First arrival wave pickup method and device based on deep learning
By constructing a first arrival picking model based on the SA-DeepLab network using deep learning, the problems of low accuracy and low efficiency in first arrival picking are solved, achieving efficient and accurate first arrival picking, and adapting to complex seismic data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA NAT PETROLEUM CORP
- Filing Date
- 2024-10-21
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies suffer from low accuracy, low efficiency, and high manpower consumption in first arrival picking, especially in complex seismic data where efficient and accurate first arrival picking is difficult to achieve.
We employ a deep learning approach based on the SA-DeepLab network to construct an initial arrival picking model, which is then trained and validated using sample data, thereby reducing manpower consumption and improving picking accuracy.
It improves the accuracy and efficiency of first arrival wave acquisition, reduces manpower consumption, enhances the acquisition accuracy of complex seismic data, and improves the network's adaptability and generalization ability in low signal-to-noise ratio and complex environments.
Smart Images
Figure CN121901593A_ABST
Abstract
Description
Technical Field
[0001] This article relates to the field of seismic data processing technology, and in particular to a method and apparatus for first arrival wave picking based on deep learning. Background Technology
[0002] In 3D seismic exploration, the "three highs" requirements of "high density, high fidelity, and high accuracy" have become the norm. First arrival picking is the foundation for near-surface velocity tomography inversion and static correction. The accuracy of first arrival picking determines the accuracy of static correction, affecting the imaging quality of time migration and depth migration. Due to increasingly complex exploration conditions, missing or bad channels, low signal-to-noise ratios, and other factors can also lead to incorrect first arrival picking. How to accurately and efficiently pick first arrivals from complex seismic data is a practical problem that cannot be ignored.
[0003] Currently, the main methods for solving the problem of first arrival picking in seismic data include manual picking, energy ratio method, and artificial neural network algorithms. Manual first arrival picking requires manual, channel-by-channel picking, offering high accuracy, but is susceptible to human subjectivity and has low efficiency due to its high manpower requirements when dealing with large amounts of real-world data. The energy ratio method uses energy changes within a time window before and after the first arrival time to determine the first arrival time. The challenge lies in selecting the appropriate time window length; too small a window is easily influenced by local sampling point values, while too large a window may overlook the true first arrival wave. Neural network algorithms require manual extraction of the first arrival wave's feature vectors, and the performance of the network model is greatly affected by the number of features, presenting significant limitations when dealing with massive amounts of real-world data.
[0004] Therefore, how to improve the accuracy and efficiency of automatic arrival acquisition, reduce the manpower consumption of arrival acquisition, and improve the accuracy of arrival acquisition for complex real-world data are urgent problems to be solved. Summary of the Invention
[0005] This application provides a first arrival wave picking method and apparatus based on deep learning. This application establishes a first arrival wave picking model based on the SA-DeepLab network and uses the first arrival wave picking model to perform accurate first arrival wave picking, which can reduce the manual consumption of first arrival wave picking and improve the accuracy of first arrival wave picking for complex real data.
[0006] In a first aspect, this application provides a deep learning-based intelligent first-arrival picking method, the method comprising: acquiring a portion of the seismic wave data to be processed as sample data; preprocessing the sample data and dividing it into a training sample set, a validation sample set, and a test sample set according to a specific ratio; setting initial parameters for a pre-constructed first-arrival picking model based on the SA-DeepLab network architecture and training it using the training set to obtain the first-arrival picking model; validating the first-arrival picking model using the test dataset, and if the conditions are met, obtaining the trained first-arrival picking model; inputting the seismic wave data to be processed into the trained first-arrival picking model to obtain the first arrival times of each corresponding seismic trace.
[0007] Secondly, embodiments of the present invention provide a first-arrival pickup device based on deep learning. The device includes a memory and a processor. The memory is used to store a deep learning-based first-arrival pickup program, and the processor is used to read and execute the deep learning-based first-arrival pickup program to perform the method described in any of the above embodiments.
[0008] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a data processing program, wherein the data processing program is executed by a processor using the deep learning-based first arrival wave picking method according to any one of claims 1-8.
[0009] Compared with related technologies, this application provides a method and apparatus for first-arrival picking based on deep learning. The method includes: acquiring a portion of the seismic wave data to be processed as sample data; preprocessing the sample data and dividing it into a training sample set, a validation sample set, and a test sample set according to a specific ratio; setting initial parameters for a pre-constructed first-arrival picking model based on the SA-DeepLab network architecture and training it using the training set to obtain the first-arrival picking model; validating the first-arrival picking model using the test dataset, and obtaining a trained first-arrival picking model if the conditions are met; and inputting the seismic wave data to be processed into the trained first-arrival picking model to obtain the first-arrival times of each corresponding seismic trace. This application establishes a first-arrival picking model based on the SA-DeepLab network and uses the first-arrival picking model for accurate first-arrival picking, which can reduce the manual labor required for first-arrival picking and improve the accuracy of first-arrival picking for complex real-world data.
[0010] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. Other advantages of this application can be realized and obtained by means of the solutions described in the description and the accompanying drawings. Attached Figure Description
[0011] The accompanying drawings are used to provide an understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.
[0012] Figure 1 This is a flowchart of the first arrival wave picking method based on deep learning according to an embodiment of this application;
[0013] Figure 2 This is a schematic diagram of a deep learning-based first-arrival wave pickup device according to an embodiment of this application;
[0014] Figure 3 This is a schematic diagram of an SA-DeepLab network in an exemplary embodiment;
[0015] Figure 4 This is a schematic diagram of DeepLab V3+ training loss in an exemplary embodiment;
[0016] Figure 5 This is a schematic diagram of DeepLab V3+ training MIoU in an exemplary embodiment;
[0017] Figure 6 This is a schematic diagram of SA-DeepLab training loss in an exemplary embodiment;
[0018] Figure 7 This is a schematic diagram of SA-DeepLab training MIoU in an exemplary embodiment;
[0019] Figure 8 This is a schematic diagram of the prediction results of DeepLab V3+ in an exemplary embodiment;
[0020] Figure 9 This is a schematic diagram of the SA-DeepLab prediction results in an exemplary embodiment;
[0021] Figure 10 This is a schematic diagram of the prediction results of DeepLab V3+ in an exemplary embodiment;
[0022] Figure 11 This is a schematic diagram of the SA-DeepLab prediction results in an exemplary embodiment;
[0023] Figure 12 This is a schematic diagram of the transfer learning training loss in an exemplary embodiment;
[0024] Figure 13 This is a schematic diagram of MIoU training during transfer learning in an exemplary embodiment;
[0025] Figure 14 This is a schematic diagram of network prediction results before transfer learning in an exemplary embodiment;
[0026] Figure 15 This is a schematic diagram of the network prediction results after transfer learning in an exemplary embodiment;
[0027] Figure 16 This is a schematic diagram of network prediction results before transfer learning in an exemplary embodiment;
[0028] Figure 17 This is a schematic diagram of the network prediction results after transfer learning in an exemplary embodiment. Detailed Implementation
[0029] This application describes several embodiments, but these descriptions are exemplary and not restrictive, and it will be apparent to those skilled in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with, or may replace, any feature or element of any other embodiment.
[0030] This application includes and contemplates combinations of features and elements known to those skilled in the art. The embodiments, features, and elements disclosed in this application may also be combined with any conventional features or elements to form a unique inventive scheme as defined by the claims. Any feature or element of any embodiment may also be combined with features or elements from other inventive schemes to form another unique inventive scheme as defined by the claims. Therefore, it should be understood that any feature shown and / or discussed in this application may be implemented individually or in any suitable combination. Therefore, the embodiments are not limited except by the limitations imposed by the appended claims and their equivalents. Furthermore, various modifications and changes may be made within the scope of the appended claims.
[0031] Furthermore, in describing representative embodiments, the specification may have presented methods and / or processes as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that it does not depend on such a specific order. As will be understood by those skilled in the art, other sequences of steps are also possible. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Moreover, the claims concerning the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of this application.
[0032] This invention provides a method for intelligent first-arrival wave picking based on deep learning, such as... Figure 1 As shown, the method includes steps S100-S140:
[0033] S100: Obtain a portion of the seismic wave data to be processed as sample data;
[0034] S110: Perform preprocessing operations on the sample data and divide it into training sample set and test sample set according to a specific ratio;
[0035] S120: Based on the pre-built first arrival wave picking model based on the SA-DeepLab network architecture, set the initial parameters of the model and train it using the training set to obtain the first arrival wave picking model.
[0036] S130: Validate the first arrival wave picking model using the test dataset. If the conditions are met, obtain the trained first arrival wave picking model.
[0037] S140: Input the seismic wave data to be processed into the trained first arrival wave picking model to obtain the first arrival time of each corresponding seismic trace.
[0038] In one exemplary embodiment, preprocessing the sample data includes:
[0039] Step 1: Perform preliminary picking of seismic data and create labeled initial arrival picking results;
[0040] Step 11: Obtain the seismic data to be processed, such as seismic data in segy format.
[0041] Step 12: Perform data segmentation on the common shot gather data after handling missing values, so that the common shot gather data after data segmentation have the same number of seismic traces and samples;
[0042] Step 13: Preprocess the acquired seismic data in seyy format by cutting it into data volumes of uniform size, such as 446 single shots of 3001x256, where 3001 is the number of longitudinal sampling points for each seismic trace and 256 is the number of seismic traces.
[0043] Step 14: Perform preliminary arrival picking for each single shot's data. The traditional STA / LTA method can be used to perform preliminary arrival picking, correct erroneous arrivals, remove outliers, and create training label data for the arrival picking model.
[0044] To avoid overfitting or underfitting caused by excessive pixel differences within sample categories, training label data is created based on the initial arrival and take-off points as the dividing points. Points before the initial arrival and take-off are set to 0, and points after the initial arrival and take-off are set to 1.
[0045] Step 2: Normalize the common shot point gather data after data segmentation.
[0046] The sample seismic data were adjusted to a uniform size and then normalized.
[0047] Step 3: Randomly flip and randomly add Gaussian noise to the normalized seismic data to obtain labeled sample data.
[0048] In this step, the normalized seismic data is randomly flipped and Gaussian noise is randomly added to increase the diversity of the data, prevent overfitting, and thus enhance the robustness and generalization ability of the model.
[0049] In one exemplary embodiment, the first-arrival picking model based on the SA-DeepLab architecture includes: a ResNet-50 feature extraction module, an ASPP module, a first self-attention module, a second self-attention module, and a concatenation module. The overall architecture of the SA-DeepLab network follows the original framework of DeepLab V3+, with the backbone network replacing the original DCNN using a ResNet-50 network, specifically as follows... Figure 3 As shown.
[0050] In this embodiment, the ResNet-50 feature extraction module is used to extract features within a first time range (shallow features) and features within a second time range (deep features);
[0051] The ASPP module is used to construct and incorporate image-level pooling features based on features in the first time range (shallow features) and features in the second time range (deep features) to obtain a feature image, which is then input into the first self-attention module.
[0052] The first and second self-attention modules use a self-attention mechanism to obtain feature maps with correlations between different paths.
[0053] The concatenation module is used to perform convolution and upsampling on the features of the first time range to obtain the same sampling interval as the second time range, and then concatenate them to obtain the final initial arrival prediction result.
[0054] The ASPP module includes one 1×1 convolutional layer, three 3×3 convolutional layers, and one image pooling layer.
[0055] Specifically, in this embodiment, the first-arrival pickup process implemented using the overall architecture of the SA-DeepLab network is as follows:
[0056] Step 1: Input seismic data into ResNet-50 for feature extraction to obtain features in the first time range (shallow features) and features in the second time range (deep features);
[0057] Step 2: The features (shallow features) within the first time range are input into the parallel ASPP module to construct and incorporate image-level pooling features to obtain the pooled shallow features.
[0058] Step 3: The shallow features after pooling are compressed through a 1x1 convolution channel and then fed into the first self-attention module. The data output from the first self-attention module is upsampled to obtain the shallow prediction results.
[0059] Step 4: Input the features (deep features) within the second time range into the second self-attention module;
[0060] Step 5: The second self-attention module is concatenated with the shallow data after the upsampling operation;
[0061] Step 6: After performing two convolutions and one upsampling operation, output the final initial prediction result.
[0062] In one exemplary embodiment, the ResNet-50 feature extraction module uses the FReLU activation function expression as follows:
[0063]
[0064] T(·) represents the funnel condition, and x represents the parameter pool window.
[0065] In this embodiment, ReLU, as the most commonly used activation function, maintains a derivative of 1, which can effectively alleviate the vanishing and exploding gradient problems. It causes some neurons to be zero, resulting in network sparsity, reducing the interdependence between parameters, and mitigating overfitting. However, when the input is less than zero, the output will be zero. The gradient of this neuron will remain zero in subsequent training iterations, and the parameter w will not be updated. Therefore, there is an activation dead zone, leading to poor robustness of the activation function during training. When faced with large gradient inputs, it can easily cause neurons to "die," preventing the network from updating.
[0066] PReLU, evolved from ReLU, adds a linear activation component to the less-than-zero portion of the input by introducing a random parameter w that changes with the data computation, effectively avoiding the "dead" phenomenon. The expressions for the two activation functions are shown below:
[0067]
[0068] To achieve pixel-level modeling capabilities, ReLU in the network is replaced with FReLU, and a funnel condition T(x) is added to extend ReLU and FReLU to 2D activation. The expression is:
[0069]
[0070] The funnel condition T(x) is a square sliding window with preset parameters, implemented through conventional convolution. It can improve the spatial dependency between pixels, capture spatially insensitive information, and thus obtain rich spatial context information. Compared with ReLU, it has a great improvement and robustness in semantic segmentation.
[0071] In one exemplary embodiment, setting initial parameters for the model and training it using the training set includes:
[0072] The training dataset is input into the first arrival wave picking model to obtain the first arrival prediction results of the training sample data;
[0073] Based on the initial prediction results and the corresponding label results of the training data, the prediction results are optimized using the Cross-Entropy Loss function;
[0074] The cross-entropy loss function is as follows:
[0075] Loss=-(y*log(p)+(1-y)*log(1-p))
[0076] Where y is the label result corresponding to the training data, and p is the initial prediction result. When y=1, y*log(p) contributes the main loss, and when y=0, (1-y)*log(1-p) contributes the main loss. This makes the loss function penalize incorrect predictions more, making the model's prediction result closer to the true label.
[0077] In one exemplary embodiment, the first-arrival picking model is validated using the test dataset. If the conditions are met, a trained first-arrival picking model is obtained, including:
[0078] The test dataset is input into the first arrival wave picking model to obtain the model's prediction results;
[0079] Using the model's prediction results and the corresponding label results of the test data, the MioU method is used to determine the model's segmentation accuracy;
[0080] If the segmentation accuracy meets the predetermined threshold, the trained first arrival wave picking model is obtained;
[0081] in,
[0082] The MIoU formula is:
[0083]
[0084] In this context, the TP part in the middle is the intersection of the true value and the predicted value, the FN+FP+TP part is the union of the true value and the predicted value, and MIoU is the average IoU of all categories on the dataset.
[0085] In one exemplary embodiment, the deep learning-based intelligent first-arrival acquisition method can also perform transfer learning. Due to the different geological environments in different work areas, the acquired signals also vary significantly. Especially in areas with complex exploration environments, there may be a series of problems such as a large thickness of shallow, low-velocity zones, significant lithological variations, complex geological structures, and various noise interferences, all of which greatly affect the signal-to-noise ratio (SNR) of seismic data. Directly applying a network model trained with a high SNR will not yield ideal results. Transfer learning uses a pre-trained model as an initial point and reapplies it to the process of training a model for a new work area. Its aim is to find similarities between different data. Transfer learning starts from data similarity and applies the learned model for solving existing problems to other relevant datasets. This allows for transfer learning of the trained network model using a small amount of new data, achieving accurate prediction of first arrivals on new datasets and improving the network's generalization and adaptability.
[0086] The first-arrival wave intelligent picking method and device based on deep learning implemented in this embodiment has the following technical effects:
[0087] (1) Accurate prediction
[0088] The introduction of the FReLU activation function and self-attention mechanism allows the network to capture more global information during training, improves the correlation between paths, and makes the predicted initial form more continuous.
[0089] (2) The method is efficient
[0090] The dilated convolution kernel (ASPP) module in the network requires fewer parameters for the same receptive field, which improves the training speed of the network while ensuring the stability of training.
[0091] (3) Generalization
[0092] Introducing transfer learning improves the generalization ability of the model, enabling the network to make predictions on another batch of data with only a small amount of data after one training.
[0093] Secondly, embodiments of the present invention provide a first-arrival wave pickup device based on deep learning, such as... Figure 2 As shown, the device includes a memory 200 and a processor 210; the memory is used to store a deep learning-based first arrival wave picking program, and the processor is used to read and execute the deep learning-based first arrival wave picking program to perform the method described in any of the above embodiments.
[0094] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a data processing program, wherein the data processing program is executed by a processor using the deep learning-based first arrival wave picking method described in any of the above embodiments.
[0095] Example 1
[0096] This example demonstrates the training process of an first-arrival picking model based on the SA-DeepLab network architecture, applied to seismic data from a two-dimensional landmass. The specific process is as follows:
[0097] Step 10: Preprocess the seismic data to be initially picked;
[0098] Step 101: Preprocess the Segy data to create labeled training sample data;
[0099] Specifically, the data is divided into 446 single-shot samples of 3001x256 pixels, where 3001 represents the number of longitudinal sampling points per seismic trace and 256 represents the number of seismic traces. In this step, the data can be initially picked using the traditional STA / LTA method, and then incorrectly picked first arrivals are corrected to create first arrival labels. To avoid overfitting or underfitting due to excessive pixel differences within sample categories, training labels are created based on the first arrival jump point as the dividing point. Points before the first arrival jump are set to 0, and points at and after the first arrival jump are set to 1.
[0100] Step 102: Normalize the label data;
[0101] First, the overall data size needs to be adjusted by dividing all the data into blocks and normalizing them. Then, the data blocks are randomly flipped and Gaussian noise is randomly added to increase the diversity of the data, prevent overfitting, and thus enhance the robustness and generalization ability of the model.
[0102] Step 11: Establish the sample dataset and label data, dividing them into training and test sets in an 8:1 ratio.
[0103] When training a model, a training set is used to train the model and adjust its hyperparameters, while a test set is used to test the model's performance.
[0104] Step 12: Establish an initial arrival picking model based on the SA-DeepLab network architecture; such as Figure 3 As shown, the model includes a ResNet-50 feature extraction module, an ASPP module, a first self-attention module, a second self-attention module, and a splicing module.
[0105] Step 13: Set the initial parameters of the model and train the initial picking model based on the SA-DeepLab network architecture using the training set;
[0106] During model training, the cross-entropy loss function is used to optimize the prediction results.
[0107] Loss=-(y*log(p)+(1-y)*log(1-p)) (4)
[0108] Here, y is the accurately picked binary label, and p is the model's predicted probability. When y = 1, y * log(p) contributes the main loss, and when y = 0, (1 - y) * log(1 - p) contributes the main loss. This makes the loss function penalize incorrect predictions more severely, making the model's prediction results closer to the true labels.
[0109] Step 14: The hyperparameters used in the final training of the initial arrival picking model based on the SA-DeepLab network architecture are shown in Table 1:
[0110] Table 1
[0111]
[0112]
[0113] Step 15: Use the test set to test the accuracy of the model;
[0114] In this step, specifically, MIoU can be used to measure the segmentation accuracy of the model:
[0115]
[0116] In this context, the TP part in the middle is the intersection of the true value and the predicted value, the FN+FP+TP part is the union of the true value and the predicted value, and MIoU is the average IoU of all categories on the dataset.
[0117] A comparative analysis of the training process of the first-arrival picking model based on Example 1 and the first-arrival picking model of a conventional DeepLab V3+ network is conducted:
[0118] like Figure 4 and Figure 5 As shown, when training with a conventional DeepLab V3+ network, the Loss function exhibits significant fluctuations at 70, 150, and 200 iterations, only stabilizing and converging after 300 iterations. Meanwhile, MIoU shows drastic fluctuations around epoch 35, converging only after approximately epoch 65, with an optimal value of 97.26%. This indicates that the conventional DeepLab V3+ network still has many problems in first-arrival prediction of seismic data, requiring network improvement.
[0119] like Figure 6 and Figure 7 As shown, after improving the activation function and adding a self-attention module, the network's Loss and MIoU are significantly improved. Specifically, the Loss function converges stably after approximately 60 iterations, and the MIoU converges after 15 epochs, with the optimal value reaching 97.97%. It is clear that the SA-DeepLab network exhibits more stable Loss and MIoU during training, converges faster, and performs better on initial arrival picking.
[0120] Example 2
[0121] This example demonstrates the test analysis of the first-arrival picking model based on the SA-DeepLab network architecture:
[0122] Step 1: Test the initial picking model of the trained DeepLab V3+ network architecture;
[0123] Tests were conducted on both high signal-to-noise ratio data without noise and noisy data.
[0124] Step 2: Test the initial picking model of the trained SA-DeepLab network architecture;
[0125] Tests were conducted on both high signal-to-noise ratio data without noise and noisy data.
[0126] Step 3: Analyze the test results of the trained DeepLab V3+ model and SA-DeepLab model.
[0127] The prediction results of the two models were compared using data with a signal-to-noise ratio (SNR) of 60.
[0128] For high signal-to-noise ratio data without noise, both models have high accuracy in predicting first arrivals. However, the DeepLab V3+ model still has some error in predicting first arrivals at mid-to-long distances, while the SA-DeepLab model has better overall prediction results, with more continuous first arrival patterns, and can basically pick them accurately.
[0129] Figure 8 The results of DeepLab V3+ picking up data without noise. Figure 10 For noisy data, DeepLab V3+ picks up the results from... Figure 8 and Figure 10 It is evident that as the signal-to-noise ratio decreases, the DeepLab V3+ model exhibits very large errors in its prediction results for medium to long offsets, with severe mispicking and discontinuous first arrivals. However, for the improved SA-DeepLab, only slight errors occur at larger offsets, and the overall first arrival shape is more consistent, resulting in a significant improvement in the first arrival prediction results.
[0130] Figure 9 The SA-DeepLab prediction results are for noisy data. Figure 11 SA-DeepLab picks up the noisy data from... Figure 9 and Figure 11 The comparative analysis shows that the overall prediction results are good, the initial arrival pattern is more continuous, and it can basically be picked accurately.
[0131] Based on the above comparative analysis, the SA-DeepLab model also shows good prediction results for data with low signal-to-noise ratio.
[0132] Example 3
[0133] This example demonstrates the transfer learning performance of the first-arrival picking model based on the SA-DeepLab network architecture, as shown below:
[0134] The first step is to select data from 47 shots in the new work area to create a dataset. Each sample size is 3501x480, meaning there are 480 tracks per shot and 3501 sampling points per track.
[0135] The second step is to retrain the SA-DeepLab model using the newly created dataset.
[0136] During training, the loss fluctuated slightly around iteration 240 times, but overall the loss and MIoU converged relatively stably, with the optimal MIoU reaching 97.44%. Figure 12 Loss and training loss for transfer learning in Example 3 Figure 13 Example 3 shows the transfer learning training of MioU.
[0137] The picking results of the SA-DeepLab model before and after transfer learning are compared. Figure 14 For network prediction results before transfer learning and Figure 16 The network prediction results before transfer learning are the initial arrival picking results before transfer learning. Under the conditions of low signal-to-noise ratio and missing or bad channels, the initial arrivals picked by our pre-trained model fluctuate greatly, are discontinuous, and have many mispicked arrivals.
[0138] pass Figure 15 For network prediction results after transfer learning and Figure 17 The network prediction results after transfer learning show that the model's data picking performance in the new work area has been significantly improved after transfer learning. This indicates that the model after transfer learning can not only accurately pick up high signal-to-noise ratio data, but also accurately predict the first arrival in complex situations such as low signal-to-noise ratio, missing channels, and bad channels.
[0139] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.
Claims
1. A first-arrival wave picking method based on deep learning, characterized in that, The method includes: A portion of the seismic wave data to be processed is obtained as sample data; The sample data is preprocessed and divided into a training sample set and a test sample set according to a specific ratio; Based on the pre-built first arrival wave picking model based on the SA-DeepLab network architecture, the initial parameters of the first arrival wave picking model are set and trained using the training sample set to obtain the first arrival wave picking model. The first arrival picking model is validated using the test sample set. If the conditions are met, the trained first arrival picking model is obtained. Input the seismic wave data to be processed into the trained first-arrival picking model to obtain the first-arrival times of each seismic trace.
2. The first-arrival wave picking method based on deep learning according to claim 1, characterized in that, Preprocessing of sample data includes: Perform preliminary picking on the seismic data and generate labeled first arrival picking results; After handling missing values, the common shot gather data is split so that the split common shot gather data have the same number of seismic traces and samples. Normalize the common shot point gather data after data segmentation; The normalized seismic data were randomly flipped and Gaussian noise was randomly added to obtain labeled sample data.
3. The first-arrival wave picking method based on deep learning according to claim 1, characterized in that, The first arrival wave picking model based on the SA-DeepLab network architecture includes: a ResNet-50 feature extraction module, an ASPP module, a first self-attention module, a second self-attention module, and a splicing module.
4. The first-arrival wave picking method based on deep learning according to claim 3, characterized in that, The ResNet-50 feature extraction module is used to extract features within a first time range and features within a second time range; The ASPP module is used to construct and incorporate image-level pooling features based on features in the first time range and features in the second time range to obtain a feature image, which is then input into the first self-attention module. The first and second self-attention modules use a self-attention mechanism to obtain feature maps with correlations between different paths. The concatenation module is used to perform convolution and upsampling on the features of the first time range to obtain the same sampling interval as the second time range, and then concatenate them to obtain the final initial prediction result.
5. The first-arrival wave picking method based on deep learning according to claim 3, characterized in that, The ResNet-50 feature extraction module uses the FReLU activation function; The FReLU activation function is: In the above function, T(·) represents the funnel condition, and x represents the parameter pool window.
6. The first-arrival wave picking method based on deep learning according to claim 1, characterized in that, The process of setting initial parameters for the first arrival picking model and training it using the training sample set is as follows: The training dataset is input into the first arrival wave picking model to obtain the first arrival prediction results of the training sample data; Based on the initial prediction results and the corresponding label results of the training data, the prediction results are optimized using the cross-entropy loss function.
7. The first-arrival wave picking method based on deep learning according to claim 6, characterized in that, The cross-entropy loss function is: Loss=-(y*log(p)+(1-y)*log(1-p)) Where y is the label result corresponding to the training data, and p is the initial prediction result.
8. The first-arrival wave picking method based on deep learning according to claim 1, characterized in that, The first-arrival picking model is validated using the test sample set. If the conditions are met, a trained first-arrival picking model is obtained, including: The test sample set is input into the first arrival wave picking model to obtain the model's prediction results; The MioU formula is used to determine the segmentation accuracy of the model by using the model's prediction results and the corresponding label results of the test sample set data. If the segmentation accuracy meets the predetermined threshold, the trained first-arrival picking model is obtained; The MIoU formula is as follows: Where TP is the intersection of the true value and the predicted value, FN+FP+TP is the union of the true value and the predicted value, and MIoU is the average IoU of all categories on the dataset.
9. A first-arrival wave pickup device based on deep learning, characterized in that, The apparatus includes a memory and a processor; the memory is used to store a deep learning-based first-arrival pickup program, and the processor is used to read and execute the deep learning-based first-arrival pickup program to perform the method according to any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data processing program, which is executed by a processor according to the deep learning-based first arrival wave picking method of any one of claims 1-8.