A method and system for suppressing random noise in pre-stack seismic data

Abstract

The present invention provides a method and system for suppressing random noise in prestack seismic data, which belongs to the field of geophysical exploration technology and deep learning. The method for suppressing random noise in prestack seismic data uses actual seismic data, prestack shot set forward simulation data and GPU parallel mode to obtain an intelligent random noise suppression network, and uses the intelligent random noise suppression network to process actual prestack seismic data to obtain prestack seismic data after random noise suppression. The present invention applies the U‑NET network to seismic data denoising, effectively avoiding the high requirements of conventional convolutional networks on the amount of training data and the scale of neural networks, and realizes multi-GPU data parallel network training, improves training efficiency, and realizes effective suppression of random noise in prestack data.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical exploration technology and deep learning, and specifically relates to a method and system for suppressing random noise in pre-stack seismic data, which is applied to random noise removal in denoising processing in petroleum geophysical exploration. Background Art

[0002] The presence of noise in seismic data seriously affects the imaging quality of seismic data. How to effectively denoise seismic data is a key step in improving the signal-to-noise ratio of seismic data in seismic data processing. Seismic data noise is mainly divided into regular noise and random noise. Regular noise has deterministic characteristics and can be effectively suppressed according to the noise formation mechanism, while random noise only has statistical characteristics and no deterministic distribution form, so noise suppression is relatively difficult. At present, common random noise suppression methods mainly include spatiotemporal domain denoising methods and transform domain denoising methods. The spatiotemporal domain denoising algorithm is directly based on spatiotemporal domain data. Representative methods mainly include non-local mean denoising algorithms based on redundant structure assumptions. This method mainly suppresses noise by grouping areas with similar local structures in seismic data. It can effectively suppress random noise without reducing the effective signal resolution, but the computational complexity is relatively high and the processing efficiency is low. The denoising method based on transform domain suppresses noise by performing data domain transform on the data to be processed to enhance the difference between noise and effective signal. The current representative methods include fk domain filtering, wavelet transform denoising and curvelet transform denoising. Although this type of method has achieved certain application effects in actual production, there are still some shortcomings in denoising performance or denoising efficiency.

[0003] In general, a variety of algorithms have been developed for suppressing random noise and have been widely used in practice. However, most of the current methods rely on manual intervention and are greatly affected by the ability of the processing personnel. The denoising process may require a lot of manual parameter adjustment, which greatly affects the denoising efficiency.

[0004] In recent years, with the rapid development of automated rapid image processing technology based on artificial intelligence, the emergence of convolutional neural network algorithms has promoted the application of deep learning in various fields, making seismic data processing based on artificial intelligence have broad application prospects. As a result, the noise removal technology of seismic data based on deep neural networks has also been rapidly developed and applied and has achieved certain results. However, the existing intelligent denoising methods are mainly based on conventional convolutional networks. The development of technology and practical application are largely limited by the size of available label data sets and the size of the neural network that can be designed. The U-NET network was published in 2015 and was originally used for medical cell segmentation. It is an expanded and optimized full convolutional network. The network consists of a downsampling path and a symmetrical upsampling path. It can support a small amount of data training model and has higher resolution than other fully connected and conventional convolutional networks. The development of seismic data random noise suppression technology based on the U-net network can effectively improve the noise suppression accuracy.

[0005] Based on the above problems, a random noise suppression technology for seismic data based on U-NET network was developed. In addition, considering that the training of the model takes the longest time in the process of intelligent processing, the conventional CPU-based training and single GPU training fail to make full use of computing resources and affect the efficiency of network training. Therefore, the present invention implements multi-GPU data parallel model training based on Tensorflow and provides a high-precision and efficient intelligent denoising process, realizes intelligent processing of noise suppression, and avoids manual intervention in the denoising process. The formed technology can better serve seismic data processing and lay the foundation for the production, application and promotion of random noise suppression technology for seismic data. There is no corresponding research literature on this method. Summary of the invention

[0006] The purpose of the present invention is to solve the difficulties existing in the above-mentioned prior art, provide a method and system for suppressing random noise in pre-stack seismic data, realize multi-GPU parallel model training, further enhance the efficiency of model training, and establish a set of intelligent denoising networks with strong generalization ability and high denoising accuracy, reduce manual intervention, improve denoising efficiency, provide high signal-to-noise ratio data for seismic data processing, and provide new technical ideas for future high-precision seismic data denoising technology and provide strong technical support for actual production applications.

[0007] The present invention is achieved through the following technical solutions:

[0008] A first aspect of the present invention provides a method for suppressing random noise in prestack seismic data. The method uses actual seismic data, prestack shot set forward simulation data, and a GPU parallel mode to obtain an intelligent random noise suppression network, and uses the intelligent random noise suppression network to process actual prestack seismic data to obtain prestack seismic data after random noise suppression.

[0009] A further improvement of the present invention is that the method comprises:

[0010] (1) Using actual seismic data and pre-stack shot gather forward modeling data to prepare a data set, and divide the data set into a training data set and a validation data set;

[0011] (2) Design a network and use a GPU and training data set to train the network in parallel to obtain a trained network;

[0012] (3) Using the validation data set to validate the trained network, a validated intelligent random noise suppression network is obtained;

[0013] (4) Collecting pre-stack seismic data and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.

[0014] A further improvement of the present invention is that the actual seismic data in step (1) is actual seismic data with high-quality denoising effect selected from denoised seismic data of a typical exploration area;

[0015] The pre-stack shot gather forward modeling data in step (1) is simulation data obtained after forward modeling of the pre-stack shot gather.

[0016] A further improvement of the present invention is that the actual seismic data accounts for 40% and the simulation data accounts for 60%.

[0017] A further improvement of the present invention is that the operation of preparing a data set using actual seismic data and pre-stack shot gather forward modeling data in step (1) comprises:

[0018] Using the simulation data as label data;

[0019] The noise data extracted from the actual seismic data is randomly matched with the simulated data and then summed to obtain the noisy data, or Gaussian random noise is added to the simulated data to obtain the noisy data.

[0020] The noisy data and the labeled data constitute the data set.

[0021] The operation of step (1) further includes: writing the data set into a compression file in the storage format of the MNIST data set.

[0022] A further improvement of the present invention is that the network designed in step (2) is a U-NET network;

[0023] The nearest neighbor interpolation algorithm is used for image upsampling in the upsampling part of the U-NET network.

[0024] A further improvement of the present invention is that the operation of using the GPU and the training data set to perform parallel training on the network to obtain a trained network in step (2) includes:

[0025] The CPU distributes the training data set to each GPU;

[0026] Each GPU trains the U-NET network using the training data set, obtains a gradient value according to the training result, and then sends the gradient value to the CPU;

[0027] The CPU reduces and averages the gradient values ​​sent by all GPUs to obtain the mean gradient, and applies the mean gradient to the loss function and optimizer to obtain the updated model.

[0028] After multiple updates, the trained network is obtained.

[0029] A further improvement of the present invention is that the loss function adopts the L2 norm;

[0030] The optimizer adopts the Adam optimizer.

[0031] A second aspect of the present invention provides a system for suppressing random noise in pre-stack seismic data, the system comprising:

[0032] A data set preparation unit is used to prepare a data set using actual seismic data and pre-stack shot gather forward simulation data, and divide the data set into a training data set and a verification data set;

[0033] A network design and training unit, connected to the data set preparation unit, is used to design a network and perform parallel training on the network using a GPU and a training data set to obtain a trained network;

[0034] A verification unit, connected to the network design training unit, for verifying the trained network using a verification data set to obtain a verified intelligent random noise suppression network;

[0035] The noise suppression unit is connected to the verification unit and is used to collect pre-stack seismic data and input the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention applies the U-NET network to seismic data denoising, effectively avoiding the high requirements of conventional convolutional networks on the amount of training data and the scale of the neural network, and realizes multi-GPU data parallel network training, improves training efficiency, and achieves effective suppression of random noise in pre-stack data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the U-NET network structure used in the present invention.

[0039] Figure 2 This is a schematic diagram of the principle of implementing multi-GPU data parallelism in the present invention.

[0040] Figure 3 This is a schematic diagram of the real-time operating status of GPU cards when multiple GPUs are running in parallel during network training in the present invention.

[0041] Figure 4 This is the noise-free single-shot data for this verification.

[0042] Figure 5 This is the single shot data after noise addition for this verification data.

[0043] Figure 6 This is a single-channel comparison of the noise-free label data and the noisy data for this verification data.

[0044] Figure 7 This is the single shot data after the trained network suppression and recovery for this verification data.

[0045] Figure 8 This is a single-channel comparison of the noise-free label data and denoised data for this verification data.

[0046] Fig. 9 is the actual noisy seismic data used in this verification.

[0047] Fig.10 This is the actual data after noise suppression of the verification data;

[0048] Fig.11 A flowchart of the steps of the method of the present invention;

[0049] Fig.12 A structural diagram of the system of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0051] Aiming at the problem of random noise suppression in pre-stack seismic data, the present invention establishes a U-NET network based on Tensorflow to perform multi-GPU data parallel training on the prepared data set, develops a computationally efficient intelligent denoising process, and verifies the denoising effect of the generated algorithm process through verification data.

[0052] The structure of the U-NET network used in the present invention is as follows Figure 1 As shown in the figure, the entire network is a fully convolutional neural network, which is divided into two parts: downsampling and upsampling. In the upsampling path of the network, a large number of channels are still maintained to ensure the high resolution of the grid. Among them, the basic structural unit of the downsampling part is two convolutional layers (the same number of channels, the convolution kernel size is 3x3) followed by a maximum pooling layer (2x2, the step size is 2x2), a total of three groups, the number of channels in the convolution layer increases from the initial 64 to 256, the number of channels in each group is twice that of the previous group, and the image is downsampled under the action of the pooling layer, and the number of horizontal and vertical sampling points of the image becomes half of that of the previous group. In the upsampling part, the nearest neighbor interpolation algorithm (NEAREST_NEIGHBOR) is used for image upsampling, and the cropping and copying method is further used to splice with the higher resolution picture in the left downsampling part, ensuring the richness and resolution of the information in the upsampling process. The number of channels in the upsampling convolution process is continuously reduced and maintained symmetric with the downsampling path, and finally a 1x1 convolution is performed to obtain the final output result.

[0053] An important structure that distinguishes the U-net network from other deep learning networks is that every time the upsampling is performed, the network layer with the same number of channels as the downsampling part must be fused. That is, when the sizes of the two sets of images are inconsistent, the corresponding images in the downsampling process need to be cropped to ensure that the image size is the same as the upsampling process, and the two sets of image arrays are further fused in the dimension where the number of channels is located, that is, copied and spliced, such as Figure 1 When the image sizes are consistent during upsampling and downsampling, the process can be simplified to simple copying and splicing, omitting the cropping process.

[0054] The deep learning network structure adopted in the present invention is based on the existing U-net network (originally used for image segmentation), and the grid structure remains basically consistent (it is a fully convolutional neural network, divided into two parts: downsampling and upsampling). Furthermore, according to the needs of the present invention itself and the characteristics of the data, the number of layers of the overall network and the number of channels of each convolutional layer are targetedly reduced, and the upsampling algorithm and other algorithms are modified to adapt to the image processing needs of the present invention as a whole.

[0055] The input and output image sizes of the U-NET network of the present invention remain consistent, ensuring an end-to-end training process. To ensure that the relative relationship between the data energy and amplitude before and after denoising remains unchanged, the loss function used is the L2 norm (the existing L2 norm formula can be used), and the Adam optimizer (the existing Adam optimizer is used) is used for iterative updating.

[0056] In addition, in order to improve the training efficiency, the present invention implements network training based on GPU data parallelism. The implementation process is as follows: Figure 2 As shown, the CPU is responsible for data distribution, gradient reduction (after data distribution, gradient updates are performed in each GPU card, and then the average of each update amount is calculated in the CPU as the current gradient update amount) and model update, and the GPU is responsible for model training.

[0057] Specifically, the CPU distributes training sample data (for the present invention, the data before denoising and the corresponding label data after denoising). The size of the training data taken by the CPU each time is batch_size*the number of GPU cards, and then the CPU distributes these data evenly to each GPU card, and the amount of data on each card is batch_size. The CPU distributes the training data taken each time evenly to each GPU card, and the amount of data on each GPU card is batch_size.

[0058] The parallel method adopted by the present invention is data parallelism, that is, each parallel unit (each GPU card) adopts the same network model for training, and performs gradient calculation on the training result (i.e., the gradient update amount) (the training process in each GPU card is the conventional deep learning network training process. By giving the corresponding loss function and optimization iterator, only the gradient calculation API of the deep learning network optimization iterator needs to be called during the network training process to obtain the current gradient), and the gradient value is transmitted to the CPU for reduction and averaging. The CPU is responsible for applying the required gradient to the optimizer to update the model (the network model update in deep learning refers to the update of the weight parameters of each network layer (specifically referring to the convolution kernel parameters here). The present invention adopts the Adam algorithm. The main idea is to update the neural network weights by continuously solving the gradient iteration, and finally minimize the target functional).

[0059] Figure 2 The loss in refers to the value of the objective function obtained based on the current network weight coefficient, that is, the L2 norm of the difference between the output result obtained by inputting the current input data into the network and the expected result (label result). The gradient at this moment can be obtained based on the existing Adam optimization iterator in the deep learning tensorflow framework.

[0060] The embodiments of the method of the present invention are as follows:

[0061] [Example 1]

[0062] like Fig.11 As shown in FIG. 1 , the random noise suppression method for prestack seismic data based on the U-NET network includes the following steps:

[0063] (1) Prepare data sets using actual seismic data and pre-stack shot forward simulation data: A key step in deep learning processing is the production of labeled data sets. To ensure the generalization ability of the training network, the source of high-quality labeled data for intelligent noise suppression technology can be composed of two parts: 1. Data analysis and extraction of high-quality denoised seismic data from typical exploration areas: Actual data labels are an important part of improving the generalization ability of deep networks. Therefore, data selection should cover a variety of actual typical exploration areas, and select data with high-quality denoising effects from them to improve the robustness of the data set; 2. Pre-stack shot forward simulation data: The denoising effect of actual data is largely affected by the denoising algorithm and the quality of the data itself, while the denoising results corresponding to the forward simulation data are clearly known, which can improve the denoising effect of the network. Of these, actual seismic data account for 40%, and pre-stack shot forward simulation data account for 60%.

[0064] The present invention can obtain a large amount of noise-free pre-stack shot data (i.e., noise-free single shot data) as the label data of the desired output after denoising through forward simulation, and the corresponding data preparation before denoising is mainly realized in two ways: 1. Only the noise data (the difference between the data before and after denoising) is extracted from the selected actual data and the random data matching is performed with the prepared forward simulation data, and then the noisy data is obtained by summing; 2. Gaussian random noise is added to the noise-free simulation data as the noisy data. The noisy data and the label data constitute the data set.

[0065] The data obtained by the above two methods each account for 50% of the model data. Finally, the above data is written into the compression file according to the magic number (magic-4 bytes), the number of images (4 bytes), the number of rows of a single image (4 bytes), the number of columns of a single image (4 bytes), and the data rules in the storage format of the MNIST data set to prepare the data set; the data set is divided into a training data set and a verification data set, and the ratio of the two can be set according to actual needs.

[0066] (2) Designing a network, and using a GPU and a training data set to perform parallel training on the network to obtain a trained network: Performing network training based on the network structure, loss function, iterative optimizer, and GPU parallel training algorithm designed in the present invention: During the training process, the noise-free single-shot data in the training data set is used as the label data, and random noise is randomly added to the label data according to the established signal-to-noise ratio during each batch (one batch is trained on each GPU) training process to enhance the training data, and two sets of data (the two sets of data are the noisy data before and after adding random noise, the noisy data is used as data input, and the noise-free data is used as the expected output) are input into the network for training, and finally a set of intelligent random noise suppression network is obtained (the entire training process is a process of continuous updating of the deep learning network, during which the network parameters are continuously optimized, and the network training stops when the expected training accuracy is reached. At this time, the corresponding network weight parameters are the final intelligent noise suppression network), and the trained network is saved;

[0067] (3) Using the validation data to validate the trained network, the validated intelligent random noise suppression network is obtained: reloading the saved network, inputting the actual seismic data to be denoised in the validation data set to perform data denoising, validating the effect of the trained network, and obtaining the validated intelligent random noise suppression network;

[0068] (4) Collecting pre-stack seismic data and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.

[0069] The present invention also provides a system for suppressing random noise of pre-stack seismic data. The embodiments of the system of the present invention are as follows:

[0070] [Example 2]

[0071] like Fig.12 As shown, the system comprises:

[0072] A data set preparation unit 10 is used to prepare a data set using actual seismic data and pre-stack shot gather forward modeling data, and divide the data set into a training data set and a verification data set;

[0073] A network design and training unit 20, connected to the data set preparation unit 10, is used to design a network and perform parallel training on the network using a GPU and a training data set to obtain a trained network;

[0074] A verification unit 30, connected to the network design training unit 20, is used to verify the trained network using a verification data set to obtain a verified intelligent random noise suppression network;

[0075] The noise suppression unit 40 is connected to the verification unit 30 and is used to collect pre-stack seismic data and input the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.

[0076] The embodiments of the present invention are as follows:

[0077] [Example 3]

[0078] The input data size for this training is 128x1000, and a data set containing 100,000 single-shot results was prepared. The above-mentioned U-NET network was established, and the learning rate was set to 0.001. During the training process, in order to further improve the generalization ability of the network, 5% Gaussian noise was further added for data enhancement. In addition, the batch_size was set to 20, and the number of learning rounds was set to 15. In addition, the cluster GPU used in this test is RTX 2080Ti with a memory size of 11G, and the number of GPU cards used for training is seven cards. The operating status of the GPU card during the model training process is as follows: Figure 3 As shown in the figure, we can see that all 7 GPU cards have high utilization rates, which effectively improves the efficiency of network training.

[0079] The network is trained based on the above parameters and the trained model is saved. In the test process, the model data is first used for testing. The forward data is tested by adding 8% noise. The input noise-free single shot data is as follows Figure 4 As shown, the single shot data with 8% noise added is as follows Figure 5 As shown, extract Figure 4 and Figure 5 The 20th single channel data comparison is as follows Figure 6 As shown in the figure, by comparing the data before and after adding noise, it can be clearly seen that the noise of the data after adding noise has largely submerged the effective signal, and conventional denoising is very difficult. At this time, the above-trained network is used to denoise the noisy data, and the denoised data obtained is as follows Figure 7 As shown by Figure 7 It can be seen that the noise is basically suppressed at this time, and the effective signal is restored. In order to verify the denoising effect of the network, the 20th single channel data of the noise-free data and the denoised data are extracted for comparison, as shown in Figure 8 As shown by Figure 8 It can be seen that the error between the denoised data and the label-free data is extremely small, and the energy and shape of the effective axis position are well restored, which verifies the denoising effect of the established network.

[0080] In order to further verify the denoising effect of the network of the present invention, actual data is further selected for testing. The input noisy data is as follows: Fig. 9As shown in the figure, it can be seen that the actual data has a low signal-to-noise ratio and serious noise development, which largely submerges the effective signal. The data is processed using the network obtained by the present invention, and the data obtained is as follows Fig.10 As shown by Fig.10 It can be seen that the random noise is well suppressed and the signal-to-noise ratio is greatly improved, which further verifies the processing effect of the intelligent denoising network obtained by the present invention.

[0081] Finally, it should be explained that the above technical solution is only one implementation method of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific implementation method of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

Claims

1. A method for suppressing random noise in prestack seismic data, characterized by: The method uses actual seismic data and pre-stack shot gather forward simulation data and GPU parallel mode to obtain an intelligent random noise suppression network, and uses the intelligent random noise suppression network to process actual pre-stack seismic data to obtain pre-stack seismic data after random noise suppression; The method comprises: (1) Using actual seismic data and pre-stack shot gather forward modeling data to prepare a data set, and divide the data set into a training data set and a validation data set; (2) designing a network, and using a GPU and a training data set to perform parallel training on the network to obtain a trained network; the network designed in step (2) is a U-NET network; (3) Using the validation data set to validate the trained network, a validated intelligent random noise suppression network is obtained; (4) collecting pre-stack seismic data, and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression; The operation of using the GPU and the training data set to perform parallel training on the network to obtain a trained network in step (2) includes: The CPU distributes the training data set to each GPU; Each GPU trains the U-NET network using the training data set, obtains a gradient value according to the training result, and then sends the gradient value to the CPU; The CPU reduces and averages the gradient values ​​sent by all GPUs to obtain the mean gradient, and applies the mean gradient to the loss function and optimizer to obtain the updated model. After multiple updates, the trained network is obtained.

2. The method for suppressing random noise in prestack seismic data according to claim 1, characterized in that: The actual seismic data in step (1) is actual seismic data with high-quality denoising effect selected from denoised seismic data of a typical exploration area; The pre-stack shot gather forward modeling data in step (1) is simulation data obtained after forward modeling of the pre-stack shot gather.

3. The method for suppressing random noise in prestack seismic data according to claim 2, characterized in that: The actual seismic data accounts for 40%, and the simulated data accounts for 60%.

4. The method for suppressing random noise in prestack seismic data according to claim 3, characterized in that: The operation of preparing a data set using actual seismic data and pre-stack shot gather forward modeling data in step (1) includes: Using the simulation data as label data; The noise data extracted from the actual seismic data is randomly matched with the simulated data and then summed to obtain the noisy data, or Gaussian random noise is added to the simulated data to obtain the noisy data; The noisy data and the labeled data constitute the data set.

5. The method for suppressing random noise in prestack seismic data according to claim 4, characterized in that: Write the dataset to a compressed file in the storage format of the MNIST dataset.

6. The method for suppressing random noise in prestack seismic data according to claim 1, characterized in that: The nearest neighbor interpolation algorithm is used for image upsampling in the upsampling part of the U-NET network.

7. The method for suppressing random noise in prestack seismic data according to claim 1, characterized in that: The loss function adopts L2 norm; The optimizer adopts the Adam optimizer.

8. A system for suppressing random noise of prestack seismic data, used to implement the method for suppressing random noise of prestack seismic data according to any one of claims 1 to 7, characterized in that: The system comprises: A data set preparation unit is used to prepare a data set using actual seismic data and pre-stack shot gather forward simulation data, and divide the data set into a training data set and a verification data set; A network design and training unit, connected to the data set preparation unit, is used to design a network and perform parallel training on the network using a GPU and a training data set to obtain a trained network; A verification unit, connected to the network design training unit, for verifying the trained network using a verification data set to obtain a verified intelligent random noise suppression network; The noise suppression unit is connected to the verification unit and is used to collect pre-stack seismic data and input the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.

Images