConvNext with a large-core attention layer.

The expansion process of the labeled pre-stack seismic data in step S2 is: obtain the overall distribution of different seismic facies on the seismic section through a clustering algorithm, and then correct and expand it by combining the accurate labels of the logging data; it includes the following steps:

A1. Cluster the post-stack seismic sections, classify the seismic sections according to the clustering results, and box out the parts of the clustering results belonging to the same cluster;

A2. In the clustering results, according to the xline and inline positions of the logging data, project the logging data into the post-stack seismic section, and then combine the accurate seismic facies information and expert experience with the logging interpretation data to cluster The results are augmented;

A3. Calculate the location range of the 3D pre-stack data xline and inline number corresponding to the expanded labeled data, and project it onto the 3D pre-stack seismic data volume to obtain the pre-stack multi-channel seismic data at the corresponding position;

A4. Draw the labeled pre-stack multi-channel seismic data at each xline and inline position as n curves, and fill the positive value area of the curve to generate a picture, and obtain a large amount of pre-stack labeled data. The value of n is determined by the number of pre-stack seismic traces, and one seismic data corresponds to one curve.

Beneficial effects of the present invention: on the model, the present invention combines the characteristics of the pre-stack seismic signal, on the basis of the traditional ConvNext model, increases the receptive field of the model by introducing a large kernel attention (LKA) mechanism, and enhances the model's ability to global information The ability to feel, make full use of and learn the spatial and lateral characteristics and reflection information of the formation. In terms of data processing, the combination of clustering algorithm and logging labels to expand the number of labels is more scientific and reasonable than the traditional cutting method. At the same time, the input is replaced by small-scale image patches with pre-stack multi-channel seismic data images corresponding to each xline and inline position, which not only expands the label data to a greater extent, but also further improves the model's ability to identify complex geological structures .

The present invention adopts the large core attention mechanism to mine the global features of the pre-stack seismic data, and combines the ConvNext image classification model to realize the intelligent seismic pattern analysis based on the pre-stack data. Compared with the traditional method of dividing pictures into small blocks for classification, the method of the present invention has a higher utilization rate of label data. Experiments show that:

The present invention makes full use of the only well logging data, and greatly expands the number of labels by combining machine learning clustering and manual correction, effectively improving the effect of model training.

The pre-stack seismic model analysis results of the present invention are better than the prediction results of traditional models (such as traditional ConvNext models and AlexNet models).

Description of drawings

Figure 1 is a frame diagram of seismic pattern analysis based on image classification;

Figure 2 is a framework diagram of seismic pattern analysis based on semantic segmentation;

Figure 3 is a pre-stack multi-channel seismic data map;

Figure 4 is a framework diagram of pre-stack seismic reflection pattern analysis based on large kernel attention and ConvNext model;

Figure 5 is a schematic diagram of the large core attention mechanism;

Figure 6 is an architecture diagram of the Attention module;

Figure 7 is an anti-bottleneck structure based on large core attention;

Figure 8 is a schematic diagram of the structure of the LKA-ConvNext Block and DowmSample modules;

Figure 9 is a network framework diagram;

Figure 10 is a clustering result map of post-stack seismic sections;

Figure 11 is a schematic diagram of post-stack seismic section correction;

Figure 12 is a schematic diagram of pre-stack seismic data;

Figure 13 is a pre-stack label filling picture;

Fig. 14 is the profile diagram of earthquake prediction by different models;

Wherein, (a) the prediction result of the model of the present invention; (b) the prediction result of the traditional ConvNext model; (c) the prediction result of the AlexNet model.

Detailed ways

In order to facilitate those skilled in the art to understand the technical content of the present invention, the content of the present invention will be further explained below in conjunction with the accompanying drawings.

The invention proposes an image classification seismic reflection mode analysis model based on the improved ConvNext framework and a corresponding seismic data processing method. In terms of the model, the present invention combines the characteristics of the pre-stack seismic signal, on the basis of the traditional ConvNext model, through the introduction of the Large Kernel Attention (LKA) mechanism to increase the receptive field of the model, enhance the model's ability to perceive global information, and fully utilize and Learn the spatial lateral properties and reflection information of the formation. In terms of data processing, the combination of clustering algorithm and logging labels to expand the number of labels is more scientific and reasonable than the traditional cutting method. At the same time, the input is replaced by small-scale image patches with pre-stack multi-channel seismic data images corresponding to each xline and inline position, which not only expands the label data to a greater extent, but also further improves the model's ability to identify complex geological structures . Concrete method and flow chart of the present invention are as shown in Figure 4, comprise the following steps:

(1) Based on a small amount of logging tag data, tag data expansion processing is performed;

(2) Input the expanded labeled pre-stack seismic data as a sample into the prediction category of the model;

(3) Calculate the loss between the real category and the predicted category of the input pre-stack data, and update the model parameters;

Model loss selects cross-entropy loss cross-entropy, the formula is as follows:

H(p,q)=-∑p(x)logq(x)

Among them, p is the real label distribution of pre-stack data, q is the normalized distribution calculated by the model, and x is the input seismic data picture;

(4) Repeat steps (2)-(3) until the loss cannot continue to decrease, indicating that the model training is completed;

(5) Call the trained model to directly predict the unlabeled pre-stack seismic images, and then insert the results into the seismic section according to the xline and inline coordinates of the seismic images, and finally obtain the overall section prediction results.

1. Network model

The network model used in the present invention is an improvement based on the ConvNext network. The present invention replaces the 7×7 convolutional layer in ConvNext with a large kernel attention layer (LKA, largekernel attention layer) that is more effective and more sensitive to global features. ). The following section will introduce in detail from two aspects of network composition and overall network architecture.

11. Network composition

The improved network model of the present invention is based on the ConvNext network architecture, and the whole is composed of LKA-ConvNext Block and down-sampling layer (down sample) alternately stacked, and each part mainly uses large-core attention, anti-bottleneck structure (Inverted bottleneck) and other operations . Next, we will first introduce the large-core attention module and the anti-bottleneck structure, and then introduce each component module in the network model in detail.

111. Large kernel attention (LKA, Large kernel attention)

In the field of computer vision, there are usually two methods to enhance the receptive field and capture long-distance information. The first method is based on transformer's self-attention mechanism, and the second method is large kernel convolution. The self-attention mechanism was originally designed for one-dimensional language processing tasks, so when processing images, it also treats two-dimensional structures as one-dimensional sequences, destroying the key two-dimensional characteristics of images. The large kernel convolution introduces a large number of parameters and calculations, which doubles the training time of the model. Based on the problems of the above two methods, the present invention introduces a large-core attention mechanism. The schematic diagram of large kernel attention is shown in Fig. 5. Similar to the depth separable convolution, it decomposes a convolution with a kernel size of k into the sum of three convolutions, which are depth convolutions with a convolution kernel size of k/d, and a convolution kernel size of 2d -1. Hole convolution with dilation rate d and channel convolution with convolution kernel size 1×1. It also has the characteristics of space adaptability, channel adaptability, and long-distance dependent learning ability. Next, based on the large-core attention mechanism, the activation function and channel convolution are supplemented to construct a complete Attention module. The structure is shown in Figure 6.

112. Inverted bottleneck structure

The anti-bottleneck structure adopts the mode of first increasing the dimension and then reducing the dimension. It is believed that this method can convert information between different dimensional feature spaces, thereby avoiding the information loss caused by dimension reduction and compression, and improving the model effect. The present invention constructs an anti-bottleneck architecture that integrates large-core attention by cascading the Attention module and two Linear full-connection layers, in which the height and width of the input and output remain unchanged, and only the number of channels changes. The principle is shown in Figure 7.

113. Network module

The network module of the present invention mainly includes LKA-ConvNext Block and down-sampling layer (downSample), LKA-ConvNext Block refers to the original ConvNext model, and adds LayerNorm regularization, GELU activation function, and Layer Scale parameter scaling on the basis of the anti-bottleneck structure and Drop Path layer to make its structure more perfect and realize the extraction of different levels of image features. In the LKA-ConvNext Block, the dimensions of the input and output remain unchanged. The main function of the downsampling layer is to change the image dimension through a 2×2 convolution with a step size of 2. The architecture of the two modules is shown in Figure 8.

12. The overall structure of the network

The overall structure of the improved network used in the present invention is shown in FIG. 9 . First, the pre-stack seismic labels are uniformly scaled to a size of 224×224×3, and then pass through a layer of 4×4 convolution and normalization layers, and then through multiple cascaded LKA-ConvNext Blocks to expand the number of channels and extract deep layers Features, add three layers of downsampling layers in the middle to change the image size, and finally convert the input length and width to 1×1 through Global Avg Pooling, and then map the number of channels to the number of categories that need to be classified through the fully connected layer (classes) to get the probability that the input is judged as each category, and the maximum probability is the classification result judged by the network.

2. Data expansion processing

For the processing of label data, the present invention learns the processing process of handwritten digit recognition in the field of computer vision, and converts multi-channel pre-stack seismic data into image form to train the network. The core idea is to obtain the overall distribution of different seismic facies on the seismic section through a clustering algorithm, and then manually correct and expand it by combining the accurate labels of the logging data. The specific process is as follows:

(1) Cluster the post-stack seismic sections, classify the seismic sections according to the clustering results, and manually frame out the part of the clustering results that belong to the same cluster, and the different clusters are shown in Figure 10;

(2) In the clustering results, the specific categories of different clusters are not clear. According to the xline and inline positions of the logging data, the logging data is projected into the post-stack seismic section, and then the clustering results are expanded by combining the accurate seismic facies information and expert experience carried in the logging interpretation data, as shown in Figure 11. The specific situations can be divided into the following two types:

a. When the well logs are distributed in the clustering results (triangle label), it can be determined that the clustering results of this class all belong to the category described in the well logging, thereby realizing the expansion of the label data;

b. When the logging distribution is outside the background or clustering results (circular label), combined with expert experience to analyze the geological structure, it can be judged that the blocks near the logging belong to the same category, and the clustering results are corrected and expansion;

(3) Calculate the position range of the 3D pre-stack data xline and inline number corresponding to the expanded large batch of labeled data, and project it onto the 3D pre-stack seismic data volume to obtain the pre-stack multi-channel seismic data at the corresponding position, such as Figure 12;

(4) Draw the labeled pre-stack multi-channel seismic data at each xline and inline position as n curves, and fill the positive value area of the curve to generate a picture (Figure 13), and obtain a large number of pre-stack labeled data;

The method of the present invention is applied to the actual pre-stack seismic data of a work area in Sichuan to test the effectiveness of the method. The Inline range of this seismic volume data is 7600-7970, and the Xline range is 7700-8250, including three seismic phases: background, channel 1 and channel 2. After the above label data processing steps, 78,594 labeled images are finally obtained, including 33,494 background labels, 19,896 river channel 1 labels, and 25,204 river channel 2 labels. In this experiment, in addition to the improved LKA-ConvNext model proposed by the present invention, a conventional ConvNext model and a small-scale AlexNet model were selected as comparative experiments to verify the effect of the improved model proposed by the present invention.

In this experiment, the common parameters of all network models are: the network training set and verification set division ratio is 8:2, the input resolution is 128×128, the loss function is cross entropy, and the optimizer is adam , the initial learning rate is 0.001.

The comparison of different model training results and time consumption is shown in Table 1. It can be seen that although the AlexNet small-scale model takes less time to train, its accuracy is significantly lower than that of the ConvNext model. Although the improved model proposed by the present invention has a certain increase in training time, it has also been further improved in accuracy compared with the traditional ConvNext model. Considering comprehensively, it can be seen that the improved model proposed by the present invention is effective.

Table 1 Comparison table of test set accuracy and training time consumption of different models

Next, apply the model to the actual unlabeled data, predict the complete seismic section, and conduct qualitative analysis based on the prediction results. The predicted results are shown in Figure 14. It can be seen that in the application of unlabeled data, the model cannot maintain the high accuracy rate during training. For areas where river channels overlap, model prediction is more difficult. The prediction results of the small-scale AlexNet model are not ideal, and continuous river channels cannot be accurately predicted. However, the improved model of the present invention and the traditional ConvNext model have better prediction effects in these areas, and the improved method of the present invention has a better prediction effect for the channel 1, and the result continuity is even better. This also reflects the advantage of the method proposed by the present invention.

Those skilled in the art will appreciate that the embodiments described here are to help readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Various modifications and variations of the present invention will occur to those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of the claims of the present invention.

Claims (6)

1. A method for analyzing a small sample prestack seismic reflection pattern based on large nuclear attention, comprising:

s1, constructing an image classification seismic reflection mode analysis model based on an improved ConvNext framework;

s2, training the model in the step S1 by adopting the expanded labeled pre-stack seismic data;

s3, predicting unlabeled pre-stack seismic pictures according to the model trained in the step S2, and then splicing single-point prediction results of all positions into a complete seismic section prediction result.

2. The method for analyzing a small sample prestack seismic reflection mode based on large nuclear attention according to claim 1, wherein the model in step S1 is specifically: the method comprises the steps of adopting four cascaded improved ConvNext modules to expand the number of channels and extract deep features of pre-stack seismic pictures, adding a downsampling module between two adjacent improved ConvNext modules for changing the size of the pre-stack seismic pictures, finally converting the length and the width of the input pre-stack seismic pictures to 1X 1 through a global pooling layer, mapping the number of channels into the number of categories to be classified through a full-connection layer, and obtaining the probability that the input pre-stack seismic pictures are respectively judged as each category, wherein the maximum probability is the final classification result corresponding to the pre-stack seismic pictures.

3. The method for analyzing the small sample prestack seismic reflection mode based on the large nuclear attention as recited in claim 2, wherein the improved ConvNext module is specifically based on an inverse bottleneck structure, adding LayerNorm regularization, GELU activation function, layer Scale parameter scaling and Drop Path Layer; the anti-bottleneck structure is a cascade structure of an attribute module and two Linear full-connection layers; the Attention module replaces the 7 x 7 convolutional layer in ConvNext with a large kernel Attention layer.

4. A method of analyzing a small sample prestack seismic reflection pattern based on large nuclear attention as in claim 3 wherein the large nuclear attention layer is embodied as a sum of three convolutions including: a depth convolution with a convolution kernel of k/d, a hole convolution with a convolution kernel of 2d-1 and an expansion rate of d, and a channel convolution with a convolution kernel of 1 x 1.

5. The method for analyzing the small sample prestack seismic reflection mode based on the large nuclear attention as recited in claim 4, wherein the expansion process of the labeled prestack seismic data in the step S2 is as follows: the integral distribution of different seismic phases on the seismic section is obtained through a clustering algorithm, and correction and expansion are carried out by combining with accurate labels of logging data; the method comprises the following steps:

a1, clustering the post-stack seismic sections, classifying the seismic sections according to clustering results, and framing out the parts of the clustering results belonging to the same class of clusters;

a2, in the clustering result, according to xline and inline positions of the logging data, the logging data are projected into a post-stack seismic section, and then the clustering result is expanded by combining accurate seismic phase information and expert experience carried by logging interpretation data;

a3, calculating the position ranges of the xline and inline numbers of the three-dimensional prestack data corresponding to the expanded labeled data, and projecting the position ranges to a three-dimensional prestack seismic data body to obtain prestack multi-channel seismic data of corresponding positions;

and A4, drawing the labeled prestack multi-channel seismic data of each xline and inline position into n curves, and filling positive value areas of the curves to generate pictures to obtain a large amount of prestack labeled data.

6. The method of claim 5, wherein the value of n is determined by the number of traces of the pre-stack earthquake.

Images