Fault constraint-based Dal-UNet stratum curved surface reconstruction method

Abstract

The invention discloses a fault constraint-based Dal-UNet stratum curved surface reconstruction method, which is applied to the field of information processing and the field of oil-gas exploration, and aims to solve the problems that an existing deep learning reconstruction algorithm is based on a supervised learning mode, a large amount of data information and manually marked training labels are needed for pre-training a network, the required economic and time cost is relatively high, and the efficiency is relatively high. The problems that in geological research with relatively rare data, an algorithm based on supervised deep learning is prone to generating an over-fitting phenomenon and is more difficult to be directly applied to a stratum curved surface reconstruction task of a complex geological structure containing a fault in the prior art are solved. According to the method, implicit features are captured from limited and sparse stratum curved surface data by using the Dal-UNet network, and a fault-containing complex stratum curved surface is reconstructed in combination with a fault constraint algorithm, so that the real landform is effectively restored.

Description

Technical Field

[0001] The present invention belongs to the fields of information processing and oil and gas exploration, and particularly relates to a formation surface reconstruction technology. Background Art

[0002] Formation surface reconstruction is a key step in carrying out geospatial visualization work, and is one of the key issues jointly concerned by the fields of information processing and oil and gas exploration. It plays an important role in promoting subsequent work such as finding solid mineral deposits, analyzing the migration and accumulation of oil, natural gas and groundwater, evaluating the stability of large engineering foundations, and monitoring and analyzing earthquake prediction. However, limited by the sparsity, inhomogeneity of data, and the complexity of geological structures, the reliability and accuracy of traditional geological horizon surface reconstruction are not high. Therefore, further expanding the research on formation surface reconstruction methods has very important significance and value.

[0003] Reconstructing the formation surface containing faults generally requires complete information such as elevation, fault throw, dip angle, etc. In most cases, only the planar projection information of the fault is available, which is not conducive to modeling the formation surface with faults and drawing contour maps. Ji Zhanhuai et al. [1] Improved the fault trace method. Only by knowing the planar position information of the fault, the shortest path algorithm can be combined with the inverse distance weighted interpolation method to effectively solve the formation interpolation problem under fault constraints.

[0004] In 2018, Ulyanov et al. found that the convolutional neural network (CNN) can directly learn the distribution characteristics of image data from the damaged two-dimensional image itself, and can achieve goals such as image restoration or super-resolution interpolation without additional labeled data for pre-training, saving time and money costs. And excellent results have been obtained in standard inverse problems (such as denoising, super-resolution), and it can be flexibly applied to various interpolation and reconstruction environments.

[0005] The U-Net network is a classic convolutional neural network, named after its unique U-shaped structure. It consists of an encoder responsible for feature extraction and downsampling and a decoder for upsampling and feature fusion. The two achieve efficient integration of multi-scale features through skip connections, and finally achieve accurate image segmentation. The U-Net network is widely used in many image segmentation fields such as medicine, biology, and remote sensing, and is known for its high data utilization efficiency, high segmentation accuracy, and strong adaptability, which has greatly promoted the development of related fields.

[0006] Traditional formation surface reconstruction methods include interpolation methods, fitting methods, methods based on partial differential equations, etc. These methods are driven by theoretical knowledge and cannot well capture the implicit features of geological data. With the continuous development of deep learning technology, many scholars have applied deep learning to the formation surface reconstruction task, resulting in many research results. Most deep learning reconstruction algorithms are based on a supervised learning mode, which requires a large amount of data and manually labeled training labels to pre-train the network. The required economic and time costs are relatively high. The quality of the data directly affects the performance of the algorithm. Moreover, in geological research with relatively scarce data, algorithms based on supervised deep learning are prone to overfitting, making it more difficult to directly apply them to the formation surface reconstruction task of complex geological structures containing faults. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a Dal-UNet formation surface reconstruction method based on fault constraints, which uses the Dal-UNet network to capture implicit features from limited and sparse formation surface data, and combines a fault constraint algorithm to reconstruct complex formation surfaces containing faults, effectively restoring the true topography.

[0008] The technical solution adopted by the present invention is as follows: A Dal-UNet formation surface reconstruction method based on fault constraints, including:

[0009] S1. Grid the survey line data;

[0010] S2. Mark the misaligned points belonging to the same fault in the gridded survey line data, and then connect the misaligned points belonging to the same fault one by one to obtain the distribution positions of each fault;

[0011] S3. Interpolate the hanging wall line and footwall line of the fault according to the distribution position of the fault to obtain sparse formation data;

[0012] S4. Replace the traditional residual convolution block in MultiResUNet with a dilated convolution residual module to obtain Dal-Unet;

[0013] S5. Train Dal-Unet with the sparse formation data obtained in step S4;

[0014] S6. Input the sparse formation data corresponding to the processed survey line data into the trained Dal-Unet to obtain the formation surface reconstruction result.

[0015] Advantages of the present invention: The method of the present invention addresses the situation where directly using survey line data to reconstruct the formation surface may lead to discrepancies between the terrain changes and the terrain trend and the real results. By integrating the reconstruction loss function and the constraint loss function, a fault surface reconstruction loss function is proposed to achieve fault constraint. Combining with the Dal-UNet model, the formation reconstruction results have been significantly improved in terms of visual effects, structural reconstruction, and fault constraint, and are closer to the real geographical features. It provides authentic and reliable geological data for subsequent geoscience research work such as searching for solid ore deposits, analyzing the migration and accumulation of oil, natural gas, and groundwater, evaluating the stability of large engineering foundations, and monitoring and analyzing earthquake prediction. Brief Description of the Drawings

[0016] Figure 1 is the method flow framework of the present invention;

[0017] Figure 2 is the three-dimensional multi-angle schematic diagram of the survey line data;

[0018] Among them, (a) is the three-dimensional scatter top view, and (b) is the three-dimensional scatter side view;

[0019] Figure 3 is the fault distribution comparison diagram;

[0020] Among them, (a) is the scatter side view, and (b) is the scatter side view with fault distribution;

[0021] Figure 4 is the three-dimensional multi-angle schematic diagram of the interpolation results of the fault hanging wall and footwall lines;

[0022] Among them, (a) is the top view of the interpolation results of the fault hanging wall and footwall lines, and (b) is the side view of the interpolation results of the fault hanging wall and footwall lines;

[0023] Figure 5 is the Dal-UNet network structure;

[0024] Figure 6 is the schematic diagram of the benchmark data construction process;

[0025] Among them, (a) is the original crater data, and (b) is the crater data with normal fault (benchmark data);

[0026] Figure 7 is the schematic diagram of the sparse observation data construction process;

[0027] Among them, (a) is the survey line sampling matrix, and (b) is the sparse observation data;

[0028] Figure 8 is the comparative experiment;

[0029] Among them, (a) is the reference data, (b) is the reconstruction effect of the survey line data interpolated with the fault hanging wall and footwall lines, (c) is the reconstruction effect of the inverse distance weighted interpolation method, (d) is the reconstruction effect of UNet + traditional reconstruction loss function, (e) is the reconstruction effect of MultiResUNet + traditional reconstruction loss function, and (f) is the reconstruction effect of the method of the present invention;

[0030] Figure 9 is the comparison between the original data and the interpolation result;

[0031] Among them, (a) is the three-dimensional display of the input data, and (b) is the three-dimensional display of the interpolation result;

[0032] Figure 10 is the reconstruction result Figure 3 three-dimensional visualization comparison;

[0033] Among them, (a) is the surface reconstruction result of the traditional reconstruction loss function at angle one, (b) is the surface reconstruction result of the fault surface reconstruction loss function at angle one, (c) is the surface reconstruction result of the traditional reconstruction loss function at angle two, and (d) is the surface reconstruction result of the fault surface reconstruction loss function at angle two. Detailed implementation manners

[0034] To facilitate those skilled in the art to understand the technical content of the present invention, the following further explains the content of the present invention in conjunction with the accompanying drawings.

[0035] As Figure 1 shown, the implementation process of the present invention includes the following content:

[0036] Standardize the acquired survey line data, map it to a 400*400 matrix grid according to the ratio of 50m:1 pixel point. If there are m survey line data points p i =(x i , y i ) that need to be mapped to the same grid point q j =(x j , y j ), in this embodiment, m is greater than or equal to 2, and perform inverse weighted averaging according to the Euclidean distance d ij from each survey line data to this grid point. The formula is as shown in (1):

[0037]

[0038] Among them: v i is the value of the survey line data point p i ; V j is the value of the grid point q j . The survey line data is as Figure 2 shown.

[0039] AsFigure 3 (a) As shown in the red box, the abrupt changes in the survey line data are due to the presence of faults. Faults are caused by localized shifts in the strata, which disrupt the continuity of the strata. If the influence of faults is ignored, the upper fault data may erroneously interfere with the reconstruction of the lower fault data, and vice versa, ultimately leading to distorted reconstruction results near the fault plane. To impose constraints on the surface reconstruction process and ensure the integrity and accuracy of the fault plane, this paper proposes an algorithm flow for fault constraints:

[0040] First, we introduce certain artificial constraints on the faults. Based on the position between the terrain trend and the displacement points (hereinafter referred to as the abrupt changes in the data in the 3D scatter plot), combined with manual interpretation, we mark which displacement points belong to the same fault, and then connect the displacement points belonging to the same fault one by one to obtain the distribution of many faults, such as Figure 3 Indicated by the black line in (b).

[0041] According to the fault distribution, the hanging wall line and the foot wall line of the fault are interpolated (the upper and lower wall lines of the fault in this invention are 2 pixels apart). The interpolation method is as follows:

[0042] Assume that the present invention needs to interpolate the upper wall line of m pixels and the lower wall line of m pixels between the fault point a and the fault point b. First, the fault value △a of the fault point a and the fault value △b of the fault point b must be calculated. The fault value is the vertical fault distance from the upper wall to the lower wall at that point. Linear interpolation between △a and △b forms a fault list D with a length of m. The set of survey line points on the upper side of the fault is denoted as P. u , the lower side is marked as P d In order to minimize the influence of faults, the upper plate line is interpolated using formula (3), and the lower plate line is interpolated using formula (4). The interpolation results of the upper and lower plate lines are as follows: Figure 4 shown.

[0043]

[0044] Where V k is the value of the interpolation point k, v i YesP u The value of point i, v j YesP d The value of point j in the middle, d is the Euclidean distance.

[0045] In order to further utilize the interpolation data of the upper and lower wall lines of the fault to achieve the final fault constraint effect, the present invention proposes a multi-task loss function - Fault Surface Reconstruction Loss. This loss function seeks a balance among the three tasks of overall reconstruction of the stratum surface, detailed reconstruction of the fault surface, and noise removal and smooth reconstruction results to achieve a good reconstruction effect of the stratum surface containing faults.

[0046]

[0047] This loss function consists of three parts, which are weighted and summed respectively, where λ C , λ R , λ TV are the weight parameters for each part:

[0048] The λ R term (reconstruction loss): responsible for the overall surface reconstruction of the sparse formation data. Among them, f(x) represents the reconstruction result, m is the mask matrix, which has a value of 1 at the positions of the valid points in the data to be repaired and 0 at the positions of the unobserved values (i.e., the positions that need to be interpolated), V obs is the data to be repaired (i.e., the known data), ⊙ is the Hadamard product, and N is the number of elements for effective calculation (i.e., the number of non-zero elements in the mask m).

[0049] The λ C term (constraint loss): By comparing the MSE error at the pixel level of the fault line positions between the formation reconstruction result and the original sparse formation data, the neural network can be made to focus more on the reconstruction of the fault line positions during the surface reconstruction process, achieving the fault constraint effect. Among them, f_m (i.e., fault mask) is also the mask matrix, which has a value of 1 at the fault line positions and 0 at other positions, V fault is the fault line data;

[0050] The λ TV term (total variation loss): This loss function is very effective in removing noise in the image while preserving edges and details. The input of the neural network in the formation reconstruction task is random noise, which is prone to generating a large number of high-frequency errors. At the same time, in order to achieve fault constraint, it is also a difficult point to retain the fault edges during the surface reconstruction. The present invention introduces the total variation loss as a regularization term to reduce these high-frequency errors through the regularized high-frequency components, which is very effective in removing high-frequency noise while retaining the fault edge data and detail textures. Among them, C refers to the number of channels, H is the height, and W is the width. x i,j refers to the pixel at the i-th row and j-th column in the reconstruction result.

[0051] The fault surface reconstruction loss function uses the known part of the sparse formation data to construct the mask matrix m, guiding the neural network for self-supervised learning. After the reconstruction result acts on the mask matrix, it is only compared point by point with the known part of the sparse formation data to minimize the numerical error. At the same time, based on the data distribution characteristics, that is, the implicit prior information, the neural network model automatically interpolates and reconstructs the missing area, improving the rationality and accuracy of the reconstruction result.

[0052] In a specific implementation, the fault surface reconstruction loss function proposed by the present invention is used to calculate the error between the model reconstruction result and the real formation data. That is, during the training process of Dal-UNet, within each training epoch, the model takes sparse formation data as input and generates the reconstruction result of the formation surface. According to formula (5), the loss function L is calculated. FSR The present invention uses the automatic differentiation technology based on the PyTorch framework for gradient backpropagation, and calculates the partial derivative of the loss function L with respect to the network parameters through the chain rule. FSR Then, the Adam optimizer is applied to perform parameter updates. Through the backpropagation process, the fault surface reconstruction loss function can effectively guide the network to update parameters, so as to reduce the overall formation surface reconstruction error, enhance the reconstruction accuracy of the fault area, optimize the surface smoothness and structural consistency, and finally achieve the purpose of accurately reconstructing the formation surface with faults.

[0053] However, traditional deep learning models often face problems such as insufficient information and difficulty in modeling spatial feature relationships when dealing with sparse data. Therefore, the present invention proposes an improved Dal-UNet structure to enhance the feature extraction ability of the model for sparse formation data and improve the accuracy of surface reconstruction.

[0054] Dal-UNet:

[0055] Figure 5 Figure shows the neural network structure of Dal-UNet, which is an improvement based on MultiResUNet. The present invention uses the dilated convolutional residual module (DalResBlock) to replace the traditional residual convolutional block, enabling the neural network to better grasp the feature connections between sparse formation data, thereby improving the accuracy and consistency of surface reconstruction; at the same time, 3×3 convolution with a stride of 2 is used for downsampling, enabling the network to learn more feature information during the dimensionality reduction process and reducing information loss; bilinear interpolation is used for upsampling to ensure the accuracy of surface reconstruction.

[0056] As Figure 5As shown in the figure, the Dal-UNet of the present invention includes: the input of the first dilated convolutional residual module is sparse formation data, the output of the first dilated convolutional residual module is downsampled and used as the input of the second dilated convolutional residual module, the output of the second dilated convolutional residual module is downsampled and used as the input of the third dilated convolutional residual module, the output of the third dilated convolutional residual module is downsampled and used as the input of the fourth dilated convolutional residual module, the output of the fourth dilated convolutional residual module is downsampled and used as the input of the fifth dilated convolutional residual module, the output of the fifth dilated convolutional residual module is upsampled and used as the input of the sixth dilated convolutional residual module, and the input of the sixth dilated convolutional residual module also includes the result of processing the output of the fourth dilated convolutional residual module through a path residual block; the output of the sixth dilated convolutional residual module is upsampled and used as the input of the seventh dilated convolutional residual module, and the input of the seventh dilated convolutional residual module also includes the result of processing the output of the third dilated convolutional residual module through a path residual block; the output of the seventh dilated convolutional residual module is upsampled and used as the input of the eighth dilated convolutional residual module, and the input of the eighth dilated convolutional residual module also includes the result of processing the output of the second dilated convolutional residual module through a path residual block; the output of the eighth dilated convolutional residual module is upsampled and used as the input of the ninth dilated convolutional residual module, and the input of the ninth dilated convolutional residual module also includes the result of processing the output of the first dilated convolutional residual module through a path residual block; the output of the ninth dilated convolutional residual module is the reconstruction result.

[0057] Train the Dal-UNet, and obtain the formation surface reconstruction result according to the trained Dal-UNet.

[0058] In practical applications, the survey line data collected through the survey line is as Figure 2 shown. By interpolating the fault hanging wall and footwall lines of the survey line data collected through the survey line, and then self-supervised training the Dal-UNet network model according to the interpolated real formation data.

[0059] The following combines specific data to illustrate the technical effects of the present invention:

[0060] 1. Simulated data

[0061] Since the real formation data (survey line data) is relatively sparse and it is difficult to directly evaluate the accuracy of the reconstruction result, therefore, to further verify the effectiveness of the Dal-UNet formation surface reconstruction method based on fault constraints proposed by the present invention, the present invention designs a set of simulated data experiments. The purpose of this experiment is to construct synthetic data with known fault structures, and compare the traditional method with the present invention in terms of reconstruction accuracy, fault preservation and surface smoothness, etc., to quantify the advantages of the present invention in fault constraint reconstruction.

[0062] During the construction of experimental data, this simulation experiment is based on crater data. A normal fault is simulated by the local subsidence of the stratum. This data serves as the benchmark (Ground Truth), as shown in Figure 6 (b). Subsequently, simulated survey line sampling is performed on the benchmark data to simulate the sparse observation situation of real stratum data. The generation method of the survey line sampling matrix is as follows: Set the minimum spacing to 10 and the maximum spacing to 40. Multiple non-uniformly spaced oblique lines are randomly generated within this range, as shown in Figure 7 (a). The finally obtained sparse observation data is as shown in Figure 7 (b).

[0063] This experiment systematically compared the stratum surface reconstruction performance of different methods under the condition of sparse observation data, and focused on evaluating their ability to preserve fault characteristics. Since the preprocessing part in the present invention involves the interpolation of the fault hanging wall and footwall lines, which will change the input data, for the sake of fairness, the survey line data after interpolating the fault hanging wall and footwall lines is used as the input in the following experiments.

[0064] The comparison methods include: (c) inverse distance weighted interpolation method; (d) UNet + traditional reconstruction loss function; (e) MultiResUNet + traditional reconstruction loss function; (f) the method proposed in the present invention - Dal-UNet + fault surface reconstruction loss function.

[0065] This experiment is carried out under the PyTorch framework. The optimizer is selected as Adam, and the initial learning rate is set to 0.001. In addition, to improve the generalization ability of the model, the experiment uses 180° rotated data for 30 rounds of pre-training, and the final total number of training rounds is 3000 rounds. The model performance evaluation adopts a combination of qualitative analysis and quantitative analysis. The indicators include mean square error (MSE), signal-to-noise ratio (SNR), structural similarity (SSIM), etc., to comprehensively analyze the performance of different methods in aspects such as fault feature preservation, surface smoothness, and overall stratum structure consistency.

[0066] Qualitative analysis:

[0067] Figure 8The reconstruction results of different methods are shown. (c) Inverse distance weighted interpolation method: obvious cuts appear in the reconstruction process, making it difficult to reconstruct the original stratigraphic structure when the data is sparse, and the spatial distribution of the fault cannot be accurately described; (d) UNet + traditional reconstruction loss function: the original stratigraphic structure is restored to a certain extent, but due to the limited receptive field of UNet, cuts along the survey line are still present in the result, accompanied by a large number of noise points; (e) MultiResUNet + traditional reconstruction loss function: the reconstruction results show obvious high-value anomalies in the edge area, that is, the errors are concentrated at the edge, and a dense abnormal high-value distribution appears, which is significantly deviated from the geological structure in the benchmark data (a). The possible reasons are: the receptive field of MultiResUNet is insufficient and the convolution layer cascade is too deep, and the traditional reconstruction loss function lacks effective noise suppression, resulting in uncontrolled boundary numerical diffusion, thus forming error accumulation; (f) The present invention: the stratified structure of the stratum is effectively maintained during the reconstruction process, and obvious step features are shown in the fault area, which is highly consistent with the benchmark data (a). At the same time, the line traces are significantly reduced, some texture details are preserved, and the overall reconstruction process is stable and less noisy. This proves that the present invention has greater advantages in the depiction of fault space morphology, boundary error control, and training stability.

[0068] Quantitative analysis:

[0069] Table 1 lists the performance of three different reconstruction methods in terms of mean square error (MSE), signal-to-noise ratio (SNR) and structural similarity (SSIM).

[0070] MSE reflects the mean square error between the reconstructed data and the real data. The smaller the value, the higher the reconstruction accuracy. Overall, the MSE of the inverse distance weighted method is 6.85, while the MSE of UNet is 4.35, indicating that both have certain reconstruction errors; in comparison, the overall MSE of MultiResUNet is as high as 19.87, indicating that there is a large deviation in its overall reconstruction. The overall MSE of the present invention is only 1.09, which significantly reduces the reconstruction error and shows a higher reconstruction accuracy. With regard to the fault area, the MSE of the inverse distance weighted method is 6.67, and that of UNet and MultiResUNet are 1.86 and 1.98 respectively, while the MSE of the present invention in the fault area is 0.86, proving that the present invention is more accurate in reconstructing the fault structure.

[0071] Table 1 Comparison of reconstruction performance of different methods

[0072] Index Inverse distance weighting method UNet MultiResUNet The method of the present invention MSE(↓) 6.85 4.35 19.87 1.09 SNR(↑) 11.64 13.62 7.01 19.62 SSIM(↑) 0.668 0.647 0.516 0.907 MSE of the fault area(↓) 6.67 1.86 1.98 0.86 SNR of the fault area(↑) 13.42 18.97 18.70 22.33 SSIM of the fault area(↑) 0.627 0.955 0.968 0.976

[0073] The SNR measures the quality of the signal. The higher the value, the less noise in the reconstruction result and the closer the data is to the true value. From the overall indicators, the SNR of the inverse distance weighting method is 11.64, that of UNet is 13.62, while the SNR of MultiResUNet is only 7.01, indicating that its reconstruction result has more noise. In contrast, the SNR of the present invention reaches 19.62, significantly improving the signal quality. In the fault area, the SNR of the inverse distance weighting method is 13.42, those of UNet and MultiResUNet are 18.97 and 18.70 respectively, while the SNR of the present invention in the fault area is 22.33, further demonstrating its advantage in noise suppression.

[0074] The SSIM is used to measure the structural similarity between the reconstruction result and the real data. The closer the value is to 1, the more accurate the structural reconstruction. Overall, the SSIM of the inverse distance weighting method is 0.668, that of UNet is 0.647, and that of MultiResUNet is even lower, only 0.516, indicating that these three methods have deficiencies in maintaining the overall formation structure consistency. The overall SSIM of the present invention is as high as 0.907, indicating its excellent performance in restoring the overall structural characteristics. In the fault area, the SSIM of the inverse distance weighting method is 0.627, those of UNet and MultiResUNet are 0.955 and 0.968 respectively, while the SSIM of the present invention in the fault area is 0.976, further reflecting the significant advantage of the present invention in accurately reconstructing the fault structure. It should be noted that the SSIM of UNet and MultiResUNet in the fault area has a significant improvement compared to the overall SSIM, which is mainly due to the interpolation of the fault hanging wall and footwall lines in the data preprocessing of the present invention, providing reliable data support for the reconstruction of the model in the fault area.

[0075] Generally speaking, the present invention performs excellently in the overall indicators (MSE, SNR, SSIM), especially in the reconstruction accuracy and structural consistency in the fault area, showing obvious advantages compared with the inverse distance weighting method, UNet and MultiResUNet, proving that the present invention has higher accuracy and better noise suppression ability in reconstructing the overall formation and fault structure.

[0076] 2. Real data

[0077] The real data experiment was carried out under the PyTorch framework. The optimizer was selected as Adam, the initial learning rate was set to 0.001, and the data was flipped 180 degrees for 30 rounds of pre-training, and the total number of training rounds was 3000 rounds. Since the preprocessing part in the present invention involves the interpolation of the fault hanging wall and footwall lines, which will change the input data, for the sake of fairness, in the following experiments, the surveyed line data after interpolating the fault hanging wall and footwall lines were used as the input, and the Dal-UNet model proposed by the present invention was used.

[0078] The final reconstructed formation surface result is as Figure 9 shown in (b), where the strike of the fault can be clearly identified, and the effect of fault constraint is better. Next, an analysis and comparison will be made on the reconstructed formation surface results of the fault surface reconstruction loss function and the traditional reconstruction loss function (lacking constraint loss and total variation loss) proposed in the present invention. Figure 10 Shown is the 3D visualization of the two reconstructed formation surface results.

[0079] In terms of visual effect. It can be seen that there are many noise points in the reconstructed formation surface result of the traditional reconstruction loss function, especially at the image edges and around the faults. Because the neural network has difficulty dealing with the spike phenomena caused by the large numerical jumps on both sides of the fault line, which is particularly prominent. On the contrary, in the reconstructed formation surface result of the fault surface reconstruction loss function, the data changes are smoother without affecting the terrain undulation, achieving a good denoising and smoothing effect, and the fault surface is more prominent.

[0080] In terms of structural reconstruction and fault constraint. According to Figure 10 the comparison of the two reconstructed formation surface effects, the result of the fault surface reconstruction loss function pays more attention to the surface reconstruction of the fault surface. The stepped faults have achieved good reconstruction effects, which are shown as cliff-like landforms in the 3D visualization, and the formation reconstruction of other terrains (such as mountains and hills) besides the faults is not ignored. Figure 10 (b), Figure 10 (d), it can also be seen that some translational faults have caused geological collapses to form rift valley landforms. In the result diagram of the traditional reconstruction loss function, due to the lack of fault constraint, the positions where there should be faults do not show the characteristics that the faults should have because the reconstruction results are too smooth, and it is naturally difficult to form landforms such as cliffs and rift valleys.

[0081] In summary, the fault surface reconstruction loss function proposed in the present invention has good improvement effects on the reconstructed formation surface results in terms of visual effect, structural reconstruction, and fault constraint. Especially, it makes the reconstructed formation surface results follow the geographical characteristics of the faults, and then presents landforms such as cliffs and rift valleys caused by the faults in the 3D visualization. The present invention has successfully achieved fault constraint, proving the effectiveness of the fault surface reconstruction loss function.

[0082] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A Dal-UNet formation surface reconstruction method based on fault constraint, characterized in that Including: S1. Grid the survey line data; S2. Mark the dislocation points belonging to the same fault in the gridded survey line data, and then connect the dislocation points belonging to the same fault one by one to obtain the distribution positions of each fault; S3. Interpolate the hanging wall line and footwall line of the fault according to the distribution position of the fault to obtain sparse formation data; S4. Replace the traditional residual convolution block in MultiResUNet with a dilated convolutional residual module to obtain Dal-Unet; S5. Train Dal-Unet with the sparse formation data obtained in step S3; S6. Input the sparse formation data corresponding to the survey line data to be processed into the trained Dal-Unet to obtain the formation surface reconstruction result.

2. The Dal-UNet formation surface reconstruction method based on tomographic constraints according to claim 1, wherein In step S1, when there are two or more survey line data points that need to be mapped to the same grid point, the value of this grid point is calculated according to the inverse weighted average of the Euclidean distances from each survey line data point to this grid point.

3. A method for reconstructing a formation surface of Dal-UNet based on fault constraint according to claim 2, characterized in that, The calculation formula for the hanging wall line difference in step S2 is: Among them, V k represents the value of the interpolation point k, v i is the value of the i-th survey line point in the survey line point set above the fault, v j is the value of the j-th survey line point in the survey line point set below the fault, D k represents the value corresponding to the interpolation point k in the dislocation list D, d ik represents the Euclidean distance between the i-th survey line point and the interpolation point k, d jk represents the Euclidean distance between the j-th survey line point and the interpolation point k.

4. The Dal-UNet formation surface reconstruction method based on tomographic constraint according to claim 2, wherein The calculation formula for the footwall line difference in step S2 is:

5. A method for reconstructing a formation surface of Dal-UNet based on tomographic constraints according to claim 4, characterized in that, The loss function used in the training process of Dal-Unet is: Among them, λ C is the reconstruction loss weight, λ R is the constraint loss weight, λ TV is the total variation loss weight, f(x) represents the reconstruction result, m is a mask matrix where the positions of the valid points in the data to be repaired are 1 and the positions of the unobserved values are 0, V obs is the data to be repaired, ⊙ is the Hadamard product, N is the number of elements for effective calculation; f_m is a mask matrix where the values at the fault line positions are 1 and the values at other positions are 0; V fault is the fault line data, C is the number of channels, H is the height, W is the width, x i,j is the pixel at the i-th row and j-th column in the reconstruction result.

Images