A seismic phase intelligent identification and detection method based on TransUNet deep learning network

Abstract

The present invention discloses a method for intelligent identification and detection of seismic phases based on the TransUNet deep learning network. The original three-dimensional seismic data is sliced ​​to obtain the original image of the two-dimensional seismic data. The original image of the two-dimensional seismic data is encoded using a convolutional neural network and a Transformer encoder structure to obtain a seismic data encoding feature map, and then the feature map is upsampled to the size of the original seismic data, thereby realizing intelligent classification of seismic phases. The present invention extracts and encodes features of seismic data using a full convolutional neural network and a Transformer encoder, and uses a U‑Net upsampling structure to decode the encoded feature map to full pixel density to realize intelligent identification and detection of seismic phases, thereby improving the problems of low efficiency and over-reliance on experience in artificial seismic phase classification; using the innate self-attention mechanism of the Transformer and the perception of local details by the U‑Net convolutional neural network, TransUNet can take into account both local detail features and global features, and can be effectively applied to intelligent classification of seismic phases, effectively improving the efficiency and accuracy of seismic phase classification.

Description

Technical Field

[0001] The present invention belongs to the fields of geophysical exploration and artificial intelligence, and specifically relates to a seismic phase intelligent identification and detection method based on the TransUNet deep learning network. Background Art

[0002] Seismic facies can be understood as the sum of the reflection characteristics of sedimentary facies in seismic data, providing a reference for analyzing geological conditions and predicting oil and gas reservoirs. Early seismic facies analysis typically required interpreters to identify facies using kinematic and geometric parameters of seismic waves, drawing on their professional knowledge and regional experience, based on human vision. Alternatively, they could identify and map seismic facies units according to specific procedures based on objective seismic attributes. These traditional seismic facies identification methods often required extensive manual effort, resulting in low efficiency and reliance on the interpreter's expertise and experience. The results were inevitably influenced by subjective factors and presented significant limitations.

[0003] In recent years, thanks to the rapid development of computer vision, the combination of semantic segmentation technology in the field of computer vision and seismic data analysis has provided new ideas for the intelligent classification of seismic phases. Among various image segmentation tasks, U-Net has become the most commonly used method and has achieved great success. Transformer is used for sequence-to-sequence prediction and has an innate self-attention mechanism. As a semantic segmentation algorithm that can take into account both local detail features and global features, the TransUNet algorithm can be effectively applied to the intelligent classification of seismic phases. Transformer encodes the labeled image patches in the convolutional neural network seismic raw data feature map into an input sequence that extracts the global context. The decoder upsamples the encoded features and then combines them with the high-resolution convolutional neural network feature map to achieve accurate seismic phase classification, effectively improving the efficiency and accuracy of seismic phase classification. Summary of the Invention

[0004] To overcome the low efficiency of traditional seismic facies classification, its overreliance on interpreter experience, and the influence of subjective factors, this paper proposes a method for intelligent seismic facies identification and detection based on the TransUNet deep learning network. This method slices the original 3D seismic data to obtain a 2D raw image. This image is then encoded using a convolutional neural network and a Transformer encoder structure to generate a seismic data encoding feature map. This feature map is then upsampled to the original seismic data size, enabling intelligent seismic facies classification.

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

[0006] A. Use sliding window strategy to flatten the 3D sample set:

[0007] (1) Divide the 2D seismic profiles along the main survey line, contact lateral line, SW-NE and NW-SE directions of the 3D seismic original data volume, select a sliding window of size n*n on each seismic profile, and set the slices taken out of each profile The number of is m, which increases the diversity of samples while ensuring that the spatial distribution law of seismic phase data is followed.

[0008] (2) The amplitude values ​​of all earthquake raw data in the training data are normalized so that their thresholds are between -1 and 1.

[0009] (3) Since there is no continuity between earthquake label values, it is necessary to convert the label data into one-hot encoding before training to enhance its sparsity. For the p-th pixel point on the earthquake profile, its one-hot encoding label is y p , can be represented by a vector of 1×C consisting of 0 and 1, where C is the number of seismic phase types in the data set. If the seismic phase of the point label is y c,p , the corresponding value is 1, and the rest are 0.

[0010] B. Constructing an intelligent seismic phase determination model based on TransUNet:

[0011] (1) Use three convolutional layers as feature extractors, and the input is , the output feature maps of the three convolutional layers are

[0012] (2) The obtained seismic data characteristic map Mapped to a D-dimensional embedding space through a trainable linear projection and adding a learnable position vector E pos , get a D-dimensional embedding vector Z0;

[0013] (3) The embedded vector is input into the Transformer module consisting of an L-layer multi-head attention module MSA and a multi-layer perceptron module MLP;

[0014] (4) The specific calculation method of the attention operation module is as follows: As Q, K, V input to each head of the self-attention mechanism, Q and K are calculated by point multiplication and normalized, that is, divided by where d k is a vector The length of the attention mechanism is then quantified into a probability distribution through the softmax function. The probability distribution is vector multiplied by V to obtain the attention mechanism calculation result Attention(Q,K,V). The calculation result is residually connected with the input embedding vector to obtain the attention operation module output Z′. l ;

[0015] (5) Attention layer output result Z′ l After a normalization operation, it is input into the fully connected neural network to obtain the calculation result and the input Z′ l Perform residual connection to obtain the final Transformer module output Z l ;

[0016] (6) Transformer module output result Z l Through three cascade upsampling modules, a slice with the same seismic data is obtained. The prediction results of full resolution of the same size, each upsampling module consists of a ReLU activation layer and a 3×3 convolution layer, and the input of each layer is the feature map output by the previous layer and the CNN feature extractor of the corresponding layer. Finally, a prediction mask with a shape of C×n×n is obtained.

[0017] C. Training the TransUNet-based seismic phase intelligent determination model:

[0018] (1) The original seismic data is divided into 3:7 ratios and the model is trained. The learning rate is set to lrate and the loss function is set to loss(·). The partial derivatives of each model parameter with respect to the loss function are calculated. and , use the gradient descent method to perform back propagation until the model converges or reaches the preset maximum number of training times to obtain the final network model;

[0019] (2) Use the test set data to test the model and calculate the model's IoU and accuracy Acc to measure the specific performance of the model.

[0020] The beneficial effects of the present invention are: using a fully convolutional neural network and a Transformer encoder to extract and encode features from seismic data, using a U-Net upsampling structure to decode the encoded feature map to full pixel density, realizing intelligent identification and detection of seismic phases, and improving the problems of low efficiency and over-reliance on experience in manual seismic phase classification; utilizing the Transformer's innate self-attention mechanism and the U-Net convolutional neural network's perception of local details, TransUNet can take into account both local detail features and global features, can be effectively applied to intelligent seismic phase classification, and effectively improve the efficiency and accuracy of seismic phase classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention DETAILED DESCRIPTION

[0022] The following combination Figure 1 The present invention is described in further detail:

[0023] A. Use sliding window strategy to flatten the 3D sample set:

[0024] (1) Setting up seismic data slices The size is 256*256, and the two-dimensional seismic profiles are divided along the main survey line, contact lateral line, SW-NE, and NW-SE directions of the three-dimensional seismic original data body. A sliding window of size 256*256 is selected on each seismic profile, and each profile is set to extract The number of samples is m, which increases the diversity of samples while ensuring that the spatial distribution law of seismic phase data is followed;

[0025]

[0026] (2) The amplitude values ​​of all the original earthquake data in the training data are normalized so that the threshold is between -1 and 1. The specific formula is as follows:

[0027]

[0028] Among them, x′ is the standardized data, x is the original data before standardization, μ is the mean of the sample data, and σ is the variance of the sample data;

[0029] (3) Since there is no continuity between earthquake label values, it is necessary to convert the label data into one-hot encoding before training to enhance its sparsity. For the p-th pixel point on the earthquake profile, its one-hot encoding label is y p , can be represented by a vector of 1×6 consisting of 0 and 1. If the earthquake phase of the point label is y c,p , the corresponding value is 1, and the rest are 0, which is as follows:

[0030] y p =[y 1,p ,…,y c,p ,…y 6,p ]

[0031] Among them, y p Represents the unique hot encoding label of the p-th pixel on the seismic profile.

[0032] B. Constructing an intelligent seismic phase determination model based on TransUNet:

[0033] (1) Use three convolutional layers as feature extractors, and the input is The output feature maps of the three convolutional layers are

[0034] (2) The obtained seismic data characteristic map Mapped to a D-dimensional embedding space through a trainable linear projection and adding a learnable position vector E pos , we get a D-dimensional embedding vector Z0, as follows:

[0035]

[0036]

[0037] E pos ∈R N×D

[0038] Among them, E represents the seismic data characteristic map A trainable linear projection to a D-dimensional embedding space;

[0039] (3) The embedding vector is input into the Transformer module consisting of an L-layer multi-head attention module MSA and a multi-layer perceptron module MLP. The output of the L-th layer is as follows:

[0040] Z′ l =MAS(LN(Z l-1 ))+Z l-1

[0041] Z l =MLP(LN(Z′ l ))+Z′ l

[0042] Among them, MAS(·) represents the attention module operation, MLP(·) represents the multi-layer perceptron module operation, and LN(·) represents the layer normalization operation;

[0043] (4) The specific calculation method of the attention operation module is as follows: As Q, K, V input to each head of the self-attention mechanism, Q and K are calculated by point multiplication and normalized, that is, divided by where d k is a vector The length of the attention mechanism is then quantified into a probability distribution through the softmax function. The probability distribution is vector multiplied by V to obtain the attention mechanism calculation result Attention(Q,K,V). The calculation result is residually connected with the input embedding vector to obtain the attention operation module output Z′. l, where the calculation formula is as follows:

[0044]

[0045] (5) Attention layer output result Z′ l After a normalization operation, it is input into the fully connected neural network to obtain the calculation result and the input Z′ l Perform residual connection to obtain the final Transformer module output Z l ;

[0046] (6) Transformer module output result Z l Through three cascade upsampling modules, a slice with the same seismic data is obtained. The prediction results of full resolution of the same size, each upsampling module consists of a ReLU activation layer and a 3×3 convolution layer, and the input of each layer is the feature map output by the previous layer and the CNN feature extractor of the corresponding layer. The final prediction mask with a shape of 6*256*256 is obtained, where is the number of target categories. The ReLU activation function formula is: ReLU(x)=max(0,x).

[0047] C. Training the TransUNet-based seismic phase intelligent determination model:

[0048] (1) The original seismic data is divided into 3:7 ratios and the model is trained. The learning rate is set to lrate and the loss function is set to loss(·). The partial derivatives of each model parameter with respect to the loss function are calculated. and , back propagation is performed using the gradient descent method until the model converges or reaches the preset maximum number of training times, and the final network model is obtained, where the learning rate lrate, loss function loss(·) and partial derivatives and The calculation formulas are:

[0049]

[0050]

[0051]

[0052]

[0053] Among them, lrate′ represents the initial learning rate, d model Indicates the model input feature dimension, warmup_step indicates the number of warmup steps, y i Represents the label value of the pixel, pi Represents the predicted value of the pixel, N represents the number of input pixels, x i Represents the input value of the pixel point, ω and bs represent the coefficient and bias term respectively;

[0054] (2) Use the test set data to test the model and calculate the model's IoU and accuracy Acc to measure the specific performance of the model. The calculation formulas for IoU and Acc are as follows:

[0055]

[0056]

[0057] Among them, GT represents the label data, Pred represents the model prediction result, TP represents the true positive, TN represents the true negative, FP represents the false positive, FN represents the false negative, IoU measures the overlap between the network prediction result and the label, and Acc represents the ratio of correctly predicted pixels to the total pixels.

[0058] The above description is merely a preferred embodiment of the present invention. Any person skilled in the art may utilize the above-described technical solution to modify or alter the above-described embodiment into equivalent embodiments with equivalent variations. Any simple modification, alteration, or modification of the above-described embodiment based on the technical solution of the present invention that does not depart from the scope of the present invention is within the scope of protection of the technical solution of the present invention.

Claims

1. A seismic phase intelligent identification and detection method based on TransUNet deep learning network, characterized in that: The following steps are involved: A. Use sliding window strategy to flatten the three-dimensional sample set: Set the size of seismic data slices, divide the two-dimensional seismic profiles along the main survey line, contact lateral line, SW-NE, and NW-SE of the three-dimensional seismic raw data volume, select a sliding window on each seismic profile, increase the diversity of samples on the basis of ensuring that the spatial distribution law of seismic phase data is followed, standardize the amplitude values ​​of all seismic raw data in the training data so that its threshold is between -1 and 1, and convert the label data into one-hot encoding to enhance its sparsity; B. Constructing an intelligent seismic phase determination model based on TransUNet: A seismic phase intelligent determination model based on TransUNet is constructed. Three convolutional layers are used as feature extractors. The obtained seismic data feature map is mapped to a D-dimensional embedding space through a trainable linear projection, and a learnable position vector is added to it to obtain a D-dimensional embedding vector. The embedding vector is input into a Transformer module composed of an L-layer multi-head attention module MSA and a multi-layer perceptron module MLP. C. Training of TransUNet-based seismic phase intelligent determination model: The original seismic data was divided into 3:7 ratios and the model was trained. The gradient descent method was used for back propagation until the model converged to obtain the final network model. The model was tested using the test set data, and the IoU and accuracy Acc of the model were calculated to measure the specific performance of the model.

Images