A Seismic Data Compressed Sensing Reconstruction Method Based on Fast Self-Supervised Learning
The self-supervised learning seismic data compressed sensing reconstruction method, utilizing deep convolutional neural networks and self-supervised learning techniques, solves the problems of data gaps and high costs in offshore oil and gas exploration, achieving efficient and low-cost data reconstruction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-12
- Publication Date
- 2026-04-03
AI Technical Summary
Existing seismic acquisition technologies suffer from high costs and serious data vulnerabilities in complex offshore oil and gas exploration, especially near drilling platforms or subsea pipelines. Traditional compressed sensing technology has insufficient reconstruction capabilities, relies on a large amount of training data, and has limited generalization capabilities.
A seismic data compressed sensing reconstruction method based on fast self-supervised learning is adopted. Self-supervised learning is performed through deep convolutional neural networks. Labeling and model training are carried out using the collected 3D seismic data with missing traces to reconstruct seismic data without missing traces. Feature extraction and reconstruction are performed using 3D gated convolutional layers, pixel rearrangement modules and self-supervised learning modules.
It achieves efficient data reconstruction without the need for additional training data, improves reconstruction quality and training efficiency, enables the acquisition of high-quality seismic data in a short time, and reduces acquisition costs and the impact of data vulnerabilities.
Smart Images
Figure CN117724152B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic acquisition technology, specifically to a method for compressed sensing reconstruction of seismic data based on fast self-supervised learning. Background Technology
[0002] Seismic exploration is an important means of offshore oil and gas exploration, and the foundation of seismic exploration is the acquisition of seismic data through field seismic acquisition. With the deepening of seismic exploration, most offshore seismic acquisition is now secondary 3D acquisition. There are two main problems in secondary 3D acquisition: (1) The refinement and complexity of targets lead to the upgrading of seismic acquisition parameter specifications and increased costs; (2) There are obstacles to acquisition and construction in the target work area, resulting in gaps in the seismic acquisition data.
[0003] Regarding the first question, using compressed sensing technology for seismic data acquisition can greatly reduce the workload of seismic acquisition, improve acquisition efficiency, and reduce costs. However, the data is irregular and needs to be reconstructed before it can be processed using conventional processing techniques.
[0004] Regarding the second issue, its main cause is that current seismic acquisition areas all contain drilling platforms or subsea pipelines to varying degrees. To avoid damaging these facilities during seismic acquisition operations, secondary 3D seismic acquisition often avoids these oilfield equipment, resulting in missing seismic data. This missing data affects the original quality of the seismic data, the suppression of multiple waves, and the migration imaging, severely hindering the seismic exploration process. Currently, a common method is to increase the density of shot points and receivers near obstacle areas to mitigate the impact of data loss to some extent, but the effect is often limited.
[0005] Compressed sensing algorithms can reconstruct regular seismic data from irregular or poorly acquired data. However, traditional compressed sensing techniques suffer from insufficient anti-spoofing reconstruction capabilities, severe reconstruction noise and artifacts; existing deep learning-based compressed sensing techniques are mostly two-dimensional algorithms and rely on large amounts of high-quality training data, resulting in limited generalization ability and practicality. Summary of the Invention
[0006] To address the aforementioned problems, the purpose of this invention is to provide a seismic data compressed sensing reconstruction method based on fast self-supervised learning. This method reconstructs randomly sampled 3D pre-stack data using compressed sensing data reconstruction technology, transforming irregular compressed sensing data into regular seismic data. This allows for subsequent processing of the seismic data using conventional seismic data processing techniques, or reduces the impact of data vulnerabilities caused by obstacles in conventional seismic acquisition on the overall acquisition quality.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A compressed sensing reconstruction method for seismic data based on fast self-supervised learning includes:
[0009] Labels were created using the collected 3D seismic data containing missing traces to form labeled data for training the network model.
[0010] Fast self-supervised learning deep convolutional neural network model training using labeled data;
[0011] Self-supervised data reconstruction was performed using a trained network model to obtain seismic data without missing data.
[0012] Furthermore, the process of creating labels using the acquired 3D seismic data containing missing traces to form labeled data for network model training includes:
[0013] Based on the acquired 3D seismic data with missing traces, a preset number of seismic label data pairs are randomly generated using the Bernoulli sampling method, including label input data and label output data.
[0014] Furthermore, the method of randomly generating a preset number of seismic label data pairs using Bernoulli sampling includes:
[0015]
[0016]
[0017] Where S is the required complete set of grid points, S a S is the set of grid point data collected under limited acquisition conditions. m For missing grid points; S i and S t These are the label input data and label output data that are divided in each iteration.
[0018] Furthermore, the training of the network model for fast self-supervised learning deep convolutional neural network using labeled data includes:
[0019] Build a deep convolutional neural network model;
[0020] Based on labeled data and a preset number of iterations, a deep convolutional neural network model is trained to obtain a deep convolutional neural network model that meets the preset requirements.
[0021] Furthermore, the deep convolutional neural network model includes:
[0022] The 3D gated convolutional layer module is used to extract features from the input data and obtain a 3D random missing feature map.
[0023] The 3D pixel rearrangement and de-rearrangement module is used to decode and encode 3D random missing feature maps, and to extract features and reconstruct missing data at different scales.
[0024] The self-supervised learning module is used to correct the network parameters in the 3D gated convolutional layer module and the 3D pixel rearrangement and de-rearrangement module, so that the output gradually converges to the label output.
[0025] Furthermore, the 3D gated convolutional layer module comprises two parts: 3D seismic data feature extraction and soft-gated feature extraction, and its formula is as follows:
[0026]
[0027]
[0028] O x,y,z =PReLU(F x,y,z )⊙σ(G x,y,z )+b
[0029]
[0030] In the formula, G x,y,z It is the 3D gated feature extraction part, W g These are its weight parameters, where I is the identity matrix; F x,y,z This is the feature extraction part of 3D seismic data, W f It is its weight parameter; O x,y,z This is the output; PreLU(y) i ) is the activation function, y i It is the input to the activation function, a i σ is the activation parameter; σ is the sigmoid function; b is the bias parameter.
[0031] Furthermore, the three-dimensional pixel rearrangement and de-rearrangement module is implemented using the permute and reshape matrix transformation functions in PyTorch.
[0032] Furthermore, the self-supervised learning module employs a masked norm 1 loss function based on PyTorch, the expression of which is:
[0033] loss = l1(o⊙M, t⊙M)
[0034] Where l1 represents the first norm, is the mean absolute error (MAE) of each element, o is the network output, t is the target seismic data, and M is the mask information matrix.
[0035] The present invention has the following advantages due to the adoption of the above technical solutions:
[0036] 1. This invention uses a self-supervised learning approach to train deep learning models, which does not require additional training data and breaks through generalization constraints;
[0037] 2. This invention uses a lightweight encoding / decoding and residual connection network structure, gated convolution, and pixel rearrangement module to improve training efficiency and achieve high performance in compressed sensing data reconstruction in a short time, which can be applied to actual production.
[0038] Therefore, this invention can be widely applied in the field of earthquake data acquisition and processing. Attached Figure Description
[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings:
[0040] Figure 1 This is a flowchart of the earthquake data compressed sensing reconstruction method based on fast self-supervised learning provided in an embodiment of the present invention;
[0041] Figure 2 This is the lightweight self-supervised learning deep learning network model structure provided in the embodiments of the present invention;
[0042] Figure 3a , Figure 3b and Figure 3c These are, respectively, the complete 'hyperbolic' data, the data after simulated random compressed sensing sampling, and the processing results of the method proposed in this invention;
[0043] Figure 4a , Figure 4b and Figure 4c These are, respectively, the complete 'zha' data, the data after simulated random compressed sensing sampling, and the processing result of the method proposed in this invention. Detailed Implementation
[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.
[0045] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0046] In some embodiments of the present invention, a method for compressed sensing reconstruction of seismic data based on fast self-supervised learning is provided, comprising: designing a self-supervised learning data processing scheme and determining the generation methods of input and target data during training; designing a three-dimensional gated convolutional layer to support the processing of three-dimensional random missing feature maps; preferentially selecting a masked L1 loss function, Adam optimizer, encoding / decoding network structure, and residual connection structure to improve the efficiency of self-supervised learning; and designing a three-dimensional pixel rearrangement and de-rearrangement module to achieve fast and highly flexible feature map decoding and encoding operations. The present invention does not require synthesized training data and can directly perform self-supervised learning on target random subsampled data, thus it can be widely applied in the field of compressed sensing reconstruction of seismic data.
[0047] Example 1
[0048] like Figure 1 As shown in the figure, this embodiment provides a seismic data compressed sensing reconstruction method based on fast self-supervised learning, which includes the following steps:
[0049] 1) Labels are created using the collected 3D seismic data containing missing traces to form label data for network model training;
[0050] 2) Utilize labeled data for fast self-supervised learning deep convolutional neural network model training;
[0051] 3) Use the trained network model to perform self-supervised data reconstruction to obtain seismic data without missing data.
[0052] Preferably, in step 1) above, the three-dimensional seismic data includes three dimensions: Xline, Inline, and Time. These three dimensions are represented by [x, y, z], and their three-dimensional size is described by [h, w, d], where z corresponds to the Time dimension. Three-dimensional seismic data with missing traces refers to seismic data that is randomly missing in the x and y dimensions. Assuming a seismic trace determined by the x and y coordinates has d sample points, then the missing data refers to the complete loss of all d sample points.
[0053] When creating labels: Based on the acquired 3D seismic data with missing traces, a preset number of seismic label data pairs are randomly generated using the Bernoulli sampling method, including the label input data and the label output data.
[0054] Specifically, Bernoulli sampling is performed on the grid points in the non-missing parts of the x and y dimensions as input data, and the remaining parts are used as the target data for this training output.
[0055]
[0056]
[0057] Where S is the required complete set of grid points, S a S is the set of grid point data collected under limited acquisition conditions. m For missing grid points; S i and S t These refer to the input data and target data, respectively, defined in each iteration. In this embodiment, one training iteration consists of 5000 iterations, i.e., 5000 corresponding S... i and S t .
[0058] Preferably, in step 2) above, training a network model for a fast self-supervised deep convolutional neural network using labeled data includes the following steps:
[0059] 2.1) Construct a deep convolutional neural network model;
[0060] 2.2) Based on the labeled data and the preset number of iterations, the deep convolutional neural network model is trained to obtain a deep convolutional neural network model that meets the preset requirements.
[0061] Preferably, in step 2.1) above, such as Figure 2 As shown, the deep convolutional neural network model in this embodiment includes: a 3D gated convolutional layer module, a 3D pixel rearrangement and de-rearrangement module, and a self-supervised learning module. The 3D gated convolutional layer module is used to extract features from the input data to obtain a 3D randomly missing feature map; the 3D pixel rearrangement and de-rearrangement module is used to achieve fast and highly malleable 3D feature map decoding and encoding operations, thereby enabling feature extraction and missing data reconstruction at different scales; the self-supervised learning module is used to improve the efficiency of self-supervised learning and continuously correct the network parameters, so that the output gradually converges to the label output.
[0062] Preferably, the 3D gated convolutional layer module comprises two parts: 3D seismic data feature extraction (corresponding to ordinary convolution) and soft-gated feature extraction, with the following formula:
[0063] G x,y,z =∑∑∑W g ·I (3)
[0064] F x,y,z=∑∑∑W f ·I (4)
[0065] O x,y,z =PReLU(F x,y,z )⊙σ(G x,y,z )+b (5)
[0066]
[0067] In the formula, G x,y,z It is the 3D gated feature extraction part, W g These are its weight parameters, where I is the identity matrix; F x,y,z This is the feature extraction part of 3D seismic data, W f It is its weight parameter; O x,y,z This is the output; PreLU(y) i ) is the activation function, y i It is the input to the activation function, a i σ is the activation parameter; σ is the sigmoid function; b is the bias parameter. The 3D seismic data feature map is processed with parameter a. i After activation by the PReLU activation function, the layer is point-by-point integrated with the 3D gated feature map (⊙) to achieve gated activation. The bias parameter b is then added to obtain the output O of this layer. x,y,z The 3D gated features are processed by the sigmoid function to ensure that the gated feature values are between 0 and 1, with 1 corresponding to activation and 0 corresponding to deactivation.
[0068] In other words, the network parameters that need to be learned in the 3D gated convolutional layer module include the weight parameters W for 3D gated feature extraction. g Weighting parameter W for feature extraction from 3D seismic data f The parameter a in the PReLU activation function i And the bias parameter b.
[0069] Preferably, the 3D pixel rearrangement and de-rearrangement module is implemented using matrix transformation functions such as permute and reshape in PyTorch.
[0070] Specifically, the 3D pixel rearrangement operation, referencing the `torch.nn.PixelShuffle` class in PyTorch, can achieve upsampling of the 3D feature map. Assuming the upsampling coefficient is `s`, the 3D pixel rearrangement operation designed in this embodiment transforms the feature map size from [b,c,h,w,d] to [b,c / s]. 3 [b,c,h,w,ds]; Correspondingly, the designed 3D pixel inverse rearrangement operation can transform the feature map size from [b,c,h,w,d] to [b,cs]. 3[h / s, w / s, d / s]. Successive pixel rearrangement and de-rearrangement operations can cancel each other out.
[0071] Preferably, the self-supervised learning module mainly includes a masked L1 loss function, an Adam optimizer, an encoder-decoder network structure, and a residual network structure. In this embodiment, a masked L1 loss function based on PyTorch is used, and its expression is as follows:
[0072] loss=l1(o⊙M,t⊙M) (7)
[0073] Here, l1 represents the L1 norm, is the mean absolute error (MAE) of each element, o is the network output, t is the target seismic data, and M is the mask information matrix. Using the mask information matrix, the reconstruction results of the target region can be optimized in a targeted manner, while other regions are masked. Thanks to the translation invariance of convolution and random training, even the masked regions can be optimized.
[0074] Preferably, in each training iteration, in the mask information matrix M, S t The pixels in the set are 1, and the rest are 0.
[0075] In step 2.2) above, the input data is S in each iteration of the model. i The training objective is S t S is fitted using a masked norm loss function. t By continuously updating the network model parameters, the output gradually moves towards the training target S. t convergence.
[0076] Thus, the seismic data compressed sensing reconstruction method based on fast self-supervised learning, as described in this invention, is realized. This method can achieve seismic data compressed sensing reconstruction without introducing additional training data through self-supervised learning. Furthermore, this invention introduces modules such as 3D gated convolution and 3D pixel rearrangement to enable rapid training of the algorithm.
[0077] Example 2
[0078] Figure 3a The complete synthetic data, hyperbolic, has dimensions [32, 32, 126]. It is randomly sampled at a 50% ratio to simulate the actual compressed sensing sampling process, yielding... Figure 3b . Figure 3b The downsampling results shown are consistent with Figure 3a The signal-to-noise ratio (SNR) of the original data shown is 3.01 dB.
[0079] The earthquake data compressed sensing reconstruction method based on fast self-supervised learning proposed in this invention is used in... Figure 3b After training on the downsampling results shown, obtain Figure 3c The compressed sensing reconstruction results are shown. The signal-to-noise ratio of the obtained results to the original data is 22.06 dB (i.e., this invention improves the SNR by 18 dB), and the total training and inference time is 56.6 seconds.
[0080] Example 3
[0081] Figure 4a The complete synthetic data zha has dimensions [256, 256, 256]. It is randomly sampled at a 50% ratio to simulate the actual compressed sensing sampling process, yielding... Figure 4b . Figure 4b The downsampling results shown are consistent with Figure 4a The signal-to-noise ratio (SNR) of the original data shown is 3.03 dB.
[0082] The seismic data compressed sensing reconstruction method based on fast self-supervised learning of this invention is used in... Figure 4b After training on the downsampling results shown, obtain Figure 4c The compressed sensing reconstruction results are shown. The signal-to-noise ratio (SNR) of the result obtained using the method of this invention compared to the original data is 21.25 dB (i.e., this invention improves the SNR by 17.2 dB). Due to the large size of the data, the total training and inference time was 49 minutes.
[0083] The two examples above use Ubuntu system servers with a 13th generation i9-13900K CPU, an NVIDIA GeForce RTX 4090 GPU, and 128GB of RAM.
[0084] In summary, the seismic data compressed sensing reconstruction method based on fast self-supervised learning disclosed in this invention can reconstruct complete data with relatively little data acquisition. This invention uses deep learning methods to achieve better reconstruction results, but employs self-supervised learning, eliminating the need for additional training data. Furthermore, this invention uses lightweight encoding / decoding and residual connection network structures, gated convolutions, and pixel rearrangement modules to improve training efficiency, achieving high performance in compressed sensing data reconstruction in a shorter time, and enabling its application in practical production.
[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for compressed sensing reconstruction of seismic data based on fast self-supervised learning, characterized in that... include: Labels were created using the collected 3D seismic data containing missing traces to form labeled data for training the network model. Fast self-supervised learning deep convolutional neural network model training using labeled data; Self-supervised data reconstruction was performed using a trained network model to obtain seismic data without missing data. The training of a network model for fast self-supervised learning of a deep convolutional neural network using labeled data includes: Build a deep convolutional neural network model; Based on labeled data and a preset number of iterations, a deep convolutional neural network model is trained to obtain a deep convolutional neural network model that meets the preset requirements. The deep convolutional neural network model includes: The 3D gated convolutional layer module is used to extract features from the input data and obtain a 3D random missing feature map. The 3D pixel rearrangement and de-rearrangement module is used to decode and encode 3D random missing feature maps, and to extract features and reconstruct missing data at different scales. The self-supervised learning module is used to correct the network parameters in the 3D gated convolutional layer module and the 3D pixel rearrangement and de-rearrangement module, so that the output gradually converges to the label output.
2. The seismic data compressed sensing reconstruction method based on fast self-supervised learning as described in claim 1, characterized in that, The process of creating labels using acquired 3D seismic data containing missing traces to form labeled data for network model training includes: Based on the acquired 3D seismic data with missing traces, a preset number of seismic label data pairs are randomly generated using the Bernoulli sampling method, including label input data and label output data.
3. The seismic data compressed sensing reconstruction method based on fast self-supervised learning as described in claim 2, characterized in that, The method of randomly generating a preset number of seismic label data pairs using Bernoulli sampling includes: in, For the required complete set of grid points, This is a collection of grid point data acquired under limited acquisition conditions. For missing grid points; and These are the label input data and label output data that are divided in each iteration.
4. The seismic data compressed sensing reconstruction method based on fast self-supervised learning as described in claim 1, characterized in that, The 3D gated convolutional layer module comprises two parts: 3D seismic data feature extraction and soft-gated feature extraction. The formula is as follows: In the formula, It is the 3D gated feature extraction part. It is its weight parameter. It is the identity matrix; This is the feature extraction part of 3D seismic data. It is its weight parameter; It is the output; It is an activation function. It is the input to the activation function. These are activation parameters; It is the sigmoid function; It is a bias parameter.
5. The seismic data compressed sensing reconstruction method based on fast self-supervised learning as described in claim 1, characterized in that, The 3D pixel rearrangement and de-rearrangement module is implemented using the permute and reshape matrix transformation functions in PyTorch.
6. The seismic data compressed sensing reconstruction method based on fast self-supervised learning as described in claim 1, characterized in that, The self-supervised learning module employs a masked norm loss function based on PyTorch, the expression of which is: in, Let represent the L1 norm, represent the mean absolute error (MAE) of each element, o represent the network output, t represent the target seismic data, and M represent the mask information matrix.