A geological layer velocity modeling method, device, medium and equipment

Abstract

The present invention discloses a geological layer velocity modeling method, device, medium and equipment, and relates to the field of geophysical exploration technology. The present invention first trains multiple preset neural network models respectively through different labeled seismic reflection waveform data, then tests multiple layer velocity modeling models through part of the labeled seismic reflection waveform data, determines the modeling accuracy of each layer velocity modeling model, and then for each type of layer velocity model, predicts the unlabeled seismic reflection waveform data through the layer velocity modeling model with the highest modeling accuracy of the layer velocity model of this type to obtain a pseudo label, mixes the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data and re-acquires the labeled seismic reflection waveform data, so as to perform multiple rounds of iterative training on multiple layer velocity modeling models, and the present invention improves the accuracy of model layer velocity modeling.

Description

Technical Field

[0001] The present invention relates to the field of geophysical exploration technology, and in particular to a geological layer velocity modeling method, device, medium and equipment. Background Art

[0002] Generally, layer velocity modeling is a key step in geophysical exploration, which is used to estimate the velocity structure of underground layers and to image and interpret underground structures and geological features. Layer velocity modeling affects the accuracy of seismic wave imaging, and high-precision layer velocity modeling is a key parameter for obtaining high signal-to-noise ratio, high fidelity and high-resolution seismic profiles. At the same time, accurate depth domain layer velocity modeling has become a key factor in the success or failure of reverse time migration.

[0003] In recent years, with the help of the nonlinear mapping ability of deep learning, supervised layer velocity modeling methods based on seismic reflection waveform data of different gather types have emerged. The method process is to first build a runnable neural network model, establish a variety of layer velocity models, and place a reasonable observation system, and then use the acoustic wave equation to forward model the seismic reflection waveform data of each layer velocity model. At this time, the seismic reflection waveform data obtained by forward modeling is used as the training data set of the built neural network model, and the established layer velocity models are used as the label data of the network. By adjusting the reasonable network hyperparameters, a neural network model with good generalization performance is obtained, which is used for layer velocity modeling of actual seismic reflection waveform data. However, the current research on the use of seismic reflection waveform data for layer velocity modeling mainly focuses on supervised learning, which has problems such as dependence on a large amount of labeled data, insufficient generalization ability, and high cost and time-consuming to obtain the actual underground layer velocity. Therefore, in order to overcome the heavy work of data labeling and the lack of labeled data in traditional supervised learning methods, semi-supervised learning methods can be applied to layer velocity modeling. By making full use of unlabeled seismic reflection waveform data, the demand for obtaining labeled data is reduced, and the modeling efficiency and accuracy of layer velocity models are improved.

[0004] Semi-supervised learning is a machine learning method that aims to train models using partly labeled data and partly unlabeled data. Compared with traditional supervised learning, semi-supervised learning can achieve good performance when sample labels are expensive or labeled data is scarce. The core idea of ​​semi-supervised learning is to use the information of unlabeled data to improve model performance. In semi-supervised learning, it is usually assumed that unlabeled data contains useful information that can help the model better learn the distribution characteristics of data. At the same time, a small amount of labeled data is used to guide the specific goals of the model learning task. The idea of ​​semi-supervised self-training method is applicable to layer velocity modeling. The principle of the original semi-supervised self-training method is: first, the network is trained with labeled data samples to obtain a supervised network model model1, and then the model1 model is used to predict unlabeled data samples to obtain pseudo labels, and the labeled data samples and pseudo-labeled data samples are mixed to form a sample data set for semi-supervised training. Next, the network is trained with this semi-supervised training sample data set to obtain a semi-supervised network model model2, and then the semi-supervised network model model2 is used to predict unlabeled data samples to obtain new pseudo labels to iteratively train the network model until the training obtains a network model with good prediction effect and stops training.

[0005] However, the current geological layer velocity modeling method based on semi-supervised learning is highly sensitive to label noise, that is, it usually relies on a small amount of labeled seismic reflection waveform data and a large amount of unlabeled seismic reflection waveform data. However, if there is noise in the unlabeled seismic reflection waveform data or incorrect pseudo-labels are given during the training process, the accuracy of the model layer velocity modeling will be reduced. Summary of the invention

[0006] Based on this, it is necessary to provide a geological layer velocity modeling method, device, medium and equipment to address the above technical problems.

[0007] The present invention adopts the following technical solutions:

[0008] The present invention provides a geological layer velocity modeling method. First, a plurality of preset neural network models are trained respectively through different labeled seismic reflection waveform data, and then a plurality of layer velocity modeling models are tested through part of the labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model. Then, for each type of layer velocity model, the layer velocity modeling model with the highest modeling accuracy for the layer velocity model of this type is used to predict the unlabeled seismic reflection waveform data, and the unlabeled seismic reflection waveform data is converted into pseudo-labeled seismic reflection waveform data. Finally, the model parameters of the remaining layer velocity modeling models are updated according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy, and the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data are mixed as the updated labeled seismic reflection waveform data, so that the updated labeled seismic reflection waveform data and the re-acquired unlabeled seismic reflection waveform data are used to perform multiple rounds of iterative training on the plurality of layer velocity modeling models, and the layer velocity modeling model with the highest final comprehensive layer velocity modeling accuracy is applied.

[0009] The present invention provides a geological layer velocity modeling device, wherein each module in the device is used to implement the geological layer velocity modeling method.

[0010] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned geological layer velocity modeling method is implemented.

[0011] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned geological layer velocity modeling method when executing the program.

[0012] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0013] The present invention trains multiple neural network models by using labeled seismic reflection waveform data, and for each layer velocity model, predicts the unlabeled seismic reflection waveform data by using the model with the best prediction performance for the layer velocity model in the current training round, thereby obtaining its pseudo-label, effectively reducing the sensitivity to noise in the model training process and improving the accuracy of the pseudo-label, thereby improving the accuracy of the model layer velocity modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0015] Figure 1 A schematic flow chart of a geological layer velocity modeling method provided by the present invention;

[0016] Figure 2 A schematic diagram of a geological layer velocity model provided by the present invention;

[0017] Figure 3 A schematic diagram of the model architecture of an improved U-Net++ network model provided by the present invention;

[0018] Figure 4 A schematic diagram of the generation and determination process of a pseudo label based on the Fold data set of a fold velocity model provided by the present invention;

[0019] Figure 5 A schematic diagram of a process of performing layer velocity modeling on a Fold data set using a semi-supervised method provided by the present invention;

[0020] Figure 6 A schematic diagram of a process of performing layer velocity modeling on a Salt data set using a semi-supervised method provided by the present invention;

[0021] Figure 7 A schematic diagram of a process of performing layer velocity modeling on a Linear data set using a semi-supervised method provided by the present invention;

[0022] Figure 8 A schematic diagram of a process of performing layer velocity modeling on the Images2 data set using a semi-supervised method provided by the present invention;

[0023] Fig. 9 A schematic diagram of a process of performing layer velocity modeling on the Images1 data set using a semi-supervised method provided by the present invention;

[0024] Fig.10 A schematic diagram of a geological layer velocity modeling device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] At present, the geological layer velocity modeling method based on semi-supervised learning is highly sensitive to label noise, that is, it usually relies on a small amount of labeled seismic reflection waveform data and a large amount of unlabeled seismic reflection waveform data. However, if there is noise in the unlabeled seismic reflection waveform data or incorrect pseudo-labels are given during the training process, the accuracy of the model layer velocity modeling will be reduced. In addition, during the semi-supervised learning process, the model will be overly dependent on the unlabeled seismic reflection waveform data, resulting in the problem of overfitting the model on the unlabeled seismic reflection waveform data and poor performance on the labeled seismic reflection waveform data, and the computing resource requirements are large. In addition, due to the complexity of using the seismic reflection waveform data in the time domain to obtain the geological velocity model in the depth domain through the neural network model, a neural network model with stronger nonlinear mapping capabilities can achieve more accurate results.

[0027] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0028] Figure 1 The present invention is a flow chart of a geological layer velocity modeling method, which specifically includes the following steps:

[0029] S101: Acquire labeled seismic reflection waveform data of annotated layer velocity models and unlabeled seismic reflection waveform data of unlabeled layer velocity models; wherein the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data correspond to multiple types of layer velocity models.

[0030] Generally, when the server of the business platform performs geological layer velocity modeling in a semi-supervised manner, it can first obtain labeled seismic reflection waveform data of the corresponding layer velocity model that has been annotated. It can be directly obtained from an existing public data set, or it can be forward modeled from the existing geological layer velocity model. Of course, unlabeled seismic reflection waveform data can also come from an existing public data set.

[0031] For example, the server can first establish layer velocity models of various geologies. The layer velocity models and seismic reflection waveform data can come from public data sets, including five velocity models of geological structures: The first type is a velocity model in which the velocity values ​​of each stratum are the same: the first horizontal layered velocity model (Linear), the fold velocity model (Fold), and the first salt velocity model (Salt). The second type is a velocity model in which the velocity values ​​at each grid point in each stratum are different: the second horizontal layered velocity model (Images1), and the second salt velocity model (Images2). The five velocity models are as follows: Figure 2 As shown, Figure 2 is a schematic diagram of a geological layer velocity model in the present invention, Figure 2In the figure, from left to right are the first horizontal layered velocity model, the fold velocity model, the first salt body velocity model, the second horizontal layered velocity model and the second salt body velocity model.

[0032] Then, the server can forward model the seismic reflection waveform data. Here, the method of middle excitation and two-side reception can be adopted. The source and the detector are evenly distributed on the surface, with the starting positions of 1.5km and 0km respectively, a total of 1 shot, 301 channels per shot, and the Ricker wavelet with a main frequency of 10Hz is selected as the excitation source function. The seismic wave field is simulated based on the acoustic wave equation using the second-order time and tenth-order space finite difference method. The time sampling interval is 1 ms, the spatial sampling is 10 m, and the sampling time is 2 s. The fully matched layer absorption boundary condition is used to suppress the non-physical reflection wave field at the boundary. Both the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data can be obtained from public data sets, or they can be forward modeled using the layer velocity model.

[0033] The server mentioned in the present invention may be a server set on a business platform, or a device such as a desktop computer, a notebook computer, etc. that can execute the solution of the present invention. For the convenience of description, the following description is only based on the server as the execution subject.

[0034] S102: training a plurality of preset neural network models respectively by using different labeled seismic reflection waveform data to obtain a plurality of layer velocity modeling models; and testing the plurality of layer velocity modeling models by using at least part of the labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model.

[0035] S103: The unlabeled seismic reflection waveform data are respectively input into each layer velocity modeling model, and for each type of layer velocity model, the prediction of the layer velocity modeling model with the highest modeling accuracy for the layer velocity model of this type is used as the pseudo-label of the corresponding unlabeled seismic reflection waveform data, so as to convert the unlabeled seismic reflection waveform data into pseudo-labeled seismic reflection waveform data.

[0036] S104: updating the model parameters of the remaining layer velocity modeling models according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy; and mixing the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data as updated labeled seismic reflection waveform data.

[0037] S105: Perform multiple rounds of iterative training on multiple layer velocity modeling models using the updated labeled seismic reflection waveform data and the re-acquired unlabeled seismic reflection waveform data, and perform layer velocity modeling using the layer velocity modeling model with the highest final comprehensive layer velocity modeling accuracy.

[0038] After obtaining the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data as described above, the server can build a deep neural network framework. In one or more embodiments of the present invention, the server can use U-Net++ as the basic framework to build a VGNET module with stronger nonlinear mapping capabilities, and embed it into each convolution node of U-Net++ to form an improved U-Net++ network model architecture, such as Figure 3 As shown, Figure 3 The schematic diagram of the model architecture of an improved U-Net++ network model in the present invention is shown in FIG. 1 , wherein the VGNET module includes an input layer, a first convolutional layer, an intermediate batch normalization layer, a first activation function, a second convolutional layer, an output batch normalization layer, a second activation function, and an output layer connected in series in sequence.

[0039] In order to realize automatic layer velocity modeling from raw seismic reflection waveform data, the present invention improves the U-Net++ network architecture to obtain an improved U-Net++ network model. Figure 3 It can be seen that the U-Net++ network will be used to capture the nodes of the seismic reflection waveform characteristics (X 0,0 ;X 1,0 ;X 2,0 ;X 3,0 ;X 4,0 ) constitutes an encoder that will be used to accurately locate the symmetrically shaped nodes (X 4,0 ;X 3,1 ;X 2,2 ;X 1,3 ;X 0,4 ) to form a decoder, and then use the idea of ​​dense connection to fuse the features of the corresponding levels in the encoder and decoder (for example, X 0,1 ;X 1,2 The present invention makes three major modifications based on the original U-Net++ network to make it suitable for the needs of geological layer velocity modeling.

[0040] First, to address the problem of low accuracy in modeling anomalies in layer velocity modeling, a nested U-Net++ network architecture was designed that better fits the nonlinear relationship between the original seismic waveform data and the more complex layer velocity model. The features of the VGG network architecture were extracted, and a VGG-style submodule was designed as an embedded convolutional module of U-Net++.

[0041] Secondly, the original U-Net++ for image segmentation reads the input image in RGB color 3 channels, and these color channels represent the information from the input image. In order to handle the different needs of seismic reflection waveform data, the present invention matches the seismic reflection waveform data with a layer velocity model as the input of the improved U-Net++ network. Before input, the SEGY format (Society of Exploration Geophysicists, Y-format) data is removed from the header to retain only the data body part, and then converted into an array format to input the improved U-Net++ network. Therefore, the input channel of the improved U-Net++ network is 1.

[0042] Finally, the seismic waveform data in the time domain is transformed into the depth domain of the geological velocity model. In order to complete the matching of the time domain and the depth domain, the original seismic waveform data and the velocity model are first stripped of the header and saved in array form, and then the input and output array dimensions are kept unchanged through resampling technology. The improved U-Net++ network automatically realizes the mapping from the time domain to the depth domain, and realizes the direct mapping of the original seismic reflection waveform data to the accurate layer velocity model.

[0043] The improvement of the nested U-Net++ network architecture enhances the nonlinear mapping relationship between the original seismic waveform data and the complex layer velocity model. By introducing a deeper network architecture, the model can learn more complex feature representations. This ability enables the network to effectively capture subtle changes in seismic waveform data, thereby improving the ability to identify velocity anomalies. The nested U-Net++ uses a more complex feature connection mechanism and effectively utilizes feature maps at different levels, thereby enhancing the learning of multi-scale information and improving the accuracy of velocity prediction values ​​at the contact between velocity anomalies and sedimentary strata.

[0044] The present invention introduces an intermediate channel in the VGNET module to increase the nonlinear expression ability of the improved U-Net++ network, so that the improved U-Net++ network is more suitable for the needs of seismic reflection waveform data mapping complex layer velocity model. In the VGNET module, two convolutional layers and two batch normalization layers are included, and each convolutional layer is followed by a ReLU activation function. The parameter setting of the convolutional layer and the batch normalization layer is to introduce nonlinear transformation and increase the expression ability of the improved U-Net++ network in the feature extraction and processing process.

[0045] In the first convolutional layer, local features in the input seismic reflection waveform data are extracted through the convolution kernels conv1(in, mid) and conv1(mid, mid). Here, by using the convolution kernel to perform a convolution operation on the seismic reflection waveform data, the patterns and structures existing in the seismic reflection waveform data can be captured, and the intermediate channels are batch normalized through the subsequent series-connected intermediate batch normalization layer BatchNorm1(mid), which helps to reduce the gradient vanishing or gradient exploding problems during the training process, thereby making the feature distribution of each channel more stable and improving the training effect of the model.

[0046] In the second convolutional layer, the extracted features are further convolved through the convolution kernels conv2(mid, out) and conv2(out, out), and the output channels are batch normalized through the subsequent series output batch normalization layer BatchNorm2(out), which helps to accelerate the convergence speed of the improved U-Net++ network, improve the training stability of the model, and enhance the model's generalization ability for input data. After that, a nonlinear activation function is applied to increase the nonlinear ability of the model, enabling it to fit more complex feature relationships.

[0047] After building the neural network model, the server can obtain multiple preset neural network models by setting different model parameters and perform semi-supervised learning training on them.

[0048] In semi-supervised learning, it is usually assumed that unlabeled data contains useful information, which can help the neural network model better learn the distribution characteristics of the data. At the same time, a small amount of labeled data is used to guide the specific goals of the neural network model learning task. Due to the complexity of using seismic reflection waveform data for layer velocity modeling, the present invention is based on the idea of ​​the original semi-supervised learning method, and is innovatively improved to realize the velocity model mapping of various geological structures using seismic waveform data.

[0049] For example, first, three different labeled feature sets and different network parameter training can be used to obtain three different initial supervised network models model1, model2 and model3. The three models are only used as examples, and the specific number is not limited in the present invention. The difference between the three different feature data sets is that the samples of the feature training sets are different, but the total number of samples of the initial models is the same.

[0050] After training three initial supervised network models with three different feature data sets, model1, model2 and model3 are used to predict the layer velocity model of the test set, and their respective root mean square error (RMSE) and mean absolute error (MAE) values ​​are calculated. Then the layer velocity model data predicted for the test set is visualized, and from a macro perspective, the ratios (error ratios) a1, a2 and a3 of the layer velocity models predicted by the three models with obvious errors in the total test set are counted. Finally, the root mean square error value, the mean absolute error and the error ratio are added to obtain the network model evaluation factor. Let b1, b2 and b3 be the model evaluation factors of model1, model2 and model3 respectively, and compare the sizes of the evaluation factors of each network model. The network model with the smallest evaluation factor is used to predict the layer velocity model of the unlabeled sample as the main pseudo-label for the next semi-supervised training. The pseudo-labels obtained by the remaining two models are used as substitutes for the velocity model with large errors in the main pseudo-label. That is, if there are individual layer velocity models with low prediction accuracy in the main pseudo-label, then they are compared with the pseudo-labels predicted by the remaining two network models. If there are velocity models with higher accuracy in the other two pseudo-labels, then the velocity model with poor accuracy in the main pseudo-label is replaced. Finally, the pseudo-labels with the highest prediction accuracy of the three supervised models and the labeled data set are mixed to form the next semi-supervised training data set. The initial value of the network hyperparameter with the lowest network model evaluation factor is used as the adjustment benchmark for the next iterative training. According to the above method, pseudo labels for semi-supervised iterative training are generated. Pseudo labels are added in an increasing number. The network is iteratively updated using the semi-supervised learning method designed by the present invention until the accuracy of layer velocity modeling decreases from high to low, and the iterative training is stopped. The layer velocity model with the highest accuracy is retained as the final result of layer velocity modeling.

[0051] In the semi-supervised layer velocity modeling method proposed in the present invention, it is crucial to determine the pseudo labels for each iterative training. Figure 4 It is a schematic diagram of the process of generating and determining pseudo labels based on the Fold dataset of the fold velocity model in the present invention. Figure 4 In the figure, the first column corresponds to the results of network model Model1 in each round of pseudo label generation, the second column corresponds to the results of network model Model2 in each round of pseudo label generation, the third column corresponds to the results of network model Model3 in each round of pseudo label generation, the first row corresponds to the results of each network model in the first round of pseudo label generation, the second row corresponds to the results of each network model in the second round of pseudo label generation, and the third row corresponds to the results of each network model in the third round of pseudo label generation.

[0052] Depend on Figure 4It can be seen that among the pseudo labels obtained in the first round of supervised training, the three pseudo labels generated by the three network models Model1, Model2 and Model3 obtained by training are compared. It can be found that each stratum generated by Model3 is smoother and has almost no abnormal velocity value area. Therefore, it can be used as the pseudo label for the second round of network training; after the second round of semi-supervised training, the pseudo labels generated by the three network models are compared. It can be found that Model2 has fewer abnormal velocity value areas and higher precision. This pseudo label has the best effect and can be used as the pseudo label for the next round of iterative training; after the third round of semi-supervised training, the pseudo labels obtained by the three network models are visualized, and it is found that their shapes are almost the same, and each pseudo label has no abnormal velocity position.

[0053] Figure 5 The figure is a flow chart of the semi-supervised method for layer velocity modeling of the Fold dataset in the present invention. When the layer velocity modeling of the Fold dataset is performed using the method of the present invention, supervised training is first performed using different datasets and parameters to obtain three different network models, each with 800 dataset samples. Then, the network model evaluation index b designed by the present invention is calculated. Figure 5 It can be seen that Model3-1 has the lowest index value, that is, the pseudo-labels generated by Model3-1 are used as the main body of pseudo-labels for the first semi-supervised training. This time, 150 pseudo-labels are obtained and mixed with labeled data for the first round of semi-supervised training. After the first round of semi-supervised training, the Model2-2 network model index is the smallest, and its prediction accuracy is better than the accuracy predicted by Model3-1. Therefore, the Model2-2 model is used to update the 150 pseudo-labels added to the training for the first time, and 150 new pseudo-labels are generated at the same time, that is, 300 pseudo-labels are mixed with the original data set for the second round of semi-supervised network training. Since the effect of the second round of network model prediction has improved little, the first 300 pseudo-labels and 100 new pseudo-labels are continued to be updated, and 400 pseudo-labels are mixed with the original data set for the third round of semi-supervised training. It is found that after the third round of semi-supervised training, the evaluation index of model Model1-4 is better and the test accuracy is significantly better than Model2-2. Subsequently, the first 400 pseudo labels were updated to generate 100 new pseudo labels for the fourth round of semi-supervised training. It was found that the evaluation index b of the obtained semi-supervised network model and the predicted speed model accuracy were close to the third round of semi-supervised results. Finally, while keeping all the network hyperparameters of the third round of semi-supervised training the same, only the number of pseudo labels was reduced, and the obtained network model evaluation index and speed model accuracy were worse than the results predicted by Model1-4. Therefore, Model1-4 is the best semi-supervised network model for layer speed modeling of the Fold dataset.

[0054] Figure 6The figure is a flow chart of layer velocity modeling of the Salt dataset using a semi-supervised method in the present invention. The velocity of each layer is [1.5, 1.7, 1.9, 2.1, 2.3, 2.5, 2.7, 2.9, 4.5] km / s. When modeling the layer velocity of the Salt dataset, supervised training is first performed with different datasets and parameters to obtain three initial network models. At this time, the number of samples of each dataset is 800. Then, the network model evaluation index b is calculated. Figure 6 It can be seen that Model3-1 has the lowest index value, that is, the pseudo-labels generated by Model3-1 are used as the main body of pseudo-labels for the first round of semi-supervised training, and 200 pseudo-labels are added this time. After the first round of semi-supervised training, the Model3-2 network model has the smallest index, and its prediction accuracy is better than that of Model3-1. Therefore, the Model3-2 model is used to update the 200 pseudo-labels added for the first time, and 200 new pseudo-labels are generated at the same time, that is, 300 pseudo-labels are mixed with the original data set for the second round of semi-supervised network training. During the iterative training process, since the prediction effect of the semi-supervised network model in the second, third and fourth rounds of training has improved little, in this iterative process, the previous pseudo-labels are updated and 50 new pseudo-labels are generated each time to join the training. After the fifth round of semi-supervised training, the speed modeling effect is significantly improved compared with before, and the total number of pseudo-labels added at this time is 550. After the fifth round of semi-supervised training, the methods of keeping the same network parameters but using different numbers of pseudo labels failed to train a model that was better than the fifth round of semi-supervised training; the method of adding the same pseudo labels and using different network parameters also failed to achieve the effect of the fifth round of velocity modeling. Therefore, the best semi-supervised network model for layer velocity modeling of the Salt dataset is Model2-6. Through this semi-supervised training process, it is concluded that for more complex salt-bearing body layer velocity models, when the number of added pseudo labels accounts for less than 70% of the label data, the layer velocity modeling accuracy gradually improves; when it exceeds 70%, the layer velocity modeling accuracy remains basically unchanged or even decreases.

[0055] Figure 7The figure is a schematic diagram of the process of layer velocity modeling for a Linear dataset using a semi-supervised method in the present invention. When the semi-supervised layer velocity modeling method proposed in the present invention is used, the Linear dataset is first subjected to one round of supervised training, and then two rounds of semi-supervised training are performed. It is found that the modeling accuracy gradually decreases. In this process, 200, 150 and 50 pseudo labels are added each time. In the three rounds of iterative training of the semi-supervised network, the supervised training Model2-1 model has the best prediction effect and the best network evaluation index. The network parameters of Model2-1 remain the same, and semi-supervised training is performed by adding or subtracting different numbers of pseudo labels. It is found that the modeling accuracy remains basically unchanged, but the overall RMSE and MAE values ​​of the test set increase. This shows that for a simple velocity model, a very high layer velocity modeling accuracy can be obtained by using the improved network of the present invention for supervised training.

[0056] Figure 8 The present invention is a schematic diagram of a process of layer velocity modeling for the Images2 dataset using a semi-supervised method. The method of the present invention first performs a round of supervised training on the Images2 dataset, and then performs two rounds of semi-supervised training to achieve the best training effect. In this process, the pseudo labels introduced each time are 200, 200 and 100 respectively. In the three rounds of iterative training of the semi-supervised network, the indicators for evaluating the network model show that Model3-1, Model1-2 and Model2-3 are the best performing models. Then, the network parameters of Model2-3 are kept unchanged, and after adding different numbers of pseudo labels, no significant improvement in modeling accuracy is observed.

[0057] Fig. 9 The figure is a schematic diagram of a process of layer velocity modeling for the Images1 dataset using a semi-supervised method in the present invention. When the semi-supervised layer velocity modeling method proposed in the present invention is used, the Images1 dataset is first subjected to a round of supervised training, and then two rounds of semi-supervised training are performed, and the modeling accuracy gradually decreases. In this process, the pseudo-labels added each time are 200, 150 and 50 respectively. In the three rounds of iterative training of the semi-supervised network, the prediction effect and network evaluation index of the Model1-1 model through supervised training are optimal. Then, the parameters of the Model1-1 network remain the same, and different numbers of pseudo-labels are added, and the modeling accuracy is still reduced.

[0058] based on Figure 1The geological layer velocity modeling method shown in the figure first trains multiple preset neural network models respectively through different labeled seismic reflection waveform data, then tests multiple layer velocity modeling models through part of the labeled seismic reflection waveform data, determines the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model, then for each type of layer velocity model, predicts the unlabeled seismic reflection waveform data through the layer velocity modeling model with the highest modeling accuracy of the layer velocity model of this type, converts the unlabeled seismic reflection waveform data into pseudo-labeled seismic reflection waveform data, and finally updates the model parameters of the remaining layer velocity modeling models according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy, and mixes the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data as the updated labeled seismic reflection waveform data, so that the multiple layer velocity modeling models are iteratively trained for multiple rounds through the updated labeled seismic reflection waveform data and the re-acquired unlabeled seismic reflection waveform data, and finally the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy is applied.

[0059] The present invention trains multiple neural network models by using labeled seismic reflection waveform data, and for each layer velocity model, predicts the unlabeled seismic reflection waveform data by using the model with the best prediction performance for the layer velocity model in the current training round, thereby obtaining its pseudo-label, effectively reducing the sensitivity to noise in the model training process and improving the accuracy of the pseudo-label, thereby improving the accuracy of the model layer velocity modeling.

[0060] The semi-supervised layer velocity field modeling method proposed in the present invention uses multiple network models to collaboratively generate pseudo labels, which helps to achieve higher prediction accuracy of velocity model pseudo labels. Since each model performs better under different data sets and hyperparameters, considering them comprehensively can improve the overall accuracy; using multiple networks for comparison each time can reduce the errors caused by the limitations of a single network model. If a network model performs poorly in some cases, other models can make up for the shortcomings, thereby improving the robustness of velocity modeling; in addition, combining the results of multiple models can reduce the risk of overfitting, better adapt to new data and improve generalization ability. Secondly, the network model evaluation factor combines three indicators, so that each iteration of training has more reliable results and more comprehensive coverage, which is more conducive to improving the overall convergence speed and accuracy of semi-supervised layer velocity modeling; more importantly, the multi-model semi-supervised layer velocity modeling method is more in line with the situation where there is no velocity model in actual seismic records.

[0061] The improvement of the nested U-Net++ network architecture enhances the nonlinear mapping relationship between the original seismic waveform data and the complex layer velocity model, and improves the accuracy of the velocity prediction value at the contact between the velocity anomaly body and the sedimentary stratum. In the VGNET submodule, in, mid, and out represent the number of input channels, the number of intermediate channels, and the number of output channels, respectively. In addition, the number of intermediate channels mid is specially introduced to increase the nonlinear expression ability of the network, making this network more suitable for the needs of mapping the complex layer velocity model from the original seismic waveform data. In the VGNET submodule, two convolutional layers and two batch normalization layers are included, and each convolutional layer is followed by a ReLU activation function. The parameter settings of the convolutional layer and the batch normalization layer are to introduce nonlinear transformations and increase the expression ability of the network during feature extraction and processing. BatchNorm1(mid) is designed to batch normalize the intermediate channels, which helps to reduce the gradient vanishing or gradient exploding problems during the training process, thereby making the feature distribution of each channel more stable and improving the training effect of the model. Designing BatchNorm2(out) to perform batch normalization on the output channels helps accelerate the convergence of the network, improve the training stability of the model, and enhance the model's generalization ability for input data.

[0062] When applying the geological layer velocity modeling method provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0063] The above is a geological layer velocity modeling method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding geological layer velocity modeling device, such as Fig.10 shown.

[0064] Fig.10 A schematic diagram of a geological layer velocity modeling device provided by the present invention includes:

[0065] The acquisition module 201 is used to acquire labeled seismic reflection waveform data of annotated layer velocity models and unlabeled seismic reflection waveform data of unlabeled layer velocity models; wherein the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data correspond to multiple types of layer velocity models;

[0066] The training and testing module 202 is used to train a plurality of preset neural network models respectively through different labeled seismic reflection waveform data to obtain a plurality of layer velocity modeling models; and to test the plurality of layer velocity modeling models through at least a portion of the labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model;

[0067] The prediction module 203 is used to input the unlabeled seismic reflection waveform data into each layer velocity modeling model respectively, and for each type of layer velocity model, use the prediction of the layer velocity modeling model with the highest modeling accuracy for the layer velocity model of this type as the pseudo label of the corresponding unlabeled seismic reflection waveform data, so as to convert the unlabeled seismic reflection waveform data into pseudo label seismic reflection waveform data;

[0068] The updating module 204 is used to update the model parameters of the remaining layer velocity modeling models according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy; and mix the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data as updated labeled seismic reflection waveform data;

[0069] The iterative training and application module 205 is used to perform multiple rounds of iterative training on multiple layer velocity modeling models through the updated labeled seismic reflection waveform data and the re-acquired unlabeled seismic reflection waveform data, and perform layer velocity modeling through the layer velocity modeling model with the highest final comprehensive layer velocity modeling accuracy.

[0070] The specific definition of the geological layer velocity modeling device can be found in the definition of the geological layer velocity modeling method above, which will not be repeated here. Each module in the above-mentioned geological layer velocity modeling device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0071] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A geological layer velocity modeling method is provided.

[0072] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A geological layer velocity modeling method is provided.

[0073] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0074] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A geological layer velocity modeling method, characterized in that: include: Acquire labeled seismic reflection waveform data of annotated layer velocity models and unlabeled seismic reflection waveform data of unlabeled layer velocity models; wherein the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data correspond to multiple types of layer velocity models; Training a plurality of preset neural network models respectively through different labeled seismic reflection waveform data to obtain a plurality of layer velocity modeling models; and testing the plurality of layer velocity modeling models through at least a portion of the labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model; The unlabeled seismic reflection waveform data is input into the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy to obtain the main pseudo labels of the unlabeled seismic reflection waveform data; and the secondary pseudo labels of the unlabeled seismic reflection waveform data are obtained through the remaining layer velocity modeling models; For each type of layer velocity model in the subject pseudo-label, when the modeling accuracy of the layer velocity modeling model corresponding to the subject pseudo-label for the layer velocity model of this type is lower than the modeling accuracy of any other layer velocity modeling model for the layer velocity model of this type, the corresponding subject pseudo-label is replaced with a secondary pseudo-label obtained by the layer velocity modeling model with the highest modeling accuracy for the layer velocity model of this type; Using the updated subject pseudo-label as the pseudo-label corresponding to the unlabeled seismic reflection waveform data to convert the unlabeled seismic reflection waveform data into pseudo-labeled seismic reflection waveform data; The model parameters of the remaining layer velocity modeling models are updated according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy; and the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data are mixed as updated labeled seismic reflection waveform data; Through the updated labeled seismic reflection waveform data and the re-acquired unlabeled seismic reflection waveform data, multiple rounds of iterative training are performed on multiple layer velocity modeling models, and layer velocity modeling is performed through the layer velocity modeling model with the highest final comprehensive layer velocity modeling accuracy.

2. The geological layer velocity modeling method according to claim 1, characterized in that: The multiple types of layer velocity models include: velocity models with the same velocity value in each formation: a first horizontal layered velocity model, a fold velocity model and a first salt body velocity model; velocity models with different velocity values ​​at each grid point in each formation: a second horizontal layered velocity model and a second salt body velocity model.

3. The geological layer velocity modeling method according to claim 1, characterized in that: The preset neural network model is an improved U-Net++ network model constructed by replacing the embedded convolution module in each node of the U-Net++ network model with a VGNET module; The VGNET module includes an input layer, a first convolutional layer, an intermediate batch normalization layer, a first activation function, a second convolutional layer, an output batch normalization layer, a second activation function and an output layer which are connected in series in sequence.

4. The geological layer velocity modeling method according to claim 1, characterized in that: The testing of multiple layer velocity modeling models by at least partially having labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model specifically includes: For each layer velocity modeling model, the layer velocity modeling model is tested by at least part of the labeled seismic reflection waveform data, a root mean square error and a mean absolute error of the layer velocity modeling model are determined, and a ratio of test data with prediction errors to total test data is counted as an error ratio; The comprehensive layer velocity modeling accuracy of the layer velocity modeling model is determined according to the sum of the root mean square error, the mean absolute error and the error ratio of the layer velocity modeling model.

5. The geological layer velocity modeling method according to claim 1, characterized in that: The amount of the re-acquired unlabeled seismic reflection waveform data increases round by round in the iterative training.

6. A geological layer velocity modeling device, characterized in that: include: An acquisition module, used to acquire labeled seismic reflection waveform data of annotated layer velocity models and unlabeled seismic reflection waveform data of unlabeled layer velocity models; wherein the labeled seismic reflection waveform data and the unlabeled seismic reflection waveform data correspond to multiple types of layer velocity models; A training and testing module is used to train multiple preset neural network models respectively through different labeled seismic reflection waveform data to obtain multiple layer velocity modeling models; and to test the multiple layer velocity modeling models through at least part of the labeled seismic reflection waveform data to determine the comprehensive layer velocity modeling accuracy of each layer velocity modeling model and the modeling accuracy of each type of layer velocity model; A prediction module is used to input the unlabeled seismic reflection waveform data into the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy to obtain a main pseudo-label of the unlabeled seismic reflection waveform data; and obtain a secondary pseudo-label of the unlabeled seismic reflection waveform data through the remaining layer velocity modeling models; for each type of layer velocity model in the main pseudo-label, when the modeling accuracy of the layer velocity modeling model corresponding to the main pseudo-label for the layer velocity model of this type is less than the modeling accuracy of any other layer velocity modeling model for the layer velocity model of this type, the corresponding main pseudo-label is replaced with the secondary pseudo-label obtained by the layer velocity modeling model with the highest modeling accuracy for the layer velocity model of this type; the updated main pseudo-label is used as the pseudo-label of the corresponding unlabeled seismic reflection waveform data to convert the unlabeled seismic reflection waveform data into pseudo-labeled seismic reflection waveform data; An updating module is used to update the model parameters of the remaining layer velocity modeling models according to the model parameters of the layer velocity modeling model with the highest comprehensive layer velocity modeling accuracy; and mix the labeled seismic reflection waveform data and the pseudo-labeled seismic reflection waveform data as updated labeled seismic reflection waveform data; The iterative training and application module is used to perform multiple rounds of iterative training on multiple layer velocity modeling models through updated labeled seismic reflection waveform data and re-acquired unlabeled seismic reflection waveform data, and perform layer velocity modeling through the layer velocity modeling model with the highest final comprehensive layer velocity modeling accuracy.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images