A fault-constrained Dal-UNet stratigraphic surface reconstruction method

Abstract

This invention discloses a fault-constrained Dal-UNet stratigraphic surface reconstruction method, applicable to the fields of information processing and oil and gas exploration. Existing deep learning reconstruction algorithms, based on supervised learning, require extensive data and manually labeled training data for pre-training, resulting in high economic and time costs. Data quality directly impacts algorithm performance, and in geological studies with relatively scarce data, supervised deep learning algorithms are prone to overfitting, making them even more difficult to apply directly to stratigraphic surface reconstruction tasks involving complex geological structures with faults. This invention utilizes the Dal-UNet network to capture hidden features from limited and sparse stratigraphic surface data and combines it with a fault-constrained algorithm to reconstruct complex stratigraphic surfaces containing faults, effectively restoring the true topography.

Description

Technical Field

[0001] This invention belongs to the fields of information processing and oil and gas exploration, and specifically relates to a stratigraphic surface reconstruction technology. Background Technology

[0002] Stratigraphic surface reconstruction is a crucial step in geospatial visualization and a key issue of common concern in the fields of information processing and oil and gas exploration. It plays an important role in advancing subsequent work such as finding solid mineral deposits, analyzing the migration and accumulation of oil, natural gas and groundwater, evaluating the foundation stability of large-scale engineering projects, and monitoring, analyzing and predicting earthquakes. However, due to the sparsity and heterogeneity of data and the complexity of geological structures, the reliability and accuracy of traditional geological stratigraphic surface reconstruction are not high. Therefore, further research on stratigraphic surface reconstruction methods is of great significance and value.

[0003] Reconstructing stratigraphic surfaces containing faults typically requires complete information such as elevation, fault displacement, and dip angle. However, in most cases, only the planar projection information of the fault is available, which is detrimental to stratigraphic surface modeling and contour mapping with faults. Ji Zhanhuai et al. improved the fault trajectory method, which only requires knowledge of the fault's planar location information. It can effectively solve the stratigraphic interpolation problem under fault constraints by using the shortest path algorithm combined with the inverse distance weighted interpolation method.

[0004] In 2018, Ulyanov et al. discovered that convolutional neural networks (CNNs) can directly learn the distribution characteristics of image data from the damaged 2D image itself, achieving goals such as image restoration or super-resolution interpolation without the need for additional labeled data pre-training, thus saving time and money. Furthermore, they achieved excellent results in standard inverse problems (such as denoising and super-resolution) and can be flexibly applied to various interpolation and reconstruction environments.

[0005] U-Net is a classic convolutional neural network, named for its unique U-shaped structure. It consists of an encoder responsible for feature extraction and downsampling, and a decoder for upsampling and feature fusion. These two components are connected via skip connections to efficiently integrate multi-scale features, ultimately achieving accurate image segmentation. U-Net is widely used in image segmentation fields such as medicine, biology, and remote sensing, and is renowned for its efficient data utilization, high segmentation accuracy, and strong adaptability, greatly promoting the development of these fields.

[0006] Traditional stratigraphic surface reconstruction methods include interpolation, fitting, and partial differential equation-based methods. These methods are driven by theoretical knowledge and cannot effectively capture the implicit characteristics of geological data. With the continuous development of deep learning technology, many scholars have applied deep learning to stratigraphic surface reconstruction tasks, resulting in numerous research achievements. However, most deep learning reconstruction algorithms are based on supervised learning models, requiring a large amount of data and manually labeled training data for pre-training the network. This incurs high economic and time costs, and the quality of the data directly affects the algorithm's performance. Furthermore, in geological studies where data is relatively scarce, supervised deep learning algorithms are prone to overfitting, making them even more difficult to directly apply to stratigraphic surface reconstruction tasks involving complex geological structures such as faults. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a Dal-UNet-based stratigraphic surface reconstruction method based on fault constraints. This method utilizes the Dal-UNet network to capture hidden features from finite and sparse stratigraphic surface data, and combines it with a fault constraint algorithm to reconstruct complex stratigraphic surfaces containing faults, effectively restoring the true topography.

[0008] The technical solution adopted in this invention is: a Dal-UNet stratigraphic surface reconstruction method based on fault constraints, comprising:

[0009] S1. Grid the survey line data;

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

[0011] S3. Based on the distribution location of the faults, interpolate the hanging wall and footwall lines of the faults to obtain sparse stratigraphic data;

[0012] S4. Replace the traditional residual convolutional blocks in MultiResUNet with dilated convolutional residual modules to obtain Dal-Unet;

[0013] S5. Use the sparse stratigraphic data obtained in step S4 to train Dal-Unet;

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

[0015] The beneficial effects of this invention are as follows: Addressing the issue that directly reconstructing stratigraphic surfaces using survey line data may lead to discrepancies between the actual terrain and topographic orientation, this invention proposes a fault surface reconstruction loss function to achieve fault constraint. Combined with the Dal-UNet model, this significantly improves the visual effects, structural reconstruction, and fault constraint aspects of the reconstructed stratigraphic results, making them closer to real geographical features. This provides reliable geological data for subsequent geological research, including the search for solid mineral deposits, the analysis of oil, natural gas, and groundwater migration and accumulation, the evaluation of the stability of large-scale engineering foundations, and the monitoring, analysis, and prediction of earthquakes. Attached Figure Description

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

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

[0018] Wherein, (a) is a top view of the three-dimensional scatter plot, and (b) is a side view of the three-dimensional scatter plot;

[0019] Figure 3 A comparison map of fault distribution;

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

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

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

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

[0024] Figure 6 A schematic diagram illustrating the process of constructing the baseline data;

[0025] Among them, (a) is the original crater data, and (b) is the crater data containing normal faults (baseline data).

[0026] Figure 7 A schematic diagram illustrating the process of constructing sparse observation data;

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

[0028] Figure 8 For comparative experiments;

[0029] Among them, (a) is the baseline data, (b) is the reconstruction effect of the survey line data with interpolated fault upper and lower plate 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 For comparison of the original data and the interpolated results;

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

[0032] Figure 10 For reconstruction results Figure 3 3D visualization comparison;

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

[0034] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

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

[0036] The acquired survey data was standardized and mapped onto a 400*400 matrix grid at a ratio of 50m:1 pixel. Data points of the survey line Need to be mapped to the same grid point In this embodiment The value is greater than or equal to 2, based on the Euclidean distance from each survey line data point to that grid point. The inverse weighted average is shown in formula (1):

[0037] (1)

[0038] (2)

[0039] in: These are survey line data points The value; Grid points The value. Survey line data as follows: Figure 2 As shown.

[0040] like Figure 3 (a) As shown in the red box, abrupt changes in the survey data originate from the presence of faults. Faults arise from local displacement of strata, thus disrupting the continuity of the strata. If the influence of faults is ignored, data from the upper fault block may incorrectly interfere with the reconstruction of data from the lower fault block, and vice versa, ultimately leading to distortion of the reconstruction results near the fault plane. To impose constraints during surface reconstruction and ensure the integrity and accuracy of the fault plane, this invention proposes an algorithm for fault constraint:

[0041] First, certain artificial constraints are introduced regarding the faults. Based on the terrain trend in the 3D scatter plot and the location of fault points (hereinafter referred to as fault points where data in the 3D scatter plot changes abruptly), combined with manual interpretation, it is determined which fault points belong to the same fault. Then, fault points belonging to the same fault are connected one by one to obtain the distribution locations of various faults, such as... Figure 3 As shown by the black line in (b).

[0042] Based on the fault distribution, the upper and lower disk lines of the fault are interpolated (in this invention, the upper and lower disk lines of the fault are 2 pixels apart), and the interpolation method is as follows:

[0043] Suppose this invention requires interpolating the upper plate line (m pixels) and lower plate line (m pixels) between fault points a and b. First, the fault value Δa at fault point a and the fault value Δb at fault point b must be calculated. The fault value is the vertical displacement from the upper plate to the lower plate at that point. A fault list D of length m is formed by linear interpolation between Δa and Δb. The set of survey points on the upper side of the fault is denoted as P. u The lower side is denoted as P. d To minimize the impact of the fault, the hanging wall line was interpolated using formula (3), and the footwall line was interpolated using formula (4). The interpolation results for the hanging wall and footwall lines are as follows: Figure 4 As shown.

[0044] (3)

[0045] (4)

[0046] in It is the value of the interpolation point k. It is P u The value of point i in the middle, It is P d The value of point j, where d is the Euclidean distance.

[0047] To further utilize the interpolation data of the hanging wall and footwall lines of the fault to achieve the final fault constraint effect, this invention proposes a multi-task loss function—Fault Surface Reconstruction Loss. This loss function seeks a balance among three tasks: overall reconstruction of the stratigraphic surface, detailed reconstruction of the fault plane, and noise removal and smoothing of the reconstruction results, in order to achieve a good reconstruction effect of the stratigraphic surface containing the fault.

[0048] (5)

[0049] The loss function consists of three parts, which are weighted and summed, where Weight parameters for each part:

[0050] Item (Reconstruction Loss): Responsible for the overall surface reconstruction of sparse stratigraphic data. Among them... Representing the reconstruction result, m is a mask matrix, where it is set to 1 at points with values ​​in the data to be repaired, and to 0 at points with unobserved values ​​(i.e., where interpolation is needed). This is the data to be repaired (i.e., known data). It is the Hadamard product, where N is the number of validly computed elements (i.e., the number of non-zero elements in the mask m).

[0051] The constraint loss term compares the pixel-level MSE error of fault line locations between the reconstructed stratigraphic data and the original sparse stratigraphic data, thereby enabling the neural network to focus more on reconstructing fault line locations during surface reconstruction, achieving a fault constraint effect. (i.e., fault mask) is a mask matrix where the value is 1 at the fault line location and 0 at other locations. It is fault line data;

[0052] The term (Total Variational Loss): This loss function is highly effective in preserving edges and details when removing noise from images. In stratigraphic reconstruction tasks, the input to the neural network is random noise, which easily generates a large number of high-frequency errors. Furthermore, preserving fault edges during surface reconstruction to achieve fault constraints is also a challenge. This invention introduces the total variational loss as a regularization term, reducing these high-frequency errors through regularization of the high-frequency components. This is highly effective in removing high-frequency noise while preserving fault edge data and detailed textures. Here, C refers to the number of channels, H is the height, and W is the width. The pixel in the i-th row and j-th column of the reconstruction result.

[0053] The fault surface reconstruction loss function utilizes the known portions of sparse stratigraphic data to construct a mask matrix m, guiding the neural network through self-supervised learning. After being processed by the mask matrix, the reconstruction result is compared point-by-point only with the known portions of the sparse stratigraphic data, minimizing numerical errors. Simultaneously, based on the data distribution characteristics, i.e., implicit prior information, the neural network model automatically performs interpolation reconstruction on missing regions, improving the reasonableness and accuracy of the reconstruction results.

[0054] In its specific implementation, the fault surface reconstruction loss function proposed in this invention is used to calculate the error between the model reconstruction result and the real stratigraphic data. That is, during the training process of Dal-UNet, in each training epoch, the model uses sparse stratigraphic data as input to generate the reconstruction result of the stratigraphic surface. The loss function is calculated according to formula (5). This invention employs automatic differentiation based on the PyTorch framework for gradient backpropagation and calculates the loss function using the chain rule. The partial derivatives of the network parameters are used, and 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, thereby reducing the overall stratigraphic surface reconstruction error, enhancing the reconstruction accuracy of the fault region, optimizing the surface smoothness and structural consistency, and ultimately achieving the goal of accurately reconstructing the fault-containing stratigraphic surface.

[0055] However, traditional deep learning models often face problems such as insufficient information and difficulty in modeling spatial feature relationships when processing sparse data. Therefore, this invention proposes an improved Dal-UNet structure to enhance the model's feature extraction capability for sparse stratigraphic data and improve the accuracy of surface reconstruction.

[0056] Dal-UNet:

[0057] Figure 5 The neural network architecture of Dal-UNet is demonstrated, which is an improvement on MultiResUNet. This invention uses a dilated convolutional residual module (DalResBlock) to replace the traditional residual convolutional block, enabling the neural network to better grasp the feature relationships between sparse stratigraphic data, thereby improving the accuracy and consistency of surface reconstruction. Simultaneously, 3×3 convolutions with a stride of 2 are used for downsampling, allowing the network to learn more feature information during dimensionality reduction and reducing information loss. Bilinear interpolation is used for upsampling to ensure the accuracy of surface reconstruction.

[0058] like Figure 5As shown, the Dal-UNet of the present invention includes: the input of a first dilated convolutional residual module is sparse stratigraphic data; the output of the first dilated convolutional residual module is downsampled and used as the input of a second dilated convolutional residual module; the output of the second dilated convolutional residual module is downsampled and used as the input of a third dilated convolutional residual module; the output of the third dilated convolutional residual module is downsampled and used as the input of a fourth dilated convolutional residual module; the output of the fourth dilated convolutional residual module is downsampled and used as the input of a fifth dilated convolutional residual module; the output of the fifth dilated convolutional residual module is upsampled and used as the input of a sixth dilated convolutional residual module; the input of the sixth dilated convolutional residual module also includes the output of the fourth dilated convolutional residual module after path residual block processing. The output of the sixth dilated convolution residual module is upsampled and used as the input of the seventh dilated convolution residual module. The input of the seventh dilated convolution residual module also includes the result of the output of the third dilated convolution residual module processed by the path residual block. The output of the seventh dilated convolution residual module is upsampled and used as the input of the eighth dilated convolution residual module. The input of the eighth dilated convolution residual module also includes the result of the output of the second dilated convolution residual module processed by the path residual block. The output of the eighth dilated convolution residual module is upsampled and used as the input of the ninth dilated convolution residual module. The input of the ninth dilated convolution residual module also includes the result of the output of the first dilated convolution residual module processed by the path residual block. The output of the ninth dilated convolution residual module is the reconstruction result.

[0059] The Dal-UNet is trained, and the resulting stratigraphic surface reconstruction is obtained based on the trained Dal-UNet.

[0060] In practical applications, the survey data collected through the survey line, such as Figure 2 As shown, the Dal-UNet network model is trained in a self-supervised manner by interpolating the fault hanging wall and footwall lines of the survey line data collected by the survey line, and then training the model based on the interpolated real stratigraphic data.

[0061] The technical effects of the present invention will be explained below with reference to specific data:

[0062] 1. Simulated data

[0063] Because real stratigraphic data (survey line data) is relatively sparse, it is difficult to directly assess the accuracy of the reconstruction results. Therefore, to further verify the effectiveness of the fault-constrained Dal-UNet stratigraphic surface reconstruction method proposed in this invention, a set of simulated data experiments was designed. This experiment aims to construct synthetic data with known fault structures and compare traditional methods with this invention in terms of reconstruction accuracy, fault preservation, and surface smoothness, in order to quantify the advantages of this invention in fault-constrained reconstruction.

[0064] In constructing the experimental data, this simulation experiment used crater data as a basis and simulated a normal fault through local subsidence. This data served as the ground truth. Figure 6 As shown in (b). Subsequently, simulated survey line sampling was performed on the baseline data to simulate the sparse observation of real stratigraphic data. The survey line sampling matrix was generated as follows: a minimum spacing of 10 and a maximum spacing of 40 were set, and multiple oblique lines with unequal spacing were randomly generated within this range, such as... Figure 7 As shown in (a), the final sparse observation data is as follows: Figure 7 As shown in (b).

[0065] This experiment systematically compared the stratigraphic surface reconstruction performance of different methods under sparse observation data conditions, focusing on evaluating their ability to preserve fault features. Since the preprocessing part of this invention involves interpolation of the hanging wall and footwall lines of the fault, which alters the input data, for fairness, the following experiments all use the survey data after interpolating the hanging wall and footwall lines as input.

[0066] The comparison methods include: (c) inverse distance weighted interpolation; (d) UNet + traditional reconstruction loss function; (e) MultiResUNet + traditional reconstruction loss function; (f) the method proposed in this invention - Dal-UNet + fault surface reconstruction loss function.

[0067] This experiment was conducted using the PyTorch framework, with Adam selected as the optimizer and an initial learning rate of 0.001. Furthermore, to improve the model's generalization ability, 30 rounds of pre-training were performed using data rotated 180°, with a final training run of 3000 rounds. Model performance was evaluated using a combination of qualitative and quantitative analysis, with metrics including mean squared error (MSE), signal-to-noise ratio (SNR), and structural similarity (SSIM) to comprehensively analyze the performance of different methods in terms of fault feature preservation, surface smoothness, and overall stratigraphic consistency.

[0068] Qualitative analysis:

[0069] Figure 8The reconstruction results of different methods are shown. (c) Inverse distance weighted interpolation: Obvious cuts appear during the reconstruction process, making it difficult to reconstruct the original stratigraphic structure under sparse data conditions and failing to accurately characterize the spatial distribution of faults; (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 still exist along the survey line direction in the results, accompanied by a large amount of noise; (e) MultiResUNet + traditional reconstruction loss function: The reconstruction results show obvious high-value anomalies in the edge region, that is, the error is concentrated at the edge, with a dense distribution of anomalous high values, which deviates significantly from the geological structure in the baseline data (a). The possible reasons are: the receptive field of MultiResUNet is insufficient and the convolutional layers are too deep, coupled with the lack of effective noise suppression in the traditional reconstruction loss function, which leads to uncontrolled diffusion of boundary values, resulting in error accumulation; (f) This invention: It effectively maintains the stratigraphic structure during the reconstruction process and shows obvious step features in the fault region, with high consistency with the baseline data (a). Meanwhile, the traces of the survey lines are significantly reduced, some texture details are preserved, and the overall reconstruction process is stable with less noise. This demonstrates that the present invention has greater advantages in characterizing fault spatial morphology, controlling boundary errors, and improving training stability.

[0070] Quantitative analysis:

[0071] Table 1 lists the performance of three different reconstruction methods on three metrics: mean squared error (MSE), signal-to-noise ratio (SNR), and structural similarity (SSIM).

[0072] MSE reflects the mean squared difference between the reconstructed data and the real data; a smaller value indicates higher reconstruction accuracy. Overall, the MSE of the inverse distance weighted method is 6.85, while UNet's MSE is 4.35, indicating that both have some reconstruction error. In contrast, MultiResUNet's overall MSE is as high as 19.87, indicating a significant deviation in overall reconstruction. The overall MSE of this invention is only 1.09, significantly reducing reconstruction error and demonstrating higher reconstruction accuracy. Regarding fault regions, the MSE of the inverse distance weighted method is 6.67, UNet and MultiResUNet are 1.86 and 1.98 respectively, while the MSE of this invention in fault regions is 0.86, proving that this invention is more accurate in reconstructing fault structures.

[0073] Table 1. Comparison of reconstruction performance of different methods

[0074] index Inverse distance weighting UNet MultiResUNet 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 in the fault region (↓) 6.67 1.86 1.98 0.86 SNR (increased) in the fault region 13.42 18.97 18.70 22.33 SSIM (increased) in the fault region 0.627 0.955 0.968 0.976

[0075] Signal strength (SNR) measures signal quality; a higher SNR indicates less noise in the reconstructed result and a closer approximation of the true value. Overall, the inverse distance weighting method has an SNR of 11.64, UNet has 13.62, while MultiResUNet's SNR is only 7.01, indicating higher noise levels in its reconstruction. In contrast, the present invention achieves an SNR of 19.62, significantly improving signal quality. In the tomographic region, the inverse distance weighting method has an SNR of 13.42, UNet and MultiResUNet have SNRs of 18.97 and 18.70 respectively, while the present invention achieves an SNR of 22.33 in the tomographic region, further demonstrating its superior noise suppression capabilities.

[0076] SSIM is used to measure the structural similarity between the reconstructed results and the actual data; the closer the value is to 1, the more accurate the structural reconstruction. Overall, the SSIM of the inverse distance weighted method is 0.668, UNet is 0.647, and MultiResUNet is even lower at only 0.516, indicating that these three methods are insufficient in maintaining the consistency of the overall stratigraphic structure. In contrast, the overall SSIM of this invention is as high as 0.907, demonstrating its excellent performance in restoring overall structural features. In fault regions, the SSIM of the inverse distance weighted method is 0.627, UNet and MultiResUNet are 0.955 and 0.968 respectively, while the SSIM of this invention in fault regions is 0.976, further demonstrating the significant advantage of this invention in accurately reconstructing fault structures. It is worth noting that the SSIM of UNet and MultiResUNet in fault regions is significantly improved compared to the overall SSIM. This is mainly due to the interpolation of the hanging wall and footwall lines of the fault in the data preprocessing of this invention, which provides reliable data support for the model's reconstruction in fault regions.

[0077] In summary, this invention performs well in all overall metrics (MSE, SNR, SSIM), especially in terms of reconstruction accuracy and structural consistency in fault regions. It has significant advantages over inverse distance weighting, UNet, and MultiResUNet, demonstrating that this invention has higher accuracy and better noise suppression capabilities when reconstructing the overall strata and fault structure.

[0078] 2. Real data

[0079] Real-world data experiments were conducted using the PyTorch framework. The optimizer was Adam, with an initial learning rate of 0.001. The data was flipped 180 degrees for 30 rounds of pre-training, resulting in a total of 3000 training rounds. Since the preprocessing in this invention involves interpolation of the fault's upper and lower plate lines, which alters the input data, for fairness, the following experiments all use the interpolated fault upper and lower plate line survey data as input, and employ the proposed Dal-UNet model.

[0080] The final stratigraphic surface reconstruction results are as follows: Figure 9 As shown in (b), the fault strike can be clearly identified, and the fault constraint effect is good. The following will analyze and compare the stratigraphic surface reconstruction results of the fault surface reconstruction loss function proposed in this invention and the traditional reconstruction loss function (lacking constraint loss and total variational loss). Figure 10 The presentation showcases three-dimensional visualizations of two types of stratigraphic surface reconstruction results.

[0081] In terms of visual effects, the traditional reconstruction loss function's stratigraphic surface reconstruction results show a lot of noise, especially at image edges and around faults. This is particularly noticeable because neural networks struggle to handle the sharp spikes caused by large numerical jumps on both sides of the fault line. In contrast, the stratigraphic surface reconstruction result using the fault surface reconstruction loss function smooths out data variations without affecting topographic relief, achieving excellent noise reduction and smoothing effects, and making the fault planes more prominent.

[0082] Regarding structural reconstruction and fault constraint. According to Figure 10 The comparison of the two types of stratigraphic surface reconstruction results shows that the result of the fault surface reconstruction loss function pays more attention to the surface reconstruction of the fault plane. The stepped faults are reconstructed very well and are represented as steep cliffs in 3D visualization. The stratigraphic reconstruction of other terrains (such as mountains and hills) besides faults is not ignored. Figure 10 (b) Figure 10 As can be seen in (d), some strike-slip faults have caused geological collapses, forming rift valley landforms. In the results of traditional reconstruction loss function, due to the lack of fault constraints, the locations where faults should exist do not show the characteristics of faults because the reconstruction results are too smooth, and naturally it is difficult to form landforms such as steep cliffs and rift valleys.

[0083] In summary, the fault surface reconstruction loss function proposed in this invention significantly improves the visual effects, structural reconstruction, and fault constraint aspects of the reconstructed geological structure. In particular, it ensures that the reconstructed geological structure conforms to the geographical characteristics of the fault, thereby presenting landforms such as steep cliffs and rift valleys caused by the fault in 3D visualization. This invention successfully achieves fault constraint, demonstrating the effectiveness of the fault surface reconstruction loss function.

[0084] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A Dal-UNet stratigraphic surface reconstruction method based on fault constraints, characterized in that, include: S1. Grid the survey line data; S2. Mark the fault points belonging to the same fault in the gridded survey data, and then connect the fault points belonging to the same fault one by one to obtain the distribution location of each fault. S3. Based on the distribution location of the faults, interpolate the hanging wall and footwall lines of the faults to obtain sparse stratigraphic data; S4. Replace the traditional residual convolutional blocks in MultiResUNet with dilated convolutional residual modules to obtain Dal-Unet; S5. Dal-Unet is trained using the sparse stratigraphic data obtained in step S3; the loss function used in the Dal-Unet training process is: ； in, To reconstruct the loss weights, To constrain the loss weights, The total variation loss weights, Representing the reconstruction result, m is a mask matrix in the data to be repaired, where the positions of the observed values ​​are 1 and the positions of the unobserved values ​​are 0. This is data to be repaired. It is the Hadamard product, where N is the number of elements that can be effectively computed; It is a mask matrix where the value is 1 at the fault line location and 0 at other locations; This is fault line data; C represents the number of channels, H represents the height, and W represents the width. The pixel in the i-th row and j-th column of the reconstruction result; S6. Input the sparse stratigraphic data corresponding to the survey line data to be processed into the trained Dal-Unet to obtain the stratigraphic surface reconstruction results.

2. The method for reconstructing Dal-UNet stratigraphic surfaces based on fault constraints according to claim 1, characterized in that, 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 the grid point is calculated by inverse weighted averaging of the Euclidean distance from each survey line data point to the grid point.

3. The method for reconstructing Dal-UNet stratigraphic surfaces based on fault constraints according to claim 2, characterized in that, The interpolation formula for the upper plate line in step S2 is: ； in, This represents the value of the interpolation point k. It is the value of the i-th survey point in the set of survey points on the upper side of the fault. It is the value of the j-th survey point in the set of survey points on the lower side of the fault. This represents the value of the corresponding interpolation point k in the error list D. This represents the Euclidean distance between the i-th survey point and the point k to be interpolated. This represents the Euclidean distance between the j-th survey point and the k-th interpolation point.

4. The method for reconstructing Dal-UNet stratigraphic surfaces based on fault constraints according to claim 3, characterized in that, The interpolation formula for the lower plate line in step S2 is: 。

Images