Coupling noise denoising network method based on monte carlo guided attention mechanism

Abstract

The application discloses a kind of coupling noise denoising network methods based on Monte Carlo guided attention mechanism, it is related to the field of oil and gas exploration technology, construct the sample data set suitable for deep learning network training and inference;Build the seismic data denoising deep learning network with U-Net as main stem, introduce Swin-Transformer module, model the global statistical characteristics of seismic signal;Introduce Monte Carlo guided attention mechanism, evaluate the feature response under different sampling results;Small block division and block-by-block denoising processing are carried out using DAS-VSP seismic data, and the complete DAS-VSP seismic data is recovered by small block reconstruction mode.The application introduces Monte Carlo sampling mechanism, randomly disturbs and evaluates the potential attention area in feature map, combines the global modeling capability of Transformer and the local feature extraction advantage of U-Net, realizes the efficient suppression of coupling noise, effectively retains the key wave field characteristics of seismic signal, and improves the accuracy and robustness of DAS-VSP seismic data denoising.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a coupled noise denoising network method based on the Monte Carlo guided attention mechanism. Background Technology

[0002] In recent years, Distributed Fiber Acoustic Vertical Seismic Profiling (DAS-VSP) has shown broad application prospects in seismic exploration and monitoring due to its high spatial resolution and adaptability to extreme downhole environments. However, limited by the complexity of wellbore structures and non-standard fiber optic cable laying, DAS-VSP data often contains strong coupling noise, which frequently exhibits characteristics such as frequency aliasing and complex spatial structures, severely affecting the imaging quality and accuracy of geological interpretation. Although deep convolutional neural networks based on U-Net have achieved some success in seismic denoising, their fixed network structure is difficult to fully adapt to the diversity of complex coupling noise. Furthermore, most existing deep learning methods rely on fitting coupling noise, which differs significantly from the highly non-stationary and aliasing noise characteristics of field data, especially when dealing with frequency aliasing and highly complex coupling noise, often making effective removal difficult.

[0003] Therefore, how to effectively suppress coupling noise in DAS-VSP data and improve the overall signal-to-noise ratio has become an important problem that urgently needs to be solved. Summary of the Invention

[0004] To address the technical challenges in DAS-VSP seismic data processing, such as insufficient coupling noise suppression, susceptibility of traditional attention mechanisms to noise interference, and difficulty in simultaneously modeling local details and global features, this invention discloses a coupling noise denoising network method based on a Monte Carlo guided attention mechanism. This method introduces a Monte Carlo sampling mechanism during the attention learning process to randomly perturb and evaluate potential regions of interest in the feature map. By combining the global modeling capabilities of Transformer with the local feature extraction advantages of U-Net, it achieves efficient suppression of coupling noise while effectively preserving key wavefield features of the seismic signal, thereby improving the accuracy and robustness of DAS-VSP seismic data denoising.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The coupled noise denoising network method based on Monte Carlo guided attention mechanism includes the following steps:

[0007] Step S1: Input the noise in the first arrival portion of the synthetic DAS-VSP seismic data and the actual DAS-VSP measured seismic data to construct a sample dataset suitable for deep learning network training and inference.

[0008] Step S2: Construct a deep learning network for denoising seismic data with U-Net as the backbone. Introduce the Swin-Transformer module into the encoder and decoder structure of U-Net and use its sliding window self-attention mechanism to model the global statistical features of seismic signals.

[0009] Monte Carlo guided attention mechanism is introduced into deep learning network. During network training, multiple potential regions of interest in the feature map are randomly sampled and perturbed multiple times, and the feature response under different sampling results is evaluated.

[0010] The network learns from the training data and updates the network accordingly.

[0011] Step S3: The trained model is used to divide the DAS-VSP seismic data to be processed into small blocks and perform block-by-block denoising. Finally, the complete DAS-VSP seismic data is restored by small block reconstruction, thereby effectively suppressing coupled noise.

[0012] Furthermore, the specific process of step S1 is as follows:

[0013] Generate a clean signal dataset of an ideal seismic response, and a pure noise dataset containing multiple types of coupling interference;

[0014] The two datasets above are then stacked in a targeted manner to generate noisy training samples.

[0015] Further in S1, the process of generating a clean signal dataset, namely the DAS-VSP seismic dataset, using forward modeling is as follows:

[0016] Construct geological models of various stratigraphic types, including typical structural types such as horizontal interface models, thin-layer models, diagonal interface models, and fault models;

[0017] By using Ricker wavelets with different dominant frequencies, the formation model was excited through the acoustic wave equation to obtain a clean DAS-VSP synthetic record.

[0018] The synthesized clean records were divided into blocks using sliding windows of different sizes, ultimately forming a clean signal dataset.

[0019] Furthermore, the Ricker wavelet formula is:

[0020] ;

[0021] in, For amplitude, This is the dominant frequency of the wavelet.

[0022] Furthermore, in S1, the process of constructing a pure noise dataset, i.e., the noise in the initial arrival portion of the actual DAS-VSP measured data, is as follows:

[0023] Extracting the coupling noise component from real DAS-VSP data;

[0024] The extracted noise is divided into blocks using sliding windows of different sizes, ultimately forming a pure noise dataset.

[0025] Using the generated clean signal dataset and the pure noise dataset, construct the noisy data samples required for training.

[0026] Furthermore, in step S2, the entire system adopts an encoder-decoder architecture, and the network structure is built based on the U-Net architecture.

[0027] In the encoder section, the input seismic data undergoes two convolution operations to extract initial feature information, which is then input into the TransConv module for feature enhancement. The output of the TransConv module is used for subsequent downsampling and attention feature enhancement. The downsampling process uses max pooling to progressively reduce the spatial resolution of the feature map, enabling high-level features to aggregate a wider range of input information, thereby effectively improving the network's receptive field.

[0028] The attention feature enhancement process includes channel attention and Monte Carlo guided attention mechanisms, used to adaptively weight the feature maps. In the channel attention mechanism, the feature map is first convolved and then activated by the GELU activation function. Subsequently, global average pooling is used to extract channel-level statistical features, and then two convolution operations are performed to generate corresponding channel attention weights, which are used to adjust the feature responses of different channels. Following the channel attention mechanism, the Monte Carlo guided attention mechanism is introduced. During network training, multiple potential regions of interest in the feature map are randomly sampled and perturbed, and the sampling results are evaluated. In the Monte Carlo attention mechanism, the feature map is first convolved to reduce the number of channels to 1. The weights of each part of the feature map are extracted using the Monte Carlo method, and then the result is multiplied element-wise with the input feature map to obtain the final attention output.

[0029] Furthermore, in each stage of the encoder, the number of channels in the network remains unchanged at 64, and each level of convolution operation uses a 3×3 kernel with a stride of 1. In the decoder, the feature map is upsampled level by level through transposed convolution. The transposed convolution kernel is 2×2 with a stride of 2 to restore the spatial resolution of the feature map. The upsampled features are then concatenated with the features of the corresponding level in the encoder.

[0030] The TransConv module includes a Transformer submodule and a convolution submodule, and adopts a dual-branch structure. One branch inputs features into the Transformer submodule, and the other branch performs multiple convolution operations on the features. The outputs of the two branches are concatenated at the end of the module to achieve the fusion of global and local features.

[0031] Furthermore, in step S2, a seismic data denoising network is used as a supervised learning network, and the training process is carried out using known noisy DAS-VSP seismic data and corresponding reference seismic data;

[0032] Assume the input noise data is Clean DAS data is The mean squared error (MSE) loss function is defined as follows:

[0033] ;

[0034] in, This refers to the neural network described in this invention. Represents training parameters, For weight parameters, This is the bias parameter.

[0035] Furthermore, in step S3, during the denoising process, the trained denoising network based on the Monte Carlo attention mechanism is used to perform forward prediction processing on the two-dimensional DAS-VSP noisy data. The processing procedure is as follows:

[0036] S31. Divide the two-dimensional DAS-VSP noisy data to be processed into multiple data blocks according to the preset window size and sliding step size. When the original data size cannot be divided by the window size, perform zero-filling processing in the data edge area.

[0037] S32. Expand the data block in the channel dimension to meet the input channel requirements of the denoising network, and organize it into input batches according to the preset batch size, and input it into the trained denoising network for forward inference.

[0038] S33. Obtain the denoising results corresponding to each data block, and reconstruct the denoised data blocks according to the spatial position and sliding step order of the data blocks in the partitioning stage. For data regions with overlap, use a weighted average method for fusion. If zero padding was performed in the data partitioning stage, remove the padding region in the data reconstruction stage to restore the original data size, and finally obtain a denoising result consistent with the original two-dimensional DAS-VSP data size.

[0039] The beneficial effects of this invention are that by introducing a Monte Carlo-guided attention mechanism, it effectively overcomes the problems of traditional attention mechanisms, such as convergent processing of different regions and susceptibility to coupling noise interference during feature weighting, significantly improving the accuracy and robustness of attention selectivity. Simultaneously, by combining the global modeling capabilities of Swin-Transformer with the local feature extraction advantages of U-Net, it achieves collaborative modeling of local details and global dependencies in seismic signals. Application results on synthetic data and actual DAS-VSP seismic data demonstrate that this invention has stronger noise suppression capabilities in complex coupled noise environments and can effectively preserve key information of the seismic wavefield, providing a reliable technical means for high-quality recovery of DAS-VSP seismic data. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the coupled noise denoising network based on the Monte Carlo guided attention mechanism of the present invention;

[0041] Figure 2 This is a diagram of the forward velocity model of the present invention;

[0042] Among them, (a) undulating layered model; (b) weakly undulating layered model; (c) inclined fault model; (d) horizontal layered model; (e) inclined layered model; (f) horizontal multilayered model;

[0043] Figure 3 This is a diagram of seismic data synthesized using a forward velocity model according to the present invention;

[0044] Among them, (a) clean DAS-VSP data; (b) real DAS-VSP data with coupling noise; (c) DAS-VSP data with added coupling noise; (d) real DAS-VSP data with coupling noise FX spectrum;

[0045] Figure 4 This is a sample dataset for the present invention;

[0046] Figure 5 The denoising effect of this invention on synthesized DAS-VSP data is shown in the figure.

[0047] Among them, (a) the original clean DAS-VSP record; (b) the record after adding coupling noise; and (c) the denoising result after processing by the method of the present invention;

[0048] Figure 6 The denoising effect of this invention on actual DAS-VSP data is shown in the figure.

[0049] Among them, (a) is the record after adding coupling noise; and (b) is the denoising result after processing by the method of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This invention discloses a coupled noise denoising network method based on a Monte Carlo guided attention mechanism, such as... Figure 1 As shown, it specifically includes:

[0052] (1) Input the noise of the first arrival part of the synthetic DAS-VSP seismic data and the actual DAS-VSP measured seismic data to construct a sample dataset suitable for training and inference of deep learning networks.

[0053] In supervised learning frameworks, neural networks learn this mapping relationship through a large number of paired noisy and clean data samples. Therefore, the quality of the training data largely determines the model's generalization ability and denoising performance during the inference phase. However, publicly available DAS-VSP training datasets are extremely limited, and researchers typically need to construct their own datasets. Furthermore, the types and distributions of coupling noise contained in actual DAS-VSP data from different regions vary significantly, further increasing the complexity of training data construction.

[0054] To enhance the model's generalization ability across different geological environments, the noiseless DAS-VSP samples in the training data must fully reflect key geophysical factors such as stratum reflection characteristics, wave propagation paths, and source locations. Ignoring these factors may hinder the model's ability to effectively identify and retain weak, valid reflection signals during denoising, leading to information loss. Therefore, this paper constructs a training dataset combining synthetic and measured data. Specifically, a clean signal dataset representing an ideal seismic response and a pure noise dataset containing various types of coupling interference are first generated. Then, the two datasets are selectively overlaid to generate noisy training samples, enriching the noise types while ensuring the geological validity and representativeness of the signals.

[0055] The clean signal dataset was generated using a forward modeling process: First, multiple geological models of different stratigraphic types were constructed. To diversify the data, 50 velocity models of different stratigraphic types were built, including typical structural types such as horizontal interface models, thin-layer models, dipping interface models, and fault models. Then, using Ricker wavelets of different dominant frequencies, these stratigraphic models were excited according to the acoustic wave equation to obtain clean DAS-VSP synthetic records. These synthesized clean records constitute the clean signal sample set.

[0056] Ricker's wavelet formula is as follows:

[0057] ;

[0058] in, For amplitude, This is the dominant frequency of the wavelet.

[0059] The parameter configurations used in the data synthesis process are detailed in Table 1 below.

[0060] Table 1 Forward Model Parameters

[0061]

[0062] Stratigraphic models such as Figure 2 As shown in the figure, six different formation velocity models are used for forward modeling of acoustic wave equations. Different colors represent different geological layers, with redder colors indicating higher wave velocities in those layers. The red inverted triangles in the figure indicate the location of the seismic source (shot point), and the thick black vertical lines indicate the locations of distributed acoustic sensors in the well.

[0063] Figure 3 (a) shows the noise-free DAS-VSP records obtained through forward modeling using the finite difference method. When constructing the pure noise dataset, considering the significant difference between the simulated coupling noise and actual observations, we used coupling noise extracted from real DAS-VSP data as part of the synthetic noise sample. For example... Figure 3 As shown in (b), this figure illustrates an example of the coupling noise used; its corresponding FX spectrum is shown below. Figure 3 (d) As can be seen from the spectrum, the energy of random noise is mainly concentrated in the low-frequency region and exhibits an irregular interference distribution, while coupled noise shows obvious frequency segmentation characteristics. The specific generation process of the noise dataset is as follows: First, the coupled noise part is extracted from real DAS-VSP data. Then, the extracted noise is divided into blocks using sliding windows of different sizes, finally forming the noise dataset.

[0064] Based on this, the noisy data samples required for training are constructed using the clean signal dataset and the pure noise dataset generated above. The specific training data construction process is shown in Algorithm 1. Figure 3 (c) provides an example of synthesized data after superimposing actual DAS-VSP coupling noise. Figure 4 This visualization shows a portion of the training dataset, with the first row showing clean signal blocks, the second row showing synthetic signal blocks with added noise (i.e., training samples), and the third row showing noisy data blocks. Figure 4 As can be observed from the three figures in the second row, the superimposed noise includes not only coupled noise but also a certain degree of random noise, further enhancing the complexity and realism of the training data.

[0065] (2) Construct a deep learning network for earthquake data denoising with U-Net as the backbone. In the encoder and decoder structure of U-Net, introduce the Swin-Transformer module and use its sliding window self-attention mechanism to model the global statistical features of earthquake signals, thereby achieving the collaborative representation of local detail features and long-distance dependencies. In the deep learning network, introduce the Monte Carlo guided attention mechanism. During the network training process, perform multiple random sampling and perturbation on multiple potential areas of interest in the feature map, and evaluate the feature responses under different sampling results, thereby selecting areas with stronger representation capabilities of real earthquake signals and suppressing feature responses dominated by coupled noise, thereby improving the ability of the attention mechanism to distinguish between noise and effective signals. The network is updated by learning from the training data.

[0066] This invention provides a seismic data denoising network structure based on Monte Carlo guided attention, employing an encoder-decoder architecture built upon the U-Net framework. In the encoder section, the input seismic data undergoes two convolutional operations to extract initial feature information, which is then fed into the TransConv module for feature enhancement. The output of the TransConv module is used for subsequent downsampling and attention-based feature enhancement. Downsampling is achieved through max pooling to progressively reduce the spatial resolution of the feature maps and expand the network's receptive field. Attention-based feature enhancement includes channel attention and Monte Carlo guided attention mechanisms for adaptive weighting of the feature maps.

[0067] In the channel attention mechanism, the feature map is first convolved and then activated by the GELU activation function to enhance its nonlinear expressive power. Subsequently, global average pooling is used to extract channel-level statistical features, and two more convolution operations are performed to generate corresponding channel attention weights, thereby weighting and adjusting the feature responses of different channels. Following the channel attention mechanism, a Monte Carlo guided attention mechanism is introduced. During network training, multiple potential regions of interest in the feature map are randomly sampled and perturbed, and the sampling results are evaluated. This enhances the response to effective seismic signal areas and suppresses feature regions dominated by coupling noise.

[0068] In the Monte Carlo attention mechanism, the feature map is first convolved to reduce the number of channels to 1. Then, the weights of each part of the feature map are extracted using the Monte Carlo method. The result is then multiplied element-wise with the input feature map to obtain the final output of the attention. To improve the robustness of traditional attention mechanisms in feature region selection during the attention feature modeling process, this invention introduces an attention guidance strategy based on Monte Carlo principles to filter out regions of interest containing effective seismic signal features from the feature map.

[0069] In the attention region selection process, the variance of the feature response is used as an evaluation index. Variance measures the dispersion of values ​​at each location in the feature map relative to the mean; the larger the variance, the more significant the change in the feature response within that region. For seismic data features, regions with large variances typically correspond to locations with significant changes in the amplitude of the in-phase axis, indicating that the region contains relatively rich and effective seismic signal information. When the variance is close to zero, it indicates that the feature response changes within that region are small, typically corresponding to background or blank areas lacking effective seismic signals. In the attention weight generation process, this invention divides the feature map into multiple local regions and generates multiple candidate attention regions within the feature map using Monte Carlo random sampling. The variance of the feature response is calculated for each candidate region. When the variance meets a preset threshold condition, the region is determined to be an effective attention region and assigned a high attention weight; when the variance is below the threshold, the attention weight of the corresponding region is reduced or suppressed.

[0070] By employing the Monte Carlo-based attention region selection method described above, the attention mechanism can be guided to focus more on regions containing effective seismic signal features, thereby improving the network's ability to distinguish between real seismic signals and coupled noise, and enhancing the robustness of feature modeling. Furthermore, for regions with more complex feature structures and richer details, their corresponding variance values ​​are usually larger. Prioritizing the enhancement of attention weights in such regions helps preserve detailed structural information in the seismic data, improving the resolution and reliability of the denoising results. In addition, considering that the correlation of seismic data in the vertical direction is usually stronger than in the horizontal direction, and that seismic wave energy gradually attenuates with increasing propagation depth, to reduce the impact of amplitude differences on variance calculation results, the feature responses within the candidate regions are normalized before performing attention region variance evaluation. This ensures that the attention weight allocation more accurately reflects the structural complexity within the region.

[0071] Throughout the encoder stages, the network's channel count remains constant at 64 to reduce model parameter complexity while maintaining feature representation capabilities. Each convolutional operation uses a 3×3 kernel with a stride of 1; each max-pooling operation halves the spatial resolution of the feature map. In the decoder, the feature map is progressively upsampled using transposed convolutions with a 2×2 kernel and a stride of 2 to restore the spatial resolution. The upsampled features are then concatenated with the corresponding layers in the encoder to fully utilize the high-resolution feature information retained during the encoding stage, thereby improving the accuracy of seismic data reconstruction. The TransConv module comprises a Transformer submodule and a convolutional submodule, employing a dual-branch structure. One branch inputs features into the Transformer submodule, modeling long-distance dependencies between features through self-attention to enhance global contextual information; the other branch performs multiple convolutional operations on the features to extract local structural features. The outputs of the two branches are concatenated at the module's end, achieving the fusion of global and local features. Through the above network structure design, this invention, while maintaining the stability of the U-Net network structure, introduces a Transformer module and a Monte Carlo guided attention mechanism to achieve collaborative modeling of global statistical features and local detailed features of seismic signals. This effectively suppresses coupling noise in DAS-VSP seismic data and preserves key wavefield information.

[0072] The seismic data denoising network used in this invention is a supervised learning network, and its training process is performed using known noisy DAS-VSP seismic data and corresponding reference seismic data. Assume the input noisy data is... Clean DAS data is Then the mean squared error (MSE) loss function can be defined as:

[0073] ;

[0074] in, This invention represents the neural network. Represents training parameters, For weight parameters, This is the bias parameter.

[0075] During model training, the mean squared error loss function is used to measure the difference between the network output and the reference seismic data. This loss function constrains the update direction of the network parameters, ensuring the network output gradually approximates the reference seismic data. The network parameters are updated using a gradient-based first-order adaptive optimization algorithm, where the first-order and second-order moment estimation parameters are set to preset values, and the learning rate is set to a preset learning rate to ensure the stability and convergence of the training process. During training, sample data is iteratively updated in batches, and the network parameters are optimized through multiple iterations. To avoid overfitting during training, this invention introduces an early stopping strategy. When the loss function no longer decreases significantly in several consecutive iterations, the training process is terminated early, thereby obtaining a network model with better generalization ability. The training of the network model of this invention can be completed on a general-purpose computing platform, including a computing environment composed of a central processing unit and a graphics processing unit. The above training environment does not constitute a limitation of this invention.

[0076] (3) The trained model is used to divide the DAS-VSP seismic data to be processed into small blocks and perform block-by-block denoising. Finally, the complete DAS-VSP seismic data is restored by small block reconstruction, so as to effectively suppress the coupled noise.

[0077] In the denoising process, a trained denoising network based on the Monte Carlo attention mechanism is used to perform forward prediction processing on the noisy 2D DAS-VSP data to suppress coupled noise and recover the effective signal. The specific processing steps are as follows:

[0078] First, the noisy 2D DAS-VSP data to be processed is divided into multiple data blocks according to the preset window size and sliding step size. When the original data size cannot be divided by the window size, zero-padding is performed in the data edge area to ensure that the size of each data block is consistent and covers the entire data area, thereby avoiding network inference anomalies caused by inconsistent input sizes and reducing the risk of edge information loss.

[0079] Secondly, the data blocks are expanded in the channel dimension to meet the input channel requirements of the denoising network, and then organized into input batches according to a preset batch size. These batches are then fed into the trained denoising network for forward inference. The denoising network suppresses noise components in the data blocks through convolutional feature extraction, multi-scale feature fusion, and attention gating modules, and enhances its ability to represent effective signals and their structural features under the guidance of the attention mechanism.

[0080] Subsequently, the denoising results for each data block are obtained, and the denoised data blocks are reconstructed according to their spatial location and sliding step order during the partitioning stage. For overlapping data regions, a weighted average method is used for fusion to ensure the spatial continuity and smoothness of the reconstruction results and avoid the generation of splicing gaps and edge artifacts. If zero-padding was performed during the data partitioning stage, the padding area is removed during the data reconstruction stage to restore the original data size, ultimately obtaining a denoising result consistent with the original 2D DAS-VSP data size.

[0081] Figure 5 The diagram shows the denoising effect of this invention on synthesized DAS-VSP data. Figure 5 (a) is clean data. Figure 5 (b) is the data with added coupling noise. Figure 5 (c) shows the denoising effect of the present invention. It can be seen that this method not only effectively suppresses DAS coupling noise, but also performs better in the recovery of overall effective signal events, especially in the reconstruction of deep weak signals. This indicates that the method achieves a high performance level in both global signal recovery and noise suppression.

[0082] Figure 6 The diagram shows the denoising effect of the present invention on actual DAS-VSP data. Figure 6 (a) presents actual data. Figure 6 (b) demonstrates the denoising effect. It can be seen that most random noise and coupling noise are effectively suppressed, shallow in-phase axes are clearly restored, and deep weak signals are reconstructed relatively completely. Overall, it exhibits the best clarity and coherence. The invention can not only effectively suppress coupling noise, but also has obvious advantages in weak signal recovery.

[0083] Through the above steps, automated denoising processing of two-dimensional DAS-VSP data is achieved, effectively suppressing coupling noise without manual intervention, while maintaining the structural features and detailed information of the signal.

[0084] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A coupled noise denoising network method based on Monte Carlo guided attention mechanism, characterized in that, Includes the following steps: Step S1: Input the noise in the first arrival portion of the synthetic DAS-VSP seismic data and the actual DAS-VSP measured seismic data to construct a sample dataset suitable for deep learning network training and inference. Step S2: Construct a deep learning network for denoising seismic data with U-Net as the backbone. Introduce the Swin-Transformer module into the encoder and decoder structure of U-Net and use its sliding window self-attention mechanism to model the global statistical features of seismic signals. Monte Carlo guided attention mechanism is introduced into deep learning network. During network training, multiple potential regions of interest in the feature map are randomly sampled and perturbed multiple times, and the feature response under different sampling results is evaluated. The network learns from the training data and updates the network accordingly. Step S3: The trained model is used to divide the DAS-VSP seismic data to be processed into small blocks and perform block-by-block denoising. Finally, the complete DAS-VSP seismic data is restored by small block reconstruction, thereby effectively suppressing coupled noise.

2. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 1, characterized in that, The specific process of step S1 is as follows: Generate a clean signal dataset of an ideal seismic response, and a pure noise dataset containing multiple types of coupling interference; The two datasets above are then stacked in a targeted manner to generate noisy training samples.

3. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 2, characterized in that, In S1, the process of generating a clean signal dataset, namely the DAS-VSP seismic dataset, using forward modeling is as follows: Construct geological models of various stratigraphic types, including typical structural types such as horizontal interface models, thin-layer models, diagonal interface models, and fault models; By using Ricker wavelets with different dominant frequencies, the formation model was excited through the acoustic wave equation to obtain a clean DAS-VSP synthetic record. The synthesized clean records were divided into blocks using sliding windows of different sizes, ultimately forming a clean signal dataset.

4. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 3, characterized in that, The Ricker wavelet formula is: ； in, For amplitude, This is the dominant frequency of the wavelet.

5. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 2, characterized in that, In S1, the process of constructing a pure noise dataset, that is, the noise in the initial arrival portion of the actual DAS-VSP measured data, is as follows: Extracting the coupling noise component from real DAS-VSP data; The extracted noise is divided into blocks using sliding windows of different sizes, ultimately forming a pure noise dataset. Using the generated clean signal dataset and the pure noise dataset, construct the noisy data samples required for training.

6. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 2, characterized in that, In step S2, the overall structure adopts an encoder-decoder configuration, and the network structure is built based on the U-Net architecture. In the encoder section, the input seismic data undergoes two convolution operations to extract initial feature information, which is then input into the TransConv module for feature enhancement. The output of the TransConv module is used for subsequent downsampling and attention feature enhancement. The downsampling process uses max pooling to progressively reduce the spatial resolution of the feature map, enabling high-level features to aggregate a wider range of input information, thereby effectively improving the network's receptive field. The attention feature enhancement process includes channel attention mechanism and Monte Carlo guided attention mechanism, which are used to adaptively weight the feature maps; In the channel attention mechanism, the feature map is first convolved and then activated by the GELU activation function. Subsequently, channel-level statistical features are extracted using global average pooling. Two more convolution operations are then performed to generate corresponding channel attention weights, which are used to weight and adjust the feature responses of different channels. Following the channel attention mechanism, a Monte Carlo guided attention mechanism is introduced. During network training, multiple potential regions of interest in the feature map are randomly sampled and perturbed, and the sampling results are evaluated. In the Monte Carlo attention mechanism, the feature map is first convolved to reduce the number of channels to 1. The weights of each part of the feature map are extracted using the Monte Carlo method, and then the result is multiplied element-wise with the input feature map to obtain the final attention output.

7. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 6, characterized in that, In each stage of the encoder, the number of channels in the network remains constant at 64. Each convolution operation uses a 3×3 kernel and a stride of 1. In the decoder, the feature map is upsampled step by step through transposed convolution. The transposed convolution kernel is 2×2 and the stride is 2 to restore the spatial resolution of the feature map. The upsampled features are then concatenated with the features of the corresponding layer in the encoder. The TransConv module includes a Transformer submodule and a convolution submodule, and adopts a dual-branch structure. One branch inputs features into the Transformer submodule, and the other branch performs multiple convolution operations on the features. The outputs of the two branches are concatenated at the end of the module to achieve the fusion of global and local features.

8. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 2, characterized in that, In step S2, a seismic data denoising network is used as a supervised learning network, and the training process is carried out using known noisy DAS-VSP seismic data and corresponding reference seismic data. Assume the input noise data is Clean DAS data is The mean squared error (MSE) loss function is defined as follows: ； in, This refers to the neural network described in this invention. Represents training parameters, For weight parameters, This is the bias parameter.

9. The coupled noise denoising network method based on Monte Carlo guided attention mechanism as described in claim 2, characterized in that, In step S3, during the denoising process, the trained denoising network based on the Monte Carlo attention mechanism is used to perform forward prediction processing on the noisy 2D DAS-VSP data. The processing procedure is as follows: S31. Divide the two-dimensional DAS-VSP noisy data to be processed into multiple data blocks according to the preset window size and sliding step size. When the original data size cannot be divided by the window size, perform zero-filling processing in the data edge area. S32. Expand the data block in the channel dimension to meet the input channel requirements of the denoising network, and organize it into input batches according to the preset batch size, and input it into the trained denoising network for forward inference. S33. Obtain the denoising results corresponding to each data block, and reconstruct the denoised data blocks according to the spatial position and sliding step order of the data blocks in the partitioning stage. For data regions with overlap, use a weighted average method for fusion. If zero padding was performed in the data partitioning stage, remove the padding region in the data reconstruction stage to restore the original data size, and finally obtain a denoising result consistent with the original two-dimensional DAS-VSP data size.

Images