A method and system for seismic data reconstruction
By randomly reordering missing earthquake data and training with deep learning, combined with optimized trace sets and transfer learning, the problem of insufficient accuracy in reconstruction of large-area missing earthquake data by traditional CNNs is solved, and higher accuracy data reconstruction is achieved.
Patent Information
- Application Number
- CN202311260827.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-27
- Publication Date
- 2025-11-18
- Estimated Expiration
- 2043-09-27
AI Technical Summary
Traditional CNNs suffer from insufficient reconstruction accuracy when processing large areas of continuously missing seismic data, especially poor interpolation results for low-amplitude missing components.
By randomly reordering missing seismic data, a random coding network and reconstruction model are constructed. The model is trained using a deep learning network, and combined with optimized trace gathers and transfer learning methods, the reconstruction distance is reduced and the reconstruction accuracy is improved.
It effectively improves the reconstruction accuracy of large-area missing seismic data, ensures the matching of reconstructed data with label data, and adapts to the geological characteristics of different work areas.
Smart Images
Figure CN119716973B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of earthquake data processing technology, and specifically relates to an earthquake data reconstruction method and system. Background Technology
[0002] In recent years, increasingly complex exploration environments, the promotion of "two-wide and one-high" seismic exploration technology, and the impact of large-area obstacles have often resulted in large-scale missing seismic data, which seriously affects the fidelity, resolution, and signal-to-noise ratio of subsequent data processing, thereby reducing the reliability of the final data interpretation. Therefore, it is necessary to reconstruct incomplete or irregular seismic data to improve the integrity of seismic data.
[0003] With the development of deep learning, many neural networks with different structures have been applied to interpolation of missing seismic data. Among them, CNNs can effectively interpolate small-gap seismic traces with both regular and irregular gaps. However, for seismic traces with large gaps and continuous gaps, the interpolation results for low-amplitude missing components need further improvement.
[0004] In the single-stage training process of traditional CNNs for interpolating continuously missing seismic traces, the model tends to interpolate the missing traces that are closer to the non-missing traces first. The reconstruction accuracy of missing traces that are farther away from the non-missing traces is worse. Therefore, it can achieve better reconstruction accuracy for seismic traces with small intervals of missing traces, but it is not enough to handle large-area missing seismic data well. Summary of the Invention
[0005] The purpose of this invention is to solve the problems existing in the prior art and provide a method for reconstructing seismic data.
[0006] This invention is achieved through the following technical solution: a seismic data reconstruction method, comprising the following steps:
[0007] Obtain complete earthquake data and corresponding missing earthquake data;
[0008] The missing seismic data is cut into several missing data pieces, and the complete seismic data is correspondingly cut into an equal number of intact data pieces.
[0009] The missing data slices are randomly reordered to obtain randomly sorted data slices;
[0010] The label data is obtained by performing the same random reordering on the complete data slices as on the missing data slices;
[0011] The model is reconstructed using samples constructed from the randomly sorted data slices and labeled data;
[0012] Generate a randomized data patch of missing seismic data to be tested, wherein the randomized data patch has the same random reordering as the sample.
[0013] Input the randomly sorted data slices to be tested into the trained reconstruction model, and output the reconstructed data slices.
[0014] Furthermore, missing earthquake data is created based on complete earthquake data: complete earthquake data is obtained through numerical simulation, and then missing traces are marked on the complete earthquake data according to the missing information of historical actual data or randomly, and the data of the missing traces are filled with 0 values, thus obtaining missing earthquake data.
[0015] Furthermore, the same random reordering is performed on the preferred trace sets of both the missing and complete seismic data. The preferred trace set is the trace set with the smallest reconstruction distance. For a single missing trace, the reconstruction distance is the number of traces between the missing trace and the nearest non-missing trace. For a trace set, the reconstruction distance is the maximum value of the reconstruction distances of all missing traces in that trace set.
[0016] Furthermore, the steps to obtain the optimal set of paths are as follows:
[0017] Missing seismic data are converted to different types of gathers, the reconstruction distances of each gather are compared, and the gather with the smallest reconstruction distance is selected as the preferred gather for missing seismic data; the preferred gather type for complete seismic data is the same as that for missing seismic data.
[0018] Furthermore, the missing seismic data is converted to gathers including common shot point gathers, common receiver gathers, common center point gathers, common offset gathers, and common signal gathers.
[0019] Furthermore, the random reordering steps include: assigning a random number to each data point, and then reordering them according to the order of the random numbers.
[0020] Furthermore, deep learning networks are used as the reconstruction model.
[0021] Furthermore, after outputting the reconstructed data fragments of the missing seismic data to be measured, the data fragments are reverse-randomly reordered to obtain complete reconstructed data fragments arranged in the original order, and then the complete reconstructed data fragments are spliced together to form complete reconstructed seismic data.
[0022] Furthermore, after training the reconstruction model using samples obtained through numerical simulation, the reconstruction model is then fine-tuned and trained using samples constructed based on actual data through transfer learning.
[0023] Furthermore, the steps for constructing samples based on actual data include: collecting actual complete seismic data, marking missing traces on the actual complete seismic data according to the missing information of historical actual data or randomly, filling the missing trace data with 0 values, and using it as the actual missing seismic data.
[0024] This invention also provides a seismic data reconstruction system for implementing the seismic data reconstruction method of this invention, including a random coding network, a reconstruction model, and a splicing module. The random coding network is used to randomly reorder the input seismic data and output randomly ordered data slices to the reconstruction model. The reconstruction model is used to output reconstructed data slices based on the randomly ordered data slices. The splicing module is used to reverse the random reordering of the reconstructed data slices to obtain complete reconstructed data slices arranged in the original order, and then splice the complete reconstructed data slices into complete reconstructed seismic data.
[0025] Compared with the prior art, the beneficial effects of the present invention include:
[0026] 1. This invention randomly reorders missing seismic data, making the missing traces more dispersed and preventing them from clustering into large areas of missing data, thereby reducing the reconstruction distance. For the reconstruction model, the smaller the reconstruction distance, the closer the reconstructed data is to the labeled data, thus improving the reconstruction accuracy.
[0027] 2. This invention creates missing earthquake data based on complete earthquake data, obtains complete earthquake data through numerical simulation, ensures the quality of complete earthquake data, and ensures the matching between complete earthquake data and missing earthquake data.
[0028] 3. Marking missing traces on complete seismic data based on the missing information in historical data can reduce the number of training samples; randomly marking missing traces on complete seismic data can enrich the training samples and cover various missing situations.
[0029] 4. By optimizing the gathers, the present invention can further reduce the reconstruction distance, thereby further improving the reconstruction accuracy.
[0030] 5. This invention uses samples constructed based on actual data to fine-tune and reconstruct the model through transfer learning, which can obtain a model that is more suitable for the current work area. Attached Figure Description
[0031] Figure 1 A data flow diagram for seismic data reconstruction methods;
[0032] Figure 2 A flowchart for training the reconstruction model;
[0033] Figure 3 This is a flowchart for reconstructing data using a trained reconstruction model. Detailed Implementation
[0034] In the single-stage training process of traditional CNNs for interpolating continuously missing seismic traces, the model tends to interpolate the missing traces that are closer to the non-missing traces first. The reconstruction accuracy of missing traces that are farther away from the non-missing traces is worse. Therefore, it can achieve better reconstruction accuracy for seismic traces with small intervals of missing traces, but it is not enough to handle large-area missing seismic data well.
[0035] This invention reduces the reconstruction distance by randomly reordering missing seismic data, thus dispersing the missing traces and preventing them from clustering into large areas of missing data. For the reconstruction model, the smaller the reconstruction distance, the closer the reconstructed data is to the labeled data, thereby improving reconstruction accuracy.
[0036] The present invention will now be described in further detail with reference to the accompanying drawings:
[0037] refer to Figure 1 As shown, an earthquake data reconstruction system is used to implement the earthquake data reconstruction method of the present invention. It includes a random coding network, a reconstruction model, and a splicing module. The random coding network is used to randomly reorder the input earthquake data and output randomly ordered data slices to the reconstruction model. The reconstruction model is used to output reconstructed data slices based on the randomly ordered data slices. The splicing module is used to reverse the random reordering of the reconstructed data slices to obtain complete reconstructed data slices arranged in the original order, and then splices the complete reconstructed data slices into complete reconstructed earthquake data.
[0038] The seismic data reconstruction method of the present invention includes the following steps:
[0039] Obtain complete earthquake data and corresponding missing earthquake data;
[0040] The missing seismic data is cut into several missing data pieces, and the complete seismic data is correspondingly cut into an equal number of intact data pieces.
[0041] The missing data slices are randomly reordered to obtain randomly sorted data slices;
[0042] The label data is obtained by performing the same random reordering on the complete data slices as on the missing data slices;
[0043] The model is reconstructed using samples constructed from the randomly sorted data slices and labeled data;
[0044] Generate a randomized data patch of missing seismic data to be tested, wherein the randomized data patch has the same random reordering as the sample.
[0045] Input the randomly sorted data slices to be tested into the trained reconstruction model, and output the reconstructed data slices.
[0046] The following is a more detailed explanation.
[0047] 1) Creating the training dataset
[0048] Missing earthquake data is created based on complete earthquake data: Complete earthquake data is obtained through numerical simulation, and then missing traces are marked on the complete earthquake data according to the missing information of historical actual data or randomly. The data of the missing traces are filled with 0 values, thus obtaining missing earthquake data.
[0049] The missing seismic data is randomly reordered to obtain a randomly sorted data patch. The steps of random reordering include: assigning a random number to each data point, and then reordering them according to the order of the random numbers.
[0050] Preferably, the preferred trace sets of the missing seismic data and the complete seismic data are randomly reordered. The preferred trace set is the trace set with the smallest reconstruction distance. For a single missing trace, the reconstruction distance is the number of traces between the missing trace and the nearest non-missing trace. For a trace set, the reconstruction distance is the maximum reconstruction distance of all missing traces in the trace set.
[0051] The steps to obtain the optimal set of paths are as follows:
[0052] Missing seismic data are converted to different types of gathers, the reconstruction distances of each gather are compared, and the gather with the smallest reconstruction distance is selected as the preferred gather for missing seismic data; the preferred gather type for complete seismic data is the same as that for missing seismic data.
[0053] The missing seismic data was converted to a collection that includes common shot point gathers, common receiver gathers, common center point gathers, common offset gathers, and common trace number gathers.
[0054] Label data is obtained by randomly reordering complete earthquake data in the same way as missing earthquake data.
[0055] Samples are constructed based on randomly sorted data slices and label data. That is, the randomly sorted data slices are used as input data for the reconstruction model, and the model parameters are continuously updated by the reconstruction error between the reconstruction data output by the reconstruction model and the label data.
[0056] II) Training and Reconstructing the Model
[0057] The sample set consisting of the above samples is divided into a training set, a validation set, and a test set.
[0058] Build a suitable reconstruction model according to the requirements and use deep learning networks, including convolutional neural networks, long short-term recurrent neural networks, etc.
[0059] Use the prepared training set to debug and determine the hyperparameters for network training;
[0060] refer to Figure 2As shown, the network is trained according to the prepared training set and hyperparameters;
[0061] Training stops when the set number of iterations is met or the reconstruction error drops to a specified value.
[0062] Save the trained reconstructed model.
[0063] Preferably, after training the reconstruction model using samples obtained through numerical simulation, the reconstruction model is then fine-tuned and trained using samples constructed based on actual data through transfer learning.
[0064] The steps for constructing samples based on actual data include: collecting actual complete seismic data, marking missing traces on the actual complete seismic data according to the missing information of historical actual data or randomly, filling the missing traces with 0 values, and using them as actual missing seismic data.
[0065] III) Data Restructuring
[0066] refer to Figure 3 As shown, a randomly sorted data patch of the missing earthquake data to be tested is generated and input into the trained reconstruction model, and the reconstructed data patch is output. The random sorted data patch of the missing earthquake data to be tested is generated in the same way as the training samples, that is, the missing earthquake data to be tested is sliced, each data point is assigned a random number, and then the data is reordered according to the order of the random numbers.
[0067] The number of randomized data pieces to be tested is equal to the number of randomized data pieces in the sample, and their sizes are the same. The randomized data pieces to be tested have the same random numbering for the corresponding channels as the randomized data pieces in the sample, which means that the randomized data pieces to be tested have the same random reordering as the sample.
[0068] After outputting reconstructed data fragments of the missing seismic data to be measured, a reverse random reordering process is performed to obtain complete reconstructed data fragments arranged in the original order. These complete reconstructed data fragments are then stitched together to form complete reconstructed seismic data. The reverse random reordering process is the inverse of the random reordering process; first, the original numbers are obtained through reverse random numbering, and then the data is sorted according to the original numbering order.
[0069] Example 1
[0070] In this example, simulated data is used for pre-training first. After training, a small amount of real data is used for fine-tuning training through transfer learning to finally obtain a model suitable for the current work area.
[0071] 1. Create the training dataset:
[0072] 1.1. Based on the geological model of the work area and the geological tasks, design an observation system, and then conduct numerical simulation to obtain a complete seismic dataset.
[0073] 1.2. Based on the distribution of obstacles in the work area, mark the corresponding locations in the complete seismic data and fill missing traces with 0 values.
[0074] 1.3 In this example, only the case of missing receiver points is handled. For large-area missing receiver points, the missing seismic data is converted to common receiver gathers, common shot gathers, common center gathers, common offset gathers, and common trace number gathers. The reconstruction distance of each gather is calculated. In this example, the reconstruction distance of the common shot gather is the smallest, so the missing seismic data is converted to the common shot gather.
[0075] 1.4. Perform the same random coding, rearrangement and recording on each gather of the common shot point gather seismic data.
[0076] 1.5 Divide the above earthquake data into training set, validation set and test set.
[0077] 1.6 The process of creating the training set for actual data is similar to steps 1.2-1.5, and will not be described in detail.
[0078] 2. Deep learning network training:
[0079] 2.1 In this example, the commonly used U-net network structure with attention mechanism is selected for training;
[0080] 2.2 Use the prepared training set to debug and determine the hyperparameters for network training;
[0081] 2.3. Train the network according to the prepared training set and hyperparameters;
[0082] 2.4 Save the trained network model.
[0083] 3. Deep learning reconstruction part
[0084] 3.1. Reconstruct the data using the trained random coding network. Preprocess the missing seismic data and create random coded data slices and corresponding masks according to the process in step 1;
[0085] 3.2 Input the data slices and corresponding masks into the trained network model to output the reconstructed seismic data;
[0086] 3.3. Perform unrandom reordering on the reconstructed data;
[0087] 3.4. Stitch the output seismic data to the original size.
[0088] Example 2
[0089] In this example, simulated data is used for pre-training first. After training, a small amount of real data is used for fine-tuning training through transfer learning to finally obtain a model suitable for the current work area.
[0090] 1. Create the training dataset:
[0091] 1.1. Based on the geological model of the work area and the geological tasks, design an observation system, and then conduct numerical simulation to obtain a complete seismic dataset.
[0092] 1.2. Based on the distribution of obstacles in the work area, mark the corresponding locations in the complete seismic data and fill missing traces with 0 values.
[0093] 1.3 In this example, only the case of missing shot points is considered. For large-area missing shot points, the missing seismic data is converted to common receiver gathers, common shot gathers, common center gathers, common offset gathers, and common trace number gathers. The reconstruction distance of each gather is calculated. In this example, the reconstruction distance of the common receiver gather is the smallest, so the missing seismic data is converted to the common receiver gather.
[0094] 1.4. Perform the same random coding, rearrangement and recording on each gather of the common shot point gather seismic data.
[0095] 1.5 Divide the above earthquake data into training set, validation set and test set.
[0096] 1.6 The process of creating the training set for actual data is similar to steps 1.2-1.5, and will not be described in detail.
[0097] 2. Deep learning network training:
[0098] 2.1 In this example, a Convolutional Neural Network (CNN) is used for training;
[0099] 2.2 Use the prepared training set to debug and determine the hyperparameters for network training;
[0100] 2.3. Train the network according to the prepared training set and hyperparameters;
[0101] 2.4 Save the trained network model.
[0102] 3. Deep learning reconstruction part
[0103] 3.1. Reconstructing data using a pre-trained random coding network, such as... Figure 3 As shown. The missing seismic data is preprocessed, and randomized coded data slices and corresponding masks are created according to the process in step 1.
[0104] 3.2 Input the data slices and corresponding masks into the trained network model to output the reconstructed seismic data;
[0105] 3.3. Perform unrandom reordering on the reconstructed data;
[0106] 3.4. Stitch the output seismic data to the original size.
[0107] Example 3
[0108] In this example, simulated data is used for pre-training first. After training, a small amount of real data is used for fine-tuning training through transfer learning to finally obtain a model suitable for the current work area.
[0109] 1. Create the training dataset:
[0110] 1.1. Based on the geological model of the work area and the geological tasks, design an observation system, and then conduct numerical simulation to obtain a complete seismic dataset.
[0111] 1.2. Based on the distribution of obstacles in the work area, mark the corresponding locations in the complete seismic data and fill missing traces with 0 values.
[0112] 1.3 In this example, only the case of missing offsets is considered. For large-area missing offsets, the missing seismic data is converted to common receiver gathers, common shot gathers, common center gathers, common offset gathers, and common trace number gathers. The reconstruction distance of each gather is calculated. In this example, the reconstruction distance of the common center gather is the smallest, so the missing seismic data is converted to the common center gather.
[0113] 1.4. Perform the same random coding, rearrangement and recording on each gather of the common shot point gather seismic data.
[0114] 1.5 Divide the above earthquake data into training set, validation set and test set.
[0115] 1.6 The process of creating the training set for actual data is similar to steps 1.2-1.5, and will not be described in detail.
[0116] 2. Deep learning network training:
[0117] 2.1 In this example, a recurrent neural network (RNN) is used for training;
[0118] 2.2 Use the prepared training set to debug and determine the hyperparameters for network training;
[0119] 2.3. Train the network according to the prepared training set and hyperparameters;
[0120] 2.4 Save the trained network model.
[0121] 3. Deep learning reconstruction part
[0122] 3.1. Reconstructing data using a pre-trained random coding network, such as... Figure 3 As shown. The missing seismic data is preprocessed, and randomized coded data slices and corresponding masks are created according to the process in step 1.
[0123] 3.2 Input the data slices and corresponding masks into the trained network model to output the reconstructed seismic data;
[0124] 3.3. Perform unrandom reordering on the reconstructed data;
[0125] 3.4. Stitch the output seismic data to the original size.
[0126] Example 4
[0127] In this example, simulated data is used for pre-training first. After training, a small amount of real data is used for fine-tuning training through transfer learning to finally obtain a model suitable for the current work area.
[0128] 1. Create the training dataset:
[0129] 1.1. Based on the geological model of the work area and the geological tasks, design an observation system, and then conduct numerical simulation to obtain a complete seismic dataset.
[0130] 1.2. Based on the distribution of obstacles in the work area, mark the corresponding locations in the complete seismic data and fill missing traces with 0 values.
[0131] 1.3 In this example, only the case of missing center points is handled. For large areas of missing center points, the missing seismic data is converted to common receiver point gathers, common shot point gathers, common center point gathers, common offset gathers, and common trace number gathers. The reconstruction distance of each gather is calculated. In this example, the reconstruction distance of the common offset gather is the smallest, so the missing seismic data is converted to the common offset gather.
[0132] 1.4. Perform the same random coding, rearrangement and recording on each gather of the common shot point gather seismic data.
[0133] 1.5 Divide the above earthquake data into training set, validation set and test set.
[0134] 1.6 The process of creating the training set for actual data is similar to steps 1.2-1.5, and will not be described in detail.
[0135] 2. Deep learning network training:
[0136] 2.1 In this example, the Long Short-Term Memory (LSTM) network is selected for training;
[0137] 2.2 Use the prepared training set to debug and determine the hyperparameters for network training;
[0138] 2.3. Train the network according to the prepared training set and hyperparameters;
[0139] 2.4 Save the trained network model.
[0140] 3. Deep learning reconstruction part
[0141] 3.1. Reconstructing data using a pre-trained random coding network, such as... Figure 3 As shown. The missing seismic data is preprocessed, and randomized coded data slices and corresponding masks are created according to the process in step 1.
[0142] 3.2 Input the data slices and corresponding masks into the trained network model to output the reconstructed seismic data;
[0143] 3.3. Perform unrandom reordering on the reconstructed data;
[0144] 3.4. Stitch the output seismic data to the original size.
[0145] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc.
[0146] The above technical solutions are merely specific embodiments of the present invention. For those skilled in the art, based on the principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the technical solutions described in the above specific embodiments of the present invention. Therefore, the foregoing descriptions are only preferred and not restrictive.
Claims
1. A seismic data reconstruction method, characterized in that, Includes the following steps: Obtain complete earthquake data and corresponding missing earthquake data; The missing seismic data is cut into several missing data pieces, and the complete seismic data is correspondingly cut into an equal number of intact data pieces. The missing data slices are randomly reordered to obtain randomly sorted data slices; The label data is obtained by performing the same random reordering on the complete data slices as on the missing data slices; The model is reconstructed using samples constructed from the randomly sorted data slices and labeled data; Generate a randomized data patch of missing seismic data to be tested, wherein the randomized data patch has the same random reordering as the sample. Input the randomly sorted data slices to be tested into the trained reconstruction model, and output the reconstructed data slices.
2. The seismic data reconstruction method according to claim 1, characterized in that, Missing earthquake data is created based on complete earthquake data: Complete earthquake data is obtained through numerical simulation, and then missing traces are marked on the complete earthquake data according to the missing information of historical actual data or randomly. The data of the missing traces are filled with 0 values, thus obtaining missing earthquake data.
3. The seismic data reconstruction method according to claim 1, characterized in that, Both the missing and complete seismic data are randomly reordered using their respective preferred trace sets. The preferred trace set is the one with the smallest reconstruction distance. For a single missing trace, the reconstruction distance is the number of traces between the missing trace and the nearest non-missing trace. For a trace set, the reconstruction distance is the maximum reconstruction distance of all missing traces in that trace set.
4. The seismic data reconstruction method according to claim 3, characterized in that: The steps to obtain the optimal set of paths are as follows: Missing seismic data are converted to different types of gathers, the reconstruction distances of each gather are compared, and the gather with the smallest reconstruction distance is selected as the preferred gather for missing seismic data; the preferred gather type for complete seismic data is the same as that for missing seismic data.
5. The seismic data reconstruction method according to claim 4, characterized in that, The missing seismic data was converted to a collection that includes common shot point gathers, common receiver gathers, common center point gathers, common offset gathers, and common trace number gathers.
6. The seismic data reconstruction method according to claim 1, characterized in that, The steps of random reordering include: assigning a random number to each data point, and then reordering them according to the order of the random numbers.
7. The seismic data reconstruction method according to claim 1, characterized in that, A deep learning network is used as the reconstruction model.
8. The seismic data reconstruction method according to claim 1, characterized in that, After outputting the reconstructed data fragments of the missing seismic data to be measured, the data fragments are reverse-randomly reordered to obtain complete reconstructed data fragments arranged in the original order. Then, the complete reconstructed data fragments are spliced together to form complete reconstructed seismic data.
9. The seismic data reconstruction method according to claim 1, characterized in that, After training the reconstruction model using samples obtained through numerical simulation, the reconstruction model is then fine-tuned and trained using samples constructed based on actual data through transfer learning.
10. The seismic data reconstruction method according to claim 9, characterized in that, The steps for constructing samples based on actual data include: collecting actual complete seismic data, marking missing traces on the actual complete seismic data according to the missing information of historical actual data or randomly, filling the missing traces with 0 values, and using them as actual missing seismic data.
11. A seismic data reconstruction system, characterized in that, The method for reconstructing seismic data as described in any one of claims 1 to 10 includes a random coding network, a reconstruction model, and a splicing module. The random coding network is used to randomly reorder the input seismic data and output randomly ordered data slices to the reconstruction model. The reconstruction model is used to output reconstructed data slices based on the randomly ordered data slices. The splicing module is used to reverse the random reordering of the reconstructed data slices to obtain complete reconstructed data slices arranged in the original order, and then splices the complete reconstructed data slices into complete reconstructed seismic data.
Citation Information
Patent Citations
Vertical seismic profiling seismic data regularization reconstruction method and device and storage medium
CN111123361A
Three-dimensional seismic data reconstruction method and device
CN113093273A