A pre-stack seismic data reconstruction method based on improved Transformer

Abstract

The present invention discloses a prestack seismic data reconstruction method based on an improved Transformer, which is characterized by establishing a prestack seismic data reconstruction model based on the improved Transformer, collecting a large amount of prestack seismic data, performing data cleaning and data slicing, converting the three-dimensional prestack seismic data into two-dimensional prestack seismic data, thereby constructing a prestack seismic data reconstruction data set, then designing an overall structure recovery module based on a Transformer basic network model, alternately using an axial attention mechanism and a standard attention mechanism, designing a structure feature screening module based on CNN and gated convolution, upsampling the structure features to a required size and screening useful structure features, then constructing a texture repair module based on a fast Fourier convolution (CNN), gradually adding the above structure features to the texture repair module, finally completing the establishment of the prestack seismic data reconstruction model, and realizing the reconstruction of the prestack seismic data.

Description

Technical Field

[0001] The present invention belongs to the field of seismic data reconstruction and deep learning, and specifically relates to a pre-stack seismic data reconstruction method based on an improved Transformer. Background Art

[0002] With the continuous improvement of oil and gas exploration and development, the object of seismic exploration is gradually shifting from structural oil and gas reservoirs to lithological oil and gas reservoirs. Therefore, pre-stack seismic data with high signal-to-noise ratio, high resolution and high fidelity are needed to characterize the fine structure of underground oil and gas reservoirs. Due to the limitations of natural conditions and the influence of human factors, the pre-stack seismic data actually collected will have problems such as missing seismic traces, insufficient spatial sampling, and strong noise. Therefore, it is necessary to suppress or reconstruct the pre-stack seismic data. Conventional pre-stack seismic data reconstruction methods are limited by the Nyquist sampling theorem and often require data to have a higher sampling rate, making the data acquisition cost higher; on the other hand, conventional pre-stack seismic data reconstruction methods are usually more dependent on the prior information of pre-stack seismic data and require a lot of manual intervention.

[0003] With the rapid development of deep learning, deep learning has been applied in many fields with remarkable results. The original pre-stack seismic data can be reconstructed by deep learning methods without the need for underground prior information, a large amount of manual intervention, or sampling rate restrictions, which can make the geophysical information contained in it more realistically reflect the geophysical characteristics of underground geological bodies. Therefore, it is urgent to use deep learning methods to establish a pre-stack seismic data reconstruction model. Summary of the invention

[0004] In order to overcome the influence of many factors such as sampling rate limitation, dependence on underground prior information, strong noise interference and so on on the reconstruction of pre-stack seismic data, which leads to poor reconstruction effect of pre-stack seismic data, so that the geophysical information contained in it cannot truly reflect the geophysical characteristics of underground geological bodies, affecting subsequent seismic data processing, the present invention proposes a pre-stack seismic data reconstruction method based on improved Transformer, which constructs a pre-stack seismic data reconstruction data set by centrally collecting a large amount of pre-stack seismic data, performing data cleaning and data standardization, and designing a structure recovery model by using the transformer basic network model, and then designing a multi-scale structure feature upsampler based on simple CNN and gated convolution, upsampling the grayscale sketch to any size and extracting multi-scale structure features, and then using CNN based on fast Fourier convolution to design a texture restoration model, and gradually adding the above multi-scale structure features to the texture restoration model, and finally completing the establishment of the pre-stack seismic data reconstruction model to achieve pre-stack seismic data reconstruction.

[0005] To achieve the above object, the technical solution of the present invention mainly includes the following steps:

[0006] A. Constructing pre-stack seismic data reconstruction dataset:

[0007] Collect a large amount of prestack seismic data and use statistical analysis methods to identify possible erroneous values ​​or outliers, such as deviation analysis, identifying values ​​that do not comply with distribution or regression equations, or using constraints between different attributes and external data to detect and clean up data. Then standardize the three-dimensional prestack seismic data from different dimensions into the form of two-dimensional prestack seismic data images, and finally realize the construction of the prestack seismic data reconstruction data set.

[0008] B. Constructing the overall structure recovery module:

[0009] Since Transformer has the ability to capture global structures, we use its ability to restore the overall structure at a relatively low resolution. For the input image data, we first use three convolutional layers to downsample it to reduce the amount of attention learning calculations, and then alternately use the axial attention module and the standard attention module to reduce the complexity of the standard attention module. For the input feature X∈R h×w×c , we assume that X ri,rj ,X ci,cj ∈R c , represents the feature vector of row i, j and column i, j of X. Then the axial attention score A based on row-level and column-level RPE row ,A col It can be written as:

[0010]

[0011]

[0012] Where W rq , W rk , W cq , W ck is the trainable parameter of the row and column keys; is the trainable RPE value between rows i and j, is the RPE value between columns i and j. Then, the attention score is processed by a softmax operation. After encoding the stacked transformer blocks, the features are upsampled by three transposed convolutions to output the structure at the input image size.

[0013] C. Constructing structural feature screening module:

[0014] In order to capture the overall structure of the 2D seismic data more completely, the generated structural features are upsampled to an arbitrary scale without significant degradation. and Line A complete convolutional network is needed to process them into a feature space. First, a CNN is trained to upsample the overall structural features to the required size, and then 3 layers of downsampling convolutions (encoders), 3 layers of dilated convolution blocks, and 3 layers of upsampling convolutions (decoders) are designed. For the encoder and decoder, gated convolutions (GCs) are used to selectively transmit useful features. Finally, 4 feature maps Sk, k∈{0, 1, 2, 3} are selected from an intermediate layer and the outputs of 3 decoder layers to pass the structural features to the texture restoration model:

[0015]

[0016] D. Build texture repair module:

[0017] The texture restoration model consists of several convolution downsampling, image upsampling, and fast Fourier convolution layers. The fast Fourier convolution layer consists of a local branch using traditional convolution and a global branch convolution after fast Fourier transform. By merging the two branches, a larger receptive field and local invariance can be obtained. Then the features of the structural information are passed to the pre-trained CNN texture restoration model based on fast Fourier convolution, and finally the pre-stack seismic data reconstruction model is established to achieve pre-stack seismic data reconstruction.

[0018] The beneficial effects of the present invention are as follows: by cleaning and standardizing the collected prestack seismic data, the problem that high-dimensional prestack seismic data is difficult to reconstruct by conventional machine learning methods can be effectively solved; a transformer-based structural recovery model is designed to capture the overall structural features of prestack seismic data; a structural feature screening module is designed to upsample the generated structural features to any proportion without obvious degradation, and gated convolution filtering can selectively output useful features; a texture repair model is designed to integrate the overall structural features with the texture features, and a larger receptive field and local invariance can be obtained. In addition, the present method also solves the problem that the reconstruction of prestack seismic data is affected by many factors such as sampling rate limitation, dependence on underground prior information, and strong noise interference, so that the geophysical information contained therein more truly reflects the geophysical characteristics of underground geological bodies. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the model structure diagram of the present invention DETAILED DESCRIPTION

[0020] Combine the following Figure 1 The present invention is described in further detail:

[0021] (1) Collect a large amount of prestack seismic data and use statistical analysis to identify possible erroneous values ​​or outliers. Then slice the three-dimensional prestack seismic data in the shape of (128, 128, 128) from the time, horizontal and vertical dimensions respectively. Then, splice the slices of three different dimensions to form two-dimensional prestack seismic data in the shape of (3, 128, 128), thereby realizing the construction of the prestack seismic data reconstruction dataset.

[0022] (2) Constructing the structure recovery module. For the seismic data with an input shape of (3, 128, 128), three convolutional layers are first used to downsample it to (3, 32, 32). Then, the axial attention module and the standard attention module are used alternately. For the input feature X∈R h×w×c , we assume that X ri,rj ,X ci,cj ∈R c , represents the feature vector of row i, j and column i, j of X. Then the axial attention score A based on row-level and column-level RPE row ,A col It can be written as:

[0023]

[0024]

[0025] Where W rq , W rk , W cq , W ck is the trainable parameter of the row and column keys; is the trainable RPE value between rows i and j, is the RPE value between columns i and j. Then, the attention score is processed by a softmax operation. After encoding the stacked transformer blocks, the features are upsampled to (3, 128, 128) through three transposed convolutions, and finally the binary cross entropy (BCE) loss is used to optimize the prediction:

[0026]

[0027] in, represents the binary real edge, Represents the generated edge.

[0028] (3) Construct a structural feature screening module: Design 3 layers of downsampling convolution (encoder), 3 layers of residual blocks with dilated convolution, and 3 layers of upsampling convolution (decoder). For the encoder and decoder, gated convolutions (GCs) are used to selectively transmit useful features. Finally, 4 feature maps Sk, k∈{0, 1, 2, 3} are selected from the middle layer and the output of the 3 decoder layers to pass the structural features to the texture restoration model:

[0029]

[0030] (4) A texture restoration module is constructed, which consists of three layers of convolution downsampling, three layers of deconvolution upsampling, and nine layers of fast Fourier convolution. The fast Fourier convolution layer consists of a local branch using traditional convolution and a global branch convolution after fast Fourier transform. The features of the structural information are then passed to the pre-trained CNN texture restoration model based on fast Fourier convolution. Finally, the pre-stack seismic data reconstruction model is established to achieve pre-stack seismic data reconstruction.

[0031] The above is only a preferred embodiment of the present invention. Any technician familiar with the profession may use the above technical solution to modify or change it into an equivalent example with equivalent changes. Any simple modification, change or modification of the above embodiment based on the technical solution of the invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the invention.

Claims

1. A pre-stack seismic data reconstruction method based on improved Transformer, characterized in that: The following steps are involved: A. Constructing pre-stack seismic data reconstruction dataset: By collecting a large amount of prestack seismic data, the data is cleaned using statistical analysis methods, and then the three-dimensional prestack seismic data is standardized into the form of two-dimensional prestack seismic data from different dimensions, and then the two-dimensional prestack seismic data is sliced ​​into a fixed size to realize the construction of the prestack seismic data reconstruction data set; B. Constructing the overall structure recovery module: Using the Transformer's capabilities, we design an overall structure recovery module that alternates between axial attention modules and standard attention modules to achieve output structures at the input image size through downsampling and upsampling operations. C. Constructing structural feature screening module: Design a structural feature screening module, upsample the overall structural features through a convolutional neural network (CNN), selectively transfer useful features using gated convolution, and finally pass the structural features to the texture restoration model; D. Build texture repair module: A texture restoration module based on fast Fourier convolution is designed, including convolution downsampling, convolution upsampling and fast Fourier convolution layers. The structural information features are transferred to the pre-trained texture restoration model based on fast Fourier convolution, and finally the pre-stack seismic data reconstruction is realized.

Images