Seismic strong background noise removal method based on dual-channel network

Abstract

The invention discloses a seismic strong background noise removal method based on a two-channel network, and relates to the field of geophysical exploration data processing, and the method comprises the steps: selecting a noiseless signal and background noise when the noiseless signal and the background noise are not excited from a single-channel seismic record after a seismic source is excited, carrying out the superposition, generating noisy data, and carrying out the normalization processing; constructing a dual-channel constraint denoising network composed of a global constraint sub-network and a denoising sub-network; optimizing network parameters through the training sample set; and performing global constraint denoising and slice local denoising on the noisy seismic data by using the trained model, and finally splicing to obtain a complete denoising result. According to the method, the splicing effect problem caused by traditional blocking processing is effectively solved, the strong background noise removal effect is remarkably improved, and seamless high-quality denoising with the complete edge is achieved.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration data processing, and in particular to a method for removing strong seismic background noise based on a dual-channel network. Background Technology

[0002] Seismic exploration is an important method for obtaining information about subsurface geological structures by exciting seismic waves and acquiring ground vibration data, and it has wide applications in the field of geophysical exploration. During seismic data acquisition, geophones, as highly sensitive ground vibration monitoring devices, are constantly in operation, recording not only valid seismic signals but also inevitably capturing background noise caused by environmental or human factors. When the energy of local noise sources is high, the valid seismic signal is severely interfered with, leading to a decrease in the signal-to-noise ratio of the seismic data, thus affecting the accuracy of subsequent data processing and interpretation. Background noise typically exhibits complex, non-stationary, and structured characteristics, and some background noise is highly similar to the valid signal in terms of spectrum, amplitude, and phase, making traditional denoising methods such as convolution, wavelet transform, and median filtering difficult to effectively remove. Therefore, how to efficiently remove strong background noise has become a significant challenge in seismic data processing.

[0003] In recent years, deep learning technology has provided a new solution for seismic data denoising and has achieved good results to some extent. However, deep learning-based denoising methods still face many challenges. Due to the high resolution of seismic data, directly processing the complete data places extremely high demands on computational resources and easily leads to the loss of detailed information. Currently, two main methods are used to address this problem: one is to reduce the seismic data through linear interpolation before inputting it into the network for denoising, and then enlarge it back to the original size, but this method often results in the loss of detailed information; the other is to divide the seismic data into multiple smaller slices and denoise them separately, but this method is prone to edge effects when stitching together to restore the original data, resulting in obvious stitching marks in the denoising results and affecting the integrity and continuity of the data. In addition, existing methods often struggle to balance global structural constraints and local detail optimization when dealing with strong background noise, further limiting the improvement of denoising performance.

[0004] To address the aforementioned problems, there is an urgent need for a method that can balance the integrity of seismic signals with effective noise removal. This method should effectively remove strong background noise while resolving edge seams and signal distortion issues caused by traditional block processing, thereby achieving high-quality seismic data denoising. This invention aims to significantly improve the removal of strong background noise from seismic data by constructing a dual-channel constraint network, combining the advantages of a global constraint subnetwork and a dual-channel denoising subnetwork. This provides technical support for high-fidelity denoising of seismic data in complex geological environments. Summary of the Invention

[0005] This invention addresses the problems of poor removal of strong background noise in seismic data and splicing effects caused by traditional block processing in existing technologies. It proposes a seismic strong background noise removal method based on a dual-channel network. This method constructs and normalizes a training sample set, designs a dual-channel constrained denoising network (including a global constraint sub-network and a denoising sub-network), and combines it with an improved deep learning architecture to effectively remove strong background noise from seismic data while simultaneously solving the splicing effect problem caused by traditional block processing.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: S1: Preprocess the seismic data to obtain preprocessed seismic data; based on the acquired seismic data, select clean seismic data without background noise interference from the single-channel seismic records after source excitation; extract data with strong background noise from the unexcited single-channel seismic records; superimpose the clean seismic data and the strong background noise data according to their intensity to generate noisy seismic data; perform maximum-minimum normalization on the noisy seismic data and its corresponding clean seismic data respectively; interpolate the normalized noisy seismic data to a fixed length and use it as a global constraint input sample to construct a sample dataset; simultaneously, divide the normalized noisy seismic data and the clean seismic data into several slices of the same length to complete the preprocessing.

[0007] The construction of the sample dataset includes the following steps: S11: After firing, a continuous segment containing effective seismic waves, free from background noise interference, is selected from the single-channel seismic record. This segment, called the signal segment, is recorded as clean seismic data. Before firing, a segment containing only background noise and the same length as the signal segment is selected from the single-channel seismic record and recorded as strong background noise data. Using the above methods, N sets of clean seismic data and M sets of strong background noise data were obtained respectively. S12: Set clean seismic data Calculate the mean of their absolute values, denoted as . and ; Calculate the noise scaling factor A scaling factor α is set to adjust the noise intensity; then, a noise superposition operation is performed. Noisy seismic data were obtained. ; S13: Using noisy seismic data After normalizing the noisy and clean seismic data by applying maximum and minimum values, the data is interpolated to a fixed length to obtain the noisy seismic data. As a global constraint sample Normalized clean seismic data is used as the label; meanwhile, the normalized, uninterpolated noisy seismic data and the clean seismic data are each divided into several slices of equal length. and .

[0008] S2: Construct a dual-channel network model, which includes a global constraint sub-network and a dual-channel denoising sub-network. The global constraint sub-network is used to extract effective signal features from seismic data, and the dual-channel denoising sub-network is used to extract noise features from seismic data. S3: Input the preprocessed seismic data into the dual-channel network for training to obtain the trained dual-channel network model; Steps S2 and S3 are as follows: The constructed denoising network is trained using the dataset. The global constraint sub-network learns and trains on the global constraint input samples and outputs globally constrained denoised seismic data. The globally constrained denoised seismic data and the noisy seismic data slices are used as the dual-channel input data of the denoising sub-network. The denoising sub-network learns and trains on the dual-channel input data and outputs denoised seismic data slices. During the training process, the mean square error function is used as the training objective function. The loss is minimized by updating the parameters, and finally the constrained denoising network model is obtained.

[0009] S4: The trained dual-channel network model is used to process seismic data containing strong background noise, removing the strong background noise to obtain processed seismic data. Specifically, based on the trained constrained denoising network model, the noisy seismic data to be processed is normalized by its own maximum and minimum values ​​and then sliced. At the same time, the normalized noisy seismic data is interpolated and reduced in size, and then input into the trained global constrained sub-network, which outputs the preliminary denoising result, global denoised seismic data. Then, the global denoised seismic data and the noisy seismic data slices are combined into dual-channel input data, which is input into the trained denoising sub-network, outputting the denoised seismic data slices. Finally, all the denoised seismic data slices are spliced ​​together in their original order to obtain the final complete denoised seismic data.

[0010] The constrained denoising network model includes a global constraint sub-network and a dual-channel denoising sub-network, which are improved Restore-RWKV network architectures. The Restore-RWKV architecture is a 4-layer U-shaped encoder-decoder architecture, improved by introducing a Cross-ChannelMix (CCM) module. By processing the output feature maps of different encoder layers, features at different scales are fused. The specific structure is as follows: The global constraint network consists of a 4-layer cascaded encoder-decoder and incorporates a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 2, 3, 3, and 4 R-RWKV blocks, respectively. Each R-RWKV block includes a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module improves the Restore-RWKV network, processing the constrained residual data as output. Finally, the global constraints are input to the samples. The denoised seismic data is obtained by adding it to the output residual data; The denoising subnetwork consists of a 4-layer cascaded encoder-decoder and introduces a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 4, 6, 6, and 8 R-RWKV blocks, respectively. The R-RWKV block contains a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module is introduced to improve the Restore-RWKV network. After processing, the residual data is output. Finally, the dual-channel input data and the output residual data are added together to obtain the denoised seismic data slice.

[0011] The improvement of the Restore-RWKV network by introducing the Cross-ChannelMix (CCM) module includes the following steps: S21: Define the output characteristics of the designed encoder in stages I, II, and III as follows: These features have different spatial dimensions and number of channels: By using bilinear interpolation Upsampling to Size And use 1×1 convolution to unify the number of channels to the maximum number of feature channels. To obtain the aligned features ; S22: Concatenate the aligned three-scale features along the channel dimension. Converted into sequence mode through the Unfolding operation. This provides a unified input dimension for global channel mixing; S23: Yes After layer normalization, the channels are mixed through two fully connected layers, and the residual connections retain the original feature information. S24: The mixed sequence is restored to 2D features and split into three parts along the channel dimension. Each part is then subjected to 1×1 convolution and spatial size adjustment to restore the number of channels and spatial resolution of the original features. S25: The reconstructed multi-scale features are concatenated with the features of the corresponding layer of the decoder through skip connections to form a fused feature containing cross-channel context information, which is then input into the subsequent deconvolution layer for noise residual estimation.

[0012] The training method for the constrained denoising network model includes the following steps: S31: Input the global constraint input samples into the global constraint subnetwork for learning and training, and output globally denoised seismic data. ; S32: Combine the obtained globally denoised seismic data with the noisy seismic data slices to form a dual-channel input. The corresponding clean data slices are used as labels. A dual-channel input denoising subnetwork is trained and outputs denoised seismic data slices. ; S33: During training, the mean squared error function is used as the objective function for training a single network, specifically including: The MSE training objective function in the globally constrained subnetwork is defined as follows: in: This represents the clean seismic data after normalization in the sample pairs. This represents the output of the global constraint subnetwork for samples with global constraints. The denoising result is the globally denoised seismic data. It is the total number of input samples as a global constraint. This represents the loss value of the globally constrained subnetwork. The MSE training objective function in the denoising sub-network is defined as follows: in: This represents a slice corresponding to the location in the clean seismic data after normalization. This represents the dual-channel input sample formed by the noisy seismic data slice output by the denoising sub-network and the globally denoised seismic data. The denoising result is the denoised seismic data slice. It is the total number of dual-channel input samples for the denoising sub-network. This represents the loss value of the denoising subnetwork.

[0013] The beneficial effects of this invention are: 1. By employing a global constraint sub-network, the consistency and integrity of the denoising results in spatial distribution are ensured, avoiding the splicing effect caused by traditional block processing. An improved CCM module is introduced to enhance the extraction and fusion capabilities of multi-scale features, significantly improving the denoising effect. It effectively removes strong background noise while preserving detailed information of the seismic signal, achieving high-quality denoising with complete edges and no gaps. Furthermore, this invention is applicable to seismic data processing in complex geological environments, and performs particularly well in denoising high-resolution seismic data.

[0014] 2. This invention solves the seam problem caused by block processing in traditional methods by combining global constraints and local denoising. The global constraint sub-network learns from global input samples to generate denoising results with continuity and consistency, providing a reliable reference for subsequent local denoising. The denoising sub-network further processes the dual-channel input data to extract more refined feature information, thereby achieving efficient removal of strong background noise.

[0015] 3. This invention enhances the network's ability to extract and fuse multi-scale features by introducing an improved CCM module. This module achieves effective fusion of cross-channel contextual information through alignment, concatenation, and blending operations on output features from different stages of the encoder, thereby improving the network's ability to model complex background noise. The design of skip connections further enhances the network's ability to constrain details by concatenating the fused features with corresponding layer features from the decoder.

[0016] 4. This invention ensures the consistency of input data distribution through normalization and slicing of training data, thereby improving the model's generalization ability. The dual-channel input design combines global denoising results with local noise information, further enhancing the denoising effect. The training process uses the mean squared error function as the objective function and gradually approaches the optimal solution through the optimizer's parameter update strategy, ultimately achieving high-quality denoising results.

[0017] In summary, this invention achieves effective removal of strong background noise in seismic data by constructing a training sample set, designing a dual-channel constrained denoising network, introducing an improved CCM module, and optimizing the training strategy. At the same time, it solves the splicing effect problem caused by traditional block processing, and finally obtains high-quality denoising results with complete edges and no seams. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the specific implementation of the present invention; Figure 2 This is a diagram of the global constraint subnetwork structure in an embodiment of the present invention; Figure 3 This is a diagram of the denoising sub-network structure in an embodiment of the present invention; Figure 4 This is a structural diagram of the CCM module in an embodiment of the present invention; Figure 5 These are comparison diagrams of the global constraint denoising results in the embodiments of the present invention; (a) noisy data, (b) global constraint sub-network output, (c) clean data; Figure 6 These are comparison diagrams of the denoising results of the denoising sub-network in the embodiments of the present invention: (a) noisy data, (b) dual-channel denoising results, and (c) clean data.

[0019] In the picture: Input Projection: Input projection layer; R-RWKV Block: Contains the Re-WKV attention mechanism and the Omni-Shift layer; Down: downsampling layer; Encoder Projector: The encoder projection layer; Channel Mix: Channel mixing module; Split: a layer that divides a data segment. Decoder Projector: The decoder projection layer; Up: Upsampling layer; Output Projection: Output projection layer; Q-Shift: Q-shift operation; Rc, Kc: Parameter adjustment; Norm: Normalization layer; Element-wise addition; Sigmoid: Sigmoid activation function; Squared ReLU: The squared ReLU activation function. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0021] This invention provides a method for removing strong background noise from earthquakes based on a dual-channel network, combined with... Figure 1 To be continued Figure 6The specific embodiments of the present invention will be described in detail below. In practice, the present invention achieves effective removal of strong background noise in seismic data by constructing a training sample set, building an improved dual-channel constrained denoising network, and optimizing the training strategy, while solving the splicing effect problem caused by traditional block processing.

[0022] like Figure 1 As shown: This embodiment of the invention provides a method for denoising dual-channel seismic data, comprising the following steps: S1: Acquire seismic data and construct a training sample set. During seismic exploration, detectors record continuous seismic waveform data, which are stored separately for each detector channel, called single-channel seismic records. A single-channel seismic record contains all waveform information received by that channel. Based on the acquired seismic data, signal segments without background noise interference are selected from the single-channel seismic records after source excitation as clean seismic data; segments of the same length as the signal segments are extracted from the unexcited single-channel seismic records as strong background noise data. Each set of clean seismic data and the corresponding strong background noise are superimposed according to their intensity at a certain ratio to generate noisy seismic data. Maximum-minimum normalization is performed on both the noisy seismic data and its corresponding clean seismic data. The normalized noisy seismic data is interpolated to a fixed length and used as global constraint input samples to construct a sample dataset. Simultaneously, the normalized noisy seismic data and the clean seismic data are divided into several slices of the same length. Dataset construction methods include: S11: In seismic exploration, geophones acquire continuous seismic waveform data. This data is recorded separately for each geophone channel, called a single-channel seismic record. A single-channel seismic record contains all waveform information received on that channel. After shot firing, a continuous segment containing valid seismic waves, free from background noise interference, is selected from the single-channel seismic record; this segment is called a signal segment and is considered clean seismic data. Without shot firing, a segment of the same length as the signal segment, containing only background noise, is selected from the single-channel seismic record; this segment is considered strong background noise data. Through this method, N sets of clean seismic data and M sets of strong background noise data are obtained.

[0023] Clean earthquake data records are Strong background noise data is recorded as ; S12: Assuming clean seismic data Calculate the mean of their absolute values, denoted as . and Calculate the noise scaling factor. A scaling factor α is set to adjust the noise intensity. Perform noise superposition operation: Noisy seismic data were obtained. .

[0024] S13: Next, we will use noisy seismic data. After normalizing the noisy seismic data to its maximum and minimum values, the noisy seismic data is then interpolated to a fixed length of 256×1024 using nearest neighbor interpolation to obtain the noisy seismic data. As a global constraint sample Normalized clean seismic data was used as labels. Simultaneously, both the normalized noisy and clean seismic data were divided into several 256×256 slices. and .

[0025] S2: Construct a dual-channel constrained denoising network based on the improved Restore-RWKV architecture, including a global constraint sub-network and a denoising sub-network. Specifically, the dual-channel constrained denoising network includes a global constraint sub-network and a denoising sub-network, which are mainly based on the improved Restore-RWKV network architecture. The Restore-RWKV architecture is a 4-layer U-shaped encoder-decoder architecture, improved by introducing a Cross-ChannelMix (CCM) module, which fuses features at different scales by processing the output feature maps of different encoder levels. Specifically, the global constraint subnetwork structure is as follows: Figure 2 As shown, the global constraint network consists of a 4-layer cascaded encoder-decoder architecture and incorporates a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 2, 3, 3, and 4 R-RWKV blocks, respectively. Each R-RWKV block includes a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module improves the Restore-RWKV network, processing the constrained residual data as output. Finally, the global constraint input samples are used. The denoised seismic data is obtained by adding it to the output residual data; Global constraint subnetwork structure as follows Figure 3 The denoising subnetwork shown contains a 4-layer cascaded encoder-decoder and introduces a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 4, 6, 6, and 8 R-RWKV blocks respectively. The R-RWKV block contains a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module is introduced to improve the Restore-RWKV network. After processing, the residual data is output. Finally, the dual-channel input data and the output residual data are added to obtain the denoised seismic data slice.

[0026] Specifically, such as Figure 4 As shown, the CCM module structure includes: S21: Define the output characteristics of the designed encoder in stages I, II, and III as follows: These features have different spatial dimensions and number of channels: , By using bilinear interpolation Upsampling to Size And use 1×1 convolution to unify the number of channels to the maximum number of feature channels. To obtain the aligned features ; S22: Concatenate the aligned three-scale features along the channel dimension. Converted into sequence mode through Unfolding operation This provides a unified input dimension for global channel mixing; S23: Yes After layer normalization, the channels are mixed through two fully connected layers, and the residual connections retain the original feature information. S24: The mixed sequence is restored to 2D features. The seismic data is split into three parts along the channel dimension, and then subjected to 1×1 convolution and spatial size adjustment respectively to restore the original number of channels and spatial resolution of the features. S25: The reconstructed multi-scale features are concatenated with the features of the corresponding layer of the decoder through skip connections to form a fused feature containing cross-channel context information, which is then input into the subsequent deconvolution layer for noise residual estimation.

[0027] S3. Input the sample data from the sample dataset into the constructed dual-channel denoising network for training. The global constraint sub-network learns and trains on the globally constrained input samples, outputting globally constrained denoised seismic data. The globally constrained denoised seismic data is interpolated, enlarged, and sliced. These slices, along with noisy seismic data of the same size, are used as the dual-channel input data for the denoising sub-network. The denoising sub-network learns and trains on the dual-channel input data and outputs denoised seismic data slices. During model training, the mean squared error function is used as the training objective function. By updating the parameters and minimizing the loss, the final model is obtained. Specifically, it includes: S31: Input the global constraint input samples into the global constraint subnetwork for learning and training, and output globally denoised seismic data. The globally denoised seismic data was then enlarged and sliced. Specifically, linear interpolation was used to enlarge the data back to its original size, and then slicing it into 256×256 pixels with a step size of 256 to obtain slices of the globally denoised seismic data. ; S32: Combine the obtained globally denoised seismic data slices with the noisy seismic data slices to form a dual-channel input. The corresponding clean data slices are used as labels. A dual-channel input denoising subnetwork is trained and outputs denoised seismic data slices. ; S33: During training, the mean squared error function is used as the objective function for training a single network, specifically including: The MSE training objective function in the globally constrained subnetwork is defined as follows: in: This represents the clean seismic data after normalization in the sample pairs. This represents the output of the global constraint subnetwork for samples with global constraints. The denoising result is the globally denoised seismic data. It is the total number of input samples as a global constraint. This represents the loss value of the global constraint subnetwork.

[0028] The MSE training objective function in the denoising sub-network is defined as follows: in: This represents a slice corresponding to the location in the clean seismic data after normalization. This represents the dual-channel input sample formed by the noisy seismic data slice output by the denoising sub-network and the globally denoised seismic data. The denoising result is the denoised seismic data slice. It is the total number of dual-channel input samples for the denoising sub-network. This represents the loss value of the denoising subnetwork. During network training, the number of training epochs was set to 100, the batch size was set to 1, the learning rate was set to 0.0001, the Adam optimizer was selected, and the best network model was obtained through iterative training and then output.

[0029] S4. Based on the trained dual-channel constrained denoising network model, the noisy seismic data to be processed is normalized by its own maximum and minimum values ​​and then sliced. Simultaneously, the normalized noisy seismic data is reduced by nearest neighbor interpolation. This is then input into the trained global constrained sub-network, and the preliminary denoising result, globally denoised seismic data, is output. The specific denoising effect is as follows: Figure 5As shown, the global denoised seismic data is then interpolated, enlarged, and sliced ​​using a linear interpolation method. This slice, along with the noisy seismic data slices, forms a dual-channel input data set. This input is then fed into the trained denoising sub-network, which outputs denoised seismic data slices. Finally, all denoised seismic data slices are stitched together in their original order to obtain the final complete denoised seismic data. The specific denoising effect is shown in the figure. Figure 6 As shown.

[0030] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for removing strong background noise from earthquakes based on a dual-channel network, characterized in that, Includes the following steps: S1: Preprocess the seismic data to obtain preprocessed seismic data; S2: Construct a dual-channel network model, which includes a global constraint sub-network and a dual-channel denoising sub-network. The global constraint sub-network is used to extract effective signal features from seismic data, and the dual-channel denoising sub-network is used to extract noise features from seismic data. S3: Input the preprocessed seismic data into the dual-channel network for training to obtain the trained dual-channel network model; S4: The trained dual-channel network model is used to process the seismic data containing strong background noise, remove the strong background noise, and obtain the processed seismic data.

2. The earthquake background noise removal method based on a dual-channel network according to claim 1, characterized in that: The preprocessing of earthquake data is based on the acquired earthquake data, and clean earthquake data without background noise interference is selected from the single-channel earthquake records after the source excitation. Strong background noise data is extracted from unexcited single-channel seismic records. Clean seismic data and strong background noise data are superimposed proportionally according to their intensity to generate noisy seismic data. Maximum-minimum normalization is performed on both the noisy seismic data and their corresponding clean seismic data. The normalized noisy seismic data is interpolated to a fixed length and used as a global constraint input sample to construct a sample dataset. At the same time, the normalized noisy seismic data and clean seismic data are divided into several slices of the same length to complete the preprocessing.

3. The earthquake background noise removal method based on a dual-channel network according to claim 2, characterized in that: Steps S2 and S3 are specifically as follows: the constructed denoising network is trained using the dataset, the global constraint sub-network learns and trains on the global constraint input samples, and outputs globally constrained denoised seismic data. The globally constrained denoised seismic data and the noisy seismic data slices are used as the dual-channel input data of the denoising sub-network. The denoising sub-network learns and trains on the dual-channel input data and outputs denoised seismic data slices. During the training process, the mean square error function is used as the training objective function. The loss is minimized by updating the parameters, and finally the constrained denoising network model is obtained.

4. The earthquake background noise removal method based on a dual-channel network according to claim 3, characterized in that: Step S4 specifically involves: based on the trained constrained denoising network model, the noisy seismic data to be processed is normalized by its own maximum and minimum values ​​and then sliced. At the same time, the normalized noisy seismic data is interpolated and reduced, and then input into the trained global constrained sub-network to output the preliminary denoising result, global denoised seismic data. Then, the global denoised seismic data and the noisy seismic data slices are combined into dual-channel input data, which is input into the trained denoising sub-network to output the denoised seismic data slices. Finally, all the denoised seismic data slices are spliced ​​together in their original order to obtain the final complete denoised seismic data.

5. The earthquake background noise removal method based on a dual-channel network according to claim 2, characterized in that: The construction of the sample dataset includes the following steps: S11: After firing, a continuous segment containing effective seismic waves, free from background noise interference, is selected from the single-channel seismic record. This segment, called the signal segment, is recorded as clean seismic data. Before firing, a segment containing only background noise and the same length as the signal segment is selected from the single-channel seismic record and recorded as strong background noise data. Using the above methods, N sets of clean seismic data and M sets of strong background noise data were obtained respectively. S12: Set clean seismic data Calculate the mean of their absolute values, denoted as . and ; Calculate the noise scaling factor A scaling factor α is set to adjust the noise intensity; then, a noise superposition operation is performed. Noisy seismic data were obtained. ; S13: Using noisy seismic data After normalizing the noisy and clean seismic data by applying maximum and minimum values, the data is interpolated to a fixed length to obtain the noisy seismic data. As a global constraint sample Normalized clean seismic data is used as the label; meanwhile, the normalized, uninterpolated noisy seismic data and the clean seismic data are each divided into several slices of equal length. and .

6. The earthquake background noise removal method based on a dual-channel network according to claim 4, characterized in that: The constrained denoising network model includes a global constraint sub-network and a dual-channel denoising sub-network, which are improved Restore-RWKV network architectures. The Restore-RWKV architecture is a 4-layer U-shaped encoder-decoder architecture, improved by introducing a Cross-ChannelMix (CCM) module. By processing the output feature maps of different encoder layers, features at different scales are fused. The specific structure is as follows: The global constraint network consists of a 4-layer cascaded encoder-decoder and incorporates a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 2, 3, 3, and 4 R-RWKV blocks, respectively. Each R-RWKV block includes a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module improves the Restore-RWKV network, processing the constrained residual data as output. Finally, the global constraints are input to the samples. The denoised seismic data is obtained by adding it to the output residual data; The denoising subnetwork consists of a 4-layer cascaded encoder-decoder and introduces a Cross-ChannelMix (CCM) module. Each layer of the encoder-decoder contains 4, 6, 6, and 8 R-RWKV blocks, respectively. The R-RWKV block contains a Re-WKV attention mechanism and an Omni-Shift layer. The Cross-ChannelMix (CCM) module is introduced to improve the Restore-RWKV network. After processing, the residual data is output. Finally, the dual-channel input data and the output residual data are added together to obtain the denoised seismic data slice.

7. The earthquake background noise removal method based on a dual-channel network according to claim 6, characterized in that: The improvement of the Restore-RWKV network by introducing the Cross-ChannelMix (CCM) module includes the following steps: S21: Define the output characteristics of the designed encoder in stages I, II, and III as follows: These features have different spatial dimensions and number of channels: , By using bilinear interpolation Upsampling to Size And use 1×1 convolution to unify the number of channels to the maximum number of feature channels. Obtain the aligned features ; S22: Concatenate the aligned three-scale features along the channel dimension. Converted into sequence mode through the Unfolding operation. This provides a unified input dimension for global channel mixing; S23: Yes After layer normalization, the channels are mixed through two fully connected layers, and the residual connections retain the original feature information. S24: The mixed sequence is restored to 2D features and split into three parts along the channel dimension. Each part is then subjected to 1×1 convolution and spatial size adjustment to restore the number of channels and spatial resolution of the original features. S25: The reconstructed multi-scale features are concatenated with the features of the corresponding layer of the decoder through skip connections to form a fused feature containing cross-channel context information, which is then input into the subsequent deconvolution layer for noise residual estimation.

8. The earthquake background noise removal method based on a dual-channel network according to claim 6, characterized in that: The training method for the constrained denoising network model includes the following steps: S31: Input the global constraint input samples into the global constraint subnetwork for learning and training, and output globally denoised seismic data. ; S32: Combine the obtained globally denoised seismic data with the noisy seismic data slices to form a dual-channel input. The corresponding clean data slices are used as labels. A dual-channel input denoising subnetwork is trained and outputs denoised seismic data slices. ; S33: During training, the mean squared error function is used as the objective function for training a single network, specifically including: The MSE training objective function in the globally constrained subnetwork is defined as follows: in: This represents the clean seismic data after normalization in the sample pairs. This represents the output of the global constraint subnetwork for samples with global constraints. The denoising result is the globally denoised seismic data. It is the total number of input samples as a global constraint. This represents the loss value of the globally constrained subnetwork. The MSE training objective function in the denoising sub-network is defined as follows: in: This represents a slice corresponding to the location in the clean seismic data after normalization. This represents the dual-channel input sample formed by the noisy seismic data slice output by the denoising sub-network and the globally denoised seismic data. The denoising result is the denoised seismic data slice. It is the total number of dual-channel input samples for the denoising sub-network. This represents the loss value of the denoising subnetwork.

Images