Three-dimensional geologic model automatic reconstruction method based on conditional spectrum normalized generative adversarial network

Abstract

The invention provides a three-dimensional geologic model automatic reconstruction method based on a conditional spectrum normalized generative adversarial network, which relates to the field of three-dimensional geologic modeling and comprises the following steps: preprocessing a sample three-dimensional geologic model to obtain a conditional data set, and setting random noise; constructing a conditional spectrum normalized generative adversarial network, and training the conditional spectrum normalized generative adversarial network through the conditional data set and the random noise to obtain a trained conditional spectrum normalized generative adversarial network; and inputting condition data of a to-be-reconstructed three-dimensional geologic model into the trained condition spectrum normalization generative adversarial network to obtain a reconstructed three-dimensional geologic model. According to the method, a discriminator loss function and a generator joint loss function are designed to train and update a discriminator and a generator, the gradient dissipation problem caused by a traditional loss measurement mode is solved, and the similarity between a reconstruction result and condition data is improved by calculating an extra loss value for a condition data position.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional geological modeling, and in particular to a method for automatically reconstructing a three-dimensional geological model based on a conditional spectral normalization generative adversarial network. Background Art

[0002] The fine reconstruction of three-dimensional geological models is a key link in geo-information science research and resource development, which can improve the understanding and description of underground stratigraphic structures and properties. A good stratigraphic model can accurately reflect the complex geological pattern characteristics of the study area and improve the scientificity and accuracy of exploration and development decisions. However, with the deepening of exploration, the stratigraphic interlacing is complex, the spatial distribution is irregular, and the sedimentary patterns are more diverse. Coupled with the typical heterogeneity and non-stationarity presented by geological models, modeling technology faces greater challenges and requirements. Modeling methods can be mainly divided into deterministic modeling and stochastic modeling, which respectively give deterministic stratigraphic distribution prediction results based on geological data and randomly simulate multiple possible prediction results for unknown areas. Although deterministic modeling can directly integrate the experience and knowledge of geological experts, it is limited by the sparsity of geological data, and there are problems such as modeling difficulties, poor characterization of results continuity and variability, and difficulty in assessing reliability. One of the most typical methods in stochastic modeling is the simulation method based on geostatistics, such as multi-point geostatistics, which can solve the problem of reconstructing the spatial distribution of underground heterogeneous structures and attributes, and can quantify the uncertainty of the results. However, since it is based on the assumption of stationarity, the simulation results only retain the repetitive patterns in the training images, which not only lacks diversity, but also makes it difficult to characterize the anisotropy and structural characteristics of the stratigraphic model, and the application effect in the field of fine modeling is poor. In addition, this method is computationally intensive, and a large number of parameters are discarded after the simulation process is completed. The parameters need to be recalculated for the next simulation, which is inefficient. Thanks to the powerful feature extraction ability of deep learning, the calculation of a large number of parameters is integrated in the calculation and optimization process of the model, and there is no need to recalculate the parameters when simulating again, which can effectively overcome the limitations of the above numerical simulation methods. Among them, the method based on generative adversarial network (GAN) can effectively fit the distribution law characteristics of the training data set and realize the automatic construction of three-dimensional geological bodies due to its special adversarial learning idea.

[0003] As a typical unsupervised learning algorithm, the generative adversarial network is mainly composed of two basic networks: the generator and the discriminator. The generator takes random noise as input, maps it to the manifold space of the training data set through a neural network, and generates the corresponding simulation results. The discriminator needs to distinguish the input data from different sources to identify their authenticity. The generator and the discriminator improve their respective performance in an adversarial learning manner, and generate realistic data through multiple rounds of iterations, back propagation and gradient updates until the "Nash equilibrium". The trained network parameters are saved, and the generation task can be completed quickly without recalculation. Based on the unknown density function, GAN can directly perform variational optimization, describe the structural characteristics of the geological body by fitting high-order statistical features, and its superiority in random simulation of strata is obvious. Specifically, it does not simply copy the geological pattern from the training image, but learns the potential distribution characteristics of the data in it, which can ensure the authenticity and generation diversity of the simulation results.

[0004] However, the highly nonlinear, unsteady and inhomogeneous characteristics of real geological structures make it difficult for GAN to reproduce the global features implicit in the data, limiting the reconstruction of complex heterogeneous structures. Specifically, for spatial attribute fields with strong non-stationarity, the fixed-size receptive field ignores the statistical feature changes within different scales during feature extraction, making it difficult to fit the more comprehensive potential distribution law of the corresponding training data set. However, the ASPP structure has multiple parallel dilated convolution modules with different sampling rates. By constructing convolution kernels with different receptive fields, the features extracted at each sampling rate are further processed in a separate branch and fused to generate the final feature map. Therefore, GAN can better perceive local and multi-scale feature change information, understand the structural attribute information of the image, enhance the correlation between local features and global features, effectively solve the problems caused by the scale changes of geological features in the reconstruction task, and improve the reconstruction performance in complex scenes.

[0005] In addition, since the training of GAN requires frequent adjustment of hyperparameters and design of network architecture, when the training between the generator and the discriminator is unbalanced, the generator is prone to fall into a local optimal state, that is, finding a small number of repeated samples to "cheat" the discriminator. In subsequent training, the generator will tend to continue to generate these samples and ignore other possibilities, resulting in a loss of diversity in results and mode collapse. Although the Wasserstein distance can measure the difference between the generated and true distributions, in order to meet the Lipschitz constraint, methods such as weight clipping and gradient penalty are computationally expensive, and the interpolation process is prone to cause training instability or divergence. Therefore, it is necessary to introduce an efficient weight normalization method to stabilize the training and control the Lipschitz constant of the discriminator, thereby reducing mode collapse, preventing its gradient anomalies, and improving generalization ability, thereby improving the performance and training stability of GAN.

[0006] At the same time, in the early stage of training, due to the large difference between the generated data distribution and the real distribution, the saturation interval of log2 is easy to exist when measuring the difference based on KL divergence and JS divergence, and the gradient dissipation or explosion phenomenon will occur in the back propagation process. In addition, the conditional data plays a decisive role in the reconstruction effect of the stratigraphic model. In order to make the reconstruction result satisfy the conditional data on the basis of satisfying the original geological knowledge, that is, it is hoped that the generated result can match the observed values ​​of pixels or voxels at known positions. However, due to the influence of the network structure, the conditional data is difficult to integrate into the simulation process and the constraint effect is poor. Directly applying sparse hard data in the modeling process may be ignored by the discriminator as noise, and it is difficult to affect the global spatial characteristics. Therefore, it is necessary to design an effective data enhancement method and establish a reasonable measurement method, which can not only punish the generation of unreasonable data distribution, but also force the learning of the geological pattern knowledge implicit in the real data distribution.

[0007] In summary, in order to realize an efficient and automatic reconstruction method of three-dimensional geological models, the following challenges exist: the stratigraphic heterogeneity is obvious, the local multi-scale characteristics vary complexly, and the spatial structure is difficult to accurately characterize; the GAN model has poor stability, and the training process is prone to mode collapse and gradient anomaly problems; the prior geological knowledge is insufficiently expressed, and the reconstruction results are difficult to be faithful to the conditional constraints. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide a method for automatic reconstruction of three-dimensional geological models based on conditional spectral normalization generative adversarial networks, which is used to solve the problems of insufficient expression of prior geological knowledge in existing model reconstruction methods and difficulty in reconstructing results to be faithful to conditional constraints.

[0009] The present invention provides a method for automatically reconstructing a three-dimensional geological model based on a conditional spectral normalization generative adversarial network, comprising:

[0010] S1: Obtain a sample 3D geological model of the study area, preprocess the sample 3D geological model to obtain a conditional data set, and set random noise;

[0011] S2: construct a conditional spectral normalization generative adversarial network, train the conditional spectral normalization generative adversarial network through a conditional data set and random noise, and obtain a trained conditional spectral normalization generative adversarial network;

[0012] S3: Input the conditional data of the three-dimensional geological model to be reconstructed into the trained conditional spectrum normalization generative adversarial network to obtain the reconstructed three-dimensional geological model.

[0013] Preferably, step S1 specifically comprises:

[0014] S11: Divide the sample 3D geological model into N sub-models based on a segmentation algorithm of 3D sliding voxels;

[0015] S12: For each sub-model, three two-dimensional sections are randomly selected along the X-axis, Y-axis and Z-axis directions, and random noise is filled between the two-dimensional sections to increase the dimension as conditional data;

[0016] S13: Repeat S12 N times to obtain a conditional data set of the same size as the sub-model.

[0017] Preferred:

[0018] The conditional spectral normalization generative adversarial network includes: a generator and a discriminator;

[0019] The generator includes: an encoder and a decoder;

[0020] The encoder includes: a DCNN module, a first Conv module, a first dilated Conv module, a second dilated Conv module, a third dilated Conv module, an Image Pooling module, a channel attention module, and a second Conv module;

[0021] The decoder includes: a third Conv module, a concat module, a fourth Conv module, a first Upsample module and a second Upsample module;

[0022] The DCNN module is connected to the first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module and the Image Pooling module;

[0023] The first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module, and the Image Pooling module are connected to the channel attention module;

[0024] The channel attention module, the second Conv module, the first Upsample module and the concat module are connected in sequence;

[0025] The DCNN module, the third Conv module, the concat module, the fourth Conv module, and the second Upsample module are connected in sequence;

[0026] The discriminator includes a first convolutional layer, a first hybrid attention module, a second convolutional layer, a second hybrid attention module, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer, which are connected in sequence.

[0027] Preferably, step S2 specifically comprises:

[0028] S21: Initialize the generator, discriminator and training hyperparameters;

[0029] S22: Obtain random noise z, conditional data set A and real data set B, where N is the batch training size;

[0030] S23: Input the random noise z and the conditional data set A into the generator G to obtain the generated data set B ′ =G(z,A), through the conditional data set A, the real data set B and the generated data set B ′ Calculate the discriminator loss function

[0031] S24: Through the discriminator loss function Perform back propagation and use the optimizer to update the parameters of the discriminator D to obtain the updated discriminator

[0032] S25: through the conditional data set A, the real data set B and the generated data set B ′ Calculate the generator joint loss function

[0033] S26: By combining the generator with the loss function Perform back propagation and use the optimizer to update the parameters of the generator G to obtain the updated generator

[0034] S27: Repeat steps S23-S26 until the training is completed to obtain a trained conditional spectral normalization generative adversarial network.

[0035] Preferred:

[0036] Discriminator loss function The calculation formula is:

[0037]

[0038] Among them, i represents the number of training times, D(B i ,A i ) represents the real data set B for the i-th training i And conditional data set A i Input the discrimination result of the discriminator D, D(B i ′ ,A i ) represents the generated data set B for the i-th training i ′ And conditional data set A i Input the discriminator’s judgment results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, and E represents the mathematical distribution expectation of the corresponding data set.

[0039] Preferred:

[0040] Generator joint loss function The calculation formula is:

[0041]

[0042] in, is the prior adversarial loss term, is the conditional loss term, λ is the weight of the conditional loss term;

[0043]

[0044] Among them, i represents the number of training times, B i represents the real data set for the i-th training, Represents the generated data set B for the i-th training i ′ And conditional data set A i Input the updated discriminator The discrimination results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, E represents the mathematical distribution expectation of the corresponding data set, represents element-wise multiplication, Dist represents Frobenius norm, and M represents a binary mask. The position M corresponding to the conditional data is marked as 1, and the other positions are marked as 0.

[0045] A storage medium, wherein the storage medium stores instructions and data for implementing the method for automatically reconstructing a three-dimensional geological model based on a conditional spectral normalization generative adversarial network.

[0046] A three-dimensional geological model automatic reconstruction device based on conditional spectral normalization generative adversarial network, comprising: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional geological model automatic reconstruction method based on conditional spectral normalization generative adversarial network.

[0047] The present invention has the following beneficial effects:

[0048] (1) The conditional spectral normalization generative adversarial network consists of a generator and a discriminator. The generator consists of an encoder-decoder. The encoder directly transfers the shallow features extracted by the DCNN module to the decoder to avoid the loss of feature information caused by multiple convolution processes. At the same time, by coupling multiple parallel expansion Conv modules, it can effectively extract and fuse the deep statistical characteristics of the multi-scale local features of the geological structure. The decoder combines deep and shallow features for step-by-step upsampling, and achieves model reconstruction based on both global and local features. The generator can achieve the fusion of multi-scale features, better perceive the characteristics of non-stationary geological bodies, enhance the ability to extract multi-scale local features, and solve the problems caused by the scale changes of geological features in the reconstruction task. The spectral normalization strategy is used on the convolutional layer in the discriminator to stabilize the network training process, reduce the risk of gradient explosion or disappearance, and improve the overall training stability and effect of the network.

[0049] (2) A channel attention module is added to the generator, and a mixed attention module is added to the discriminator. By introducing the attention module, we can focus on the more important parts of the feature image, which not only improves the accuracy and efficiency of feature extraction, but also improves the quality of the reconstruction results.

[0050] (3) A discriminator loss function and a generator joint loss function are designed to train and update the discriminator and generator, overcoming the gradient dissipation problem caused by the traditional loss measurement method. The additional loss value is calculated for the conditional data position to improve the similarity between the reconstruction result and the conditional data, thereby realizing the automatic reconstruction of the three-dimensional geological model under conditional constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;

[0052] Figure 2 Generate adversarial network structure diagram for conditional spectral normalization;

[0053] Figure 3 This is the structure diagram of the hybrid attention module;

[0054] Figure 4 Multiple automatic reconstruction models for three-dimensional lithofacies models of fold structures;

[0055] Figure 5It is the phase ratio statistical diagram of the lithofacies reconstruction model and its corresponding reference model;

[0056] Figure 6 The reconstructed variogram and connected function diagram of the lithofacies reconstruction model;

[0057] Figure 7 Multiple automatic reconstruction models for underground three-dimensional aquifer models;

[0058] Figure 8 It is the quantitative statistical results of the aquifer model;

[0059] Fig. 9 The reconstructed variogram and connected function diagram of the aquifer model;

[0060] Fig.10 This is a structural diagram of the device according to an embodiment of the present invention;

[0061] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0062] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0063] Reference Figure 1 The present invention provides a method for automatically reconstructing a three-dimensional geological model based on a conditional spectral normalization generative adversarial network, comprising:

[0064] S1: Obtain a sample 3D geological model of the study area, preprocess the sample 3D geological model to obtain a conditional data set, and set random noise;

[0065] Furthermore, step S1 is specifically as follows:

[0066] S11: Divide the sample 3D geological model into N sub-models based on a segmentation algorithm of 3D sliding voxels;

[0067] Specifically, geological model preparation: preparing a three-dimensional geological model, i.e., a training image built by combining survey data and expert experience, and obtaining its corresponding attribute information;

[0068] Dataset construction: The segmentation method based on 3D sliding voxels is used to divide it into a small model of 64×64×64 (unit: voxel);

[0069] S12: For each sub-model, three two-dimensional sections are randomly selected along the X-axis, Y-axis and Z-axis directions, and random noise is filled between the two-dimensional sections to increase the dimension as conditional data;

[0070] S13: Repeat S12 N times to obtain a conditional data set of the same size as the sub-model.

[0071] Specifically, three two-dimensional sections at different positions are randomly selected along the XYZ directions as the corresponding conditional data x con , fill random noise between sections to upgrade the dimension of two-dimensional section data, and establish a data set mapping relationship with the label model.

[0072] S2: construct a conditional spectral normalization generative adversarial network, train the conditional spectral normalization generative adversarial network through a conditional data set and random noise, and obtain a trained conditional spectral normalization generative adversarial network;

[0073] Furthermore, the conditional spectral normalization generative adversarial network includes: a generator and a discriminator;

[0074] The generator includes: an encoder and a decoder;

[0075] The encoder includes: a DCNN module, a first Conv module, a first dilated Conv module, a second dilated Conv module, a third dilated Conv module, an ImagePooling module, a channel attention module, and a second Conv module;

[0076] The decoder includes: a third Conv module, a concat module, a fourth Conv module, a first Upsample module and a second Upsample module;

[0077] The DCNN module is connected to the first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module and the ImagePooling module;

[0078] The first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module, and the ImagePooling module are connected to the channel attention module;

[0079] The channel attention module, the second Conv module, the first Upsample module and the concat module are connected in sequence;

[0080] The DCNN module, the third Conv module, the concat module, the fourth Conv module, and the second Upsample module are connected in sequence;

[0081] The discriminator includes a first convolutional layer, a first hybrid attention module, a second convolutional layer, a second hybrid attention module, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer, which are connected in sequence.

[0082] Specifically, the overall design of the generator is an encoding-decoding architecture, using convolutional layers and batch normalization layers to replace the fully connected layers and pooling layers in traditional three-dimensional convolutional neural networks. In the information extraction process, shallow features are extracted through the DCNN module (shallow network module), and the encoder is responsible for transferring the shallow feature information obtained by preliminary downsampling to the corresponding position of the decoder to reduce the loss of detail information caused by multi-layer convolution. In the deep convolution process, the ASPP structure is introduced, which consists of five parallel branches, namely a 1×1 ordinary convolution module, three 3×3 dilated convolution modules, and a global average pooling module. The first four branches are composed of convolutional layers, batch normalization layers, and activation layers. The pooling module adds an additional pooling layer, which can effectively extract high-order statistical features of the data set. In the parallel module, different values ​​of the dilation rate setting represent receptive fields of different sizes. The results of the five branches are fused and passed into the coupled channel attention module to adaptively adjust the weight of each feature channel, so that the model can pay more attention to the features that are useful for the task. The decoder completes the upsampling reconstruction process under the guidance of deep and shallow feature information to ensure the quality of output features;

[0083] The discriminator uses a discriminative method based on PatchGAN, which enhances the stability of the training process by using convolutional layers, while removing the pooling layer and batch normalization layer used in traditional downsampling. After the first two convolutional layers of the discriminator, a hybrid attention module (CBAM) of space and channel is embedded, which can simultaneously focus on the important information of the input data in the spatial and channel dimensions, and improve the discriminator's ability to discriminate the input data. The structure of the hybrid attention module is as follows: Figure 3 shown.

[0084] At the same time, in order to satisfy the Lipschitz continuity condition, it is ensured that the local change of the discriminator function f(x) in the input space is not too drastic, that is, any two points x and x in the input space ′ ,satisfy:

[0085] |f(x)-f(x′)|≤K||xx′||

[0086] Using spectral normalization technology on each convolutional layer can process the spectral norm of the weight matrix and limit the gradient change of the discriminator function. The spectral norm represents the maximum singular value of the weight matrix W, that is, W T The square root value of the maximum eigenvalue λ1 of W can be written as:

[0087]

[0088] In order to satisfy the Lipschitz constraint, let σ(W) = 1, any input weight matrix x can satisfy the following after matrix multiplication:

[0089] |Wx|<σ(W)|x|

[0090] Therefore, during the training process, it is only necessary to normalize the weight matrix. The new weight matrix W SM =W / σ(W) not only satisfies the Lipschitz continuity constraint, but also avoids sensitivity to specific directions in the hidden layer. In addition, the LeakyRelu activation function is used after the convolutional layer other than the final output layer.

[0091] Furthermore, step S2 is specifically as follows:

[0092] S21: Initialize the generator, discriminator and training hyperparameters;

[0093] Specifically, initialize the generator to G θ , initialize the discriminator to D θ , define the hyperparameter value, for example: the number of training rounds train num , batch training size M, learning rate η G and η D , conditional loss measurement method, conditional loss item weight coefficient λ con and optimizer, etc. The settings of these parameters have an important impact on the efficiency of the training process and the performance of the model;

[0094] S22: Obtain random noise z, conditional data set A and real data set B, where N is the batch training size;

[0095] S23: Input the random noise z and the conditional data set A into the generator G to obtain the generated data set B ′ =G(z,A), through the conditional data set A, the real data set B and the generated data set B ′ Calculate the discriminator loss function

[0096] Furthermore, the discriminator loss function The calculation formula is:

[0097]

[0098] Among them, i represents the number of training times, D(B i ,A i ) represents the real data set B for the i-th training i And conditional data set A i Input the discrimination result of the discriminator D, D(B i ′ ,A i ) represents the generated data set B for the i-th training i′ And conditional data set A i Input the discriminator’s judgment results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, and E represents the mathematical distribution expectation of the corresponding data set.

[0099] S24: Through the discriminator loss function Perform back propagation and use the optimizer to update the parameters of the discriminator D to obtain the updated discriminator

[0100] S25: through the conditional data set A, the real data set B and the generated data set B ′ Calculate the generator joint loss function

[0101] Furthermore, the generator joint loss function The calculation formula is:

[0102]

[0103] in, is the prior adversarial loss term, is the conditional loss term, λ is the weight of the conditional loss term;

[0104]

[0105] Among them, i represents the number of training times, B i represents the real data set for the i-th training, Represents the generated data set B for the i-th training i ′ And conditional data set A i Input the updated discriminator The discrimination results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, E represents the mathematical distribution expectation of the corresponding data set, represents element-wise multiplication, Dist represents Frobenius norm, and M represents a binary mask. The position M corresponding to the conditional data is marked as 1, and the other positions are marked as 0.

[0106] Specifically, the binary mask can quickly filter out the attribute information of the conditional profile location between the simulation results and the reference model and perform loss measurement.

[0107] S26: By combining the generator with the loss function Perform back propagation and use the optimizer to update the parameters of the generator G to obtain the updated generator

[0108] S27: Repeat steps S23-S26 until the training is completed to obtain a trained conditional spectral normalization generative adversarial network.

[0109] Further, the performance evaluation of the trained conditional spectral normalization generative adversarial network is:

[0110] The test data set is used for reconstruction verification and evaluation. If the reconstruction performance of the model meets the standard, the model parameters will be saved and these parameters can be used for subsequent conditional reconstruction of the 3D geological model. Otherwise, the network parameter values ​​are reset and training starts again.

[0111] like Figure 4 It is a variety of automatic reconstruction models of three-dimensional lithofacies models of fold structures; Figure 5 is the quantitative statistical result of the lithofacies reconstruction model. Figure 5 (a) is a statistical diagram of the phase ratios of 50 randomly selected reconstructed models and their corresponding reference models (the specific values ​​are displayed in the form of a scatter plot next to the box plot); Figure 5 (b) is the MDS plot between the above 50 realizations and their corresponding references, and the Euclidean distance is used to visualize the difference between the two distributions; Figure 6 Figure 2 is the reconstructed variogram and connectivity function diagram of the lithofacies reconstruction model. (a) is a distribution diagram of the variogram curve of a reference model and 50 random reconstruction results; (b) and (c) are the connectivity function diagrams of the 50 reconstructed models and the corresponding references in the X and Y directions. It can be seen that the reference model is basically consistent with the simulation results.

[0112] Figure 7 It is a variety of automatic reconstruction models of underground three-dimensional aquifer models; Figure 8 It is the quantitative statistical result of the aquifer model. Figure 8 (a) is the phase ratio statistics of 50 randomly selected reconstructed models and their corresponding reference models; Figure 8 (b) is the MDS plot between the above 50 realizations and their corresponding references, and the Euclidean distance is used to visualize the difference between the two distributions; Fig. 9 The figure shows the reconstructed variogram and connectivity function of the aquifer model. (a) is a distribution diagram of the variogram curve of a reference model and 50 random reconstruction results; (b) and (c) are the X- and Y-direction connectivity function diagrams of the 50 reconstructed models and the corresponding references. It can be seen that the reference model is basically consistent with the simulation results.

[0113] In summary, the conditional spectral normalization generative adversarial network of the present invention has excellent reconstruction effect.

[0114] S3: Input the conditional data of the three-dimensional geological model to be reconstructed into the trained conditional spectrum normalization generative adversarial network to obtain the reconstructed three-dimensional geological model.

[0115] Specifically, the random noise z and the conditional data A are input into the generator of the encoding-decoding architecture together. The encoder extracts feature information of different scales through several parallel atrous convolution modules with receptive fields of different sizes, and processes them separately in the corresponding branches. The decoder fuses the deep and shallow feature information to jointly guide the upsampling process to obtain the generated data, which is the reconstructed three-dimensional geological model.

[0116] See also Fig.10 , Fig.10 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a three-dimensional geological model automatic reconstruction device 401 based on conditional spectral normalization generative adversarial network, a processor 402 and a storage medium 403.

[0117] A three-dimensional geological model automatic reconstruction device 401 based on conditional spectral normalization generative adversarial network: The three-dimensional geological model automatic reconstruction device 401 based on conditional spectral normalization generative adversarial network implements the three-dimensional geological model automatic reconstruction method based on conditional spectral normalization generative adversarial network.

[0118] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the three-dimensional geological model automatic reconstruction method based on conditional spectral normalization generative adversarial network.

[0119] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the three-dimensional geological model automatic reconstruction method based on conditional spectral normalization generative adversarial network.

[0120] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0121] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0122] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for automatic reconstruction of three-dimensional geological models based on conditional spectral normalization generative adversarial network, characterized in that: include: S1: Obtain a sample 3D geological model of the study area, preprocess the sample 3D geological model to obtain a conditional data set, and set random noise; S2: construct a conditional spectral normalization generative adversarial network, train the conditional spectral normalization generative adversarial network through a conditional data set and random noise, and obtain a trained conditional spectral normalization generative adversarial network; S3: Input the conditional data of the three-dimensional geological model to be reconstructed into the trained conditional spectrum normalization generative adversarial network to obtain the reconstructed three-dimensional geological model.

2. The method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network according to claim 1, characterized in that: Step S1 is specifically as follows: S11: Divide the sample 3D geological model into N sub-models based on a segmentation algorithm of 3D sliding voxels; S12: For each sub-model, three two-dimensional sections are randomly selected along the X-axis, Y-axis and Z-axis directions, and random noise is filled between the two-dimensional sections to increase the dimension as conditional data; S13: Repeat S12 N times to obtain a conditional data set of the same size as the sub-model.

3. The method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network according to claim 1, characterized in that: The conditional spectral normalization generative adversarial network includes: a generator and a discriminator; The generator includes: an encoder and a decoder; The encoder includes: a DCNN module, a first Conv module, a first dilated Conv module, a second dilated Conv module, a third dilated Conv module, an ImagePooling module, a channel attention module, and a second Conv module; The decoder includes: a third Conv module, a concat module, a fourth Conv module, a first Upsample module and a second Upsample module; The DCNN module is connected to the first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module and the ImagePooling module; The first Conv module, the first dilated Conv module, the second dilated Conv module, the third dilated Conv module, and the ImagePooling module are connected to the channel attention module; The channel attention module, the second Conv module, the first Upsample module and the concat module are connected in sequence; The DCNN module, the third Conv module, the concat module, the fourth Conv module, and the second Upsample module are connected in sequence; The discriminator includes a first convolutional layer, a first hybrid attention module, a second convolutional layer, a second hybrid attention module, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer, which are connected in sequence.

4. The method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network according to claim 3 is characterized in that: Step S2 is specifically as follows: S21: Initialize the generator, discriminator and training hyperparameters; S22: Obtain random noise z, conditional data set A and real data set B, where N is the batch training size; S23: Input the random noise z and the conditional data set A into the generator G to obtain the generated data set B ′ =G(z,A), through the conditional data set A, the real data set B and the generated data set B ′ Calculate the discriminator loss function S24: Through the discriminator loss function Perform back propagation and use the optimizer to update the parameters of the discriminator D to obtain the updated discriminator S25: through the conditional data set A, the real data set B and the generated data set B ′ Calculate the generator joint loss function S26: By combining the generator with the loss function Perform back propagation and use the optimizer to update the parameters of the generator G to obtain the updated generator S27: Repeat steps S23-S26 until the training is completed to obtain a trained conditional spectral normalization generative adversarial network.

5. The method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network according to claim 4 is characterized in that: Discriminator loss function The calculation formula is: Among them, i represents the number of training times, D(B i ,A i ) represents the real data set B for the i-th training i And conditional data set A i Input the discrimination result of the discriminator D, D(B i ′ ,A i ) represents the generated data set B for the i-th training i ′ And conditional data set A i Input the discriminator’s judgment results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, and E represents the mathematical distribution expectation of the corresponding data set.

6. The method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network according to claim 4, characterized in that: Generator joint loss function The calculation formula is: in, is the prior adversarial loss term, is the conditional loss term, λ is the weight of the conditional loss term; Among them, i represents the number of training times, B i represents the real data set for the i-th training, Represents the generated data set B for the i-th training i ′ And conditional data set A i Input the updated discriminator The discrimination results, P(A), P(B) and P(B ′ ) represent the distribution characteristics of the conditional data set, the real data set and the generated data set respectively, E represents the mathematical distribution expectation of the corresponding data set, represents element-wise multiplication, Dist represents Frobenius norm, and M represents a binary mask. The position M corresponding to the conditional data is marked as 1, and the other positions are marked as 0.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing the method for automatic reconstruction of a three-dimensional geological model based on a conditional spectral normalization generative adversarial network as described in any one of claims 1 to 6.

8. A three-dimensional geological model automatic reconstruction device based on conditional spectral normalization generative adversarial network, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the method for automatic reconstruction of three-dimensional geological models based on conditional spectral normalization generative adversarial network as described in any one of claims 1 to 6.

Images