Method and system for suppressing seismic data noise based on deep convolutional neural network

Abstract

The invention provides a method and system for suppressing seismic data noise based on a deep convolutional neural network, and the method comprises the steps: dividing a training set and a test set based on a set seismic data sample, and constructing a denoising convolutional neural network structure which comprises an input layer, a hidden layer and an output layer; further introducing an attention mechanism and a cavity convolution module to form an ADCNN network structure comprising a plurality of convolution layers; training and verifying the formed ADCNN network structure model through continuous iteration based on the training set until the training error of the model reaches a set requirement, and storing the corresponding ADCNN network structure model; and further performing trial verification on the stored ADCNN network structure model by using the test set data. By adopting the scheme, the problems of limited operation conditions and complicated operation process in the prior art can be solved, and high-efficiency and high-quality seismic data noise suppression processing is realized based on simple operation.

Description

Technical Field

[0001] This invention relates to the field of data noise reduction and optimization technology, and in particular to a method and system for suppressing noise in seismic data based on a deep convolutional neural network. Background Technology

[0002] Seismic data plays a crucial role in geological research and engineering. However, geological data often contains significant random noise, which severely impacts subsequent data processing, interpretation, and reservoir prediction. Suppressing random noise is a critical step in pre-stack seismic data processing; however, because random noise is formed by the combined effects of various factors, is distributed throughout the entire time and frequency band, and lacks a fixed frequency and propagation direction, it is difficult to effectively separate it from seismic records.

[0003] To address the problem of random noise suppression, scholars have conducted relevant research and proposed different suppression methods based on the characteristics of effective signals and random noise. Typical methods include KL transform filtering, adaptive filtering, Radial transform, radial prediction filtering, wavelet transform, singular value decomposition, polynomial fitting, and S-transform. While these methods have achieved certain practical application results, they are all subject to certain assumptions or limitations, resulting in significant variations in denoising performance across different work areas. Their adaptability and fidelity need further improvement. Furthermore, methods with better denoising effects typically have higher computational complexity and lower computational efficiency, requiring extensive parameter tuning and failing to adequately meet the challenges of large-scale data processing in the era of broadband, wide-azimuth, and high-density single-point acquisition. Therefore, researching fidelity-preserving and efficient denoising methods remains a relentless pursuit.

[0004] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] To address the above problems, this invention provides a method for suppressing seismic data noise based on a deep convolutional neural network. In one embodiment, the method includes:

[0006] Step S10: Divide the earthquake data samples into training and testing sets to use as input and output data to build a network model; the earthquake data samples include linear in-phase axis synthetic earthquake data and curved in-phase axis synthetic earthquake data of a set size;

[0007] Step S20: Construct a denoising convolutional neural network structure including an input layer, hidden layers, and an output layer;

[0008] Step S30: Based on the denoising convolutional neural network structure, an attention mechanism and a dilated convolution module are introduced to form an ADCNN network structure containing multiple convolutional layers;

[0009] Step S40: Train and validate the ADCNN network structure model by iteratively training it based on the training set until the training error of the model reaches the set requirement, and save the corresponding ADCNN network structure model.

[0010] Step S50: Use the test set as input and output data to perform trial calculations and verification on the saved ADCNN network structure model.

[0011] In an optional embodiment, in step S10, the seismic data sample is subjected to seismic record forward modeling using Ricker wavelet data that conforms to the dominant frequency of the actual data, and random noise that meets the set requirements is added.

[0012] Furthermore, in one embodiment, in step S10, the linear and bending in-phase axis synthetic seismic records in the test set are set to a 1:1 distribution.

[0013] Preferably, in one embodiment, the method further includes step S11: before constructing the denoising convolutional neural network structure, normalizing the training sample data to make the input and output data distribution of the subsequent BN operation more reasonable and improve the convergence effect of the network model.

[0014] Optionally, in one embodiment, in step S20, in the DnCNN network structure, the input layer sets parameters for the convolution kernel to perform inner product operations to obtain input features, fills edges with zeros, and uses ReLU as the activation function; the hidden layer sets parameters for the convolution kernel to avoid covariate shifting, fills edges with zeros, and then performs BN operations, using ReLU as the activation function; the output layer sets parameters for the convolution kernel to reconstruct residual data, and does not perform BN operations to prevent errors in the results.

[0015] In one embodiment, in step S30, a dilation parameter is introduced into the convolution operation in the neural network to characterize the distance between data points when performing convolution kernel operations.

[0016] Furthermore, in one embodiment, in step S30, a dilated convolution module is introduced, the downsampling operation is eliminated, and dilated convolution is achieved by zero padding and increasing the dilation coefficient; the ADCNN network structure contains 10 convolutional layers to avoid setting too many dilated convolutional layers and destroying the continuous features of the data.

[0017] Preferably, in one embodiment, the relationship between the dilated convolution kernel parameters and the dilation coefficient is as follows:

[0018] kd size=(γ-1)(k) size -1)+k size

[0019] In the formula, k size Where γ is the kernel size, kd is the dilation coefficient, and γ is the kernel size. size This represents the size of the dilated convolution kernel.

[0020] In an optional embodiment, an attention mechanism is introduced, and the number of channels in the convolution kernel is set to 1, so that the feature information is effectively compressed, and then the effective features learned are enhanced through multiplication operations.

[0021] Furthermore, in one embodiment, step S50 includes verifying the network learning effect using linear in-phase axis synthetic seismic data and curved in-phase axis synthetic seismic data from the test set, and performing residual analysis on the denoising results.

[0022] Optionally, in one embodiment, step S50 further includes selecting single-channel data from the model for analysis to form a single-channel time-domain comparison display diagram; and using multiple noise evaluation indicators to perform numerical analysis on the noise suppression effect, intuitively verifying the denoising performance.

[0023] Based on the application aspects of the methods described in any one or more of the above embodiments, the present invention also provides a system for suppressing seismic data noise based on a deep convolutional neural network, which performs the methods described in any one or more of the above embodiments.

[0024] Compared with the closest prior art, the present invention also has the following beneficial effects:

[0025] This invention provides a method and system for suppressing seismic data noise based on a deep convolutional neural network (DCNN). The method divides seismic data samples into training and testing sets, constructs a denoising DCNN structure including an input layer, hidden layers, and an output layer, and then introduces an attention mechanism and a dilated convolution module to form an ADCNN network structure with multiple convolutional layers. The ADCNN network structure model is trained and validated iteratively based on the training set until the training error reaches a set requirement, at which point the corresponding ADCNN network structure model is saved. This approach, by introducing an attention mechanism module, focuses the rich seismic information contained in the input data onto the key information of random noise, thereby reducing the network's attention to other information and allowing it to focus more on useful features, improving learning efficiency. The application of dilated convolution expands the receptive field; by changing the dilation rate, the range of the receptive field can be altered, thereby capturing multi-scale information. Based on simple computation, this method achieves efficient and high-quality seismic data noise suppression, offering better timeliness and practicality.

[0026] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0028] Figure 1 This is a flowchart illustrating a method for suppressing seismic data noise based on a deep convolutional neural network, provided in an embodiment of the present invention.

[0029] Figure 2 This is an example diagram of random noise training samples for the method of suppressing seismic data noise provided in the embodiments of the present invention;

[0030] Figure 3 This is a schematic diagram illustrating the DnCNN network denoising process of the method for suppressing seismic data noise provided in this embodiment of the invention.

[0031] Figure 4 This is a schematic diagram illustrating the ADCNN network denoising process of the method for suppressing seismic data noise provided in this embodiment of the invention.

[0032] Figure 5 This is a schematic diagram of the holed convolution method for suppressing seismic data noise provided in the embodiments of the present invention;

[0033] Figure 6 This is an example diagram illustrating the denoising process of linear phase axis synthesized seismic data in the method for suppressing seismic data noise provided in this embodiment of the invention;

[0034] Figure 7 This is an example diagram of the denoising process of synthesized seismic data with bent phase axes in the method for suppressing seismic data noise provided in the embodiments of the present invention;

[0035] Figure 8 This is an example diagram illustrating the denoising process of complex model-synthesized seismic data in the method for suppressing seismic data noise provided in this embodiment of the invention;

[0036] Figure 9 This is a single-channel time-domain schematic diagram of the test data verification results in the method for suppressing seismic data noise provided in the embodiments of the present invention;

[0037] Figure 10 This is an enlarged display example of the actual pre-stack gather denoising results in the method for suppressing seismic data noise provided in the embodiments of the present invention. Detailed Implementation

[0038] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. Those skilled in the art will then fully understand how the present invention uses technical means to solve technical problems and achieve technical effects, and will be able to implement the present invention specifically based on the above-described implementation process. It should be noted that, as long as there is no conflict, the various embodiments and features of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0039] Although the flowchart describes the operations as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. The order of the operations can be rearranged. A process can terminate when its operation is complete, but it may also have additional steps not included in the diagram. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0040] Computer equipment includes user equipment and network equipment. User equipment or clients include, but are not limited to, computers, smartphones, and PDAs (Personal Digital Assistants); network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Computer equipment can operate independently to implement this invention, or it can connect to a network and implement this invention through interaction with other computer devices within the network. The network in which the computer equipment resides includes, but is not limited to, the Internet, wide area networks (WANs), metropolitan area networks (MANs), local area networks (LANs), and VPN networks.

[0041] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0042] Seismic data plays a crucial role in geological research, but it often contains significant random noise, which negatively impacts subsequent data processing, interpretation, and reservoir prediction. Suppressing random noise is a critical step in pre-stack seismic data processing; however, because random noise is formed by the combined effects of various factors, is distributed across all times and frequencies, and lacks a fixed frequency and propagation direction, it is difficult to effectively separate it from seismic records.

[0043] To address the problem of random noise suppression, scholars have conducted relevant research and proposed different suppression methods based on the characteristics of effective signals and random noise. Typical methods include KL transform filtering, adaptive filtering, Radial transform, radial prediction filtering, wavelet transform, singular value decomposition, polynomial fitting, and S-transform. While these methods have achieved certain practical application results, they are all subject to certain assumptions or limitations, resulting in significant variations in denoising performance across different work areas. Their adaptability and fidelity need further improvement. Furthermore, methods with better denoising effects typically have higher computational complexity and lower computational efficiency, requiring extensive parameter tuning and failing to adequately meet the challenges of large-scale data processing in the era of broadband, wide-azimuth, and high-density single-point acquisition. Therefore, researching fidelity-preserving and efficient denoising methods remains a relentless pursuit.

[0044] Deep learning technology, a branch of machine learning, has attracted widespread attention in recent years due to its application and development. Conventional deep convolutional neural networks suffer from limitations in the receptive field of feature maps due to the convolutional kernel, resulting in insufficient ability to capture information from adjacent seismic traces when used for seismic data denoising. To address this, this invention proposes a method for suppressing random noise in pre-stack seismic signals using a deep convolutional neural network that combines dilated convolution and an attention mechanism. The attention mechanism module focuses the rich seismic information contained in the input data onto the key information of random noise, thereby reducing the network's attention to other information, i.e., focusing on useful features and improving network learning efficiency. Dilated convolution expands the receptive field by changing the dilation rate, achieving the effect of capturing multi-scale information. Practical applications show that this method outperforms traditional methods such as fx filtering and curvelet filtering, making it a high-fidelity and efficient denoising method.

[0045] To address the aforementioned problems, this invention provides a method and system for suppressing seismic data noise based on a deep convolutional neural network (ADCNN). Specifically targeting the intelligent and efficient suppression of random noise in pre-stack seismic records, this invention first proposes a method for suppressing pre-stack seismic signal noise using a deep convolutional neural network (ADCNN) that combines dilated convolution and attention mechanisms. Comparative analysis using simple and complex theoretical models, as well as actual pre-stack data processing, shows that this method outperforms traditional methods such as fx filtering and curvelet filtering, demonstrating strong effectiveness and practicality. This scheme, as an effective method for suppressing pre-stack seismic noise, can lay the foundation for refined tectonic interpretation and high-precision post-stack inversion in the study area.

[0046] The following describes the detailed flow of the method according to an embodiment of the present invention with reference to the accompanying drawings, the steps of which can be executed in a computer system containing, for example, a set of computer-executable instructions. Although the logical order of the steps is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0047] Example 1:

[0048] This invention aims to focus on the "amplitude preservation, frequency preservation, and AVO preservation" characteristics of seismic data, and to use deep learning for pre-stack gather denoising to improve the quality of pre-stack gathers, thus laying a data foundation for pre-stack reservoir inversion and prediction research.

[0049] Traditional denoising neural networks (DnCNN) suffer from limited receptive fields in feature maps when processing pre-stack seismic data noise due to the limitations of convolutional kernels, failing to effectively capture information from adjacent seismic traces. This invention addresses this issue by proposing a Deep Convolutional Neural Network (ADCNN) combining dilated convolution and attention mechanisms to suppress random noise in pre-stack seismic signals. The attention mechanism in this method focuses the rich seismic information contained in the input data onto the crucial random noise, reducing the network's focus on other information and allowing it to concentrate on useful features, thus improving learning efficiency. Dilated convolution expands the receptive field; by changing the dilation rate, the receptive field's extent can be altered, capturing multi-scale information. Comparative analysis using simple theoretical models, complex theoretical models, and actual pre-stack data processing demonstrates that this method significantly outperforms traditional methods such as fx filtering and curvelet filtering in suppressing noise.

[0050] Figure 1 This diagram illustrates a flowchart of a method for suppressing seismic data noise based on a deep convolutional neural network, as provided in Embodiment 1 of the present invention. (Refer to...) Figure 1 As can be seen, the method includes the following steps.

[0051] Step S10: Divide the earthquake data samples into training and testing sets to use as input and output data to build a network model; the earthquake data samples include linear in-phase axis synthetic earthquake data and curved in-phase axis synthetic earthquake data of a set size;

[0052] Step S20: Construct a denoising convolutional neural network structure including an input layer, hidden layers, and an output layer;

[0053] Step S30: Based on the denoising convolutional neural network structure, an attention mechanism and a dilated convolution module are introduced to form an ADCNN network structure containing multiple convolutional layers;

[0054] Step S40: Train and validate the ADCNN network structure model by iteratively training the model based on the training set until the training error of the model reaches the set requirements. Save the corresponding ADCNN network structure model for noise suppression calculation of the required seismic data.

[0055] Step S50: Use the test set as input and output data to perform trial calculations and verification on the saved ADCNN network structure model.

[0056] This invention utilizes an ADCNN network constructed with an attention mechanism and dilated convolution modules for denoising pre-stack seismic data, resulting in a more optimized network structure and higher training efficiency. The network model built using this invention can extract complex noise features more effectively. Actual denoising results demonstrate that this technique significantly suppresses noise in pre-stack seismic data, outperforming traditional methods such as fx filtering and curvelet filtering.

[0057] Preferably, in one embodiment, the seismic data sample is obtained by performing forward modeling of seismic records using Ricker wavelet data that conforms to the dominant frequency of actual data, and by adding random noise that meets set requirements.

[0058] In one alternative embodiment, the linear and bending in-phase composite seismic records in the test set are distributed in a 1:1 ratio.

[0059] In practical applications, Ricker wavelet data with the dominant frequency matching the actual data was used to perform forward modeling of seismic records for different models, with 25% random noise added. A total of 800 records were compiled, including 400 linear phase axis synthetic seismic data and 400 curved phase axis synthetic seismic data, forming the training samples and training labels, with the training and test sets set at a 720:80 ratio.

[0060] Based on the above steps, earthquake data is obtained through forward modeling, and then noise is added to obtain noisy earthquake data. These data are used to form training samples and training labels, which serve as the input and output values ​​for building the network.

[0061] Specifically, Figure 2 A partial random noise training sample image is shown according to an embodiment of the present invention. Figure 2As shown, seismic data and noise-added data were obtained through forward modeling to prepare training samples, with the Ricker wavelet set as the dominant frequency of 25Hz. The seismic records used for training included 400 linear in-phase axis synthetic seismic data and 400 curved in-phase axis synthetic seismic data, totaling 800 records. Each synthetic record had 80 traces, a duration of 2000ms, and a sampling interval of 4ms. Random noise was then added to these synthetic records, with a parameter set to 25%. These data were used to form training samples and training labels. The training and test sets were then set at a ratio of 720:80, with the linear and curved in-phase axis synthetic seismic records in the test set distributed in a 1:1 ratio. Some training samples are shown below. Figure 2 As shown, Figure 2 a, Figure 2 b represents clean synthetic seismic data. Figure 2 c. Figure 2 d is Figure 2 a, Figure 2 b corresponds to the noisy synthetic seismic data.

[0062] The method further includes normalizing the training sample data before constructing the denoising convolutional neural network structure, and then determining the relevant parameters of the input layer, hidden layer, and output layer of the network. This makes the input and output data distribution of the subsequent Batch Normalization (BN) operation more reasonable, thereby improving the convergence performance of the network model. Normalizing the input data improves the numerical distribution and ensures good network convergence.

[0063] Specifically, Figure 3 A flowchart of the DnCNN network denoising process according to an embodiment of the present invention is shown. Figure 3 As shown, the input seismic data needs to be preprocessed. The first step is to normalize the data so that the numerical distribution of the input after the BN operation is more reasonable, thereby improving the convergence effect of the network.

[0064] The data used for training and testing must be normalized to ensure consistent data standards throughout the denoising process and guarantee network accuracy.

[0065] Further, step S20 is performed to construct a denoising convolutional neural network structure including an input layer, a hidden layer, and an output layer.

[0066] First, inner product operations are performed using 64 convolutional kernels with ReLU activation function. Second, hidden layers are configured with parameters to avoid covariate shifts, and Batch Normalization (BN) and ReLU activation are performed. Finally, convolutional kernels are used to reconstruct the residual data output layer.

[0067] Therefore, preferably, in one embodiment, a DnCNN network structure is constructed, including an input layer, 15 hidden layers, and an output layer. The DnCNN network structure has 17 layers, including an input layer, 15 hidden layers, and an output layer. The input layer sets the parameters of the convolution kernel to perform inner product operations to obtain input features, pads edges with zeros, and uses ReLU as the activation function. The hidden layers set the parameters of the convolution kernel to avoid covariate shift, pad edges with zeros, and then perform Batch Normalization (BN) operations, using ReLU as the activation function. The output layer sets the parameters of the convolution kernel to reconstruct residual data, but does not perform BN operations to prevent errors in the results.

[0068] After training, the residuals (i.e., the learned random noise) are subtracted from the original seismic data to obtain the denoising result.

[0069] Furthermore, through step S30, an attention mechanism and a dilated convolution module are introduced on the basis of the convolutional neural network DnCNN to form an ADCNN network structure containing 10 convolutional layers. This achieves the purpose of denoising by learning random noise features through multiplication operations and residuals, including:

[0070] Step S31: Introduce attention mechanism; Spatial attention mechanism is introduced by using convolution and multiplication, and it is integrated into DnCNN to make feature extraction of complex noise networks more focused, network structure more optimized, and training efficiency higher. It can effectively deal with the problem of complex noise extraction and improve the efficiency of network feature extraction.

[0071] Step S32: Introduce dilated convolution; Dilated convolution is a computational method that uses a dilation rate parameter to represent the data distance between convolution kernels. Compared to traditional convolution, dilated convolution eliminates downsampling while preserving information integrity. Dilated convolution is implemented using zero padding and an increased dilation coefficient.

[0072] Figure 4 A flowchart of the ADCNN network denoising process according to an embodiment of the present invention is shown. Figure 4 As shown, the ADCNN network structure used contains 10 convolutional layers, specifically as follows: Figure 5 As shown, since setting too many dilated convolutional layers would destroy the continuous features of the data, 10 convolutional layers are used. Dilated convolution is represented by DilatedConv. The first layer's dilated convolution kernel parameters (3×3×1×16,1) specifically mean: 3×3 is the kernel size, the number of channels is 1, the number of kernels is set to 16, and the 1s separated by commas are the dilation factor. The meaning of the convolution parameters for each layer in the following diagram is the same. The second layer's dilated convolution kernel parameters are (3×3×16×32,2), and so on. The fifth layer's convolution kernel is set to 1, i.e., (3×3×128×1,5), and then from the sixth layer onwards, it is symmetrical to the previous ones.

[0073] Up to the 10th layer, as shown in the dashed box in the figure, an attention mechanism module is introduced. To enable the network to better learn random noise features, this module operates as follows: First, the number of channels in the convolutional kernel is set to 1, allowing for effective compression of feature information; then, the result is multiplied by the output of the previous layer to enhance the effective features learned; finally, the learned random noise is subtracted from the original seismic data to obtain the denoising result. (The symbols in the figure are...) Defined as multiplication.

[0074] Specifically, Figure 5 A schematic diagram of dilated convolution according to an embodiment of the present invention is shown. Figure 5 As shown, the convolution operation in neural networks introduces a dilation parameter to characterize the distance between data points during the convolution kernel operation. The advantage of dilated convolution is that it eliminates downsampling while preserving information integrity. Dilated convolution is implemented by padding the kernel with zeros during the convolution kernel operation, thus significantly increasing the dilation coefficient, as shown below. Figure 4 Optionally, the relationship between the dilated convolution kernel parameters and the dilation coefficient is as follows: (1)

[0075] kd size =(γ-1)(k) size -1)+k size

[0076] In the formula, k size Where γ is the kernel size, kd is the dilation coefficient, and γ is the kernel size. size This refers to the size of the dilated convolution kernel; when γ is 1, kd size Same as a regular convolution kernel.

[0077] Based on this, step S40 is executed, which involves iteratively training and validating the ADCNN network structure model based on the training set until the training error of the model reaches the set requirement, and then saving the corresponding ADCNN network structure model. In step S40, the trained ADCNN model is saved, and through continuous iteration, the training model is made to meet the error requirements.

[0078] When training a model, the training error generally decreases with each iteration based on the training samples. By setting a value for the error, such as using a loss function to represent the training error, the loss function usually decreases with each iteration during training. In practical applications, the value of the loss function can be set. Optionally, the training can be stopped when the loss function is less than a first set value (such as 0.1%), and the trained model can be saved.

[0079] Optionally, in one embodiment, the loss function of the deep learning neural network constructed in this invention uses mean squared error (MSE), which calculates the average of the squares of the differences between the predicted and true values. In this operation, training is stopped when the model's training error reaches a satisfactory range through continuous iteration. The model that meets the error requirement is saved for subsequent suppression of pre-stack noise.

[0080] Furthermore, in step S50, the saved ADCNN network structure model is tested and verified using the test set as input and output data.

[0081] When using the test set to test and validate the model, a loss function is also used to characterize the training error. During the test, the value of the current loss function is compared and analyzed with the value of the set loss function. Optionally, it is set that when the loss function is less than a second set value (such as 0.1%), it indicates that the model has passed the validation. The second set value and the first set value can be flexibly set according to the needs, and the values ​​can be different.

[0082] Step S50 includes verifying the network learning effect using linear in-phase axis synthetic seismic data and curved in-phase axis synthetic seismic data from the test set, and performing residual analysis on the denoising results.

[0083] Furthermore, embodiments of the present invention also use complex synthesis models and actual pre-stack data of the study area for denoising verification.

[0084] Step S50 also includes selecting the open-source complex synthetic seismic model from the SEG website as the training dataset and verifying the network learning effect.

[0085] Step S50 also includes selecting single-channel data from the model for analysis to form a single-channel time-domain comparison display diagram; and using multiple noise evaluation indicators to perform numerical analysis on the noise suppression effect, intuitively verifying the denoising performance.

[0086] Specifically, in practical applications, denoising verification is performed using theoretical models and actual pre-stack data of the study area. Furthermore, step S50 may also include steps S501, S502, S503, S504, and S505.

[0087] In step S501, the learning effect of the network is tested using linear in-phase axis synthesized seismic data, and a comparative analysis is performed with traditional methods such as fx filtering and curvelet transform. Figure 6 As shown. Figure 6 This is a denoised image of linear in-phase axis synthesized seismic data according to an embodiment of the present invention. Figure 6 'a' represents a noise-free model. Figure 6 b-noise-adding model, Figure 6 c represents the denoising result of fx filtering. It can be seen that the added random noise residue is quite severe. Figure 6 d shows the result of curve wave denoising, indicating that the noise removal is relatively clean, but some noise remains. Figure 6 e represents the ADCNN denoising result, showing a significant improvement in the sharpness of the in-phase axis, with background noise essentially eliminated, and is consistent with... Figure 6 The model without noise has the highest consistency.

[0088] Simultaneously, residual analysis was performed on the denoising results to obtain... Figure 6 f- Figure 6 h, it was found that although the curve wave method was effective in removing random signals, it also lost a lot of effective signals, while deep learning denoising basically did not lose any effective signals.

[0089] In step S502, the learning effect of the network is verified using synthetic seismic data of bending phase axes in the test set, and compared with traditional methods. Figure 7 As shown. Figure 7 A denoised image of composite seismic data with curved phase axes according to an embodiment of the present invention. Figure 7 'a' represents a noise-free model. Figure 7 b-noise-adding model, Figure 7 c represents the denoising result of fx filtering; background noise is still noticeable. Figure 7 Figure d shows the result of curvelet denoising, indicating that the method can effectively suppress random noise and significantly improve the signal-to-noise ratio. Figure 7 e represents the denoising result of ADCNN, and is similar to... Figure 7 The noise-free model showed the highest fit, the cleanest noise removal, and the best preservation of the effective signal. Residual analysis of the denoising results also yielded... Figure 7 f- Figure 7 Similarly, it was found that while curve denoising effectively removed random signals, it also resulted in significant loss of valid signals, while deep learning denoising achieved virtually no loss of valid signals. Therefore, different morphological in-phase axis models all verified that the ADCNN method has a stronger ability to suppress random effects.

[0090] In step S503, the open-source complex synthetic seismic model from the SEG website is selected as the training dataset. The network learning effect is verified, and a comparative analysis is performed with traditional methods. Figure 8 As shown. Figure 8 This is a denoised image of a complex model according to an embodiment of the present invention. Figure 8 'a' represents a noise-free model. Figure 8 b-noise-adding model, Figure 8 c represents the denoising result of fx. Figure 8 f represents the residual of this method, indicating that the method has poor ability to suppress random noise in complex models and suffers significant loss of effective signals. Figure 8 d represents the result of curve wave denoising, which shows good results, but from... Figure 8 The residual results between g and the original signal show that a significant amount of effective signal is lost. Figure 8 e represents the denoising result of ADCNN, and is similar to... Figure 8 The noise-free model showed the highest fit, the cleanest noise removal, and the best preservation of the effective signal. Residual analysis was also performed on the denoising results. Figure 8 h, ADCNN denoises with virtually no loss of effective signal. Therefore, complex models also demonstrate that the ADCNN method is more effective at suppressing random effects.

[0091] In step S504, to better demonstrate the denoising effect, the 60th data point of the model is selected for single-channel analysis, such as... Figure 9 As shown. Figure 9 This is a comparison diagram of single-channel time domain noise reduction according to an embodiment of the present invention. The diagram shows the result after deep learning denoising and its fit with the original seismic data. Particularly at the location indicated by the rectangle, when the original effective signal is mixed with random noise, the deep learning method can effectively extract the effective signal and suppress random noise. Finally, four noise evaluation metrics were used to numerically analyze the noise suppression effects of various methods, allowing for a more intuitive comparison of denoising performance. As shown in Table 1, it can be seen that among the four metrics—PSNR, SNR, MSE, and SSIM—the ADCNN network denoising method performs best (where a smaller MSE value indicates better denoising performance, and the other three are the opposite). The curve wave denoising method also performs well in suppressing random noise, only slightly worse than the ADCNN method.

[0092] Table 1 Comparison of noise reduction performance of seismic data

[0093]

[0094]

[0095] In step S505, after model testing, ADCNN is used to suppress random noise in the actual pre-stack data of this work area, and two traditional algorithms are compared, such as... Figure 10 As shown. Figure 10 This is a partially enlarged view of the actual pre-stack gather denoising result according to an embodiment of the present invention. Figure 10 a represents the original noisy seismic data. It can be seen that random noise interference is quite severe, the phase axes are relatively blurred, and the signal-to-noise ratio is low, which interferes with subsequent processing and research. Figure 10 b represents the result of fx filtering. As can be seen from the figure, some noise is suppressed, but the signal-to-noise ratio is still not high, the interference of the in-phase axis is strong, and the overall effect is not good. Figure 10 c represents the result of the curve transform, and the profile is significantly cleaner than fx, with clearer in-phase axes, but background noise is still clearly visible. Figure 10 d represents the deep learning result. Compared with the results of the previous two algorithms, the random noise of this seismic profile is significantly suppressed more effectively.

[0096] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0097] It should be noted that, in other embodiments of the present invention, the method can also be combined with one or more of the above embodiments to obtain a new method for suppressing seismic data noise, so as to efficiently achieve reliable noise suppression processing of pre-stack seismic data.

[0098] Example 2:

[0099] It should be noted that, based on the methods in any one or more embodiments of the present invention described above, the present invention also provides a storage medium storing program code that can implement the methods described in any one or more embodiments. When the program code is executed by the operating system, it can implement the method for suppressing seismic data noise based on a deep convolutional neural network as described above.

[0100] Example 3:

[0101] The methods described in detail in the above-disclosed embodiments of the present invention can be implemented using various forms of devices or systems. Therefore, based on other aspects of the methods described in any one or more of the above embodiments, the present invention also provides a system for suppressing seismic data noise based on a deep convolutional neural network. This system is used to execute the method for suppressing seismic data noise based on a deep convolutional neural network described in any one or more of the above embodiments. Specific embodiments are given below for detailed description.

[0102] Specifically, the system for suppressing seismic data noise based on a deep convolutional neural network provided in this embodiment of the invention includes:

[0103] The sample data preparation module is configured to divide the seismic data samples into training and test sets based on the settings, so as to use them as input and output data to build a network model; the seismic data samples include linear phase axis synthetic seismic data and curved phase axis synthetic seismic data of a set size;

[0104] The neural network building module is configured to build a denoising convolutional neural network structure that includes an input layer, hidden layers, and an output layer.

[0105] The network model optimization module is configured to introduce an attention mechanism and a dilated convolution module based on the denoising convolutional neural network structure, forming an ADCNN network structure containing multiple convolutional layers.

[0106] The optimized model training module is configured to train and validate the ADCNN network structure model by iteratively training it based on the training set until the training error of the model reaches the set requirements, and then save the corresponding ADCNN network structure model to realize the noise suppression operation of the required seismic data.

[0107] The method also includes a trial calculation and verification module, which is configured to use the test set as input and output data to perform trial calculations and verifications on the saved ADCNN network structure model.

[0108] In one optional embodiment, the sample data preparation module sets the seismic data samples to perform forward modeling of seismic records using Ricker wavelet data that conforms to the dominant frequency of the actual data, and adds random noise that meets the set requirements.

[0109] Furthermore, in one embodiment, the sample data preparation module sets the linear and bending in-phase axis synthetic seismic records in the test set to a 1:1 distribution.

[0110] Preferably, in one embodiment, the method further includes a sample processing module, which is configured to normalize the training sample data before constructing the denoising convolutional neural network structure, so that the input and output data distribution of the subsequent BN operation is more reasonable and the convergence effect of the network model is improved.

[0111] Optionally, in one embodiment, in the DnCNN network structure constructed by the neural network construction module, the input layer sets parameters for the convolution kernel to perform inner product operations to obtain input features, fills edges with zeros, and uses ReLU as the activation function; the hidden layer sets parameters for the convolution kernel to avoid covariate shifting, fills edges with zeros, and then performs BN operations, using ReLU as the activation function; the output layer sets parameters for the convolution kernel to reconstruct residual data, and does not perform BN operations to prevent errors in the results.

[0112] In one embodiment, the network model optimization module introduces a dilation parameter in the convolution operation of the neural network to characterize the distance between data points when performing convolution kernel operations.

[0113] Furthermore, in one embodiment, the network model optimization module introduces a dilated convolution module, which eliminates the downsampling operation and achieves dilated convolution through zero padding and increasing the dilation coefficient; the ADCNN network structure contains 10 convolutional layers to avoid setting too many dilated convolutional layers and destroying the continuous features of the data.

[0114] Preferably, in one embodiment, the relationship between the dilated convolution kernel parameters and the dilation coefficient is as follows:

[0115] kd size =(γ-1)(k) size -1)+k size

[0116] In the formula, k size Where γ is the kernel size, kd is the dilation coefficient, and γ is the kernel size. size This represents the size of the dilated convolution kernel.

[0117] In an optional embodiment, an attention mechanism is introduced, and the number of channels in the convolution kernel is set to 1, so that the feature information is effectively compressed, and then the effective features learned are enhanced through multiplication operations.

[0118] Furthermore, in one embodiment, the trial calculation and verification module is configured to verify the network learning effect using linear in-phase axis synthetic seismic data and curved in-phase axis synthetic seismic data in the test set, respectively, and to perform residual analysis on the denoising results.

[0119] Optionally, in one embodiment, the trial calculation and verification module is further configured to select single-channel data of the model for analysis to form a single-channel time-domain comparison display diagram; and to use multiple noise evaluation indicators to perform numerical analysis on the noise suppression effect, so as to intuitively verify the noise reduction performance.

[0120] In the system for suppressing seismic data noise based on deep convolutional neural networks provided in this embodiment of the invention, each module or unit structure can operate independently or in combination according to the actual sample data processing requirements and model operation requirements, so as to achieve the corresponding technical effects.

[0121] It should be understood that the embodiments disclosed herein are not limited to the specific structures, processing steps, or materials disclosed herein, but should be extended to equivalent substitutions of these features as understood by those skilled in the art. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting.

[0122] The phrase "an embodiment" in the specification means that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0123] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for suppressing noise in seismic data based on a deep convolutional neural network, characterized in that, The method includes: Step S10: Divide the earthquake data samples into training and testing sets to use as input and output data to build a network model; the earthquake data samples include linear in-phase axis synthetic earthquake data and curved in-phase axis synthetic earthquake data of a set size; Step S20: Construct a denoising convolutional neural network structure including an input layer, hidden layers, and an output layer; Step S30: Based on the denoising convolutional neural network structure, an attention mechanism and a dilated convolution module are introduced to form an ADCNN network structure containing multiple convolutional layers; Step S40: Train and validate the ADCNN network structure model by iteratively training it based on the training set until the training error of the model reaches the set requirement, and save the corresponding ADCNN network structure model. Step S50: Use the test set as input and output data to perform trial calculations and verification on the saved ADCNN network structure model.

2. The method according to claim 1, characterized in that, In step S10, the seismic data sample is subjected to forward modeling of seismic records using Ricker wavelet data that conforms to the dominant frequency of the actual data, and random noise that meets the set requirements is added.

3. The method according to claim 1, characterized in that, In step S10, the linear and bending in-phase composite seismic records in the test set are set to be distributed in a 1:1 ratio.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes step S11: before constructing the denoising convolutional neural network structure, the training sample data is normalized to make the input and output data distribution of the subsequent BN operation more reasonable and improve the convergence effect of the network model.

5. The method according to claim 1, characterized in that, In step S20, in the DnCNN network structure, the input layer sets parameters for the convolution kernel to perform inner product operations to obtain input features, fills edges with zeros, and uses ReLU as the activation function; the hidden layer sets parameters for the convolution kernel to avoid covariate shifting, fills edges with zeros, and then performs BN operations, using ReLU as the activation function; the output layer sets parameters for the convolution kernel to reconstruct residual data, but does not perform BN operations to prevent errors in the results.

6. The method according to claim 1, characterized in that, In step S30, the dilation rate parameter is introduced into the convolution operation in the neural network to characterize the distance between data when performing convolution kernel operation.

7. The method according to claim 1, characterized in that, In step S30, a dilated convolution module is introduced, the downsampling operation is eliminated, and dilated convolution is achieved by zero padding and increasing the dilation coefficient; the ADCNN network structure contains 10 convolutional layers to avoid setting too many dilated convolutional layers and destroying the continuous features of the data.

8. The method according to claim 7, characterized in that, The relationship between the dilated convolution kernel parameters and the dilation coefficient is as follows: kd size =(γ-1)(k size -1)+k size In the formula, k size Where γ is the kernel size, kd is the dilation coefficient, and γ is the kernel size. size This represents the size of the dilated convolution kernel.

9. The method according to claim 1, characterized in that, An attention mechanism is introduced, and the number of channels in the convolution kernel is set to 1, which effectively compresses the feature information and then enhances the effective features learned through multiplication operations.

10. The method according to claim 1, characterized in that, Step S50 includes verifying the network learning effect using linear in-phase axis synthetic seismic data and curved in-phase axis synthetic seismic data from the test set, and performing residual analysis on the denoising results.

11. The method according to claim 1, characterized in that, Step S50 also includes selecting single-channel data from the model for analysis to form a single-channel time-domain comparison display diagram; and using multiple noise evaluation indicators to perform numerical analysis on the noise suppression effect, intuitively verifying the denoising performance.

12. A system for suppressing noise in seismic data based on a deep convolutional neural network, characterized in that, The system performs the method as described in any one of claims 1 to 11.

Images