Three-dimensional seismic data random noise suppression method based on spatial regularization feature fusion network

Abstract

The invention provides a three-dimensional seismic data random noise suppression method based on a spatial regularization feature fusion network. The seismic data denoising method comprises the following steps: firstly, dividing a data set into training data and test data in proportion, and adding random noise containing Gaussian noise and multiplicative noise at different levels into the training data; random noise containing different levels of Gaussian noise, multiplicative noise and salt and pepper noise is added to the training data; in the training process, the data are sent into a multi-level feature extraction module to adaptively acquire different feature information of the seismic data; a parallel attention gating fusion module is used for separating signals and noise, more effective information is reserved, and more noise is removed; structural features are corrected through a spatial structure correction module, and the data reconstruction capability of the model is enhanced. The network provided by the invention has high generalization and accuracy, and especially has good performance in the aspect of denoising three-dimensional seismic data containing various types of random noise.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, in particular to the field of three-dimensional seismic data noise blind suppression technology, and relates to a method for noise suppression of multiple types of random noise data, in particular to a three-dimensional seismic data random noise suppression method based on a spatial regularization feature fusion network. Background Art

[0002] Traditional seismic noise suppression models are divided into four categories: spatial domain, transform domain, natural statistics, and sparse representation. The spatial domain-based method directly filters the elements of the data. For example, the median filter method reduces noise by calculating the median of the neighborhood around the pixel. The method is simple and computationally efficient, but it may cause problems such as blurred data details and boundary fusion. The transform domain-based method converts the data to the frequency domain or wavelet domain for processing. For example, BLS-GSM models and reconstructs the wavelet coefficients to remove noise, but it is not effective in dealing with complex noise scenes. The natural statistics-based method uses the natural statistical characteristics of the data to suppress noise. For example, BM3D estimates and suppresses noise based on the similarity between data blocks, but it can only handle specific types of noise. The sparse representation method assumes that the data has a sparse representation in a certain domain. For example, NLSC decomposes the data into a sparse representation form and then removes noise through sparse coding, but it has high computational complexity and high training cost.

[0003] The denoising method based on deep learning uses deep neural networks to learn the mapping relationship between noisy data and clean data, accurately reconstructs the details and structure of noisy data by training the network, and minimizes the difference between the real noisy data and the network output data to achieve the purpose of data noise suppression. Among them, generative adversarial networks, convolutional neural networks, and encoders are widely used in data noise suppression. Based on the generative adversarial network, the encoder maps the input to a latent representation with reduced dimension, and the decoder remaps the simplified representation back to the original space and the probability model estimates the number of bits representing the compressed subvolume. The composite loss function considers the loss values ​​of the three dimensions respectively, and combines the Huber, L1, L2, LogCosh and Mean functions, which is conducive to better reconstruction of effective signals, maintaining brightness changes, and retaining the structure of the original data. However, the training process is unstable and produces artifacts and over-smoothing. Based on the convolutional neural network, SURE-TCNN is based on stein's unbiased risk estimation. The weight parameters are determined by implementing a matrix-based SURE convolutional neural network (CNN) on each slice in the time-frequency domain, but the implementation process is relatively complicated. The encoder-based ANN is implemented by an end-to-end U-Net encoder-decoder 3-DCNN. The encoder extracts features and the decoder reconstructs effective information. It can accurately reconstruct seismic signals and suppress noise in seismic data at the same time, but the computational cost is high.

[0004] At present, most seismic data noise suppression technologies require manual design and feature selection, which leads to long model derivation time and modeling and assumptions based on specific noise types, thus limiting their application scenarios. Therefore, it is urgent to design a denoising scheme that can accurately reconstruct 3D seismic data, effectively suppress random noise, and have generalization capabilities, improve the model's ability to extract features of detailed information, and enhance denoising performance. Summary of the invention

[0005] 1. Purpose of the Invention

[0006] The purpose of the present invention is to provide a three-dimensional seismic data random noise suppression method based on a spatial regularization feature fusion network for seismic data containing multiple types of random noise, so that the model can better perceive the subtle texture information and three-dimensional structural information of the seismic data, so as to better restore the shallow texture features and overall structural information of the seismic data, thereby achieving the purpose of improving the quality of seismic noise suppression data.

[0007] 2. Technical solution

[0008] 1. A method for suppressing random noise of three-dimensional seismic data based on a spatial regularization feature fusion network, characterized by comprising the following steps:

[0009] Step 1: Use m seismic images of the actual work area with original size of S1×S2×S3 and containing different types of random noise to divide the training set and the test set according to a certain ratio;

[0010] Where m∈Z + ; S1, S2, S3∈Z + ;

[0011] Step 2: Construct a module based on multi-level feature extraction, parallel attention gate fusion and spatial structure correction, which includes the following structures:

[0012] Step 2.1: Construct a multi-level feature extraction module to extract new 3D features from existing 3D features and original data. The module includes the following structure:

[0013] First, a 1×1×1 convolutional layer is used to process the noisy data to obtain the original feature F1; then, F1 is passed through an M1×M1×M2 convolutional layer and a LeakyReLU activation function to obtain feature information F2, F1 and F2 are connected by residual connection to obtain F3, F3 is connected by an M2×M1×M1 convolutional layer and a LeakyReLU activation function to obtain feature information F4, F2 and F4 are connected by residual connection to obtain F5, and F5 is connected by an M1×M2×M1 convolutional layer and a LeakyReLU activation function to obtain feature information F6; then, feature information F2, F4 and F6 are concatenated in the channel dimension, and the structural feature information of the three dimensions is comprehensively focused on. A 1×1×1 convolutional layer is used for feature mapping to obtain feature information F7, and F7 is separated into feature tensors in the channel dimension to obtain feature information F 7x 、F 7y and F 7z , respectively use M2×M2×M1 convolution layer + LeakyReLU activation function, M1×M2×M2 convolution layer + LeakyReLU activation function and M2×M1×M2 convolution layer + LeakyReLU activation function to further extract structural information in different directions, and then obtain feature information F 7x' 、F 7y' and F 7z' , concatenate in the channel dimension, use 1×1×1 convolution for feature mapping to obtain feature information F 7c ; At the same time, F1 and F7 are concatenated in the channel dimension to obtain feature information F8; Finally, F 7c And F8 through residual connection, 1×1×1 convolution layer + LeakyReLU activation function, get the comprehensive output feature F O1 ;

[0014] in, a is a fixed coefficient; M1, M2∈Z + ;

[0015] Step 2.2: Construct a parallel attention gated fusion module to fuse multi-channel feature information. The module includes the following structure:

[0016] First, the channel attention mechanism and spatial attention mechanism are used to process the seismic data to obtain the channel feature F of the seismic data. C and spatial feature F S , F C and F S Concatenate in the channel dimension to get F CS ; Then, F CSF' is obtained by performing feature mapping through a 1×1×1 convolution layer, filtering unnecessary information through a Dropout layer with a loss rate of p, and ensuring the stability of the output results through a layer normalization layer and a M3×M3×M3 convolution layer. CS ; then, F' CS The feature weight matrix M in the tensor is extracted through the GeLU activation function w , M w With F' CS Multiply, weaken the non-important feature information and strengthen the important feature information, extract the key structural information of the seismic data, and obtain the feature information F G ; Finally, F G Use M3×M3×M3 convolutions with dilation=d1, dilation=d2, and dilation=d3 respectively. The different effective receptive fields of the convolution kernels can focus on a wider range of detail information and obtain feature information F d1 、F d2 and F d3 ; F d1 and F d2 Concatenate in the channel dimension and use 1×1×1 convolution layer to get feature information F d12 ; F d12 and F d3 Perform concatenation and 1×1×1 convolution layer on the channel dimension to obtain feature information F O2 ;

[0017] Among them, the channel attention mechanism obtains global features through global average pooling, maps the global features through 1×1×1 convolution layer, PReLU activation function and 1×1×1 convolution, and then uses Sigmoid activation function to map to probability, multiply with input features at pixel level, and assign different attention weights to each channel; the spatial attention mechanism uses global average pooling and global maximum pooling respectively, splices them in the channel dimension, combines the overall background information and important information, maps the global features through M4×M4×M4 convolution, uses Sigmoid activation function to map to probability, multiplies with input features at pixel level, and assigns attention weights to different spatial positions; a is a learnable parameter; d1, d2, d3, M3, M4∈Z + ;

[0018] Step 2.3: Construct a spatial structure correction module to extract the structural information of the seismic data, combine it with the original feature information for correction, and reconstruct a more accurate three-dimensional structure. This module includes the following structures:

[0019] First, the input information X' is passed through the M5×M5×M6 convolution layer + M6×M6×M5 convolution layer + PReLU activation function to obtain the structural feature information X s1 , X' and X s1 After the residual connection, the feature information X is obtained s1r , X', X s1 and X s1r The corrected structural feature information X1 is obtained by splicing in the channel dimension; then, X1 passes through the M5×M5×M6 convolution layer + M6×M6×M5 convolution layer + PReLU activation function to obtain the structural feature information X s2 , X' and X s2 After residual connection, we get X s2r , X', X s2 and X s2r The corrected structural feature information X2 is obtained by splicing in the channel dimension; finally, X1 and X2 are connected through residuals to obtain the feature information X 2r , X1, X2 and X 2r Concatenate in the channel dimension and perform feature mapping in the 1×1×1 convolution layer to obtain feature information F r , F r and X' to obtain the final corrected structural feature information F O3 ;

[0020] Among them, M5, M6∈Z + ;

[0021] Step 3: Construct a hybrid loss function to train the multi-attention feature adaptive fusion network. The calculation formula is as follows:

[0022] Network training loss function: Loss = MSE + α × RMSE + β × Perception_loss, including mean square error MSE focuses on the difference at the pixel level, root mean square error RMSE can directly reflect the difference at the pixel level, and perceptual loss Perception_loss focuses on the high-level features of the data through pre-trained VGG16;

[0023] in, N is the number of elements, y i is the element value of the label data, is the element value of the denoised data; N is the number of elements, y i is the element value of the label data, is the element value of the denoised data; Perception_loss is calculated by sending the denoised data and label data to the pre-trained VGG16 network using MSE;

[0024] Step 4: Introduce the improved feature extraction module constructed in step 2 into the loss function of step 3 and design and combine it into a complete feature fusion network based on spatial regularization for training:

[0025] Step 4.1: Construct a feature fusion network based on spatial regularization. This module includes the following structures:

[0026] First, the input feature information X is passed through a 1×1×1 convolutional layer to obtain the feature information X C , the feature information F is obtained through the multi-level feature extraction module O1 , F O1 After the parallel attention gate fusion module, the feature information F is obtained O2 , F O2 After the spatial structure correction module, the feature information F is obtained O3 ; Then, F O1 、F O2 and F O3 The feature information F is obtained through residual connection OR , F O1 、F O2 、F O3 and F OR Splicing is performed on the channel dimension, and a 1×1×1 convolution layer is used to obtain the comprehensive feature information F OC ; Then, F O1 、F O2 、F O3 and F OC The feature information F is obtained through residual connection OCR , F O1 、F O2 、F O3 、F OC and F OCR Perform concatenation and global average pooling operations on the channel dimension to obtain the global feature F OCG , F OCG The global features are mapped through a 1×1×1 convolution layer, a GeLU activation function, and an M7×M7×M7 convolution layer, and then mapped to a probability P through a Softmax activation function. s Finally, P s Respectively with F O1 、F O2 、F O3 and F OC Multiply to get F S1 、F S2 、F S3 and F OS , F S1 、F S2 、F S3 and F OSAfter concatenation in the channel dimension, F is obtained after dimensionality reduction through M7×M7×M7 convolutional layer and 1×1×1 convolutional layer. O ;

[0027] Among them, Softmax function: M7∈Z + ;P s ∈(0,1);

[0028] Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam;

[0029] Step 5.3: Put the seismic noisy data of the training set S1×S2×S3 and its corresponding seismic clean data into the model for training, and save the model after a complete training cycle;

[0030] Step 6: Call the model after a complete training cycle for testing: put the seismic noisy data of the test set with a size of S1×S2×S3 into the model to obtain the denoising result, and obtain the seismic noise suppression data of S1×S2×S3.

[0031] The innovative features of the present invention are as follows:

[0032] (1) Compared with the traditional information extraction module, the multi-level feature extraction module uses new feature information extracted by multiple feature mapping convolution layers, which is concatenated with the original data and original feature information through residual connection and channel dimension, so as to minimize the loss of detailed features of seismic data. The components are then divided into different feature extraction channels through channel dimension, and finally feature fusion is performed, so that the model can dynamically adjust the degree of attention to different features under different input data.

[0033] (2) Compared with the traditional CBAM module, the parallel attention gated fusion module adopts channel attention and spatial attention mechanisms to form a parallel channel spatial attention module; gated convolution is added to filter out unimportant feature information and capture more important channel and spatial feature information; dilated convolution with different expansion rates is used to extract subtle information, and then channel dimension feature fusion is performed step by step, so that the model can retain more effective information and remove more noise.

[0034] (3) Compared with the traditional structure extraction module, the spatial structure correction module extracts the structural information of the data multiple times, and then performs residual connection and channel dimension feature fusion with the original data and original feature information, so as to correct the structural information and enhance the data reconstruction capability of the model.

[0035] 3. Beneficial effects

[0036] The present invention discloses a method for suppressing the noise of seismic data, which is a method for suppressing the noise of multiple types of random noise data. A method for suppressing the random noise of three-dimensional seismic data based on a spatial regularization feature fusion network is provided. Firstly, a multi-level feature extraction module is used to adaptively extract different feature information; a parallel attention gated fusion module is used to fuse the feature information, focusing on recovering more detail information and filtering out noise; finally, a spatial structure correction module is used to fuse richer features and correct the original structural feature information. This method not only ensures the effective noise suppression purpose of the model, but also improves the data recovery accuracy and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is an algorithm flow chart based on the overall spatial regularization feature fusion network. First, different levels of Gaussian noise and salt and pepper noise are added to the pure seismic data as training noisy data and the pure seismic data are sent to the denoising network for denoising, and the training denoised data is output; the difference between the denoised data and the pure seismic data is calculated using a mixed loss function, and the denoising network is guided to suppress noise through back propagation; different levels of Gaussian noise, salt and pepper noise, and multiplicative noise are added to the pure seismic data as test noisy data, and sent to the trained denoising network to obtain test denoised data;

[0038] Figure 2 This is the structure diagram of the multi-level feature extraction module. The input data size is unified as 100×100×100. First, the new feature information extracted by multiple feature mapping convolution layers is concatenated with the original data and original feature information through residual connection and channel dimension, so as not to lose the detailed features of the seismic data as much as possible. Then, the components are divided into different feature extraction channels through the channel dimension, and finally the features are fused.

[0039] Figure 3 This is the structure diagram of the parallel attention gated fusion module. First, the input data of size 100×100×100 uses channel attention and spatial attention mechanisms to form a parallel channel spatial attention module; gated convolution is added to filter out unimportant feature information and capture more important channel and spatial feature information; dilated convolution with different expansion rates is used to extract subtle information, and then channel dimension feature fusion is performed step by step;

[0040] Figure 4 This is the structure diagram of the spatial structure correction module. The input data with a size of 100×100×100 extracts the structural information of the data multiple times, and then performs residual connection and channel dimension feature fusion with the original data and original feature information to correct the structural information and improve the data reconstruction capability.

[0041] Figure 5 It is the structure of the complete spatial regularization feature fusion network;

[0042] Figure 6 This is the effect diagram after denoising using the spatial regularization feature fusion network. The first column is the real sub-data after artificial denoising, the second column is the sub-data with random noise, and the third column is the model denoising; DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, and the embodiments use partial fault data of oil fields. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments.

[0044] Step 1: Use 1196 seismic images of the actual work area with an original size of 100×100×100 and containing different types of random noise to divide the training set and the test set according to a certain ratio;

[0045] Step 2: Construct a module based on multi-level feature extraction, parallel attention gate fusion and spatial structure correction, which includes the following structures:

[0046] Step 2.1: Construct a multi-level feature extraction module to extract new 3D features from existing 3D features and original data

[0047] Step 2.2: Construct a parallel attention gate fusion module to fuse multi-channel feature information

[0048] Step 2.3: Construct a spatial structure correction module to extract the structural information of the seismic data, combine it with the original feature information for correction, and reconstruct a more accurate three-dimensional structure;

[0049] Step 3: Construct a hybrid loss function to train the spatial regularization feature fusion network;

[0050] Step 4: Introduce the improved feature extraction module constructed in step 2 into the loss function of step 3 and design and combine it into a complete feature fusion network based on spatial regularization for training:

[0051] Step 4.1: Construct a feature fusion network based on spatial regularization. This module includes the following structures:

[0052] First, the input feature information X is passed through a 1×1×1 convolutional layer to obtain the feature information X C , the feature information F is obtained through the multi-level feature extraction module O1 , F O1 After the parallel attention gate fusion module, the feature information F is obtained O2 , F O2 After the spatial structure correction module, the feature information F is obtainedO3 ; Then, F O1 、F O2 and F O3 The feature information F is obtained through residual connection OR , F O1 、F O2 、F O3 and F OR Splicing is performed on the channel dimension, and a 1×1×1 convolution layer is used to obtain the comprehensive feature information F OC ; Then, F O1 、F O2 、F O3 and F OC The feature information F is obtained through residual connection OCR , F O1 、F O2 、F O3 、F OC and F OCR Perform concatenation and global average pooling operations on the channel dimension to obtain the global feature F OCG , F OCG The global features are mapped through a 1×1×1 convolution layer, a GeLU activation function, and a 3×3×3 convolution layer, and then mapped to a probability P through a Softmax activation function. s Finally, P s Respectively with F O1 、F O2 、F O3 and F OC Multiply to get F S1 、F S2 、F S3 and F OS , F S1 、F S2 、F S3 and F OS After concatenation in the channel dimension, F is obtained after dimensionality reduction through 3×3×3 convolutional layers and 1×1×1 convolutional layers. O ;

[0053] Among them, Softmax function: P s ∈(0,1);

[0054] Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam;

[0055] Step 5.3: Put the 100×100×100 training set of seismic noisy data and its corresponding seismic clean data into the model for training, and save the model after a complete training cycle;

[0056] Step 6: Call the model after a complete training cycle for testing: put the test set seismic noisy data of size 100×100×100 into the model to obtain the denoising result, and obtain 100×100×100 seismic noise suppression data.

[0057] The following table compares the denoising effect between this method and different methods. The evaluation index is the peak signal-to-noise ratio PSNR, and the calculation formula is: Unit: decibel (dB); Structural similarity SSIM, the calculation formula is Where x and y represent the original data and the denoised data respectively, μ x and μ y They represent the element means of the two data sets, and represents the variance of two data sets, σ xy Represents the covariance of the two images; c1 and c2 are constants greater than zero to avoid the denominator being zero;

[0058] Table 1 Comparison of recognition accuracy between the method of the present invention and different models

[0059] Tab.1 Our method improves the recognition rate of various models

[0060] Model PSNR SSIM TSVD 26.85 0.6709 3D-DnCNN 28.14 0.7558 SURE-TCNN 29.01 0.7957 3DSC-GAN 29.86 0.8315 Algorithm of the present invention 32.31 0.8948

Claims

1. A random noise suppression method for 3D seismic data based on spatial regularization feature fusion network, characterized in that The following steps are involved: Step 1: Use m seismic images of the actual work area with original size of S1×S2×S3 and containing different types of random noise to divide the training set and the test set according to a certain ratio; Where m∈Z + ; S1, S2, S3∈Z + ; Step 2: Construct a module based on multi-level feature extraction, parallel attention gate fusion and spatial structure correction, which includes the following structures: Step 2.1: Construct a multi-level feature extraction module to extract new 3D features from existing 3D features and original data. The module includes the following structure: First, a 1×1×1 convolutional layer is used to process the noisy data to obtain the original feature F1; then, F1 is passed through an M1×M1×M2 convolutional layer and a LeakyReLU activation function to obtain feature information F2, F1 and F2 are connected by residual connection to obtain F3, F3 is connected by an M2×M1×M1 convolutional layer and a LeakyReLU activation function to obtain feature information F4, F2 and F4 are connected by residual connection to obtain F5, and F5 is connected by an M1×M2×M1 convolutional layer and a LeakyReLU activation function to obtain feature information F6; then, feature information F2, F4 and F6 are concatenated in the channel dimension, and the structural feature information of the three dimensions is comprehensively focused on. A 1×1×1 convolutional layer is used for feature mapping to obtain feature information F7, and F7 is separated into feature tensors in the channel dimension to obtain feature information F 7x 、F 7y and F 7z , respectively use M2×M2×M1 convolution layer + LeakyReLU activation function, M1×M2×M2 convolution layer + LeakyReLU activation function and M2×M1×M2 convolution layer + LeakyReLU activation function to further extract structural information in different directions, and then obtain feature information F 7x' 、F 7y' and F 7z' , concatenate in the channel dimension, use 1×1×1 convolution for feature mapping to obtain feature information F 7c ; At the same time, F1 and F7 are concatenated in the channel dimension to obtain feature information F8; Finally, F 7c And F8 through residual connection, 1×1×1 convolution layer + LeakyReLU activation function, get the comprehensive output feature F O1 ; in, a is a fixed coefficient; M1, M2∈Z + ; Step 2.2: Construct a parallel attention gated fusion module to fuse multi-channel feature information. The module includes the following structure: First, the channel attention mechanism and spatial attention mechanism are used to process the seismic data to obtain the channel feature F of the seismic data. C and spatial feature F S , F C and F S Concatenate in the channel dimension to get F CS ; Then, F CS F' is obtained by performing feature mapping through a 1×1×1 convolution layer, filtering unnecessary information through a Dropout layer with a loss rate of p, and ensuring the stability of the output results through a layer normalization layer and a M3×M3×M3 convolution layer. CS ; then, F' CS The feature weight matrix M in the tensor is extracted through the GeLU activation function w , M w With F' CS Multiply, weaken the non-important feature information and strengthen the important feature information, extract the key structural information of the seismic data, and obtain the feature information F G ; Finally, F G Use M3×M3×M3 convolutions with dilation=d1, dilation=d2, and dilation=d3 respectively. The different effective receptive fields of the convolution kernels can focus on a wider range of detail information and obtain feature information F d1 、F d2 and F d3 ; F d1 and F d2 Concatenate in the channel dimension and use 1×1×1 convolution layer to get feature information F d12 ; F d12 and F d3 Perform concatenation and 1×1×1 convolution layer on the channel dimension to obtain feature information F O2 ; Among them, the channel attention mechanism obtains global features through global average pooling, maps the global features through 1×1×1 convolution layer, PReLU activation function and 1×1×1 convolution, and then uses Sigmoid activation function to map to probability, multiply it with the input features at the element level, and assign different attention weights to each channel; the spatial attention mechanism uses global average pooling and global maximum pooling respectively, splices them in the channel dimension, combines the overall background information and important information, maps the global features through M4×M4×M4 convolution, maps it to probability using Sigmoid activation function, multiplies it with the input features at the element level, and assigns attention weights to different spatial positions; a is a learnable parameter; d1, d2, d3, M3, M4∈Z + ; Step 2.3: Construct a spatial structure correction module to extract the structural information of the seismic data, combine it with the original feature information for correction, and reconstruct a more accurate three-dimensional structure. This module includes the following structures: First, the input information X' is passed through the M5×M5×M6 convolution layer + M6×M6×M5 convolution layer + PReLU activation function to obtain the structural feature information X s1 , X' and X s1 After the residual connection, the feature information X is obtained s1r , X', X s1 and X s1r The corrected structural feature information X1 is obtained by splicing in the channel dimension; then, X1 passes through the M5×M5×M6 convolution layer + M6×M6×M5 convolution layer + PReLU activation function to obtain the structural feature information X s2 , X' and X s2 After residual connection, we get X s2r , X', X s2 and X s2r The corrected structural feature information X2 is obtained by splicing in the channel dimension; finally, X1 and X2 are connected through residuals to obtain the feature information X 2r , X1, X2 and X 2r Concatenate in the channel dimension and perform feature mapping in the 1×1×1 convolution layer to obtain feature information F r , F r and X' to obtain the final corrected structural feature information F O3 ; Among them, M5, M6∈Z + ; Step 3: Construct a hybrid loss function to train the spatial regularization feature fusion network. The calculation formula is as follows: Network training loss function: Loss = MSE + α × RMSE + β × Perception_loss, including mean square error MSE focuses on the difference at the element level, root mean square error RMSE can directly reflect the difference at the element level, and perception loss Perception_loss focuses on the high-level features of the data through pre-trained VGG16; in, N is the number of elements, y i is the element value of the label data, is the element value of the denoised data; N is the number of elements, y i is the element value of the label data, is the element value of the denoised data; Perception_loss is calculated by sending the denoised data and label data to the pre-trained VGG16 network using MSE; Step 4: Introduce the improved feature extraction module constructed in step 2 into the loss function of step 3 and design and combine it into a complete feature fusion network based on spatial regularization for training: Step 4.1: Construct a feature fusion network based on spatial regularization. This module includes the following structures: First, the input feature information X is passed through a 1×1×1 convolutional layer to obtain the feature information X C , the feature information F is obtained through the multi-level feature extraction module O1 , F O1 After the parallel attention gate fusion module, the feature information F is obtained O2 , F O2 After the spatial structure correction module, the feature information F is obtained O3 ; Then, F O1 、F O2 and F O3 The feature information F is obtained through residual connection OR , F O1 、F O2 、F O3 and F OR Splicing is performed on the channel dimension, and a 1×1×1 convolution layer is used to obtain the comprehensive feature information F OC ; Then, F O1 、F O2 、F O3 and F OC The feature information F is obtained through residual connection OCR , F O1 、F O2 、F O3 、F OC and F OCR Perform concatenation and global average pooling operations on the channel dimension to obtain the global feature F OCG , F OCG The global features are mapped through a 1×1×1 convolution layer, a GeLU activation function, and an M7×M7×M7 convolution layer, and then mapped to a probability P through a Softmax activation function. s Finally, P s Respectively with F O1 、F O2 、F O3 and F OC Multiply to get F S1 、F S2 、F S3 and F OS , F S1 、F S2 、F S3 and F OS After concatenation in the channel dimension, F is obtained after dimensionality reduction through M7×M7×M7 convolutional layer and 1×1×1 convolutional layer. O ; Among them, Softmax function: M7∈Z + ;P s ∈(0,1); Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam; Step 5.3: Put the seismic noisy data of the training set S1×S2×S3 and its corresponding seismic clean data into the model for training, and save the model after a complete training cycle; Step 6: Call the model after a complete training cycle for testing: put the seismic noisy data of the test set with a size of S1×S2×S3 into the model to obtain the denoising result, and obtain the seismic noise suppression data of S1×S2×S3.

Images