Deep learning-based fault intelligent identification and three-dimensional geologic model restoration method

Abstract

The invention discloses a fault intelligent identification and three-dimensional geological model restoration method based on deep learning, and belongs to the field of geological exploration and artificial intelligence. According to the fault intelligent recognition method, multi-source geological data are collected, a sample library is constructed, a deep neural network fused with an attention mechanism is designed, a fault probability distribution diagram is output through multi-scale feature extraction and attention weight distribution, the problem of sample imbalance is solved by adopting a mixed loss function, and finally a fault recognition result is output. Based on a fault identification result, the three-dimensional geological model restoration method comprises the steps of constructing a fault geometric model, determining a restoration area, performing local reconstruction on the model by adopting a self-adaptive algorithm constrained by a geological law, and evaluating a restoration effect through cross validation. According to the method, through deep fusion of deep learning and geological modeling, the complex fault recognition precision and the model repair efficiency are remarkably improved, and reliable technical support is provided for oil-gas exploration and mineral resource development.

Description

Technical Field

[0001] This invention relates to the fields of geological exploration and artificial intelligence, specifically to a method for intelligent fault identification and three-dimensional geological model repair based on deep learning. Background Technology

[0002] Fault identification is a crucial technical step in geological exploration, significantly impacting oil and gas migration, mineral enrichment, and early warning of geological hazards. Traditional methods primarily rely on manual interpretation of seismic profiles by geological experts, using features such as interruptions and fault shifts in reflected wave groups for annotation. However, this approach has significant limitations. First, manual interpretation is inefficient, requiring several hours to analyze a single profile, and is influenced by subjective experience, with different experts showing discrepancies of 20%-30%, leading to insufficient consistency. Second, the scale and format differences of multi-source data (such as seismic data, well logging curves, and outcrop images) make effective fusion difficult using traditional methods, weakening the reliability of multi-evidence constraints. Furthermore, for complex structures such as low-order faults and concealed faults, traditional techniques like coherence analysis and curvature attributes have an accuracy rate of less than 50%, easily resulting in missed or false identifications, thus limiting the accuracy of deep resource exploration.

[0003] In the area of ​​3D geological model restoration, existing technologies are limited by the accuracy of fault identification, exhibiting significant shortcomings. The restoration process heavily relies on manual adjustment of the stratigraphic interfaces on both sides of the fault, lacking quantitative constraints and causing significant deviations between the model and the actual geological structure. Furthermore, simple mathematical interpolation methods (such as linear interpolation) ignore geological constraints such as the continuity of stratigraphic deposition and the distribution patterns of lithology, often resulting in restoration results that do not conform to geological logic, impacting the reliability of engineering decisions. Efficiency is also a prominent issue; the restoration of large 3D models (e.g., those covering a 100km × 100km × 5km area) requires several days to weeks, failing to meet the rapid iterative demands of exploration and development.

[0004] In recent years, machine learning techniques such as convolutional neural networks (CNNs) have been introduced into tomography, but there are still problems such as the limited ability of models to capture tomographic features, the lack of multi-source data fusion mechanisms, and the failure to form a closed-loop system from identification to repair.

[0005] Therefore, traditional methods for fault identification and 3D geological model restoration suffer from technical bottlenecks such as high reliance on manual labor, low efficiency, difficulty in fusion of multi-source data, and insufficient accuracy. These shortcomings highlight the urgency of developing an integrated intelligent solution, providing room for technological improvement in this invention. Summary of the Invention

[0006] The purpose of this invention is to provide a method for intelligent fault identification and three-dimensional geological model repair based on deep learning, which solves the above-mentioned problems of the prior art through the deep integration of deep learning and geological modeling.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based intelligent tomography identification method includes the following steps: S11: Collect multi-source geological data, perform standardized preprocessing on the data, and construct a sample database with fault labels. The multi-source geological data includes seismic profile data, drilling logging data, and geological outcrop images. S12: Design a deep neural network model that integrates attention mechanism, using seismic profile data as the main input, and outputs a fault probability distribution map through multi-scale feature extraction and attention weight allocation; S13: Use the sample database to train and optimize the deep neural network model, and use a hybrid loss function to handle the class imbalance problem; S14: The multi-source geological data collected from the target work area is preprocessed and used as the data to be identified as input into the trained model. The output includes fault boundary boxes, confidence scores and geometric parameters.

[0008] Furthermore, in S11: The seismic profile data is a three-dimensional seismic post-stack migration profile, which includes P-wave reflection amplitude and instantaneous frequency attributes; The drilling logging data, including sonic transit time and resistivity curves, is used to calibrate formation interfaces and fault locations. The geological outcrop images, specifically aerial photographs of fault outcrops taken by drones, are used to extract the surface morphological features of the faults.

[0009] Furthermore, the deep neural network model described in S12 employs an improved U-Net++ architecture, including: The encoder consists of 5 convolutional blocks, each containing 2 convolutional layers, a BatchNorm layer, and a ReLU activation function, achieving feature dimensionality reduction through max pooling. The attention gating module, located on the skip connection path between the encoder and decoder, enhances the representation of fault boundary features by calculating the spatial attention weights of the feature maps. The decoder gradually restores spatial resolution by upsampling and fusing the corresponding scale feature maps from the encoder. Auxiliary branches are added to the drilling and logging data input branches, and logging curve features are integrated into the main network through fully connected layers.

[0010] Furthermore, the attention gating module adopts the self-attention mechanism of Transformer to capture the continuity features of fault direction by calculating the long-distance dependencies of feature maps.

[0011] Furthermore, the specific expression for the hybrid loss function described in S13 is as follows: Loss = 0.5×CrossEntropyLoss + 0.5×DiceLoss; CrossEntropyLoss is used to optimize class determination, while DiceLoss is used to improve boundary recognition accuracy.

[0012] Furthermore, the training and optimization of the deep neural network model described in S13 specifically involves: pre-training on a public dataset based on a cosine annealing strategy and transfer learning method, and introducing an early stopping mechanism to prevent overfitting.

[0013] Another objective of this invention is to provide a method for repairing a three-dimensional geological model based on fault identification results, comprising the following steps: S21: Based on the fault identification results, construct a fault geometric model, wherein the fault identification results are implemented based on a deep learning-based intelligent fault identification method according to any one of claims 1-6; S22: Spatial matching of the fault geometric model with the original three-dimensional geological model to determine the fault-affected area; S23: Adaptive algorithms constrained by geological laws are used to repair the model in the fault-affected area; S24: Evaluate the repair effect through cross-validation. If the requirements are not met, return to S22 to adjust the parameters and repair again.

[0014] Furthermore, as described in S21, the fault geometry model is constructed by fitting the fault planar morphology with B-spline curves and combining Kriging interpolation to correct the vertical morphology of the fault surface.

[0015] Furthermore, the adaptive repair algorithm described in S23 specifically includes: For stratigraphic interface repair, an improved Poisson equation interpolation method is adopted. The unaffected stratigraphic interfaces on both sides of the fault are used as boundary conditions, and stratigraphic thickness constraints are introduced. By solving the Poisson equation, stratigraphic interface morphology that conforms to geological laws is generated. Lithological property restoration was performed using phase-controlled kriging interpolation based on lithological distribution data on both sides of the fault. Global model optimization involves applying Gaussian filtering to the boundary between the repaired region and the original model to eliminate abrupt interface changes.

[0016] Furthermore, the assessment of the repair effect described in S24 is conducted through a combination of quantitative and qualitative assessments, specifically as follows: The quantitative assessment uses cross-validation to compare the consistency between the model and the actual drilling data before and after the repair. It compares the calculated root mean square error with a preset threshold, including spatial error and attribute consistency. The qualitative assessment involved blind evaluation by senior geological experts, using visual inspection and evaluation to assess the relationship between faults and strata, the distribution of strata thickness, and special geological phenomena.

[0017] The deep learning-based intelligent fault identification and 3D geological model repair method provided by this invention achieves breakthrough progress in fault identification and 3D model repair through the deep integration of deep learning and geological modeling. Compared with existing technologies, it has the following significant advantages: The U-Net++ network architecture, based on a fusion attention mechanism and combined with the long-range dependency capture capability of Transformer, improves the accuracy of complex tomography identification from less than 50% in traditional methods to over 85%. The attention mechanism effectively focuses on tomographic boundary features, solving the problem of missed detection in low signal-to-noise ratio data, and achieving a dual improvement in accuracy and efficiency. By fusing well logging data with auxiliary branches and outcrop image features, a collaborative identification mechanism for multi-source data from seismic, well logging, and outcrop sources was established. This mechanism overcomes the ambiguity problem inherent in traditional single-source data sources, increasing the consistency rate between the identification results and actual drilling data to 89%. The collaborative processing of multi-source data also solves the difficulty of multi-source data fusion. The restoration effect has a high degree of geological consistency. An adaptive restoration algorithm constrained by geological laws is adopted, which incorporates quantitative indicators such as the coefficient of variation of stratum thickness (≤0.2) and spatial correlation of lithology (≥0.8) into the Poisson equation interpolation, reducing the depth error between the restoration model and the actual geological structure to 4.2m and improving the geological consistency by 40%. Breakthroughs in engineering application efficiency have overcome the problem of low efficiency due to reliance on manual labor. The fully automated technology has reduced the time for single profile identification from 2 hours to 1 minute and the model repair time from 5 days to 3 hours, improving efficiency by 40-120 times and providing feasible technical support for large-scale exploration projects. Attached Figure Description

[0018] Figure 1 This is a technical roadmap of the present invention; Figure 2 This is a diagram of the super-resolution reconstruction model under the teacher-student network architecture of the present invention; Figure 3 This is a diagram of the deep neural network architecture of the present invention; Figure 4 This is a diagram of the deep neural network architecture for the input-output of three-dimensional seismic data in this invention; Figure 5 This is a diagram of the deep neural network architecture for the super-resolution prediction results output of the present invention. Figure 6This is a comparative diagram of the fault identification results of the present invention, wherein (a) is the original three-dimensional seismic profile data, (b) is the manually interpreted fault result, (c) is the fault probability distribution diagram output by the model, and (d) is the final fault identification result. Figure 7 This is a schematic diagram comparing the normal vectors of the three-dimensional geological model before and after the normal vector repair of the present invention, wherein (a) is the model before normal vector repair and (b) is the model after normal vector repair; Figure 8 The diagram shows the visualization results of the attention mechanism of the present invention, wherein (a) is a visualization diagram of the weight of the inline direction attention mechanism, (b) is a visualization diagram of the weight of the crossline direction attention mechanism, (c) is a visualization diagram of the attention of fault strike continuity, and (d) is a visualization diagram of the attention of fault boundary enhancement. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Traditional fault identification methods primarily rely on manual interpretation by geological experts, which suffers from drawbacks such as low efficiency, high subjectivity, and difficulty in fusing multi-source data. Especially in complex tectonic regions, the accuracy rate of traditional methods is often less than 50%. Therefore, this invention proposes a complete intelligent solution, such as... Figure 1 The diagram illustrates the overall technical approach of this invention, showcasing the entire process from multi-source data acquisition, preprocessing, intelligent fault identification, 3D model repair to effect verification.

[0021] This embodiment provides a deep learning-based intelligent fault identification method, comprising: collecting multi-source geological data; performing standardized preprocessing on the data; constructing a sample database containing fault labels; the multi-source geological data including seismic profile data, well logging data, and geological outcrop images; designing a deep neural network model incorporating an attention mechanism, using seismic profile data as the main input, and outputting a fault probability distribution map through multi-scale feature extraction and attention weight allocation; training and optimizing the deep neural network model using the sample database, and employing a hybrid loss function to handle the class imbalance problem; inputting the preprocessed multi-source geological data collected from the target work area into the trained model as the data to be identified, and outputting fault identification results including fault bounding boxes, confidence scores, and geometric parameters. The specific process of this method is described in detail below.

[0022] S11: Collect multi-source geological data, perform standardized preprocessing on the data, and construct a sample database with fault labels. The multi-source geological data includes seismic profile data, drilling logging data, and geological outcrop images.

[0023] In this embodiment, the multi-source geological data used to construct the training sample database comes from multiple historical work areas that have been drilled and proven, specifically including: seismic profile data, drilling logging data, and geological outcrop images, wherein: The seismic profile data is a three-dimensional seismic post-stack migration profile with a resolution of 25m×25m×5m, which includes P-wave reflection amplitude and instantaneous frequency attributes. Drilling logging data, including sonic transit time and resistivity curves, are used to determine formation interfaces and fault locations. Geological outcrop images, aerial photographs of fault outcrops taken by drones, with a resolution of 0.1m, used to extract fault surface morphological features.

[0024] The data preprocessing stage begins with data standardization, normalizing seismic attributes to the [0,1] interval, performing depth correction and smoothing on well logging curves, and correcting distortion in outcrop images. Then, seismic profiles are cropped into 512×512 pixel sub-blocks, and a sample database is constructed using manually labeled faults. To alleviate class imbalance caused by the low pixel ratio in fault areas, data augmentation methods such as rotation, scaling, and noise addition are employed. Through these processes, a standardized sample database containing 300,000 positive samples and 700,000 negative samples is finally constructed, serving as the training sample database for deep neural network models.

[0025] S12: Design a deep neural network model that integrates an attention mechanism, using seismic profile data as the main input, and outputs a fault probability distribution map through multi-scale feature extraction and attention weight allocation.

[0026] like Figures 2-5 As shown, the deep neural network model adopts an improved U-Net++ architecture, integrating an attention mechanism and a multi-scale feature fusion module, including: The encoder consists of 5 convolutional blocks, each containing two 3×3 convolutional layers (stride 1), a BatchNorm layer, and a ReLU activation function. It achieves feature dimensionality reduction through max pooling. The attention gating module, located on the skip connection path between the encoder and decoder, suppresses the feature response of non-fault regions and enhances the feature representation of fault boundaries by calculating the spatial attention weights of the feature map. The attention gating module adopts the self-attention mechanism of Transformer and captures the continuity features of fault strike by calculating the long-distance dependencies of the feature map. The decoder gradually restores the spatial resolution by upsampling and fusing the corresponding scale feature maps of the encoder, and finally outputs a tomographic probability map with the same size as the input. Auxiliary branches are added to the drilling and logging data input branches. Through the fully connected layer, logging curve features are integrated into the main network to achieve collaborative identification of multi-source data. S13: The deep neural network model is trained and optimized using the sample database, and a hybrid loss function is used to handle the class imbalance problem.

[0027] The model training and optimization strategies include using cosine annealing with an initial learning rate of 0.001, pre-training on a public dataset (such as the SEG Fault Dataset) using transfer learning, fine-tuning with target region data, and introducing an early stopping mechanism to prevent overfitting. These measures improved the model's convergence speed by 2 times, significantly increasing training efficiency. The early stopping mechanism terminates training when the validation set loss does not decrease for 10 consecutive rounds, thus avoiding overfitting. Transfer learning enhances the model's adaptability to small sample data.

[0028] A deep neural network model is trained and optimized using a sample database. A hybrid loss function is employed, combining the advantages of cross-entropy loss and Dice loss to effectively address the class imbalance problem. The specific expression of the hybrid loss function is as follows: Loss = 0.5×CrossEntropyLoss + 0.5×DiceLoss; Among them, CrossEntropyLoss is the cross-entropy loss used to optimize class judgment, and DiceLoss is the Dice loss used to improve boundary recognition accuracy, effectively solving the class imbalance problem when the pixel ratio in the fault region is low (usually <5%).

[0029] S14: The multi-source geological data collected from the target work area is preprocessed and used as the data to be identified as input into the trained model. The output includes fault boundary boxes, confidence scores and geometric parameters.

[0030] After the model is trained, it can be applied to a new target area to be studied. The target area is a newly explored area that has not been drilled or has very few wells. 3D seismic data of this area is collected. At the same time, well logging data of any sporadic wells that may exist in the area are collected (if any). For this newly acquired target area data, the same preprocessing procedure as in S11 is used (such as normalization to the [0,1] interval and cropping to 512x512 pixel sub-blocks) to form the data to be identified.

[0031] The preprocessed data to be identified is input into the final model trained by S3 to generate a fault probability map (probability value 0-1). Fault candidate regions are generated by setting an appropriate confidence threshold (e.g., 0.7), and then a non-maximum suppression algorithm is used to remove overlapping candidate regions, retaining the optimal fault bounding box. Based on the optimized fault bounding box, geometric parameters such as strike, dip, and elevation difference are extracted. The final output of the fault identification result includes the fault bounding box, confidence score, and geometric parameters.

[0032] Figure 6 This diagram illustrates the comparison of fault identification results according to the present invention. (a) shows the original 3D seismic profile data, (b) shows the manually interpreted fault results, (c) shows the model-output fault probability distribution, and (d) shows the final fault identification results. The identification results of the method described in this invention are highly consistent with the manually interpreted results. Based on the identified fault boundary boxes, geometric parameters such as strike, dip, and displacement are further extracted. The strike parameter is obtained by detecting the straight line direction using Hough transform, the dip angle is calculated using the seismic profile depth and horizontal distance, and the displacement is converted from the amplitude of reflected wave group displacement. These parameters provide accurate geometric constraints for subsequent model repair.

[0033] Example 1: The implementation process of the deep learning-based intelligent fault identification method of the present invention is illustrated by taking the intelligent identification of faults in an oil and gas field as an example.

[0034] 1. Data Preparation and Processing In a practical application at an oil and gas field, 3D seismic data (dominant frequency 30Hz) covering an area of ​​50km × 30km was acquired. Simultaneously, well logging curves from 100 wells and images of 20 fault outcrops were collected. After standardized preprocessing, a database containing 500,000 samples was constructed. Seismic profiles were uniformly cropped into 512×512 pixel sub-blocks, including 150,000 positive samples (containing faults) and 350,000 negative samples (without faults). All samples were precisely manually labeled, including fault bounding boxes and geometric parameter labels, serving as ground truth references for model training.

[0035] 2. Model Training and Performance The model was trained on an NVIDIA A100 GPU (80GB VRAM) for approximately 48 hours. The model performed exceptionally well on the test set, achieving an accuracy of 87.6%, a recall of 82.3%, and an F1 score of 84.9%, significantly outperforming traditional CNN models (F1 score 68.5%). In practical applications, it successfully identified 12 major faults and 36 low-order faults, achieving an 89% agreement rate with newly revealed well findings.

[0036] 3. Parameter extraction verification results Five faults were randomly selected for parameter accuracy verification, and the specific comparison results are shown in Table 1. The average error analysis shows that the strike identification error is 2.1°, the dip angle error is 2.5°, and the elevation difference error is 4.8m, which fully meets the accuracy requirements for engineering applications.

[0037] Table 1 Comparison of Fault Identification Parameters

[0038] This embodiment also provides a method for repairing a three-dimensional geological model based on fault identification results. This method is implemented based on the output of a deep learning-based intelligent fault identification method, i.e., the fault identification results. A method for repairing a three-dimensional geological model based on fault identification results includes the following steps: constructing a fault geometric model based on the fault identification results, wherein the fault identification results are implemented using the deep learning-based intelligent fault identification method; spatially matching the fault geometric model with the original three-dimensional geological model to determine the fault-affected area; using an adaptive algorithm constrained by geological laws to repair the model of the fault-affected area; evaluating the repair effect through cross-validation, and returning to adjust parameters and repair again if the requirements are not met. The specific process of this invention is described below.

[0039] S21: Based on the fault identification results, a fault geometric model is constructed. The fault identification results are implemented using a deep learning-based intelligent fault identification method. The fault geometric model is constructed by fitting the fault planar morphology with B-spline curves and combining Kriging interpolation to correct the vertical morphology of the fault surface.

[0040] Based on the fault parameters (strike, dip, and displacement) obtained from the identification results, a systematic approach was used to construct an accurate fault geometric model. First, B-spline curves were used to fit the fault planar morphology, accurately describing the fault's distribution characteristics in a two-dimensional plane through parametric modeling. Then, a three-dimensional stretching operation was performed to generate an initial fault surface model, establishing a mathematical expression of the fault's spatial morphology. To improve model accuracy, multi-source data was introduced for collaborative constraints. Combining the fault depth points calibrated in drilling data, a Kriging interpolation algorithm was used to correct the vertical morphology of the fault surface. This process significantly improved the model's vertical accuracy, keeping the average error between the surface and the actual drilled points within 3 meters, effectively overcoming the depth uncertainty problem caused by relying solely on seismic data. In the model optimization stage, the fault surface was smoothed to eliminate local distortions caused by data noise, ensuring sufficient continuity and smoothness of the surface. This processing not only conforms to geological laws but also lays a good geometric foundation for subsequent model repair work.

[0041] S22: Spatial matching of the fault geometric model with the original three-dimensional geological model to determine the fault-affected area.

[0042] After the fault geometry model is constructed, the affected area of ​​the fault to be repaired is determined. Using spatial coordinate matching technology, the constructed fault geometry model is precisely overlaid with the original 3D geological model. Based on the WGS84 coordinate system, Boolean operations are used to accurately identify the stratigraphic units cut by the fault, establishing a spatial relationship model between the fault and the strata. The repair affected area is dynamically defined according to the fault scale characteristics, extending a certain distance (e.g., 50-200 meters, dynamically adjusted according to the fault scale) outwards from the fault surface to form the 3D target area to be repaired. The determination of this area comprehensively considers the impact range of the fault and the actual needs of model repair, ensuring complete coverage of the affected area while avoiding unnecessary waste of computational resources.

[0043] After location is established, the system extracts key model data such as original stratigraphic interfaces and lithological properties within the affected area, serving as initial conditions for subsequent repair operations. This step ensures that the repair process fully utilizes existing information from the original model, maintaining the consistency of geological features. This process achieves a seamless transition from fault identification to model repair area location, providing a precise spatial framework and data foundation for subsequent adaptive repair algorithms. The entire location process fully embodies the organic combination of geological laws and mathematical methods, ensuring the scientific rigor and reliability of the 3D geological model repair work.

[0044] S23: Adaptive algorithm based on geological constraints for model repair of fault-affected areas. An adaptive repair algorithm based on geological constraints is used to accurately reconstruct the fault-affected areas. This algorithm achieves local optimization of the 3D geological model by using physical constraint interpolation methods while maintaining stratigraphic continuity and lithological consistency. Figure 7 This is a schematic diagram comparing the normal vectors of the three-dimensional geological model before and after restoration according to the present invention. (a) shows the model before restoration, and (b) shows the model after restoration. The model before restoration exhibits a "suspended" stratum phenomenon, while the model after restoration maintains stratum continuity through interpolation filling. The adaptive restoration algorithm specifically includes: For stratigraphic interface restoration, a modified Poisson equation interpolation method is employed. The unaffected stratigraphic interfaces on both sides of the fault are used as boundary conditions, and stratigraphic thickness constraints are introduced (requiring the coefficient of variation of the same stratigraphic thickness to be ≤0.2). Solving the Poisson equation generates stratigraphic interface morphologies that conform to geological laws. For special geological contact relationships such as unconformities and pinch-outs, corresponding abrupt change in attitude constraints are set to ensure the geological rationality of the restoration results.

[0045] For lithological property restoration, phase-controlled kriging interpolation is used based on lithological distribution data on both sides of the fault. Lithological facies is used as a covariate to ensure that the spatial correlation between the restored area and the original data is ≥0.8. For key physical properties such as porosity and permeability, co-kriging interpolation is used, with seismic properties (such as wave impedance) as auxiliary variables, which significantly improves the accuracy and reliability of parameter interpolation.

[0046] Global model optimization involves applying Gaussian filtering (standard deviation 5-10m) to the boundary between the repaired area and the original model to effectively eliminate abrupt interface changes. Simultaneously, a geological knowledge base is incorporated to logically verify the repair results based on regional tectonic evolution models, correcting anomalous areas that do not conform to geological laws (such as inverted stratigraphic ages) to ensure the overall consistency of the model.

[0047] S24: Evaluate the repair effect through cross-validation. If the requirements are not met, return to S22 to adjust the parameters and repair again. Establish a comprehensive repair effect verification system, using a combination of quantitative and qualitative assessments to ensure that the repair quality meets engineering application standards.

[0048] Quantitative evaluation and verification indicators employ cross-validation to compare the agreement between the model and actual drilling data before and after repair. The root mean square error (RMSE) is calculated and compared with a preset threshold to objectively assess the repair accuracy. Specific quantitative standards include: Spatial error: The average deviation of the formation depth between the repaired model and the validation well is ≤5m.

[0049] Property consistency: The lithological consistency rate between the repair area and the neighboring area is ≥80%, and the coefficient of variation of physical property parameters is ≤0.15.

[0050] Qualitative assessment: Organize a blind review by senior geological experts (e.g., 3 or more) to assess the following through visual inspection: Does the relationship between the fault and the strata conform to geological laws? Does the distribution of stratigraphic thickness conform to sedimentary patterns? Whether the characterization of special geological phenomena is reasonable.

[0051] Figure 8 The diagram shows the visualization results of the attention mechanism of the present invention, wherein (a) is a visualization diagram of the weight of the inline direction attention mechanism, (b) is a visualization diagram of the weight of the crossline direction attention mechanism, (c) is a visualization diagram of the attention of fault strike continuity, and (d) is a visualization diagram of the attention of fault boundary enhancement.

[0052] A robust feedback and optimization mechanism is established to automatically adjust repair parameters when verification results fail to meet threshold requirements. Specific optimization measures include: expanding the fault-affected area, strengthening geological constraints, and adjusting interpolation algorithm parameters. Through multiple iterative optimizations, the repair results are optimized until they simultaneously meet both quantitative and qualitative assessment requirements, ensuring that the final model achieves a geological consistency rate of ≥80%.

[0053] The algorithm design and rigorous verification process of this method have enabled the standardization and refinement of the three-dimensional geological model repair process, providing reliable technical support for resource exploration and engineering decision-making under complex geological conditions.

[0054] Example 2: This example illustrates the implementation process of a three-dimensional geological model repair method based on fault identification results, using the application of three-dimensional geological model repair.

[0055] 1. Model Defect Analysis Before Repair The three-dimensional geological model of a certain mining area has two obvious defects due to inaccurate fault identification in the early stage: firstly, the direction of strata displacement on both sides of the F3 fault is obviously wrong; secondly, the concealed fault F5 was not identified, resulting in the formation of a continuous illusion that does not conform to geological laws.

[0056] 2. Repair process and parameter settings The method of this invention was used to re-identify the F3 fault (strike 210°, dip 70°) and the F5 fault (strike 30°, dip 55°). When constructing the fault surface model, the radius of the influence area of ​​the F3 fault was set to 100m, and the radius of the influence area of ​​the F5 fault was set to 80m. Strict geological constraints were imposed during the restoration process, including maintaining the sandstone thickness on both sides of the F3 fault at 30-50m and maintaining the continuity of the mudstone distribution in the hanging wall of the F5 fault. The entire restoration process took 2.5 hours of computation, and the processed model size was 50km × 50km × 3km.

[0057] 3. Verification of repair effect Quantitative assessment showed that the formation depth error between the repaired model and the validation well significantly decreased from 15m to 4.2m, and the lithological consistency rate increased from 65% to 88%. After blind review by five geological experts, they unanimously agreed that the cutting relationship between the fault and the formation in the repaired model was reasonable, and the formation contact relationship conformed to the regional tectonic evolution law.

[0058] Example 3: This example compares the performance of the deep learning-based intelligent fault identification method and the three-dimensional geological model repair method based on fault identification results described in this invention with traditional methods.

[0059] A comprehensive performance comparison test was conducted between the present invention and traditional methods in a unified hardware testing environment (Intel Xeon Gold 6248 CPU, NVIDIA A100 GPU).

[0060] The test results are shown in Table 2, which fully demonstrate the significant advantages of the method of the present invention.

[0061] Table 2 Performance Comparison of the Invention and Traditional Methods

[0062] Comprehensive test results show that the method of the present invention is significantly superior to traditional methods in terms of recognition accuracy, processing efficiency and result reliability, fully meets the accuracy and efficiency requirements of industrial applications, and is feasible for large-scale engineering promotion.

[0063] In summary, this invention, through the deep integration of deep learning and geological modeling, constructs a complete technical system from data acquisition to results verification. Compared with traditional methods, this invention significantly improves fault identification accuracy, model repair efficiency, and result reliability. Particularly under complex geological conditions, the method of this invention demonstrates clear technical advantages, providing reliable technical support for oil and gas exploration, mineral resource development, and other fields.

[0064] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A deep learning-based intelligent tomography identification method, characterized in that, Includes the following steps: S11: Collect multi-source geological data, perform standardized preprocessing on the data, and construct a sample database with fault labels. The multi-source geological data includes seismic profile data, drilling logging data, and geological outcrop images. S12: Design a deep neural network model that integrates attention mechanism, using seismic profile data as the main input, and outputs a fault probability distribution map through multi-scale feature extraction and attention weight allocation; S13: Use the sample database to train and optimize the deep neural network model, and use a hybrid loss function to handle the class imbalance problem; S14: The multi-source geological data collected from the target work area is preprocessed and used as the data to be identified as input into the trained model. The output includes fault boundary boxes, confidence scores and geometric parameters.

2. The intelligent tomography identification method based on deep learning according to claim 1, characterized in that, In S11: The seismic profile data is a three-dimensional seismic post-stack migration profile, which includes P-wave reflection amplitude and instantaneous frequency attributes; The drilling logging data, including sonic transit time and resistivity curves, is used to calibrate formation interfaces and fault locations. The geological outcrop images, specifically aerial photographs of fault outcrops taken by drones, are used to extract the surface morphological features of the faults.

3. The intelligent tomography identification method based on deep learning according to claim 1, characterized in that, The deep neural network model described in S12 employs an improved U-Net++ architecture, including: The encoder consists of 5 convolutional blocks, each containing 2 convolutional layers, a BatchNorm layer, and a ReLU activation function, achieving feature dimensionality reduction through max pooling. The attention gating module, located on the skip connection path between the encoder and decoder, enhances the representation of fault boundary features by calculating the spatial attention weights of the feature maps. The decoder gradually restores spatial resolution by upsampling and fusing the corresponding scale feature maps from the encoder. Auxiliary branches are added to the drilling and logging data input branches, and logging curve features are integrated into the main network through fully connected layers.

4. The intelligent tomography identification method based on deep learning according to claim 3, characterized in that, The attention gating module uses the Transformer's self-attention mechanism to capture the continuity features of fault strike by calculating the long-distance dependencies of the feature map.

5. The intelligent tomography identification method based on deep learning according to claim 1, characterized in that, The specific expression for the hybrid loss function described in S13 is as follows: Loss = 0.5×CrossEntropyLoss + 0.5×DiceLoss; CrossEntropyLoss is used to optimize class determination, while DiceLoss is used to improve boundary recognition accuracy.

6. The intelligent tomography identification method based on deep learning according to claim 1, characterized in that, The training and optimization of the deep neural network model described in S13 specifically involves: pre-training on a public dataset based on a cosine annealing strategy and transfer learning method, and introducing an early stopping mechanism to prevent overfitting.

7. A method for repairing a three-dimensional geological model based on fault identification results, characterized in that, Includes the following steps: S21: Based on the fault identification results, construct a fault geometric model, wherein the fault identification results are implemented based on a deep learning-based intelligent fault identification method according to any one of claims 1-6; S22: Spatial matching of the fault geometric model with the original three-dimensional geological model to determine the fault-affected area; S23: Adaptive algorithms constrained by geological laws are used to repair the model in the fault-affected area; S24: Evaluate the repair effect through cross-validation. If the requirements are not met, return to S22 to adjust the parameters and repair again.

8. The method for repairing a three-dimensional geological model based on fault identification results according to claim 7, characterized in that, As described in S21, the fault geometry model is constructed by fitting the fault planar morphology with B-spline curves and combining it with Kriging interpolation to correct the vertical morphology of the fault surface.

9. The method for repairing a three-dimensional geological model based on fault identification results according to claim 7, characterized in that, The adaptive repair algorithm described in S23 specifically includes: For stratigraphic interface repair, an improved Poisson equation interpolation method is adopted. The unaffected stratigraphic interfaces on both sides of the fault are used as boundary conditions, and stratigraphic thickness constraints are introduced. By solving the Poisson equation, stratigraphic interface morphology that conforms to geological laws is generated. Lithological property restoration was performed using phase-controlled kriging interpolation based on lithological distribution data on both sides of the fault. Global model optimization involves applying Gaussian filtering to the boundary between the repaired region and the original model to eliminate abrupt interface changes.

10. A method for repairing a three-dimensional geological model based on fault identification results according to claim 7, characterized in that, The assessment of remediation effectiveness described in S24 combines quantitative and qualitative assessments. The quantitative assessment uses cross-validation to compare the consistency between the model and the actual drilling data before and after the repair. It compares the calculated root mean square error with a preset threshold, including spatial error and attribute consistency. The qualitative assessment involved blind evaluation by senior geological experts, using visual inspection and evaluation to assess the relationship between faults and strata, the distribution of strata thickness, and special geological phenomena.

Images