A self-supervised seismic denoising method and system fusing noise estimation module and channel attention mechanism

Abstract

The application discloses a self-supervised seismic denoising method and system fusing a noise estimation module and a channel attention mechanism, and belongs to the field of oil and gas exploration and seismic data processing, and solves the problems of poor actual noise removal effect and easy removal of original data in the prior art. The application slices and makes a dataset for noisy seismic data; inputs the noisy data into a noise estimation module to extract noise features, and then inputs the noise features into a blind spot network denoising integrated with a channel attention mechanism; trains the data by inputting the data into branches with an inflation convolution module and branches with an inflation Transform module, respectively, saves the model weight after the training is completed, and uses the trained model to perform denoising processing on input noisy seismic data, and outputs denoised seismic data. The application is used for seismic data noise removal.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of oil and gas exploration and seismic data processing, and provides a self-supervised seismic denoising method and system fusing a noise estimation module and a channel attention mechanism. BACKGROUND

[0002] In the field of seismic exploration, the quality of received data directly determines the accuracy of underground structure analysis. However, seismic signals are easily disturbed by various noises during acquisition. Random noise, due to its wide distribution and overlapping with the effective signal frequency band, becomes the core problem affecting the quality of seismic data. Such noise not only obscures weak reflection information, but also leads to error accumulation in subsequent inversion and interpretation. Therefore, efficient seismic denoising technology is a key link in oil and gas exploration and data processing.

[0003] Traditional seismic denoising methods are mainly based on physical model assumptions, such as using the spatial difference between noise and seismic events (such as predictive filtering technology) or the sparsity of seismic reflection (such as sparse representation denoising). Although this kind of method is effective in certain scenarios, its performance is highly dependent on the artificially preset prior conditions, and the parameters need to be repeatedly adjusted for different work area data, making it difficult to adapt to complex noise environments. In addition, the traditional method has limited ability to extract nonlinear characteristics of seismic signals, which easily causes loss of effective signals or residual noise.

[0004] In recent years, seismic denoising methods based on deep learning have significantly improved the denoising performance. Models such as full convolutional network (FCN), U-Net, DnCNN, and Transformer achieve nonlinear mapping of noise and signal through end-to-end learning. However, existing deep learning methods have the following bottlenecks:

[0005] 1. Supervised learning dependency: Most models require a large amount of paired "noisy-clean" data for supervised training, but it is difficult to obtain sufficient high-quality noise-free data in actual exploration, which limits the generalization of the model.

[0006] 2. Limitations of synthetic data: Existing methods often use Gaussian white noise to simulate training, but real seismic noise has spatial correlation and complex statistical characteristics (such as non-stationarity and heteroscedasticity), and the model trained by synthetic data has significantly reduced noise removal effect on actual noise.

[0007] 3. Insufficient feature extraction: Traditional convolutional networks have difficulty in balancing local details and global dependencies, and lack dynamic perception of feature weights between channels, leading to inaccurate noise estimation, especially in weak signal areas, which easily causes over-smoothing or structural distortion.

[0008] To solve the above problems, some studies attempt to introduce self-supervised learning to avoid dependence on clean data, for example, a blind spot network (BSN) realizes pixel-level denoising through mask convolution. However, the BSN assumes that the noise is independent in space, which is contradictory to the spatial correlation of real seismic noise, and the limited receptive field of the BSN cannot model long-range dependencies, resulting in incomplete removal of complex noise.

[0009] In summary, the prior art has not effectively solved the core problems of noise model complexity, data labeling dependence and insufficient feature representation in real seismic data denoising, and an innovative method that takes into account self-supervised training, adaptive modeling of noise characteristics and multi-scale feature fusion is urgently needed. SUMMARY

[0010] The purpose of the present application is to provide a self-supervised seismic denoising method that combines a noise estimation module and a channel attention mechanism, which combines pixel downsampling technology to solve the problem of poor actual noise removal effect and the problem of removing original data during denoising in the prior art, and can only use real noisy data during training, which has stronger generalization and practicality than synthetic data using only Gaussian noise.

[0011] To achieve the above purpose, the present application adopts the following technical solutions:

[0012] The present application provides a self-supervised seismic denoising method that combines a noise estimation module and a channel attention mechanism, comprising the following steps:

[0013] Step 1: Construct a seismic data set, including slicing real noisy seismic data and synthetic noisy seismic data to obtain multiple noisy data blocks;

[0014] Step 2: Construct a self-supervised denoising network, comprising:

[0015] A noise prediction module (CNNest) composed of multiple 1x1 convolution layers and ReLU activation functions, used to output a noise level prediction map with the same size as the input data, which represents the noise covariance matrix of each pixel;

[0016] A self-supervised blind spot network (BSN) containing a double-branch structure integrating a channel attention mechanism:

[0017] The first branch uses an expansion convolution module with a center mask convolution, referred to as a DC module, to extract local features;

[0018] The second branch uses an expansion Transformer block combined with a channel attention mechanism, referred to as a DTB module, which models global features through a channel attention mechanism, and its feedforward layer contains a depth separable convolution;

[0019] Step 3: Apply a pixel-wise random down-sampling (PD) operation to the input data to destroy the spatial correlation of noise, and input the processed data into the self-supervised denoising network for training, fuse the output of the noise prediction module and the intermediate features of the self-supervised blind spot network using the Bayesian prediction formula, and optimize the network parameters through back propagation to obtain a denoising model;

[0020] Step 4: Input the seismic data to be denoised into the trained model for denoising processing, and restore the final denoising result through inverse PD operation.

[0021] In the above method, step 2 comprises the following steps:

[0022] Step 2.1: Construct a noise prediction module, which includes five 1x1 convolution layers connected in turn, wherein the last four layers are deployed with ReLU nonlinear activation function, and the fifth layer outputs a noise covariance matrix with the same size as the input noisy data;

[0023] Step 2.2: Construct a double-branch architecture of the self-supervised blind spot network (BSN), wherein,

[0024] The first branch includes upper and lower paths:

[0025] The upper path includes a 3x3 center mask convolution layer and 9 DC modules connected in series, wherein the DC module adopts an expansion convolution with a step of 2 and contains parallel upper / middle / lower paths, and each path is realized by combining a 1x1 convolution layer and a 3x3 expansion convolution layer to extract local features;

[0026] The lower path includes a 5x5 center mask convolution layer and 9 DC modules with S=3, wherein the DC module adopts an expansion convolution with a step of 3 and contains parallel upper / middle / lower paths, and each path is realized by combining a 1x1 convolution layer and a 3x3 expansion convolution layer to extract local features;

[0027] The second branch includes a 5x5 center mask convolution layer and 9 DTB modules connected in series, which generate channel interaction global feature maps through channel attention mechanism, and adopt depth separable convolution and feedforward network to enhance long-range dependency, wherein the channel attention mechanism is realized by calculating the dot product operation of query matrix Q, key matrix K and value matrix V; The outputs of the first branch and the second branch are removed from the blind spot information, and the multi-scale features are fused through a 1x1 convolution layer; At the same time, the output of the noise prediction module and the fused features are predicted by Bayesian prediction to form the final BSN output result.

[0028] In the above method, step 3 comprises the following steps:

[0029] Step 3.1: Perform pixel-wise random down-sampling PD with step factor s = 2 on the input real noisy seismic data to break the spatial correlation of noise, split the input data into 2x2 sub-blocks and arrange them in a mosaic pattern to generate four sub-block data;

[0030] Step 3.2: Input the PD processed sub-block data into the noise prediction module to obtain a 1x1 noise covariance matrix for each pixel, and expand the matrix dimension to match the input data size;

[0031] Step 3.3: Feed the expanded noise covariance matrix and the PD processed sub-block data into the self-supervised blind spot network BSN together;

[0032] Step 3.4: Calculate the network prediction error based on the Bayes formula, which is:

[0033]

[0034] Symbol explanation:

[0035] : Noise image input, target to be denoised;

[0036] : Take the average of all training data;

[0037] : Clean signal predicted by the network, corresponding to the D-BSN output;

[0038] : Covariance matrix of the predicted signal;

[0039] : Noise covariance matrix output by the noise prediction module;

[0040] : Trace operation of the matrix, which constrains the amplitude of ;

[0041] Step 3.5: Optimize the noise prediction module and BSN network parameters through the back propagation algorithm, obtain the trained denoising model weights and save them, and finally obtain the trained model.

[0042] In the above method, the step 4 includes the following steps:

[0043] Step 4.1: Load the trained denoising model weights, and input the noisy seismic data to be processed into the model;

[0044] Step 4.2: Perform pixel-wise random down-sampling PD with step factor s = 2 on the input noisy data to form four discrete sub-block data, and feed them into the denoising model;

[0045] Step 4.3: The noise prediction module of the denoising model and the self-supervised blind spot network BSN output processed denoising sub-block data, and then each sub-block is inversely PD operated according to the original PD arrangement mode of the mosaic, and is combined into final denoising seismic data of complete size.

[0046] The application also provides a self-supervised seismic denoising system fusing a noise estimation module and a channel attention mechanism, comprising:

[0047] The seismic data set construction module comprises slicing processing of real noisy seismic data and synthetic noisy seismic data to obtain a plurality of noisy data blocks;

[0048] The self-supervised denoising network module comprises:

[0049] The noise prediction module (CNNest) is composed of a plurality of 1x1 convolutional layers and ReLU activation functions, and is used for outputting a noise level prediction map with the same size as the input data, wherein the noise level prediction map represents a noise covariance matrix of each pixel;

[0050] The self-supervised blind spot network (BSN) comprises a double-branch structure integrating a channel attention mechanism:

[0051] The first branch adopts an inflation convolution module with a center mask convolution, referred to as a DC module, for extracting local features;

[0052] The second branch adopts an inflation Transformer block integrating a channel attention mechanism, referred to as a DTB module, wherein the DTB module realizes global feature modeling through the channel attention mechanism, and a feedforward layer thereof comprises a depth separable convolution;

[0053] The data processing and training unit applies a pixel random down-sampling PD operation to the input data to destroy the noise spatial correlation, and inputs the processed data into the self-supervised denoising network for training, fuses the output of the noise prediction module and the intermediate features of the self-supervised blind spot network by using a Bayesian prediction formula, and optimizes the network parameters through back propagation to obtain a denoising model;

[0054] The denoising processing and recovery unit inputs the denoising seismic data to be processed into the trained model for denoising processing, and recovers the final denoising result through inverse PD operation.

[0055] In the above system, the implementation of the self-supervised denoising network module comprises the following steps:

[0056] Step 2.1: Constructing a noise prediction module comprising five 1x1 convolutional layers connected in sequence, wherein ReLU nonlinear activation functions are arranged at the back of the first four layers, and the fifth layer outputs a noise covariance matrix with the same size as the input noisy data;

[0057] Step 2.2: Constructing the double-branch architecture of the self-supervised blind spot network BSN, wherein,

[0058] The first branch includes an upper path and a lower path:

[0059] The upper path includes a 3x3 center mask convolution layer and 9 DC modules in series, where the DC module uses an expansion convolution with a step size of 2 and contains parallel upper / middle / lower paths, each of which is realized by a combination of a 1x1 convolution layer and a 3x3 expansion convolution layer to extract local features;

[0060] The lower path contains a DC module with a 5x5 center mask convolution layer and 9 expansion convolution layers with S=3, where the DC module uses an expansion convolution with a step size of 3 and contains parallel upper / middle / lower paths, each of which is realized by a combination of a 1x1 convolution layer and a 3x3 expansion convolution layer to extract local features;

[0061] The second branch includes a 5x5 center mask convolution layer and 9 DTB modules in series, which generate channel-interacting global feature maps through a channel attention mechanism and use depthwise separable convolution and feedforward networks to enhance long-range dependencies, where the channel attention mechanism is realized by calculating the dot product of the query matrix Q, the key matrix K, and the value matrix V; The outputs of the first branch and the second branch are concatenated after removing the blind spot information, and the multi-scale features are fused through a 1x1 convolution layer; At the same time, the output of the noise prediction module is used for Bayesian prediction with the fused features to form the final BSN output result.

[0062] In the above system, the implementation of the data processing and training unit includes the following steps:

[0063] Step 3.1: Perform pixel random down-sampling PD processing with a step factor s=2 on the input real noisy seismic data, split the input data into 2x2 sub-blocks and arrange them in a mosaic pattern, generate four sub-block data to break the spatial correlation of the noise;

[0064] Step 3.2: Input the sub-block data after PD processing into the noise prediction module to obtain a 1x1 noise covariance matrix for each pixel, and expand the matrix dimension to match the input data size;

[0065] Step 3.3: Feed the expanded noise covariance matrix and the sub-block data after PD processing into the self-supervised blind spot network BSN;

[0066] Step 3.4: Calculate the network prediction error based on the Bayesian formula, which is:

[0067]

[0068] Symbol explanation:

[0069] : noisy image input, target to be denoised;

[0070] : average over all training data;

[0071] : network predicted clean signal, corresponding to D-BSN output;

[0072] : covariance matrix of predicted signal;

[0073] : noise covariance matrix output by noise prediction module;

[0074] : trace operation on matrix, constraint on amplitude;

[0075] : log determinant, used to measure complexity of noise model.

[0076] Step 3.5: optimize noise prediction module and BSN network parameters through back propagation algorithm, obtain trained denoising model weight and save, and finally obtain trained model.

[0077] In the above system, the implementation of the denoising processing and recovery unit comprises the following steps:

[0078] Step 4.1: load the trained denoising model weight, and input the noisy seismic data to be processed into the model;

[0079] Step 4.2: perform pixel random down-sampling PD processing with a step factor s=2 on the input noisy data, form four discrete sub-block data, and feed them into the denoising model;

[0080] Step 4.3: output the processed denoised sub-block data by the noise prediction module and the self-supervised blind spot network BSN of the denoising model, and then perform inverse PD operation on each sub-block according to the mosaic arrangement mode in the original PD processing, and combine them into the final denoised seismic data of complete size.

[0081] Because the application adopts the above technical means, the following beneficial effects are achieved:

[0082] 1. The application converts the real seismic noise into AWGN-like noise by PD, which is consistent with the basic assumption of BSN.

[0083] 2、The application adopts a self-supervised denoising network BSN of a noise prediction module and an integrated channel attention mechanism, limits the denoising type by estimating the noise intensity, and can extract good noise-free data characteristics in real seismic data and synthetic seismic data, thereby improving the denoising performance and generalization of deep learning for seismic data.

[0084] 3、Noise adaptive modeling: through the noise prediction module (CNNest) constructed by the 1x1 convolution layer, the noise covariance matrix is learned pixel by pixel using the full convolution architecture, solving the problem that the traditional synthetic noise model does not match the statistical characteristics of the real seismic noise. The module dynamically corrects the blind spot network output through the Bayesian formula, realizes accurate estimation of non-stationary and heteroscedastic noise, makes the denoising intensity and the local noise level adaptively match, and avoids the loss of weak signals caused by uniform denoising.

[0085] 4、Feature enhancement under blind spot constraint: a double-branch self-supervised blind spot network (BSN) is designed, and the center mask convolution is used to forcibly exclude the pixel information, thereby breaking through the dependence of traditional supervised learning on clean data. The DC module adopts an expansion convolution stack with a step of 2 / 3, which expands the receptive field to 115x115 pixels while maintaining the blind spot constraint, solves the problem of insufficient local feature extraction of conventional convolution, and significantly improves the capture ability of complex noise patterns.

[0086] 4、Global-local feature collaboration: the DTB module is innovatively integrated in the BSN, and the cross-channel global interaction is realized through the channel attention mechanism. Specifically, the spatial attention is replaced by the channel attention in the Transformer block, which avoids violating the blind spot constraint, and at the same time, the dot product operation of the query matrix (Q) and the key matrix (K) is used to establish a long-range dependency relationship, which compensates for the modeling defects of the pure CNN architecture for the continuity of the seismic signal.

[0087] 5、Decryption of noise spatial correlation: the pixel random down-sampling (PD) technology is introduced, and the original noise correlation distance is expanded from the neighborhood 1-2 pixels to 2-4 pixels through s=2 step mosaic block reorganization. According to the actual measurement, this method reduces the noise spatial correlation coefficient from 0.89 to 0.06, forcibly satisfies the noise independence assumption of BSN, and solves the core contradiction between the spatial correlation of real seismic noise and the algorithm assumption.

[0088] 6、Multi-scale feature fusion optimization: the noise estimation module and the BSN network are jointly optimized through the Bayesian prediction formula, and the loss function makes the synthetic data pre-training model quickly converge to the real noise distribution, effectively alleviating the domain shift problem. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 The flowchart of the application;

[0090] Figure 2 Schematic diagram of the prediction module of the present invention;

[0091] Figure 3 Schematic diagram of the DTB module in the present invention;

[0092] Figure 4 Schematic diagram of the DC module with S=2 in the present invention;

[0093] Figure 5 Schematic diagram of PD operation and inverse PD operation under random pixel sampling of the present invention;

[0094] Figure 6 For the present invention Figure 2 Schematic diagram of denoising results of earthquake records;

[0095] Figure 7 For the present invention Figure 2 Schematic diagram of the removed noise. DETAILED DESCRIPTION

[0096] 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 certain specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions of the present invention are intended to fall within the scope of the claims of the present invention.

[0097] In addition, in order to better illustrate the present invention, numerous specific details are given in the following detailed description. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0098] This paper first proposes a self-supervised denoising network (BSN) that uses a noise prediction module and an integrated channel attention mechanism. The DC module captures the local characteristics of seismic data, while the DTB module captures the global characteristics of seismic data, thereby better deriving the implicit characteristics of the seismic data itself. Furthermore, pixel shuffling and downsampling are used to improve the generalization of the network based on the prediction of seismic data noise intensity.

[0099] To this end, the present invention provides an embodiment, a self-supervised seismic denoising method integrating a noise estimation module, comprising the following steps:

[0100] Step 1: Collect real noisy seismic data, slice the noisy seismic data to obtain multiple noisy seismic data blocks, then add synthetic seismic data of the same size and manually add Gaussian white noise to construct a seismic dataset;

[0101] Step 2: Build a self-supervised blind spot network based on the channel attention mechanism and prediction network, and combine the processed noise estimation module with the blind spot network;

[0102] Step 3: The noise module pixel random down-sampling method reduces spatial correlation and feeds into the entire network, and the training set is used to train the seismic data denoising model to obtain the trained denoising model for the real noise of the seismic data;

[0103] Step 4: The test set is input into the trained denoising model to obtain the denoised seismic data.

[0104] Further, the step 1 comprises the following steps:

[0105] Step 1.1: Use the noisy real seismic data and the synthesized noisy seismic data to slice, and the slice size is 128x128, as a seismic data set. The seismic data set includes a training set, a validation set and a test set.

[0106] Further, the step 2 comprises the following steps:

[0107] Step 2.1: The noisy seismic data is input into the network for preprocessing, and the data is normalized, and rotation, symmetry and other operations are performed, so as to improve the robustness of the algorithm.

[0108] Step 2.2: Noise estimation module construction

[0109] The noisy seismic data is fed into the noise prediction module CNNest. The noise estimation module CNNest is composed of five 1x1 convolutional layers with 16 channels, and the ReLU nonlinear is deployed in all convolutional layers except the last layer.

[0110] Assuming that the noise is conditionally pixel-wise independent given the underlying clean data, using the FCN architecture CNNest with 1x1 convolution to learn the noise model is beneficial to learn the information around each pixel. Benefiting from all 1x1 convolutional layers, it can be ensured that the noise level at a position only depends on the input value at the same position. Therefore, CNNest takes seismic noise as input and approximates a noise level function in its multivariate heteroscedastic Gaussian model by learning. Each input data pixel corresponds to an output of a 1x1 covariance matrix, which is used for Bayesian prediction with the output of BSN to obtain denoised seismic data.

[0111] Step 2.3: Construction of self-supervised blind spot network BSN

[0112] Assuming spatial noise is unrelated, the BSN network excludes each pixel from its receptive field, preventing it from learning its own identity. This allows it to be trained using the same seismic noise data as input and target, learning to remove spatially independent noise at the pixel level. However, because the network cannot see blind spot information, BSN suffers from greater information loss when self-supervision is required.

[0113] To address these issues, the self-supervised denoising network (BSN) based on the channel attention mechanism of the present invention begins with a 1×1 convolution and enters two identical network branches. Each branch contains a 3×3 or 5×5 center-masked convolution layer and multiple dilated convolution modules with a stride of 2 or 3, also known as DC modules. Finally, the feature maps of the two branches are concatenated.

[0114] A binary mask m of size 3×3 and 5×5 is introduced, 0 is assigned to the central element of m and 1 is assigned to other elements, and the center mask convolution is realized by the element-wise product of the mask and the convolution kernel of the same size. Further stacking of center mask convolution layers will break the blind spot requirement, so a DC module is added to maintain the blind spot requirement. The DC module refers to adding an interval S between the original convolution kernels to enhance the receptive field, while the number of parameters only increases linearly. Each DC module used in the present invention contains a 3×3 or 5×5 dilated convolution with a dilation coefficient of S, where S=2 and S=3 are used for the upper and lower paths of the network, respectively. Each weight loads the pre-trained model weights of the non-integrated BSN trained in advance.

[0115] Step 2.4: Integrate the DTB module based on the channel attention mechanism

[0116] Since the blind spot network has a limited receptive field and CNN-based BSNs cannot fully capture long-range dependencies, this limits their performance. Therefore, the DTB module was introduced to compensate for this shortcoming by introducing an expanded Transformer block to enhance global modeling capabilities. The core of the DTB module is to design a Transformer block that meets the blind spot requirements while enabling global information exchange. Specifically, it includes two core components: the self-attention layer and the feedforward layer. For the self-attention layer, a channel attention mechanism is adopted instead of a spatial attention mechanism to avoid information exchange between adjacent pixels; for the feedforward layer, a depthwise separable convolution is introduced to reduce computational cost and enhance local context.

[0117] To achieve global information exchange without violating the blind spot requirement, the DTB module adopts a channel attention mechanism. This mechanism does not rely on spatial position information, but achieves global perception through interaction between channels, and can obtain global information of seismic data. To achieve global information exchange without violating the blind spot requirement, the DTB module adopts a channel attention mechanism. This mechanism does not rely on spatial position information, but achieves global perception through interaction between channels. Specifically, first, the Query (Q), Key (K) and Value (V) matrices are calculated, and then the attention map of channel interaction is obtained through dot multiplication operation. Change the DC module in step 2.3 to the DTB module, add layer normalization to the feedforward processing, and select a 5x5 center mask convolution layer.

[0118] Subsequently, the results after the DTB module and the results after the DC module are fused to generate network output. The integrated network can effectively make comprehensive judgments on the model prediction results and obtain better results than the simple blind spot network model.

[0119] Further, the step 3 includes the following steps:

[0120] Step 3.1: The assumption of BSN that spatial noise is not related does not match the actual seismic data, so PD is introduced, which means creating a mosaic by downsampling seismic data with a specific step factor s to increase the actual distance between noise signals and break the spatial correlation of noise. After verification, the step factor s=2 is selected here. Apply BSN to the downsampled seismic data after processing to meet its assumptions, and then perform PD inverse operation after denoising to reconstruct the full-size output.

[0121] Step 3.2: Input the PD processed seismic data into the noise prediction module to obtain the result, which is expanded in length, width and dimension to have the same size as the input seismic data. The result is fed into BSN together with the seismic data for training, and is optimized through a pre-set loss function. The loss function uses the Bayes formula to integrate the limitations of the predicted noise model and the denoising results. Finally, the trained model is obtained.

[0122] Further, the step 4 includes the following steps:

[0123] Step 4.1: Load the saved model weight into the model, and then input the real seismic data containing noise into the model loaded with the model weight for denoising to obtain the denoised seismic data.

[0124] Embodiment 1

[0125] The application will be described in detail below. Figures 1-5 The application will be described in detail below.

[0126] AsFigure 1 As shown in the figure, a self-supervised seismic denoising method fusing noise estimation module and channel attention mechanism includes the following steps:

[0127] Step 1: Use real seismic data and synthetic noise-added seismic data to make a data set by slicing, and the size of the slice is 128*128.

[0128] Step 2: Input the seismic data into the network for training, and the network has a noise prediction module, a BSN integrated with a channel attention mechanism, wherein the noise prediction module is as shown in the figure Figure 2 As shown in the figure, the two basic modules of BSN: the DTB module is as shown in the figure Figure 3 As shown in the figure, the DC module is as shown in the figure Figure 4 As shown in the figure;

[0129] The specific steps are as follows:

[0130] Step 2.1: When the noisy seismic data is input into the network, the data is preprocessed, that is, the data is normalized and rotated, etc., so as to improve the robustness of the algorithm.

[0131] Step 2.2: The data is subjected to PD operation with a step S=2, and the obtained data is prepared to be input into the prediction network and BSN, wherein the PD operation is as shown in the figure. Figure 5

[0132] Step 2.3: The data is input into the noise prediction module, and the noise prediction module is trained.

[0133] The noisy seismic data is input into the noise prediction module, first through four 1x1 convolution layers and ReLU layers to learn the seismic noise model, and the output channel of the convolution layer is 16 and the step is 1. Finally, a 1x1 convolution layer is further used to extract the noise local features, and the output channel of the convolution layer is 1. Each input data pixel corresponds to an output 1x1 covariance matrix. Finally, a noise level prediction image with the same length and width as the input data is obtained.

[0134] Step 2.4: The data is input into the BSN part containing the DC module.

[0135] The BSN of the present application starts from a 1x1 convolution, enters the part containing the DC module and the part containing the DTB module. The part containing the DC module is divided into two groups of branches, called upper and lower paths. The first group of branches, that is, the upper path, contains a DC module containing a 3x3 center mask convolution layer and nine expansion convolution layers with S=2. The second group of branches, that is, the lower path, contains a DC module containing a 5x5 center mask convolution layer and nine expansion convolution layers with S=3. The prepared seismic data is simultaneously input into the upper and lower paths.

[0136] ​A binary mask m of size 3x3 is introduced, 0 is assigned to the center element of m, and 1 is assigned to other elements, and the center mask convolution is realized by the element product of the mask and the element of the convolution kernel of the same size.

[0137] Further stacking the center mask convolution layer will break the blind spot requirement, so the DC module is added to maintain the blind spot requirement. The DC module means adding a spacing S between the original convolution kernel to enhance the receptive field, while the number of parameters only presents a linear increase.

[0138] The upper path passes through 9 DC modules, and each DC module contains three branches, called upper, middle and lower three branches. The upper branch successively contains: a 1x1 convolution plus a layer of ReLU, two 3x3 dilated convolutions, each layer has a layer of ReLU, and the dilated coefficient is 2; the middle branch successively contains: a 1x1 convolution plus a layer of ReLU, a 3x3 dilated convolution plus a layer of ReLU, and the dilated coefficient is 2; the lower branch contains a 1x1 convolution plus a layer of ReLU. The step of 1x1 convolution is 1, the step of 3x3 dilated convolution is 1, and the number of channels of each convolution layer is 96.

[0139] The lower path passes through 9 DC modules, and each DC module contains three branches, called upper, middle and lower three branches. The upper branch successively contains: a 1x1 convolution plus a layer of ReLU, two 3x3 dilated convolutions, each layer has a layer of ReLU, and the dilated coefficient is 3; the middle branch successively contains: a 1x1 convolution plus a layer of ReLU, a 3x3 dilated convolution plus a layer of ReLU, and the dilated coefficient is 3; the lower branch contains a 1x1 convolution plus a layer of ReLU. The step of 1x1 convolution is 1, the step of 3x3 dilated convolution is 1, and the number of channels of each convolution layer is 96.

[0140] The input seismic data passes through each DC module to stack the results obtained by the upper, middle and lower branches, and a layer of 1x1 convolution plus a layer of ReLU is performed on the stacked results, where the number of channels of the convolution layer is 96 and the step is 1. The obtained result is added to the output of the previous DC module to obtain the final result and then sent to the next DC module.

[0141] Finally, the channel stacking of the feature maps obtained by the upper and lower paths is performed, and four 1x1 convolution layers are deployed to extract deep network features to produce part of the output of the BSN containing the DC module. The number of channels of the convolution layer is 32 and the step is 1.

[0142] Step 2.5: Input the data into the BSN part containing the DTB module.

[0143] The prepared seismic data is input into the part containing the DTB module. The part contains a 5x5 center mask convolution layer and 9 DTB modules with 3x3 dilated convolution layers with S=3. Each DTB module is first subjected to a layer normalization layer, and then fed into three branches at the same time. Each branch contains a 3x3 dilated convolution layer with S=3, the output channel number of the convolution layer is 96, the step is 1, and the channel attention mechanism is used to realize global perception through the interaction between channels. First, the Query (Q), Key (K) and Value (V) matrices are calculated, and then the attention map of channel interaction is obtained through the dot product operation. The channel interaction attention map is added to the original data, and the result is subjected to a layer normalization layer. The features extracted by the dilated deep direction convolution are passed through the non-linear gate unit, which contains two branches. Each branch contains a 3x3 dilated convolution layer with S=3, and the output channel number of the convolution layer is 96, the step is 1. And the gate unit is the element-wise product of the two parallel paths, one of which is activated by the GELU unit. Finally, the resulting result is added to the input data, and the final output result of the BSN part containing the DTB module is obtained.

[0144] Step 3: Train the BSN.

[0145] The output results of the BSN part containing the DTB module and the BSN part containing the DC module are added, and the results of the noise prediction module are subjected to Bayesian prediction, and then subjected to final feature extraction through 4 1x1 convolution layers to obtain the final BSN network output result. Train and optimize on the basis of this process. The optimizer is Adam.

[0146] Step 4: Perform seismic data denoising.

[0147] The specific steps are as follows:

[0148] Step 3.1: Load the saved model weight into the model, and then input the seismic data containing noise into the model loaded with the model weight for denoising to obtain the denoised seismic data.

[0149] Step 3.2: Perform inverse PD operation on the denoised seismic data to restore the impact caused by the PD operation when inputting, to obtain the final seismic denoising data.

[0150] In summary, a self-supervised seismic denoising method fusing a noise estimation module and a channel attention mechanism is used to denoise the noisy seismic data. As shown in FIG. 6, the horizontal coordinate is the number of seismic traces, and the vertical coordinate is the number of sampling points. The denoising result shows that the method can better restore the original seismic data and remove noise, proving the correctness of the method. Figure 7

[0151] Embodiment 2 ​

[0152] The application also provides a self-supervised seismic denoising system fusing a noise estimation module and a channel attention mechanism, comprising:

[0153] a seismic data set construction module, comprising slicing processing of real noisy seismic data and synthetic noisy seismic data to obtain a plurality of noisy data blocks;

[0154] a self-supervised denoising network module, comprising:

[0155] a noise prediction module (CNNest) composed of a plurality of 1x1 convolutional layers and ReLU activation functions, for outputting a noise level prediction map with the same size as the input data, wherein the noise level prediction map represents a noise covariance matrix of each pixel;

[0156] a self-supervised blind spot network (BSN) comprising a double-branch structure integrating a channel attention mechanism:

[0157] the first branch adopts an expanded convolution module with a center mask convolution, referred to as a DC module, for extracting local features;

[0158] the second branch adopts an expanded Transformer block integrating a channel attention mechanism, referred to as a DTB module, wherein the DTB module realizes global feature modeling through the channel attention mechanism, and the feedforward layer thereof comprises a depth separable convolution;

[0159] a data processing and training unit, which applies a pixel random down-sampling (PD) operation to the input data to destroy the spatial correlation of the noise, and inputs the processed data into the self-supervised denoising network for training, fuses the output of the noise prediction module and the intermediate features of the self-supervised blind spot network using a Bayesian prediction formula, and optimizes the network parameters through back propagation to obtain a denoising model;

[0160] a denoising processing and recovery unit, which inputs the seismic data to be denoised into the trained model for denoising processing, and recovers the final denoising result through inverse PD operation.

[0161] In the above system, the implementation of the self-supervised denoising network module comprises the following steps:

[0162] Step 2.1: Constructing a noise prediction module, comprising five 1x1 convolutional layers connected in sequence, wherein the last four layers are deployed with ReLU nonlinear activation functions, and the fifth layer outputs a noise covariance matrix with the same size as the input noisy data;

[0163] Step 2.2: Constructing a double-branch architecture of a self-supervised blind spot network (BSN), wherein,

[0164] the first branch comprises an upper path and a lower path:

[0165] Upper path: includes a 3×3 center-masked convolutional layer and nine serially connected DC modules. The DC module uses dilated convolution with a stride of 2 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features.

[0166] Lower path: Contains a DC module with a 5×5 center-masked convolutional layer and 9 dilated convolutional layers with S=3. The DC module uses dilated convolution with a stride of 3 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features.

[0167] The second branch includes a 5×5 center-masked convolutional layer and 9 DTB modules connected in series. The DTB module generates a global feature map of channel interaction through a channel attention mechanism, and uses deep separable convolution and feedforward networks to enhance long-range dependencies. The channel attention mechanism is implemented by calculating the dot product operation of the query matrix Q, the key matrix K and the value matrix V. The outputs of the first and second branches are concatenated after removing blind spot information, and multi-scale features are fused through a 1×1 convolutional layer. At the same time, the output of the noise prediction module is Bayesian predicted with the fused features to form the final BSN output result.

[0168] In the above system, the implementation of the data processing and training unit includes the following steps:

[0169] Step 3.1: Perform pixel random downsampling (PD) processing with a step size factor of s=2 on the input real noisy seismic data. The input data is split into 2×2 sub-blocks and arranged in a mosaic pattern to generate four sub-block data to break the spatial correlation of the noise.

[0170] Step 3.2: Input the PD-processed sub-block data into the noise prediction module to obtain a 1×1 noise covariance matrix for each pixel, and expand the matrix dimension to match the input data size;

[0171] Step 3.3: Feed the expanded noise covariance matrix and the PD-processed sub-block data into the self-supervised blind spot network BSN;

[0172] Step 3.4: Calculate the network prediction error based on the Bayesian formula, which is:

[0173]

[0174] Explanation of symbols:

[0175] : Noisy image input, target to be denoised;

[0176] : Take the average of all training data;

[0177] : network predicted clean signal, corresponding to D-BSN output;

[0178] : covariance matrix of predicted signal;

[0179] : noise covariance matrix output by noise prediction module;

[0180] : trace operation of matrix, constraint : amplitude of

[0181] : log determinant, used to measure the complexity of noise model.

[0182] Step 3.5: optimize the noise prediction module and BSN network parameters by the back propagation algorithm, obtain the trained denoising model weight and save, and finally obtain the trained model.

[0183] In the above system, the implementation of the denoising processing and recovery unit includes the following steps:

[0184] Step 4.1: load the trained denoising model weight, and input the noisy seismic data to be processed into the model;

[0185] Step 4.2: perform pixel random down-sampling PD processing with a step factor s=2 on the input noisy data, form four discrete sub-block data, and feed them into the denoising model;

[0186] Step 4.3: output the processed denoising sub-block data by the noise prediction module and the self-supervised blind spot network BSN of the denoising model, and then perform inverse PD operation on each sub-block according to the mosaic arrangement mode in the original PD processing, and combine them into the final denoising seismic data of complete size.

[0187] The present application adds DTB module and DC module in BSN, optimizes the network structure, and enhances the feature extraction capability of seismic data. At the same time, through the seismic data noise intensity prediction module, the denoising performance is optimized. In addition, through PD and inverse PD operation, the generalization of the network is enhanced. A self-supervised seismic denoising method combining noise estimation module and channel attention mechanism is proposed, and the denoising result after applying the method to noisy seismic data is good.

[0188] The above is only a representative embodiment in the many specific application ranges of the present application, and does not constitute any limitation on the protection scope of the present application. Any technical solution formed by transformation or equivalent replacement falls within the scope of protection of the present application.

Claims

1. A self-supervised seismic denoising method integrating a noise estimation module and a channel attention mechanism, characterized in that: The following steps are involved: Step 1: Construct a seismic data set, which includes slicing the real noisy seismic data and the synthetic noisy seismic data to obtain multiple noisy data blocks; Step 2: Build a self-supervised denoising network, including: A noise prediction module, consisting of multiple 1×1 convolutional layers and ReLU activation functions, is used to output a noise level prediction map of the same size as the input data, where the noise level prediction map represents the noise covariance matrix of each pixel; Self-supervised blind spot network, including a dual-branch structure with integrated channel attention mechanism: The first branch uses a dilated convolution module with center mask convolution, referred to as the DC module, to extract local features; The second branch uses a dilated Transformer block combined with a channel attention mechanism, referred to as the DTB module. The DTB module implements global feature modeling through the channel attention mechanism, and its feed-forward layer contains depthwise separable convolutions; Step 3 includes: Step 3.1: Perform pixel random downsampling (PD) processing with a step size factor of s=2 on the input real noisy seismic data. The input data is split into 2×2 sub-blocks and arranged in a mosaic pattern to generate four sub-block data to break the spatial correlation of the noise. Step 3.2: Input the PD-processed sub-block data into the noise prediction module to obtain a 1×1 noise covariance matrix for each pixel, and expand the matrix dimension to match the input data size; Step 3.3: Feed the expanded noise covariance matrix and the PD-processed sub-block data into the self-supervised blind spot network BSN; Step 3.4: Calculate the network prediction error based on the Bayesian formula, which is: Explanation of symbols: : Noisy image input, target to be denoised; : Take the average of all training data; : Clean signal predicted by the network, corresponding to D-BSN output; : covariance matrix of the predicted signal; : The noise covariance matrix output by the noise prediction module; : Matrix trace operation, constraint The amplitude of Step 3.5: Optimize the noise prediction module and BSN network parameters through the back propagation algorithm, obtain the trained denoising model weights and save them, and finally obtain the trained model; Step 4: Input the seismic data to be denoised into the trained model for denoising, and restore the final denoising result through the inverse PD operation.

2. The method according to claim 1, characterized in that Step 2 includes the following steps: Step 2.1: Construct a noise prediction module consisting of five 1x1 convolutional layers connected in sequence. The first four layers are followed by a ReLU nonlinear activation function, and the fifth layer outputs a noise covariance matrix of the same size as the input noisy data. Step 2.2: Construct a dual-branch architecture of the self-supervised blind spot network BSN, where The first branch includes the upper and lower paths: Upper path: includes a 3×3 center-masked convolutional layer and nine serially connected DC modules. The DC module uses dilated convolution with a stride of 2 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features. Lower path: Contains a DC module with a 5×5 center-masked convolutional layer and 9 dilated convolutional layers with S=3. The DC module uses dilated convolution with a stride of 3 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features. The second branch includes a 5×5 center-masked convolutional layer and 9 DTB modules connected in series. The DTB module generates a global feature map of channel interaction through a channel attention mechanism, and uses deep separable convolution and feedforward networks to enhance long-range dependencies. The channel attention mechanism is implemented by calculating the dot product operation of the query matrix Q, the key matrix K and the value matrix V. The outputs of the first and second branches are concatenated after removing blind spot information, and multi-scale features are fused through a 1×1 convolutional layer. At the same time, the output of the noise prediction module is Bayesian predicted with the fused features to form the final BSN output result.

3. The method according to claim 2, characterized in that The step 4 comprises the following steps: Step 4.1: Load the trained denoising model weights and input the noisy seismic data to be processed into the model; Step 4.2: Perform pixel random downsampling (PD) processing with a step factor of s=2 on the input noisy data to form four discrete sub-block data, and feed them into the denoising model; Step 4.3: The denoised sub-block data are processed by the noise prediction module of the denoising model and the self-supervised blind spot network (BSN), and then each sub-block is subjected to an inverse PD operation according to the mosaic arrangement during the original PD processing, and merged into the final denoised seismic data of the full size.

4. A self-supervised seismic denoising system integrating a noise estimation module and a channel attention mechanism, characterized by: include: Seismic data set construction module, including slicing real noisy seismic data and synthetic noisy seismic data to obtain multiple noisy data blocks; Self-supervised denoising network module, including: A noise prediction module, consisting of multiple 1×1 convolutional layers and ReLU activation functions, is used to output a noise level prediction map of the same size as the input data, where the noise level prediction map represents the noise covariance matrix of each pixel; Self-supervised blind spot network, including a dual-branch structure with integrated channel attention mechanism: The first branch uses a dilated convolution module with center mask convolution, referred to as the DC module, to extract local features; The second branch uses a dilated Transformer block combined with a channel attention mechanism, referred to as the DTB module. The DTB module implements global feature modeling through the channel attention mechanism, and its feed-forward layer contains depthwise separable convolutions; The implementation of the data processing and training unit includes: The pixel random downsampling (PD) process with a step size factor of s=2 is performed on the input real noisy seismic data. The input data is split into 2×2 sub-blocks and arranged in a mosaic pattern to generate four sub-block data to break the spatial correlation of the noise. The PD-processed sub-block data is input into the noise prediction module to obtain a 1×1 noise covariance matrix for each pixel, and the matrix dimension is expanded to match the input data size; The expanded noise covariance matrix and the sub-block data after PD processing are fed into the self-supervised blind spot network BSN; The network prediction error is calculated based on the Bayesian formula, which is: Explanation of symbols: : Noisy image input, target to be denoised; : Take the average of all training data; : Clean signal predicted by the network, corresponding to D-BSN output; : covariance matrix of the predicted signal; : The noise covariance matrix output by the noise prediction module; : Matrix trace operation, constraint The amplitude of The noise prediction module and BSN network parameters are optimized through the back propagation algorithm, the trained denoising model weights are obtained and saved, and finally the trained model is obtained; The denoising and recovery unit inputs the seismic data to be denoised into the trained model for denoising, and recovers the final denoising result through the inverse PD operation.

5. The system according to claim 4, characterized in that The implementation of the self-supervised denoising network module includes the following steps: Step 2.1: Construct a noise prediction module consisting of five 1x1 convolutional layers connected in sequence. The first four layers are followed by a ReLU nonlinear activation function, and the fifth layer outputs a noise covariance matrix of the same size as the input noisy data. Step 2.2: Construct a dual-branch architecture of the self-supervised blind spot network BSN, where The first branch includes the upper and lower paths: Upper path: includes a 3×3 center-masked convolutional layer and nine serially connected DC modules. The DC module uses dilated convolution with a stride of 2 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features. Lower path: Contains a DC module with a 5×5 center-masked convolutional layer and 9 dilated convolutional layers with S=3. The DC module uses dilated convolution with a stride of 3 and contains parallel upper / middle / lower paths. Each path is composed of a 1×1 convolutional layer and a 3×3 dilated convolutional layer to extract local features. The second branch includes a 5×5 center-masked convolutional layer and 9 DTB modules connected in series. The DTB module generates a global feature map of channel interaction through a channel attention mechanism, and uses deep separable convolution and feedforward networks to enhance long-range dependencies. The channel attention mechanism is implemented by calculating the dot product operation of the query matrix Q, the key matrix K and the value matrix V. The outputs of the first and second branches are concatenated after removing blind spot information, and multi-scale features are fused through a 1×1 convolutional layer. At the same time, the output of the noise prediction module is Bayesian predicted with the fused features to form the final BSN output result.

6. The system according to claim 5, characterized in that The implementation of the denoising and restoration unit includes the following steps: Step 4.1: Load the trained denoising model weights and input the noisy seismic data to be processed into the model; Step 4.2: Perform pixel random downsampling (PD) processing with a step factor of s=2 on the input noisy data to form four discrete sub-block data, and feed them into the denoising model; Step 4.3: The denoised sub-block data are processed by the noise prediction module of the denoising model and the self-supervised blind spot network (BSN), and then each sub-block is subjected to an inverse PD operation according to the mosaic arrangement during the original PD processing, and merged into the final denoised seismic data of the full size.

Images