Compressed excitation ResUnet network seismic data denoising method and system based on spectral normalization

Abstract

The present invention belongs to the technical field of oil and gas exploration and seismic data processing, and provides a method and system for denoising seismic data based on a compressed excitation ResUnet network with spectral normalization. The instability problem and the problem of processing detailed information in the training of the traditional ResUnet network are solved, and the signal-to-noise ratio of seismic data denoising is effectively improved. The system includes an acquisition module, a network generation module, a training module and a denoising module. Seismic records are generated by forward numerical simulation based on wave equations, and a training data set with random noise added is constructed. The network structure includes an encoder and a decoder, wherein the encoder includes a residual module and a compressed excitation module, as well as a spectral normalization layer to enhance the stability and generalization ability of the network. The compressed excitation module enhances key features and suppresses noise by weighting the features of different channels. The present invention can effectively process complex noise, retain more details of the original data, and improve denoising performance.

Description

Technical Field

[0001] The invention belongs to the technical field of oil and gas exploration and seismic data processing, and provides a spectral normalization-based compression excitation ResUnet network seismic data denoising method and system. Background Art

[0002] In the process of collecting seismic data, various interferences are inevitable, resulting in noise in the seismic data, which affects the accuracy of subsequent interpretation results. Random noise will affect the quality of seismic records and have an adverse effect on subsequent steps such as AVO analysis and interpretation. To obtain reliable interpretation results, random noise must be suppressed. Therefore, noise suppression is of great significance.

[0003] The seismic data denoising method based on the ResUnet network is one of the important methods for seismic data denoising based on deep learning. This method uses noisy seismic data to create a training set to train the network, and uses the trained network to denoise the missing seismic data. The ResUnet network not only helps prevent overfitting, but also removes random noise more effectively. However, the traditional ResUnet network has problems such as unstable training and poor processing of detailed information, which will affect the training effect of the network and thus affect the denoising results.

[0004] Therefore, in the prior art, the seismic data denoising method based on the ResUnet network has the following technical problems:

[0005] The ResUnet network has problems with unstable training and inability to process detailed information well. The trained neural network cannot process data stably, that is, the quality of the processing results is difficult to guarantee. Summary of the invention

[0006] In response to the above research problems, the purpose of the present invention is to provide a seismic data denoising method and system based on spectral normalization and compression-excitation ResUnet network, so as to solve the problems of unstable training and inability to process detail information well in the ResUnet network in the prior art.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] The present invention provides a spectral normalization-based compression excitation ResUnet network seismic data denoising method, comprising the following steps:

[0009] Step 1: Obtain seismic data and form a training set and a validation set, where the seismic data is seismic shot records;

[0010] Step 2: Obtain the preset spectral normalization-based compressed excitation ResUnet network;

[0011] Step 3: Train and verify the spectral normalization based compression excitation ResUnet network based on the training set and the validation set;

[0012] Step 4: De-noise the seismic data based on the trained compression-excitation ResUnet network.

[0013] In the above technical solution, step 1 comprises the following steps:

[0014] Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records;

[0015] Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data;

[0016] Step 1.3: Use a preset grid size to randomly select multiple pairs of data from replica 1 and replica 2 to form a training pair set A, and finally use training pair set A as the final training set and validation set.

[0017] In the above technical solution, the spectral normalization-based compression excitation ResUnet network preset in step 2 includes an input layer, an encoder module, a decoder module and an output layer connected in sequence,

[0018] The encoder module includes an encoder composed of five residual modules and a downsampling layer. Each encoder is a layer, and each layer includes a residual module. The residual module includes two convolutional layers, two batch normalization layers, an activation function, and a residual connection for adding the input feature map directly to the output of the second convolutional layer. The residual connection includes a convolutional layer for capturing the feature image. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder. A spectral normalization layer is added to the two convolutional layers of the fifth layer.

[0019] The residual module performs convolution operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while retaining important information.

[0020] The compression excitation module is placed after each residual module and is used to weight the features of different channels to enhance key features while suppressing noise.

[0021] In the fifth layer, two convolutional layers are added: a spectrum normalization layer is added outside each convolutional layer to enhance the stability of the network, improve the generalization performance of the network, and achieve a better balance between model fitting and generalization;

[0022] The decoder module includes: a decoder including five residual modules and an upsampling layer connected, each decoder including a residual module and an upsampling layer.

[0023] In the above technical solution, the compression excitation module processes the weight information of different channels, and the specific steps of the compression excitation processing are:

[0024] First, the input feature map X is transformed to generate feature maps U, F tr is the convolution operation:

[0025] F tr : X→U,X∈R W′*H′*C′ ,U∈R W*H*C (1)

[0026] Where R represents a real number set, W, W' represents width, H, H' represents height, C, C' represents the number of channels, and the convolution formula is as follows:

[0027]

[0028] where v c represents the cth convolution kernel, x s represents the sth input covered by the current convolution kernel, u c represents the sth output, K represents the number of convolution kernels, this operation produces a feature map U with a dimension of W*H*C, and then, through the squeezing operation (F sq ) obtains global information, F sq The operation is a global average pooling operation, and its formula is as follows:

[0029]

[0030] where z c Represents the average value of the C channels of this layer. In order to obtain the weight of each channel, two linear layers are needed to process z c ,Right now:

[0031] s=sigmoid(W2*ReLU(W1z)) (4)

[0032] Where z represents the channel average of each layer, W1 and W2 represent linear layers, and s is used to represent the weight of each channel. The function of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weight s is used to weight the feature map U, and F scalc Operation to obtain feature map Right now:

[0033]

[0034] where sc represents the weight of the channel, and u represents the output feature map of the convolution operation;

[0035] A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are:

[0036] For each convolutional layer g:h in →h out The spectral norm of is constrained, h represents the number of nodes in the hidden layer, and h in represents the number of hidden nodes in the input layer, h out represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm Represents the gradient, sup represents the least upper bound, σ(A) is the spectral norm of the matrix A, A represents the weight matrix of a specific layer, and is equal to the maximum singular value of the matrix A in formula (6), which is applicable to an equation with a linear relationship, where σ represents the spectral norm, W represents the weight matrix, and ‖·‖2 represents the L2 norm:

[0037] σ(A):=max ‖h‖2≤1 ∥Ah∥2 (6)

[0038]

[0039] Among them, := means the definition is;

[0040] The activation function in the spectral normalization layer satisfies ‖a‖ Lip =1, where a represents the output of the activation function, and the activation function is the ReLU function:

[0041]

[0042] Among them, W l represents the weight matrix of the lth layer, l represents the lth layer of the network, L represents the number of layers of the neural network, and f represents the output of the activation function ReLU. According to equations (6)-(9), the spectral norm of the weight matrix W is normalized to satisfy the Lipschitz constraint σ(W) = 1, and is the normalization of W, that is:

[0043]

[0044] The spectral norm σ(A) of each layer of the discriminator used in the regularization technique is defined as the maximum singular value of the weight matrix W, that is, σ(W), and the power iteration method is used to solve σ(W) so that it satisfies the Lipschitz constraint σ(W) = 1;

[0045]

[0046] in, represents the left singular vector, represents the right singular vector and T represents the transpose.

[0047] The present invention provides a seismic data denoising system based on a spectral normalized compression-excitation ResUnet network, comprising:

[0048] Acquisition module: obtain seismic data and form training set and verification set, where the seismic data is seismic shot collection records;

[0049] Network generation module: obtain the preset spectral normalization-based compressed excitation ResUnet network;

[0050] Training module: training and verifying the spectral normalization based compression-excitation ResUnet network based on training and validation sets;

[0051] Denoising module: Denoise the seismic data based on the trained spectral normalization-based compression-excitation ResUnet network.

[0052] In the above system, the acquisition module is specifically implemented as follows:

[0053] Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records;

[0054] Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data;

[0055] Step 1.3: Randomly select 8000 pairs of data from replica 1 and replica 2 using a 256×512 grid to form training pair set A, and finally use training pair set A as the final training set and validation set.

[0056] In the above system, the spectral normalization-based compression excitation ResUnet network includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence.

[0057] The encoder module includes an encoder composed of five residual modules and a downsampling layer. Each encoder is a layer, and each layer includes a residual module. The residual module includes two convolutional layers, two batch normalization layers, an activation function, and a residual connection for adding the input feature map directly to the output of the second convolutional layer. The residual connection includes a convolutional layer for capturing the feature image. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder. A spectral normalization layer is added to the two convolutional layers of the fifth layer.

[0058] The residual module performs convolution operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while retaining important information.

[0059] The compression excitation module is placed after each residual module and is used to weight the features of different channels to enhance key features while suppressing noise.

[0060] In the fifth layer, two convolutional layers are added: a spectrum normalization layer is added outside each convolutional layer to enhance the stability of the network, improve the generalization performance of the network, and achieve a better balance between model fitting and generalization;

[0061] The decoder module includes: a decoder including five residual modules and an upsampling layer connected, each decoder including a residual module and an upsampling layer.

[0062] In the above system, the compression excitation module processes the weight information of different channels. The specific steps of the compression excitation processing are:

[0063] First, the input feature map X is transformed to generate feature maps U, F tr is the convolution operation:

[0064] F tr : X→U,X∈R W′*H′*C′ ,U∈R W*H*C (1)

[0065] Where R represents a real number set, W, W' represents width, H, H' represents height, C, C' represents the number of channels, and the convolution formula is as follows:

[0066]

[0067] where v c represents the cth convolution kernel, x s represents the sth input covered by the current convolution kernel, u c represents the sth output, K represents the number of convolution kernels, this operation produces a feature map U with a dimension of W*H*C, and then, through the squeezing operation (F sq ) obtains global information, F sq The operation is a global average pooling operation, and its formula is as follows:

[0068]

[0069] where z c Represents the average value of the C channels of this layer. In order to obtain the weight of each channel, two linear layers are needed to process z c ,Right now:

[0070] s=sigmoid(W2*ReLU(W1z)) (4)

[0071] Where z represents the channel average of each layer, W1 and W2 represent linear layers, and s is used to represent the weight of each channel. The function of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weights are used

[0072] s weights the feature map U through F scale Operation to obtain feature map Right now:

[0073]

[0074] where s c represents the weight of the channel, and u represents the output feature map of the convolution operation;

[0075] A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are:

[0076] For each convolutional layer g:h in →h out The spectral norm of is constrained, h represents the number of nodes in the hidden layer, and h in represents the number of hidden nodes in the input layer, h out represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm Represents the gradient, sup represents the least upper bound, σ(A) is the spectral norm of the matrix A, A represents the weight matrix of a specific layer, and is equal to the maximum singular value of the matrix A in formula (6), which is applicable to an equation with a linear relationship, where σ represents the spectral norm, W represents the weight matrix, and ‖·‖2 represents the L2 norm:

[0077] σ(A):=max ‖h‖2≤1 ∥Ah∥2 (6)

[0078]

[0079] Among them, := means the definition is;

[0080] The activation function in the spectral normalization layer satisfies ‖a‖ Lip =1, where a represents the output of the activation function, and the activation function is the ReLU function:

[0081]

[0082] Among them, W lrepresents the weight matrix of the lth layer, l represents the lth layer of the network, L represents the number of layers of the neural network, and f represents the output of the activation function ReLU. According to equations (6)-(9), the spectral norm of the weight matrix W is normalized to satisfy the Lipschitz constraint σ(W) = 1, and is the normalization of W, that is:

[0083]

[0084] The spectral norm σ(A) of each layer of the discriminator used in the regularization technique is defined as the maximum singular value of the weight matrix W, that is, σ(W), and the power iteration method is used to solve σ(W) so that it satisfies the Lipschitz constraint σ(W) = 1;

[0085]

[0086] in, represents the left singular vector, represents the right singular vector and T represents the transpose.

[0087] Compared with the prior art, the present invention has the following beneficial effects:

[0088] The seismic data is denoised using a compressed excitation ResUnet network based on spectral normalization. On the basis of ResUnet, a spectral normalization layer and a compressed excitation module are added. The compressed excitation module explicitly simulates the interdependence between feature channels by learning the importance of each channel. It weights the features on each channel according to their respective importance, and then highlights the key features, thereby achieving the effect of improving the signal-to-noise ratio of seismic data denoising. The spectral normalization layer decomposes the eigenvalues ​​of the weight matrix and then normalizes it. In the process of optimizing the neural network, the parameter changes are also more stable, and gradient explosion is not prone to occur, which significantly improves the denoising performance of the generative adversarial network. That is, the improved spectral normalized compressed excitation ResUnet network technology can maximize the preservation of weight matrix information; the compressed excitation module is used to retain the detailed information of the image while realizing the segmented image extraction function. By improving the network performance, the present invention ultimately achieves the effect of improving the signal-to-noise ratio of seismic data denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of a spectral normalization-based compression excitation ResUnet network seismic data denoising method of the present invention;

[0090] Figure 2 A compression-excitation ResUnet network architecture based on spectral normalization according to the present invention;

[0091] Figure 3A seismic profile of target data which is the test data of the present invention;

[0092] Figure 4 A seismic profile of the noisy seismic data of the test data of the present invention;

[0093] Figure 5 This is the denoising result of the present invention. DETAILED DESCRIPTION

[0094] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.

[0095] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0096] The technical problem solved by the present invention is to solve the problem of unstable training and inability to process detail information well in the traditional ResUnet network, thereby achieving the effect of improving the signal-to-noise ratio of seismic data denoising.

[0097] A seismic data denoising method based on spectral normalization compression-excitation ResUnet network comprises the following steps:

[0098] Step 1: Obtain seismic data and form a training set and a validation set, where the seismic data is seismic shot records; including the following steps:

[0099] Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records;

[0100] Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data;

[0101] Step 1.3: Randomly select 8000 pairs of data from replica 1 and replica 2 using a 256×512 grid to form training pair set A, and finally use training pair set A as the final training set and validation set.

[0102] Step 2: Obtain the preset spectral normalization-based compressed excitation ResUnet network; the preset spectral normalization-based compressed excitation ResUnet network has a main structure of a Unet network, which includes an encoder and a decoder. The encoder adopts a residual module and a compressed excitation module, and adds a spectral normalization layer outside the two convolutional layers of the fifth layer of the network.

[0103] The specific steps are:

[0104] Step 2.1: Use the Unet network to construct the main network structure, including: five encoders and five decoders, that is, five encoders connected to downsampling and five decoders connected to upsampling, an input layer connected to the first encoder and an output layer connected to the fifth decoder, the first encoder to the fifth encoder are respectively connected to the first decoder to the fifth decoder, each encoder includes two convolutional layers, each decoder includes two deconvolutional layers, and the convolution kernel size is 3×3. The number of channels of the five encoders is 32, 64, 128, 256, 256, and 512 respectively, and the number of channels of the five decoders is 512, 256, 128, 64, and 32 respectively. In the encoder, each layer is composed of a residual module, including: two convolutional layers and a residual connection. The residual connection includes a convolutional layer with a convolution kernel size of 1×1. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder to process the weight information of different channels. The specific steps of compression excitation processing are:

[0105] First, the input feature map X is transformed to generate the feature map U. tr is the convolution operation:

[0106] F trt :X→U,X∈R W′*H′*C′ ,U∈R W*H*C (1)

[0107] Where R represents a set of real numbers, W represents width, H represents height, and C represents the number of channels. The convolution formula is as follows:

[0108]

[0109] where v c represents the cth convolution kernel, x s represents the sth input covered by the current convolution kernel, u c represents the sth output, and K represents the number of convolution kernels. This operation produces a feature map U with a dimension of W*H*C. Then, through the squeezing operation (F sq ) to obtain global information. sq The operation is a global average pooling operation, and its formula is as follows:

[0110]

[0111] Where z represents the channel average of each layer, z c Represents the average value of the C channels of this layer. In order to obtain the weight of each channel, two linear layers are needed to process z c ,Right now:

[0112] s=sigmoid(W2*ReLU(W1z)) (4)

[0113] Among them, W1 and W2 represent linear layers, s is used to represent the weight of each channel, and the role of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weight s is used to weight the feature map U. scale Operation to obtain feature map Right now:

[0114]

[0115] where u c represents a channel in u, s c Represents the weight of the channel.

[0116] Step 2.2: A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are:

[0117] For each convolutional layer g:h in →h out The spectral norm of is constrained, h represents the number of nodes in the hidden layer in represents the number of hidden nodes in the input layer, h out represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm represents the gradient, sup represents the minimum upper bound, σ(A) is the spectral norm of the matrix A, which is equal to the maximum singular value of the matrix A in equation (6), and is applicable to an equation with a linear relationship, where σ represents the spectral norm, W represents the weight matrix, Represents the L2 norm:

[0118] σ(A):=max ‖h‖2≤1 ∥Ah∥2 (6)

[0119]

[0120] Among them, := means that it is defined as

[0121] The activation function in the spectral normalization layer satisfies ‖a‖ Lip =1, where the activation function is the ReLU function:

[0122]

[0123] Where l represents the lth layer of the network. According to equations (6)-(9), the spectral norm of the weight matrix W is normalized to satisfy the Lipschitz constraint σ(W) = 1, and is the normalization of W, that is:

[0124]

[0125] The spectral norm σ(A) of each layer of the discriminator used in the regularization technique is defined as the maximum singular value of the weight matrix W, that is, σ(W), and the power iteration method is used to solve σ(W) so that it satisfies the Lipschitz constraint σ(W) = 1;

[0126]

[0127] in, represents the left singular vector, represents the right singular vector, T represents the transpose;

[0128] Step 3: Train and verify the spectral normalization based compression excitation ResUnet network based on the training set and the validation set;

[0129] Step 4: De-noise the denoised seismic data based on the trained spectral normalization-based compression-excitation ResUnet network.

[0130] Step 4.1: Select some earthquake records of the model and add noise to them to produce test data;

[0131] Step 4.2: Input the seismic data to be denoised into the optimal model selected in step 3.2 for testing;

[0132] Example 1

[0133] The present invention provides a seismic data denoising system based on a spectral normalized compression-excitation ResUnet network, comprising:

[0134] Acquisition module: obtain seismic data and form training set and verification set, where the seismic data is seismic shot collection records;

[0135] Network generation module: obtain the preset spectral normalization-based compressed excitation ResUnet network;

[0136] Training module: training and verifying the spectral normalization based compression-excitation ResUnet network based on training and validation sets;

[0137] Denoising module: Denoise the seismic data based on the trained spectral normalization-based compression-excitation ResUnet network.

[0138] In the above system, the acquisition module is specifically implemented as follows:

[0139] Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records;

[0140] Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data;

[0141] Step 1.3: Randomly select 8000 pairs of data from replica 1 and replica 2 using a 256×512 grid to form training pair set A, and finally use training pair set A as the final training set and validation set.

[0142] In the above system, the spectral normalization-based compression excitation ResUnet network includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence.

[0143] The encoder module includes an encoder composed of five residual modules and a downsampling layer. Each encoder is a layer, and each layer includes a residual module. The residual module includes two convolutional layers, two batch normalization layers, an activation function, and a residual connection for adding the input feature map directly to the output of the second convolutional layer. The residual connection includes a convolutional layer for capturing the feature image. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder. A spectral normalization layer is added to the two convolutional layers of the fifth layer.

[0144] The residual module performs convolution operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while retaining important information.

[0145] The compression excitation module is placed after each residual module and is used to weight the features of different channels to enhance key features while suppressing noise.

[0146] In the fifth layer, two convolutional layers are added: a spectrum normalization layer is added outside each convolutional layer to enhance the stability of the network, improve the generalization performance of the network, and achieve a better balance between model fitting and generalization;

[0147] The decoder module includes: a decoder including five residual modules and an upsampling layer connected, each decoder including a residual module and an upsampling layer.

[0148] In the above system, the compression excitation module processes the weight information of different channels. The specific steps of the compression excitation processing are:

[0149] First, the input feature map X is transformed to generate feature maps U, F tr is the convolution operation:

[0150] F tr : X→U,X∈R W′*H′*C′ ,U∈R W*H*C (1)

[0151] Where R represents a real number set, W, W' represents width, H, H' represents height, C, C' represents the number of channels, and the convolution formula is as follows:

[0152]

[0153] where v c represents the cth convolution kernel, x s represents the sth input covered by the current convolution kernel, u c represents the sth output, K represents the number of convolution kernels, this operation produces a feature map U with a dimension of W*H*C, and then, through the squeezing operation (F sq ) obtains global information, F sq The operation is a global average pooling operation, and its formula is as follows:

[0154]

[0155] where z c Represents the average value of the C channels of this layer. In order to obtain the weight of each channel, two linear layers are needed to process z c ,Right now:

[0156] s=sigmoid(W2*ReLU(W1z)) (4)

[0157] Where z represents the channel average of each layer, W1 and W2 represent linear layers, and s is used to represent the weight of each channel. The function of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weight s is used to weight the feature map U, and F scale Operation to obtain feature map Right now:

[0158]

[0159] where s c represents the weight of the channel, and u represents the output feature map of the convolution operation;

[0160] A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are:

[0161] For each convolutional layer g:h in →h out The spectral norm of is constrained, h represents the number of nodes in the hidden layer, and h in represents the number of hidden nodes in the input layer, h out represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm Represents the gradient, sup represents the least upper bound, σ(A) is the spectral norm of the matrix A, A represents the weight matrix of a specific layer, and is equal to the maximum singular value of the matrix A in formula (6), which is applicable to an equation with a linear relationship, where σ represents the spectral norm, W represents the weight matrix, and ‖·‖2 represents the L2 norm:

[0162] σ(A):=max ‖h‖2≤1 ∥Ah∥2 (6)

[0163]

[0164] Among them, := means the definition is;

[0165] The activation function in the spectral normalization layer satisfies ‖a‖ Lip =1, where a represents the output of the activation function, and the activation function is the ReLU function:

[0166]

[0167] Among them, W l represents the weight matrix of the lth layer, l represents the lth layer of the network, L represents the number of layers of the neural network, and f represents the output of the activation function ReLU. According to equations (6)-(9), the spectral norm of the weight matrix W is normalized to satisfy the Lipschitz constraint σ(W) = 1, and is the normalization of W, that is:

[0168]

[0169] The spectral norm σ(A) of each layer of the discriminator used in the regularization technique is defined as the maximum singular value of the weight matrix W, that is, σ(W), and the power iteration method is used to solve σ(W) so that it satisfies the Lipschitz constraint σ(W) = 1;

[0170]

[0171] in, represents the left singular vector, represents the right singular vector and T represents the transpose.

[0172] Because the above embodiments are adopted, several key technical effects are summarized as follows:

[0173] 1. Improve denoising performance: This technical solution improves the performance of seismic data denoising by introducing spectral normalization and compression excitation modules into the ResUnet network. The spectral normalization layer enhances the stability and generalization ability of the network by constraining the spectral norm of the weight matrix, while the compression excitation module enhances the key features and suppresses noise by weighting the features of different channels, thereby improving the denoising effect.

[0174] 2. Improve network training stability: Traditional ResUnet networks have the problem of unstable training. By adding a compression excitation module and a spectrum normalization layer to the network, this technical solution effectively improves the training stability of the network, reduces the risk of gradient explosion, and enables the network to learn the characteristics of seismic data more stably.

[0175] 3. Preserve detail information: When processing feature maps, the compression excitation module not only enhances key features but also preserves image detail information by weighting the features of different channels. This allows the denoised seismic data to not only effectively suppress noise but also retain more original data details, which is beneficial for subsequent seismic data analysis and interpretation.

[0176] 4. Effectively handle complex noise: The present invention uses seismic records generated by forward numerical simulation based on wave equations, combined with seismic records with random noise added and noise-free data, to construct a training set. This method can effectively handle complex noise that may be encountered in the actual acquisition process, and improves the adaptability and practicality of the denoising method.

[0177] 5. Systematic processing flow: The present invention not only includes the design of the network structure, but also covers the entire process from seismic data acquisition, training set construction to network training and denoising. This systematic processing flow ensures the coherence and consistency from data preparation to the final denoising results.

[0178] Compared with existing technologies, the advantage of this technical solution is that it achieves more stable and efficient seismic data denoising through improved network structure and training process, while retaining more data details, which is of great significance for improving the accuracy and efficiency of seismic data processing.

[0179] like Figure 4-5As shown in the figure, the denoising result suppresses the noise well and is close to the target data, which proves the correctness of the method. The present invention uses the compressed excitation ResUnet network with improved spectral normalization to denoise the seismic data. The compressed excitation ResUnet network based on spectral normalization can extract more information from the data and solve the problem of unstable network training. The denoising performance of the traditional ResUnet network is significantly improved, achieving the effect of improving the signal-to-noise ratio of seismic data denoising.

[0180] The above are only representative embodiments of the present invention in many specific application scopes, and do not constitute any limitation on the protection scope of the present invention. Any technical solutions formed by transformation or equivalent replacement fall within the protection scope of the present invention.

Claims

1. A spectral normalization-based compressed excitation ResUnet network seismic data denoising method, characterized in that: The following steps are involved: Step 1: Obtain seismic data and form a training set and a validation set, where the seismic data is seismic shot records; Step 2: Obtain the preset spectral normalization-based compressed excitation ResUnet network; The spectral normalization-based compression excitation ResUnet network preset in step 2 includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence. The encoder module includes an encoder composed of five residual modules and a downsampling layer. Each encoder is a layer, and each layer includes a residual module. The residual module includes two convolutional layers, two batch normalization layers, an activation function, and a residual connection for adding the input feature map directly to the output of the second convolutional layer. The residual connection includes a convolutional layer for capturing the feature image. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder. A spectral normalization layer is added to the two convolutional layers of the fifth layer. The residual module performs convolution operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while retaining important information. The compression excitation module is placed after each residual module and is used to weight the features of different channels to enhance key features while suppressing noise. In the fifth layer, two convolutional layers are added: a spectrum normalization layer is added outside each convolutional layer to enhance the stability of the network, improve the generalization performance of the network, and achieve a better balance between model fitting and generalization; The decoder module includes: a decoder including five residual modules and an upsampling layer connected, each decoder including a residual module and an upsampling layer; Step 3: Train and verify the spectral normalization based compression excitation ResUnet network based on the training set and the validation set; Step 4: De-noise the seismic data based on the trained compression-excitation ResUnet network.

2. The method for denoising seismic data based on spectral normalization compression-excitation ResUnet network according to claim 1 is characterized in that: The step 1 comprises the following steps: Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records; Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data; Step 1.3: Use a preset grid size to randomly select multiple pairs of data from replica 1 and replica 2 to form a training pair set A, and finally use training pair set A as the final training set and validation set.

3. The method for denoising seismic data based on spectral normalization compression-excitation ResUnet network according to claim 2 is characterized in that: The compression excitation module processes the weight information of different channels. The specific steps of compression excitation processing are: First, the input feature map Generate feature maps through transformation operations is the convolution operation: （1） in represents the set of real numbers, Indicates width, Indicates height, Represents the number of channels, and the convolution formula is as follows: （2） in Indicates convolution kernels, Indicates the number of layers covered by the current convolution kernel. Inputs, Indicates Output, Represents the number of convolution kernels. This operation produces a dimension of Feature map , then, by extrusion operation Get global information, The operation is a global average pooling operation, and its formula is as follows: （3） in Indicates the layer The average value of the channels. In order to obtain the weight of each channel, two linear layers are needed to process ,Right now: （4） in represents the channel average of each layer, represents a linear layer, Used to represent the weight of each channel. The role of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weights are used For feature maps Weighted, through Operation to obtain feature map ,Right now: （5） in represents the weight of the channel, Represents the output feature map of the convolution operation; A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are: For each convolutional layer The spectral norm of is constrained, represents the number of nodes in the hidden layer, represents the number of hidden nodes in the input layer, represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm , represents the gradient, represents the minimum upper bound, is a matrix The spectral norm of represents the weight matrix of a specific layer, which is equal to the matrix The largest singular value of is applicable to an equation with a linear relationship, where represents the spectral norm, represents the weight matrix, Represents the L2 norm: （6） （7） in, Meaning is defined as; The activation function in the spectral normalization layer satisfies ,in, Represents the output of the activation function, the activation function is the ReLU function: （8） in, Indicates The weight matrix of the layer, Represents the network layer, represents the number of layers of the neural network, Represents the output of the activation function ReLU. According to equations (6)-(9), the weight matrix The spectral norm of is normalized, so that the Lipschitz constraint needs to be satisfied ,and yes Normalization of , that is: （9） The spectral norm of each layer of the discriminator used in the regularization technique Defined as a weight matrix The maximum singular value of is , and use the power iteration method to solve , so that it satisfies the Lipschitz constraint ; （10） （11） in, represents the left singular vector, represents the right singular vector, Indicates transpose.

4. A spectral normalization-based compressed excitation ResUnet network seismic data denoising system, characterized in that: include: Acquisition module: obtain seismic data and form training set and verification set, where the seismic data is seismic shot collection records; Network generation module: obtain the preset spectral normalization-based compressed excitation ResUnet network; The compressed excitation ResUnet network based on spectral normalization includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence. The encoder module includes an encoder composed of five residual modules and a downsampling layer. Each encoder is a layer, and each layer includes a residual module. The residual module includes two convolutional layers, two batch normalization layers, an activation function, and a residual connection for adding the input feature map directly to the output of the second convolutional layer. The residual connection includes a convolutional layer for capturing the feature image. A compression excitation module is added after the residual module in the second, third, and fourth layers of the encoder. A spectral normalization layer is added to the two convolutional layers of the fifth layer. The residual module performs convolution operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while retaining important information. The compression excitation module is placed after each residual module and is used to weight the features of different channels to enhance key features while suppressing noise. In the fifth layer, two convolutional layers are added: a spectrum normalization layer is added outside each convolutional layer to enhance the stability of the network, improve the generalization performance of the network, and achieve a better balance between model fitting and generalization; The decoder module includes: a decoder including five residual modules and an upsampling layer connected, each decoder including a residual module and an upsampling layer; Training module: training and verifying the spectral normalization based compression-excitation ResUnet network based on training and validation sets; Denoising module: Denoise the seismic data based on the trained spectral normalization-based compression-excitation ResUnet network.

5. The spectral normalization-based compressed excitation ResUnet network seismic data denoising system according to claim 4, characterized in that: The specific implementation steps of the acquisition module are: Step 1.1: Perform forward numerical simulation based on wave equation to obtain earthquake records; Step 1.2: All earthquake records obtained in step 1.1 are combined to obtain a large earthquake data collection, and then two copies are generated, where copy 1 is the earthquake record with random noise added, and copy 2 is the corresponding noise-free data, i.e., the target data; Step 1.3: Randomly select 8000 pairs of data from replica 1 and replica 2 using a 256×512 grid to form training pair set A, and finally use training pair set A as the final training set and validation set.

6. The spectral normalization-based compressed excitation ResUnet network seismic data denoising system according to claim 5, characterized in that: The compression excitation module processes the weight information of different channels. The specific steps of compression excitation processing are: First, the input feature map Generate feature maps through transformation operations is the convolution operation: （1） in represents the set of real numbers, Indicates width, Indicates height, Represents the number of channels, and the convolution formula is as follows: （2） in Indicates convolution kernels, Indicates the number of layers covered by the current convolution kernel. Inputs, Indicates Output, Represents the number of convolution kernels. This operation produces a dimension of Feature map , then, by extrusion operation Get global information, The operation is a global average pooling operation, and its formula is as follows: （3） in Indicates the layer The average value of the channels. In order to obtain the weight of each channel, two linear layers are needed to process ,Right now: （4） in represents the channel average of each layer, represents a linear layer, Used to represent the weight of each channel. The role of these two fully connected layers is to fuse the feature map information of the channel. Finally, the generated weights are used For feature maps Weighted, through Operation to obtain feature map ,Right now: （5） in represents the weight of the channel, Represents the output feature map of the convolution operation; A spectrum normalization layer is added outside each convolutional layer of the fifth layer to normalize the spectrum of the fifth layer. The specific steps of spectrum normalization are: For each convolutional layer The spectral norm of is constrained, represents the number of nodes in the hidden layer, represents the number of hidden nodes in the input layer, represents the number of hidden nodes in the output layer. By definition, the Lipschitz norm , represents the gradient, represents the minimum upper bound, is a matrix The spectral norm of represents the weight matrix of a specific layer, which is equal to the matrix The largest singular value of is applicable to an equation with a linear relationship, where represents the spectral norm, represents the weight matrix, Represents the L2 norm: （6） （7） in, Meaning is defined as; The activation function in the spectral normalization layer satisfies ,in, Represents the output of the activation function, the activation function is the ReLU function: （8） in, Indicates The weight matrix of the layer, Represents the network layer, represents the number of layers of the neural network, Represents the output of the activation function ReLU. According to equations (6)-(9), the weight matrix The spectral norm of is normalized, so that the Lipschitz constraint needs to be satisfied ,and yes Normalization of , that is: （9） The spectral norm of each layer of the discriminator used in the regularization technique Defined as a weight matrix The maximum singular value of is , and use the power iteration method to solve , so that it satisfies the Lipschitz constraint ; （10） （11） in, represents the left singular vector, represents the right singular vector, Indicates transpose.

Images