A seismic denoising method and system based on improved generative adversarial network

Abstract

The present invention discloses a seismic denoising method and system based on an improved generative adversarial network, which is used for seismic denoising and belongs to the field of oil and gas exploration and seismic data processing. It solves the problems of slow operation speed and inability to fully extract seismic record features when performing seismic denoising on existing generative adversarial networks. The present invention slices noise-free seismic data and adds Gaussian noise to produce a data set; inputs the noisy data into a generator, which extracts features and generates denoised seismic data; inputs real data into a discriminator, which determines the similarity between the denoised seismic data generated by the generator and the noise-free seismic data, and then returns a probability value indicating whether the seismic data is real seismic data. The discriminator returns the result of the judgment to the generator, optimizes the generator, and performs adversarial training in this way. After the training is completed, the model weights are saved; the trained model is used to denoise the input noisy seismic data and output the denoised seismic data. The present invention is used for removing noise from seismic data.

Description

Technical Field

[0001] A seismic denoising method and system based on an improved generative adversarial network are used for seismic denoising and belong to the fields of oil and gas exploration and seismic data processing. Background Art

[0002] Field seismic data collection is often influenced by the external environment, resulting in random noise interference, which complicates subsequent interpretation. Therefore, seismic data denoising is a crucial process in oil and gas exploration. Deep learning-based seismic data denoising technology is a key approach to seismic data denoising. It utilizes deep learning to learn the characteristics of noise in seismic data and then applies this information to denoise data, achieving effective seismic data denoising.

[0003] Traditional seismic denoising methods are mostly based on classic signal processing techniques, such as wavelet transforms, filter design, independent component analysis (ICA), and principal component analysis (PCA). These methods mathematically model and transform the signal to remove noise components and restore the authenticity of the seismic signal. However, these traditional methods also have obvious limitations. First, traditional signal processing methods often rely on preset noise models and are difficult to handle complex and variable noise conditions. Second, traditional methods may introduce errors into the valid signal, resulting in data distortion or information loss.

[0004] With the rapid development of deep learning technology, particularly the successful application of generative adversarial networks (GANs) in fields such as images and audio, researchers have begun to apply deep learning methods to seismic data processing, particularly in the task of denoising. Generative adversarial networks, through a game of negotiation between a generator and a discriminator, can automatically learn the underlying distribution of seismic data, remove noise components from the generated data, and restore the signal's authenticity. However, existing GANs suffer from training difficulties, low training speed, and computational efficiency.

[0005] Existing generative adversarial network seismic denoising methods have the following technical problems:

[0006] 1. Low computational efficiency: Since the GAN network needs to train both the generator and the discriminator network simultaneously, the structure of the deep network is complex and the computational efficiency is low. Although the GAN network that introduces nonlinear activation functions enhances the expressive power, it also increases the computational complexity.

[0007] 2. Insufficient feature extraction: The multidimensionality (such as spatial-temporal coupling characteristics) and high-frequency details of seismic data place higher demands on network modeling capabilities. The conventional convolution module of traditional GANs does not adequately model the global structural correlations of seismic data, making it difficult to distinguish subtle differences between noise and valid signals, thus affecting denoising accuracy.

[0008] In summary, traditional signal processing methods in the existing technology are limited by noise model assumptions and distortion problems, while deep learning methods still do not meet actual needs in terms of efficiency and feature extraction capabilities. There is an urgent need for a solution that takes into account both high-speed computing and high-fidelity denoising. Summary of the Invention

[0009] In response to the above research problems, the purpose of the present invention is to provide a seismic denoising method and system based on an improved generative adversarial network. The method adds a simplified nonlinear activation function convolution module (NAFB), omits the activation function, and replaces it with a simple multiplication, which simplifies the network structure and improves the efficiency of program operation, thereby solving the shortcomings of traditional generative adversarial networks, such as complex network structure, long running time, and high requirements for computing equipment.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] The present invention provides a seismic denoising method based on an improved generative adversarial network, comprising the following steps:

[0012] Step 1: Prepare a dataset for seismic record denoising from seismic shot gather data;

[0013] Step 2: Build a generative adversarial network model and perform adversarial training, including:

[0014] Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data;

[0015] Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data;

[0016] Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights;

[0017] Step 3: Use the trained model to denoise the input noisy seismic data and output the denoised seismic data.

[0018] In the above method, step 1 comprises the following steps:

[0019] Step 1.1: Use the noise-free seismic data for slicing, with a slice size of 128*128 as the noise-free sample;

[0020] Step 1.2: Adding Gaussian noise to the noise-free samples to generate noisy data, forming a paired noisy-noise data set;

[0021] In the above method, the processing flow of the generator in step 2 includes:

[0022] The network structure of the generator includes: a preliminary convolutional layer, a batch normalization layer, a NAFB module and a final convolutional output layer;

[0023] The processing process of the NAFB module includes:

[0024] a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction;

[0025] b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features;

[0026] c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map.

[0027] d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including:

[0028] Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel;

[0029] Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task.

[0030] e) Residual fusion layer: The original input feature map and the processed feature map are weighted and summed to fuse the original input and processed features, including:

[0031] Residual connection: The original input features are weighted summed with the processed features to retain the information of the original feature map;

[0032] Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between;

[0033] Final residual fusion: The processed features and input features are weighted summed again.

[0034] In the above method, the processing flow of the discriminator in step 2 includes:

[0035] The network structure of the discriminator includes: a preliminary convolutional downsampling layer, at least 4 NAFB modules connected in series, and a result output layer;

[0036] The resulting output layer includes: a global average pooling layer, a linear layer (fully connected classification layer), and a sigmoid layer;

[0037] The input data passes through a preliminary convolutional downsampling layer, which extracts primary spatial features while reducing the spatial resolution to 1 / 4 and providing multi-dimensional feature expression for subsequent processing by increasing the number of channels. Subsequently, at least four cascaded NAFB modules perform deep processing on the features in turn. Each NAFB module completes nonlinear feature transformation through feature expansion (1×1 and 3×3 depthwise separable convolutions), SimpleGate interaction (element-wise multiplication to achieve channel screening), and channel attention (global average pooling with dynamic weighting). At the same time, convolutions with a stride of 2 are gradually downsampled between modules, so that the feature map size is gradually reduced and the number of channels is gradually increased, gradually focusing on low-frequency and global structural features. Afterwards, the feature map is input into the result output layer, where the global average pooling layer compresses the final high-dimensional feature map into a global feature vector of a single channel, retaining key statistical information and eliminating redundant spatial details. Finally, the fully connected classification layer maps the global vector to a one-dimensional probability value and outputs the confidence level that the input data is true noise-free data through the Sigmoid function.

[0038] In the above method, step 3 includes the following steps:

[0039] Step 3: Load the model weights saved after training into the model, and then input the noisy seismic data into the model loaded with the model weights for denoising to obtain denoised seismic data.

[0040] The present invention also provides a seismic denoising system based on an improved generative adversarial network, comprising:

[0041] A dataset construction module is used to create a dataset for seismic record denoising from seismic shot gather data;

[0042] Generative adversarial network model building blocks, including:

[0043] Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data;

[0044] Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data;

[0045] Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights;

[0046] The seismic data denoising processing module is used to denoise the input noisy seismic data using the trained model and output the denoised seismic data.

[0047] In the above system, the dataset construction module implementation includes the following steps:

[0048] Use noise-free seismic data for slicing, with a slice size of 128*128 as noise-free samples;

[0049] Gaussian noise is added to the noise-free samples to generate noisy data, forming a paired noisy-noise-free data set.

[0050] In the above system, the NAFB module comprises:

[0051] a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction;

[0052] b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features;

[0053] c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map.

[0054] d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including:

[0055] Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel;

[0056] Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task.

[0057] e) Residual fusion layer: Perform weighted summation of the original input feature map and the processed feature map, including:

[0058] Residual connection: The original input features are weighted summed with the processed features, retaining the information of the original feature map;

[0059] Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between;

[0060] Final residual fusion: The processed features and input features are weighted summed again.

[0061] In the above system, the processing flow of the discriminator includes:

[0062] The network structure of the discriminator includes: a preliminary convolutional downsampling layer, at least 4 NAFB modules connected in series, and a result output layer;

[0063] The resulting output layer includes: a global average pooling layer, a linear layer (fully connected classification layer), and a sigmoid layer;

[0064] The input data passes through a preliminary convolutional downsampling layer, which extracts primary spatial features while reducing the spatial resolution to 1 / 4 and providing multi-dimensional feature expression for subsequent processing by increasing the number of channels. Subsequently, at least four NAFB modules in series perform deep processing on the features. Each NAFB module completes nonlinear feature transformation through feature expansion (1×1 and 3×3 depthwise separable convolutions), SimpleGate interaction (element-wise multiplication to achieve channel screening), and channel attention (global average pooling with dynamic weighting). At the same time, convolutions with a stride of 2 are gradually downsampled between modules, which gradually reduces the size of the feature map and increases the number of channels, gradually focusing on low-frequency and global structural features. Afterwards, the global average pooling layer compresses the final high-dimensional feature map into a global feature vector of a single channel, retaining key statistical information and eliminating redundant spatial details. Finally, the fully connected classification layer maps the global vector to a one-dimensional probability value and outputs the confidence level that the input data is true noise-free data through the Sigmoid function.

[0065] The present invention also provides a storage medium, and when a processor executes a program in the storage medium, the method described above is implemented.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. The present invention introduces a seismic denoising method based on an improved generative adversarial network, adopts a convolution module (NAFB) with simplified nonlinear activation function, and replaces the traditional activation function with a simple multiplication, thereby solving the problems of slow training speed, low computational efficiency and insufficient feature extraction of generative adversarial networks in the prior art.

[0068] 2. By simplifying the network structure, the computational efficiency of the denoising model is significantly improved, enabling efficient seismic data denoising on devices with lower computing power.

[0069] 3. The present invention optimizes the feature extraction capability, fully extracts the space-time coupling characteristics in seismic data, effectively removes noise and retains valid signals, and improves the denoising accuracy.

[0070] In summary, the present invention reduces the running time while ensuring the denoising effect, significantly improves the efficiency and accuracy of seismic data processing, and has strong practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flowchart of the present invention;

[0072] Figure 2 This is the network structure diagram of the present invention, in which Conv Represents a convolutional layer, NAFB represents a convolutional module, BN represents batch normalization, Endlayer represents the result output layer, Sigmoid represents the activation function, Linear represents a linear layer, LeakyReLU represents a leaky linear unit, Flatten represents a flattening operation layer, AdaptiveAvgPool represents an adaptive average pooling layer, Max-Pool represents a maximum pooling layer, SimpleGate represents a simple gating mechanism, SCA represents spatial channel attention, Dconv represents depthwise separable convolution, and LayerNorm represents layer normalization;

[0073] Figure 3 This is the structure diagram of the convolution module NAFB;

[0074] Figure 4 This is a schematic diagram of the Endlayer structure of the result output layer;

[0075] Figure 5 SCA structure diagram for spatial channel attention mechanism

[0076] Figure 6 Schematic diagram of earthquake recording containing noise in the present invention;

[0077] Figure 7 For the present invention Figure 2 Schematic diagram of denoising results of earthquake records;

[0078] Figure 8 For the present invention Figure 2 Schematic diagram of the removed noise. DETAILED DESCRIPTION

[0079] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0080] The present invention provides a seismic denoising method based on an improved generative adversarial network, comprising the following steps:

[0081] Step 1: Prepare a dataset for seismic record denoising from seismic shot gather data;

[0082] Step 2: Build a generative adversarial network model and perform adversarial training, including:

[0083] Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data;

[0084] Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data;

[0085] Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights;

[0086] Step 3: Use the trained model to denoise the input noisy seismic data and output the denoised seismic data.

[0087] In the above method, step 1 comprises the following steps:

[0088] Step 1.1: Use the noise-free seismic data for slicing, with a slice size of 128*128 as the noise-free sample;

[0089] Step 1.2: Adding Gaussian noise to the noise-free samples to generate noisy data, forming a paired noisy-noise data set;

[0090] In the above method, the processing flow of the generator in step 2 includes:

[0091] The network structure of the generator includes: a preliminary convolutional layer, a batch normalization layer, a NAFB module and a final convolutional output layer;

[0092] The processing process of the NAFB module includes:

[0093] a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction;

[0094] b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features;

[0095] c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map.

[0096] d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including:

[0097] Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel;

[0098] Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task.

[0099] e) Residual fusion layer: The original input feature map and the processed feature map are weighted and summed to fuse the original input and processed features, including:

[0100] Residual connection: The original input features are weighted summed with the processed features to retain the information of the original feature map;

[0101] Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between;

[0102] Final residual fusion: The processed features and input features are weighted summed again.

[0103] In the above method, the processing flow of the discriminator in step 2 includes:

[0104] The network structure of the discriminator includes: a preliminary convolutional downsampling layer, at least 4 NAFB modules connected in series, and a result output layer;

[0105] The resulting output layer includes: a global average pooling layer, a linear layer (fully connected classification layer), and a sigmoid layer;

[0106] The input data passes through a preliminary convolutional downsampling layer, which extracts primary spatial features while reducing the spatial resolution to 1 / 4 and providing multi-dimensional feature expression for subsequent processing by increasing the number of channels. Subsequently, at least four cascaded NAFB modules perform deep processing on the features. Each NAFB module completes nonlinear feature transformation through feature expansion (1×1 and 3×3 depthwise separable convolutions), SimpleGate interaction (element-wise multiplication to achieve channel screening), and channel attention (global average pooling with dynamic weighting). At the same time, convolutions with a stride of 2 are gradually downsampled between modules, which gradually reduces the size of the feature map and increases the number of channels, gradually focusing on low-frequency and global structural features. Afterwards, the feature map is input into the result output layer, where the global average pooling layer compresses the final high-dimensional feature map into a global feature vector of a single channel, retaining key statistical information and eliminating redundant spatial details. Finally, the fully connected classification layer maps the global vector to a one-dimensional probability value and outputs the confidence level that the input data is true noise-free data through the Sigmoid function.

[0107] In the above method, step 3 includes the following steps:

[0108] Step 3: Load the model weights saved after training into the model, and then input the noisy seismic data into the model loaded with the model weights for denoising to obtain denoised seismic data.

[0109] The present invention also provides a seismic denoising system based on an improved generative adversarial network, comprising:

[0110] A dataset construction module is used to create a dataset for seismic record denoising from seismic shot gather data;

[0111] Generative adversarial network model building blocks, including:

[0112] Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data;

[0113] Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data;

[0114] Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights;

[0115] The seismic data denoising processing module is used to denoise the input noisy seismic data using the trained model and output the denoised seismic data.

[0116] In the above system, the dataset construction module implementation includes the following steps:

[0117] Use noise-free seismic data for slicing, with a slice size of 128*128 as noise-free samples;

[0118] Gaussian noise is added to the noise-free samples to generate noisy data, thereby forming a paired noisy-noise-free data set.

[0119] In the above system, the NAFB module comprises:

[0120] a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction;

[0121] b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features;

[0122] c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map.

[0123] d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including:

[0124] Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel;

[0125] Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task.

[0126] e) Residual fusion layer: Perform weighted summation of the original input feature map and the processed feature map, including:

[0127] Residual connection: The original input features are weighted summed with the processed features, retaining the information of the original feature map;

[0128] Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between;

[0129] Final residual fusion: The processed features and input features are weighted summed again.

[0130] In the above system, the processing flow of the discriminator includes:

[0131] The network structure of the discriminator includes: a preliminary convolutional downsampling layer, at least 4 NAFB modules connected in series, and a result output layer;

[0132] The resulting output layer includes: a global average pooling layer, a linear layer (fully connected classification layer), and a sigmoid layer;

[0133] The input data passes through a preliminary convolutional downsampling layer, which extracts primary spatial features while reducing the spatial resolution to 1 / 4 and providing multi-dimensional feature expression for subsequent processing by increasing the number of channels. Subsequently, at least four NAFB modules in series perform deep processing on the features in turn. Each NAFB module completes nonlinear feature transformation through feature expansion (1×1 and 3×3 depthwise separable convolutions), SimpleGate interaction (element-wise multiplication to achieve channel screening), and channel attention (global average pooling with dynamic weighting). At the same time, convolutions with a stride of 2 are gradually downsampled between modules, so that the feature map size is gradually reduced and the number of channels is increased, gradually focusing on low-frequency and global structural features. After that, the global average pooling layer compresses the final high-dimensional feature map into a global feature vector of a single channel, retaining key statistical information and eliminating redundant spatial details. Finally, the fully connected classification layer maps the global vector to a one-dimensional probability value and outputs the confidence level that the input data is true noise-free data through the Sigmoid function.

[0134] The present invention also provides a storage medium, and when a processor executes a program in the storage medium, the method described above is implemented.

[0135] Example 1

[0136] The following combination Figure 1-5 The present invention is described in detail.

[0137] 1. Dataset Generation

[0138] Data selection and slicing: Select noise-free seismic shot data and cut the original seismic data into 128 pixel × 128 pixel slices (see attached Figure 6 The sizes corresponding to the horizontal and vertical axes are shown), as noise-free samples.

[0139] 2. Data Preprocessing

[0140] Add a noise-free sample with a mean of 0 and a standard deviation of σ 2 The noise intensity σ is dynamically adjusted according to the actual exploration environment, and the typical value range is 0.1 to 0.3 (based on the amplitude after data normalization).

[0141] Geometric transformation: Perform random horizontal flipping, 90° rotation, and mirroring operations on noisy samples to expand data diversity.

[0142] Normalization: Min-Max normalization is used to linearly map the data to the interval [-1, 1];

[0143] Dataset division: The dataset is divided into training set and test set in a ratio of 8:2 to ensure the generalization ability of the model.

[0144] 3. Generator Network Implementation Details

[0145] The structure of the generator is as follows Figure 2 As shown, the specific processing flow includes the following:

[0146] (1) Preliminary convolutional layer: The input noisy seismic data (size 128×128×1) is passed through a 7×7 convolutional layer with 64 channels, stride 1, and padding 3 to extract primary features; it is then connected to the BatchNorm layer.

[0147] (2) NAFB module stacking: The preliminary features are input into multiple NAFB modules connected in series (the number is adjusted as needed, for example, 6), and the output of each module is connected to a batch normalization layer.

[0148] (3) Final convolutional layer: After NAFB processing, the denoised seismic data is generated through a 7×7 convolutional layer (with 1 channel, 1 stride, and 3 padding).

[0149] 3. Discriminator Network Implementation Details

[0150] The processing flow of the discriminator is as follows:

[0151] (1) Preliminary convolution and downsampling: The input noise-free or generated data is passed through a 7×7 convolutional layer (96 channels, stride 2, padding 3) for feature extraction and spatial downsampling.

[0152] (2) NAFB module and feature compression: Four NAFB modules are connected in sequence, and each module is followed by a convolution with a step size of 2 to achieve downsampling and gradually extract multi-scale low-frequency features.

[0153] (3) Classification output: The features are compressed into a fixed-length vector through global average pooling (GAP). After passing through two fully connected layers, the Sigmoid function outputs the probability that the data is true noise-free data.

[0154] 5. Specific implementation of NAFB module

[0155] The NAFB module structure is as follows Figure 3 As shown, the core steps are as follows:

[0156] (1) Feature expansion: The input features are expanded through 1×1 convolution to expand the number of channels, and then the spatial features are extracted through 3×3 depthwise separable convolution.

[0157] (2) SimpleGate operation: Split the feature map into two groups by channel, and replace the traditional nonlinear activation (such as ReLU) with element-by-element multiplication to reduce the amount of calculation.

[0158] (3) Channel attention mechanism: Perform global average pooling on the SimpleGate output features, dynamically adjust the weight of each channel through a fully connected layer including Sigmoid, and strengthen important features.

[0159] (4) Residual connection: Add the module output to the original input to alleviate the gradient vanishing problem.

[0160] 6. Denoising application stage

[0161] (1) Model loading: Load the trained generator model parameters to the target device.

[0162] (2) Data input and processing: The noisy seismic data (which needs to be normalized to the same numerical range) is input into the generator, processed by the convolutional layer and multiple NAFB modules in sequence, and finally the denoised data is output.

[0163] (3) Result verification: Figure 3-5 As shown, the noisy data ( Figure 6 )After inputting the model, high-quality denoising results are output ( Figure 7 ), and can accurately separate the noise components ( Figure 8 ), proving the effectiveness of the method.

[0164] 7. Key Parameter Examples

[0165] Number of NAFB modules: 4 for the generator and 4 for the discriminator.

[0166] Gaussian noise parameter: σ 2 =0.1, adjusted based on actual data.

[0167] Training batch size: Set to 8 to balance memory usage and convergence speed.

[0168] 8. Implementation effect verification

[0169] Through comparative experiments (not shown), the training speed of the NAFB module of the present invention is increased by about 30% compared with traditional convolutional modules (such as ResNet), and the peak signal-to-noise ratio (PSNR) of the denoising result is improved by more than 2.5dB, demonstrating its advantages in efficiency and effect.

[0170] In summary, the present invention has the following characteristics:

[0171] 1. By adopting a convolutional module with simplified nonlinear activation function (NAFB) to replace the activation function in traditional generative adversarial networks, and utilizing a combined structure of feature expansion layers and depthwise separable convolution kernels, this paper solves the technical problems of high computational complexity and slow training speed caused by complex nonlinear activation functions in existing GAN networks, thereby reducing the number of network parameters and shortening the time of single training iterations.

[0172] 2. By introducing the feature block multiplication mechanism of the gated interaction layer (SimpleGate) and the channel attention mechanism of the dynamic weighted layer, the key problem of insufficient modeling of the spatiotemporal coupling characteristics of seismic data in the traditional convolutional module is solved, thereby improving the retention rate of high-frequency signals and the average signal-to-noise ratio (SNR) indicator.

[0173] 3. By constructing a residual fusion structure of multi-level NAFB modules and adopting a dual residual connection method of weighted summation of input features and processing features, the technical difficulties of gradient vanishing and information attenuation in deep network training are solved.

[0174] 4. By combining the progressive downsampling architecture designed in the discriminator with the global average pooling layer, and combining the cascade processing of four serially connected NAFB modules, the problem of insufficient perception of global structural features by the existing discriminator is solved, thereby improving the accuracy of distinguishing true from false samples.

[0175] 5. By building an end-to-end lightweight network architecture and adopting the collaborative design of depth-separable convolution kernels and channel expansion units, the problem of traditional methods' dependence on high-computing power equipment is solved.

[0176] 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 scope of protection of the present invention.

Claims

1. A seismic denoising method based on an improved generative adversarial network, comprising the following steps: Step 1: Prepare a dataset for seismic record denoising from seismic shot gather data; Step 2: Build a generative adversarial network model and perform adversarial training, including: Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data; Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data; Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights; Step 3: Use the trained model to denoise the input noisy seismic data and output the denoised seismic data; The processing flow of the generator in step 2 includes: The network structure of the generator includes: a preliminary convolutional layer, a batch normalization layer, a NAFB module and a final convolutional output layer; The processing process of the NAFB module includes: a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction; b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features; c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map; d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including: Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel; Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task; e) Residual fusion layer: Performs weighted summation of the original input feature map and the processed feature map to fuse the original input and processed features, including: Residual connection: The original input features are weighted summed with the processed features to retain the information of the original feature map; Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between; Final residual fusion: The processed features and input features are weighted summed again; The processing flow of the discriminator in step 2 includes: The discriminator network includes: a preliminary convolutional downsampling layer, at least 4 NAFB modules connected in series, and a result output layer; The result output layer includes: global average pooling layer, linear layer and sigmoid layer; The input data passes through the preliminary convolution downsampling layer, which reduces the spatial resolution to 1 / 4 while extracting primary spatial features, and provides multi-dimensional feature expression for subsequent processing by increasing the number of channels; Subsequently, at least four NAFB modules connected in series perform deep processing on the features in turn. Each NAFB module completes nonlinear feature transformation through feature expansion, SimpleGate interaction and channel attention. At the same time, convolution is gradually downsampled between NAFB modules, so that the feature map size is gradually reduced and the number of channels is gradually increased, gradually focusing on low-frequency and global structural features. Finally, the feature map passes through the result output layer, which contains: The final high-dimensional feature map is compressed into a single-channel global feature vector through a global average pooling layer, preserving key statistical information and eliminating redundant spatial details; The linear layer maps the global vector to a one-dimensional probability value and outputs the confidence that the input data is real noise-free data through the Sigmoid function.

2. The seismic denoising method based on an improved generative adversarial network according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Use the noise-free seismic data for slicing, with a slice size of 128*128 as the noise-free sample; Step 1.2: Add Gaussian noise to the noise-free samples to generate noisy data, forming a paired noisy-noise data set.

3. The seismic denoising method based on an improved generative adversarial network according to claim 1, characterized in that: Step 3 includes the following steps: Step 3: Load the model weights saved after training into the model, and then input the noisy seismic data into the model loaded with the model weights for denoising to obtain denoised seismic data.

4. A seismic denoising system based on an improved generative adversarial network, characterized by: A dataset construction module is used to create a dataset for seismic record denoising from seismic shot gather data; Generative adversarial network model building blocks, including: Generator training: Noisy data is input into the generator and processed through convolutional layers and multiple NAFB modules to generate preliminary denoised seismic data; Discriminator training: The real noise-free data and the denoised data output by the generator are input into the discriminator. The discriminator extracts features through the multi-level NAFB module and calculates the probability that both are real data; Based on the probability of the discriminator output, the generator and discriminator are alternately optimized to complete the adversarial training and save the model weights; The seismic data denoising module is used to denoise the input noisy seismic data using the trained model and output the denoised seismic data; The NAFB module includes: a) Feature expansion layer: A channel expansion unit composed of 1×1 convolution kernels, used to increase the channel dimension of the input feature map and provide more information for subsequent deep feature extraction; b) Deep feature extraction layer: a subsequent 3×3 depthwise separable convolution kernel is used to extract spatial features; c) SimpleGate: The gated interaction layer divides the previous feature map into two parts and then performs element-by-element multiplication to achieve interaction with the feature map; d) Dynamic weighting layer: recalibrates the importance of feature channels through the channel attention mechanism, including: Global feature statistics module: A global pooling layer is used to perform global average pooling on each channel of the input feature to obtain the average value of each channel; Weight generation module: A 1×1 convolution module is used to generate weights that reflect the contribution of each channel to the task; e) Residual fusion layer: Performs a weighted sum of the original input feature map and the processed feature map, including: Residual connection: The original input features are weighted summed with the processed features, retaining the information of the original feature map; Post-fusion feature processing: The fused features are processed through two convolutional layers with a SimpleGate in between; Final residual fusion: The processed features and input features are weighted summed again.

5. The system according to claim 4, characterized in that The dataset building block implementation includes: Use noise-free seismic data for slicing, with a slice size of 128*128 as noise-free samples; Gaussian noise is added to the noise-free samples to generate noisy data, thereby forming a paired noisy-noise-free data set.

6. A storage medium, characterized in that When the processor executes the program in the storage medium, it implements the method according to any one of claims 1 to 3.

Images