Surface wave noise suppression method and device based on multi-information fusion

Abstract

The invention relates to the technical field of geophysical exploration, and particularly discloses a surface wave noise suppression method and device based on multi-information fusion, and the method comprises the steps: screening a sample label data set from actual data; based on the sample label data set, generating multi-information input data, and preparing a training data set; training a neural network model by using the training data set, and storing parameters of the trained neural network model; and inputting the multi-information input data into the trained neural network model to obtain de-noised data. According to the method, the deep neural network architecture with multi-modal input is established, and the physical information conforming to the surface wave characteristics of the current work area is introduced on the basis of the seismic data information, so that the ability of the network to identify the surface waves of different work areas is enhanced, the generalization ability of the network is improved, and intelligent processing of pre-stack data seismic surface wave noise suppression is further realized.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, specifically to a method and apparatus for suppressing surface wave noise based on multi-information fusion. Background Technology

[0002] Suppression of surface wave noise is a crucial step in seismic data processing. Surface waves, as regular noise, are characterized by low speed, low frequency, and high energy, and are generally distributed within the near offset range in common shot point ensembles. Currently, the more mature surface wave suppression algorithms in commercial software mainly include: (1) based on the low-frequency and high-energy characteristics of surface waves, surface wave energy suppression is achieved by statistically analyzing the energy difference with the effective signal through methods such as frequency division filtering; (2) based on the low-speed characteristics of surface waves, surface wave noise suppression is achieved by constructing a frequency spatial domain filter; and (3) based on the dispersive properties of surface waves, surface wave reconstruction is achieved, followed by matched subtraction to achieve the effect of surface wave suppression. In summary, these algorithms in commercial software have been industrialized, but the quality of processing depends on the experience and ability of the processing personnel.

[0003] In recent years, deep learning algorithms for noise suppression in seismic data have been a key research focus for scholars both domestically and internationally. Mainstream algorithms target random noise suppression and have achieved some practical results. The main research approaches for intelligent surface wave noise suppression algorithms focus on network model optimization and the introduction of physical information such as signal characteristics into the loss function. These approaches have largely driven the practical application of intelligent surface wave suppression algorithms. However, most of these algorithms rely solely on seismic data as input data, and the networks they build can only rely on this single piece of information for noise identification. The diversity of seismic noise data characteristics and distribution can lead to multiple solutions in the network's identification process, thus affecting the noise suppression effect.

[0004] Based on this technical background, this invention studies a surface wave noise suppression method and device based on multi-information fusion. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a surface wave noise suppression method and apparatus based on multi-information fusion. This method establishes a deep neural network architecture with multi-modal input, introduces physical information that conforms to the surface wave characteristics of the current work area on the basis of seismic data information, enhances the network's ability to identify surface waves in different work areas, improves the network's generalization ability, and thus realizes intelligent processing of pre-stack data seismic surface wave noise suppression.

[0006] To achieve the above objectives, a first aspect of the present invention provides a surface wave noise suppression method based on multi-information fusion, comprising:

[0007] Filter sample label datasets from actual data;

[0008] Based on the sample label dataset, generate multi-information input data and prepare a training dataset;

[0009] The neural network model is trained using the training dataset, and the parameters of the trained neural network model are saved.

[0010] The multi-information input data is input into the trained neural network model to obtain the denoised data.

[0011] A second aspect of the present invention provides a surface wave noise suppression device based on multi-information fusion, comprising:

[0012] The filtering module is used to filter sample label datasets from actual data;

[0013] The generation and preparation module is used to generate multi-information input data and prepare training datasets based on the sample label dataset;

[0014] The training module is used to train the neural network model using the training dataset and save the parameters of the trained neural network model.

[0015] The application module is used to input the multi-information input data into the trained neural network model to obtain the denoised data.

[0016] A third aspect of the present invention provides an electronic device, the electronic device comprising:

[0017] Memory, which stores executable instructions;

[0018] A processor that executes the executable instructions in the memory to implement the surface wave noise suppression method based on multi-information fusion as described in the first aspect.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the surface wave noise suppression method based on multi-information fusion described in the first aspect.

[0020] The beneficial effects of this invention include:

[0021] (1) The surface wave noise suppression method based on multi-information fusion proposed in this invention establishes a deep neural network architecture with multi-modal input, introduces physical information that conforms to the surface wave characteristics of the current work area on the basis of seismic data information, enhances the network's ability to identify surface waves in different work areas, improves the network's generalization ability, and thus realizes intelligent processing of pre-stack data seismic surface wave noise suppression.

[0022] (2) The surface wave noise suppression method based on multi-information fusion proposed in this invention achieves high-precision identification and suppression of surface wave noise by establishing a multi-modal input network architecture based on multi-information fusion, thereby improving the denoising accuracy and efficiency of seismic data. It can be applied to the noise suppression process in the preprocessing of Earth seismic data and solves the problem of insufficient denoising accuracy of traditional intelligent surface wave suppression algorithms.

[0023] (3) The surface wave noise suppression method based on multi-information fusion proposed in this invention introduces surface wave distribution area information and low-frequency characteristics of seismic data to further enhance surface wave characteristics and improve the accuracy of neural network recognition of surface waves, thereby achieving efficient and high-precision suppression of surface wave noise. This method can not only better serve the noise suppression process in seismic data processing, but also provide new ideas for the practical application of intelligent noise suppression.

[0024] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0025] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating the surface wave noise suppression method based on multi-information fusion proposed in this invention.

[0027] Figure 2 This is a schematic diagram of the input and output data required by the neural network in a specific implementation of the surface wave noise suppression method based on multi-information fusion proposed in this invention.

[0028] Figure 3 This is a schematic diagram of the neural network architecture used in a specific implementation of the surface wave noise suppression method based on multi-information fusion proposed in this invention.

[0029] Figure 4 This is a schematic diagram comparing test results based on actual data in a specific implementation of the surface wave noise suppression method based on multi-information fusion proposed in this invention. Detailed Implementation

[0030] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0031] This invention provides a surface wave noise suppression method based on multi-information fusion, such as... Figure 1 As shown, it includes:

[0032] Filter sample label datasets from actual data;

[0033] Based on the sample label dataset, generate multi-information input data and prepare the training dataset;

[0034] The neural network model is trained using the training dataset, and the parameters of the trained neural network model are saved.

[0035] The multi-information input data is fed into the trained neural network model to obtain the denoised data.

[0036] According to the present invention, the multi-information input data is obtained by low-pass filtering the original data according to the current surface wave frequency distribution range of the work area;

[0037] The training dataset consists of surface wave distribution area information data obtained from the apparent velocity distribution of surface waves in the current work area.

[0038] In this invention, by establishing a deep neural network architecture with multimodal input, physical information that conforms to the surface wave characteristics of the current work area is introduced on the basis of seismic data information, which enhances the network's ability to identify surface waves in different work areas and improves the network's generalization ability, thereby realizing intelligent processing of pre-stack data seismic surface wave noise suppression.

[0039] According to the present invention, the neural network model is a UNET network structure;

[0040] The neural network model includes a downsampling encoding part and an upsampling decoding part. The input end is a three-channel input layer and the output layer is a single-channel output layer.

[0041] During neural network training, an L2 norm loss function and the Adam optimizer are used for iterative updates.

[0042] According to the present invention, the input dimensions of the three-channel input layers are equal;

[0043] The data output by a single-channel output layer is equal in size to the corresponding original input data.

[0044] According to the present invention, the basic network structure of the encoding stage of the neural network model consists of two convolutional layers with the same number of channels and a max pooling layer.

[0045] In this invention, by establishing a multi-modal input network architecture that integrates multiple information, high-precision identification and suppression of surface wave noise is achieved, thereby improving the denoising accuracy and efficiency of seismic data. It can be applied to the noise suppression process in the preprocessing of Earth seismic data, solving the problem of insufficient denoising accuracy in traditional intelligent surface wave suppression algorithms.

[0046] Preferably, the downsampling encoding part includes a pooling layer and a dropout deactivation layer;

[0047] The number of horizontal and vertical sampling points in the input image after downsampling via the downsampling encoding part is half that before sampling.

[0048] The inactivation rate of the dropout deactivation layer is 0.5%.

[0049] According to the present invention, the upsampling decoding part is implemented using the nearest neighbor interpolation algorithm;

[0050] The upsampling decoding process decreases the number of channels sequentially, and finally outputs the final result by a 1x1 convolutional layer.

[0051] In this invention, based on the input seismic data, information such as surface wave distribution area information and low-frequency characteristics of seismic data are introduced to further enhance the surface wave characteristics and improve the accuracy of neural network recognition of surface waves. This achieves efficient and high-precision suppression of surface wave noise, which can not only better serve the noise suppression process in seismic data processing, but also provide new ideas for the practical application of intelligent noise suppression.

[0052] The present invention will be described in more detail below through embodiments.

[0053] Example 1:

[0054] like Figure 1 As shown, this embodiment proposes a surface wave noise suppression method based on multi-information fusion. Addressing the problem of surface wave noise suppression in seismic data, this method fully utilizes the low-frequency and low-speed characteristics of surface waves as regular noise. A multi-information fusion intelligent noise suppression network is established based on a deep learning framework. The network is trained by preparing a dataset of actual data samples and labels, forming a highly efficient and accurate intelligent surface wave noise suppression technology. The effectiveness of the proposed algorithm is verified through practical data testing.

[0055] This embodiment employs a multi-information input network structure. The network input data includes three types of data: raw seismic data, low-frequency seismic data, and surface wave distribution area. The raw seismic data is the common shot point data before surface wave suppression. The low-frequency seismic data is obtained by low-pass filtering the raw data based on the current surface wave frequency distribution range of the work area. The surface wave distribution area is surface wave distribution area information data obtained based on the current surface wave apparent velocity distribution of the work area. In this embodiment, the network is trained using the common shot point data after surface wave suppression as labels. The input and output data diagram is shown below. Figure 2 (in Figure 2 (a) represents the raw seismic data input from the network. Figure 2 (b) represents the low-frequency seismic data input from the network. Figure 2 (c) represents the surface wave distribution range of the network input. Figure 2 (d) represents the denoised data output by the network; the neural network structure used is as shown in the figure. Figure 3 As shown, the network uses a UNET network structure, including a downsampling encoding part and an upsampling decoding part. The input is a three-channel input layer, receiving the three data points of equal size mentioned above. The output is a single-channel output layer, outputting the denoised result data of equal size to the original data. The basic network structure in the encoding stage consists of two convolutional layers with the same number of channels and a max-pooling layer. Downsampling is achieved by the pooling layer, resulting in half the number of horizontal and vertical sampling points of the input image after downsampling. To improve the network's generalization ability and avoid overfitting, a dropout deactivation layer with a deactivation rate of 0.5 is introduced in the downsampling stage. In the decoding stage, the nearest neighbor interpolation algorithm is used for upsampling. The number of channels decreases continuously during the upsampling process, and the final output result is obtained after a 1x1 convolution. During network training, this embodiment uses an L2 norm loss function and an Adam optimizer for iterative updates.

[0056] The specific process of this method can be summarized as follows:

[0057] (1) Collection of actual data sample labels: Since the surface wave characteristics obtained from the simulated data are significantly different from the actual data, in order to enable the network model to effectively identify surface waves during the actual surface wave noise suppression process, actual exploration data is selected as sample label data. The denoising effect of the data before and after denoising obtained from the mature surface wave denoising algorithms in existing commercial software is evaluated, and high-quality denoised data is selected as the label dataset.

[0058] (2) Input data multi-information data generation and training set preparation: Based on the original data before denoising in the selected sample label data pairs, the low-frequency data required for the input end is prepared according to the frequency characteristics of the surface wave in the current work area. The surface wave distribution area information data index is obtained according to the surface wave apparent velocity distribution characteristics, thereby generating the input and output datasets required for network training.

[0059] (3) Network training: The network is trained using the prepared dataset to obtain a high-precision and high-efficiency surface wave noise suppression network and save it.

[0060] (4) Network promotion test application: Load the trained network model onto the work area data that has not been trained and apply surface wave noise suppression to obtain the denoised data.

[0061] In this embodiment, the training input data size is 256*1024, and a total of 110,000 samples were prepared. The data sample label data was prepared using TFRecord. The above-mentioned network was established, and the learning rate was set to 0.0001, the batch size was set to 5, and the number of training rounds was set to 50. In addition, the cluster GPU used in this test is an RTX 2080Ti with 11G of video memory. Eight GPUs were used for training, and the data parallelism method was adopted to improve the training efficiency.

[0062] The network was trained based on the above parameters, and the trained model was saved. During testing, the data size was 140*3001, and the input data before processing was as follows: Figure 4 As shown in (a), the network structure based on a single raw seismic data input yields the following results: Figure 4 As shown in (b), the results obtained based on this technical method are as follows: Figure 4 As shown in (c), the denoising results obtained by the multi-information fusion network model proposed in this embodiment are significantly better than those obtained by the single-channel input network, indicating that the multi-information network architecture has a significant advantage in improving the accuracy of network noise recognition.

[0063] Example 2:

[0064] This embodiment provides a surface wave noise suppression method based on multi-information fusion, such as... Figure 1 As shown, it includes:

[0065] Filter sample label datasets from actual data;

[0066] Based on the sample label dataset, generate multi-information input data and prepare the training dataset;

[0067] The neural network model is trained using the training dataset, and the parameters of the trained neural network model are saved.

[0068] The multi-information input data is fed into the trained neural network model to obtain the denoised data.

[0069] In this embodiment, the multi-information input data is obtained by low-pass filtering the original data according to the current work area surface wave frequency distribution range;

[0070] The training dataset consists of surface wave distribution area information data obtained from the current surface wave apparent velocity distribution in the work area;

[0071] In this embodiment, the neural network model is a UNET network structure;

[0072] The neural network model includes a downsampling encoding part and an upsampling decoding part. The input end is a three-channel input layer and the output layer is a single-channel output layer.

[0073] During the training of the neural network, an L2 norm loss function and the Adam optimizer are used for iterative updates.

[0074] In this embodiment, the input dimensions of the three-channel input layer are equal;

[0075] The data output by a single-channel output layer is equal in size to the corresponding original input data.

[0076] In this embodiment, the basic network structure of the encoding stage of the neural network model consists of two convolutional layers with the same number of channels and a max pooling layer.

[0077] In this embodiment, the downsampling encoding part includes a pooling layer and a dropout deactivation layer;

[0078] The number of horizontal and vertical sampling points in the input image after downsampling via the downsampling encoding part is half that before sampling.

[0079] The deactivation rate of the dropout deactivation layer is 0.5;

[0080] In this embodiment, the upsampling decoding part is implemented using the nearest neighbor interpolation algorithm;

[0081] The upsampling decoding process decreases the number of channels sequentially, and finally outputs the final result by a 1x1 convolutional layer.

[0082] Example 3:

[0083] This embodiment provides a surface wave noise suppression device based on multi-information fusion, comprising:

[0084] The filtering module is used to filter sample label datasets from actual data;

[0085] The generation and preparation module is used to generate multi-information input data and prepare training datasets based on the sample label dataset;

[0086] The training module is used to train the neural network model using the training dataset and save the parameters of the trained neural network model.

[0087] The application module is used to input multi-information input data into a trained neural network model to obtain denoised data;

[0088] In this embodiment, the multi-information input data is obtained by low-pass filtering the original data according to the current work area surface wave frequency distribution range;

[0089] The training dataset consists of surface wave distribution area information data obtained from the current surface wave apparent velocity distribution in the work area;

[0090] In this embodiment, the neural network model is a UNET network structure;

[0091] The neural network model includes a downsampling encoding part and an upsampling decoding part. The input end is a three-channel input layer and the output layer is a single-channel output layer.

[0092] During the training of the neural network, an L2 norm loss function and the Adam optimizer are used for iterative updates.

[0093] In this embodiment, the input dimensions of the three-channel input layer are equal;

[0094] The data output by a single-channel output layer is equal in size to the corresponding original input data.

[0095] In this embodiment, the basic network structure of the encoding stage of the neural network model consists of two convolutional layers with the same number of channels and a max pooling layer.

[0096] In this embodiment, the downsampling encoding part includes a pooling layer and a dropout deactivation layer;

[0097] The number of horizontal and vertical sampling points in the input image after downsampling via the downsampling encoding part is half that before sampling.

[0098] The deactivation rate of the dropout deactivation layer is 0.5;

[0099] In this embodiment, the upsampling decoding part is implemented using the nearest neighbor interpolation algorithm;

[0100] The upsampling decoding process decreases the number of channels sequentially, and finally outputs the final result by a 1x1 convolutional layer.

[0101] Example 4:

[0102] This invention provides an electronic device including a memory and a processor.

[0103] Memory, which stores executable instructions;

[0104] The processor executes executable instructions in memory to implement a surface wave noise suppression method based on multi-information fusion.

[0105] This memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0106] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of the invention, the processor is used to execute computer-readable instructions stored in the memory.

[0107] Those skilled in the art should understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this invention.

[0108] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0109] Example 5:

[0110] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a surface wave noise suppression method based on multi-information fusion.

[0111] A computer-readable storage medium according to embodiments of the present invention stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the methods described in the foregoing embodiments of the present invention are performed.

[0112] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0113] The surface wave noise suppression method based on multi-information fusion proposed in the embodiments of the present invention establishes a deep neural network architecture with multi-modal input, introduces physical information that conforms to the surface wave characteristics of the current work area on the basis of seismic data information, enhances the network's ability to identify surface waves in different work areas, improves the network's generalization ability, and thus realizes intelligent processing of pre-stack data seismic surface wave noise suppression.

[0114] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A surface wave noise suppression method based on multi-information fusion, characterized in that, include: Filter sample label datasets from actual data; Based on the sample label dataset, generate multi-information input data and prepare a training dataset; The neural network model is trained using the training dataset, and the parameters of the trained neural network model are saved. The multi-information input data is input into the trained neural network model to obtain the denoised data.

2. The method according to claim 1, characterized in that, The multi-information input data is obtained by low-pass filtering the original data according to the current surface wave frequency distribution range of the work area. The training dataset is surface wave distribution area information data obtained based on the current surface wave apparent velocity distribution in the work area.

3. The method according to claim 1, characterized in that, The neural network model is a UNET network structure; The neural network model includes a downsampling encoding part and an upsampling decoding part, with a three-channel input layer and a single-channel output layer. During the training of the neural network, an L2 norm loss function and an Adam optimizer are used for iterative updates.

4. The method according to claim 3, characterized in that, The input dimensions of the three-channel input layers are equal; The data output by the single-channel output layer is equal in size to the corresponding original input data.

5. The method according to claim 3, characterized in that, The basic network structure of the encoding stage of the neural network model consists of two convolutional layers with the same number of channels and a max pooling layer.

6. The method according to claim 3, characterized in that, The downsampling encoding part includes a pooling layer and a dropout deactivation layer; The number of horizontal and vertical sampling points in the input image after downsampling by the downsampling encoding part is half that before sampling; The inactivation rate of the dropout deactivation layer is 0.

5.

7. The method according to claim 3, characterized in that, The upsampling decoding part is implemented using the nearest neighbor interpolation algorithm; The upsampling decoding process decreases the number of channels sequentially, and finally outputs the final result by a 1x1 convolutional layer.

8. A surface wave noise suppression device based on multi-information fusion, characterized in that, include: The filtering module is used to filter sample label datasets from actual data; The generation and preparation module is used to generate multi-information input data and prepare training datasets based on the sample label dataset; The training module is used to train the neural network model using the training dataset and save the parameters of the trained neural network model. The application module is used to input the multi-information input data into the trained neural network model to obtain the denoised data.

9. An electronic device, characterized in that, The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the surface wave noise suppression method based on multi-information fusion according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the surface wave noise suppression method based on multi-information fusion as described in any one of claims 1-7.

Images