Grid tomography and velocity optimization model training method based on well-seismic fusion

Abstract

The present application relates to a kind of well seismic fusion-based grid tomography and depth learning's speed optimization model training method, comprising: based on seismic, well logging, rock physics and geology layered data constructs multisource data set, obtains multisource feature data set by standardization and feature extraction;By improving grey correlation analysis, the target feature that is strongly related with longitudinal wave velocity is filtered, and target feature data set is constructed;Adaptive grid tomography inversion is carried out with well constraint, and velocity-feature correlation label data set is constructed;Based on target feature data set and label data set, well seismic fusion velocity optimization depth learning model is trained, and speed prediction quality control factor is embedded in whole process.The quality control factor of the present application is composed of three factors of well point constraint confidence, physical mechanism coincidence degree and feature matching degree by weighting, realizes the double breakthrough of velocity field high-precision optimization and prediction result reliability quantization, greatly improves the precision and industrial application value of velocity modeling in complex structure area.

Description

Technical Field

[0001] This invention relates to the field of oil and gas seismic exploration velocity modeling technology, and more specifically, to a velocity optimization model training method based on well-seismic fusion using grid tomography and deep learning. Background Technology

[0002] In oil and gas exploration and development, high-precision velocity models are the core foundation of pre-stack seismic depth migration imaging, directly determining the accuracy of subsurface structural imaging and serving as the core basis for well trajectory optimization, reservoir prediction, fracturing for production enhancement, and well location deployment. Especially in complex exploration areas such as the Permian thin-layer shale in the Sichuan Basin and the ultra-deep clastic rocks in the Tarim Basin, where stratigraphic deformation is severe and reservoir heterogeneity is high, extremely high demands are placed on the accuracy and reliability of velocity modeling results.

[0003] Currently, the mainstream velocity modeling techniques in the industry mainly include three categories: stacked velocity analysis, grid tomography inversion, and well-seismic joint velocity modeling. Among them, grid tomography inversion, which inverts the subsurface velocity field by minimizing the seismic travel time residual, is the core technology for velocity modeling in complex tectonic zones. However, it has significant drawbacks in practical applications: First, the inversion effect is highly dependent on the initial velocity model. When the initial model deviates significantly from the actual velocity field, the inversion is prone to getting trapped in local optima, resulting in slow iterative convergence and an inability to effectively characterize small-scale velocity anomalies within 100m. Second, the inversion grid often uses fixed-scale subdivision, which cannot balance the modeling accuracy of complex tectonic zones with the computational efficiency of flat strata, making it difficult to balance inversion accuracy and computational efficiency. Third, well constraints are mostly single-point hard constraints, failing to realize the geological regularity constraints between wells and layers. The well-seismic matching error generally exceeds 5%, and the velocity modeling accuracy of deep and ultra-deep strata is seriously insufficient, ultimately leading to structural distortion and inaccurate layer positioning in the migration imaging profile.

[0004] In recent years, deep learning technology has been gradually introduced into the field of velocity modeling. By using neural networks to learn the nonlinear mapping relationship between seismic data and velocity fields, rapid reconstruction and optimization of velocity fields can be achieved. For example, Chinese patent CN117492079A discloses a seismic velocity model reconstruction method based on the TDS-Unet network, which only uses seismic data as input and lacks hard constraints from well data, resulting in extremely poor generalization in areas with few wells. Chinese patent CN112308206A discloses a magnetotelluric-seismic wave joint inversion method, but this patent focuses on the joint inversion framework itself, making it difficult to quantitatively classify and dynamically constrain the spatial reliability of the velocity field. The model's generalization ability and result controllability still have significant shortcomings. Overall, existing deep learning speed modeling methods still have core bottlenecks: First, most are end-to-end black-box models, lacking constraints from the physical mechanisms of seismic wave propagation, resulting in inversion results that are prone to outliers that do not conform to geological laws, poor model generalization ability, and extremely unstable application effects in new work areas; Second, the degree of well-seismic fusion is low, achieving only simple fusion at the data level and failing to achieve deep fusion at the geological law and physical mechanism levels; Third, in small sample scenarios with a limited number of effective well points in oil and gas field work areas, the models are prone to overfitting and cannot meet the accuracy requirements of industrial applications. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for training a velocity optimization model based on well-seismic fusion grid tomography and deep learning, which improves the accuracy of velocity modeling while realizing the quantification of prediction results and full-space quality control.

[0006] The technical solution adopted by this invention to solve its technical problem is: constructing a velocity optimization model training method based on well-seismic fusion and deep learning, including the following steps: S1. Construct a multi-source dataset based on the acquired seismic data, well logging data, rock physical property data and geological stratification data of the target work area. Perform standardization processing and feature extraction on the multi-source dataset to construct a multi-source feature dataset. S2. Determine the correlation between each type of feature data and P-wave velocity in the multi-source feature dataset, and select target feature data from the feature data based on the correlation to construct a target feature dataset; S3. Based on the target feature dataset, perform adaptive grid tomography inversion with well constraints to obtain an initial high-precision velocity model. With the measured velocity at well points and the tomography inversion velocity as the core labels, construct a velocity-feature association label dataset. S4. Based on the target feature dataset and the velocity-feature association label dataset, train a well-seismic fusion velocity optimization deep learning model to obtain the trained velocity optimization model. The deep learning model has a built-in three-dimensional spatial velocity prediction quality control factor throughout the entire process of training, encoding-decoding, and inference.

[0007] According to the above scheme, the quality control factor includes a weighted sub-factor of well point constraint confidence, a sub-factor of physical mechanism fit, and a sub-factor of feature matching, which are used for loss constraints in the training process, feature weighting in the encoding-decoding process, and reliability quantification in the inference process.

[0008] According to the above scheme, the velocity optimization model includes an encoder module with quality control constraints, a multi-scale attention feature fusion module, and a decoder module with well-seismic dual constraints.

[0009] According to the above scheme, the calculation formula for the three-dimensional spatial velocity prediction quality control factor is as follows: (1) In the formula, Q(x,y,z) is the velocity prediction quality control factor of the grid at coordinates (x,y,z) in three-dimensional space. well (x,y,z) is the wellpoint constraint confidence factor, Q phy (x,y,z) is the physical mechanism fit factor, Q feat (x,y,z) are feature matching degree factors; w1, w2, and w3 are weight coefficients, and w1+w2+w3=1.

[0010] According to the above scheme, a quality control weighted joint loss function is constructed based on the quality control factors to constrain the model training process. The calculation formula of the quality control weighted joint loss function is as follows: (2) In the formula, N is the number of training samples, and Q i Loss is the quality control factor for the grid corresponding to the i-th sample. fit Loss is the velocity fitting loss. well Loss is the wellpoint constraint loss. phy The loss is determined by the physical mechanism constraint, and α, β, and γ are weighting coefficients.

[0011] According to the above scheme, in the encoder module with quality control constraints, the channel and spatial dimensions of the input features are dynamically weighted by the quality control factor to enhance feature extraction in high quality control value regions and suppress invalid features and noise in low quality control value regions. In the decoder module with well-seismic dual constraints, the deconvolution decoding process is constrained by a quality control factor. The accuracy of velocity prediction results in high quality control value areas is enhanced, and well point and physical mechanism constraints are added to low quality control value areas to avoid geological anomalies in the prediction results.

[0012] According to the above scheme, during model inference, the optimized three-dimensional velocity field and the corresponding three-dimensional quality control factor field are output simultaneously to quantify the reliability of the velocity prediction results for each grid. Based on the quality control factor field, the velocity prediction results are subjected to hierarchical quality control. The prediction results are directly adopted in the high reliability region with a quality control factor ≥ 0.8, the tomographic inversion constraint is added to the medium reliability region with a quality control factor of 0.5~0.8, and the low reliability region with a quality control factor < 0.5 is subject to manual quality control and parameter correction, so as to achieve full-process quality control of velocity modeling results.

[0013] According to the above scheme, step S2 includes: An improved grey relational analysis method is obtained by setting a sample confidence weight factor term in the grey relational analysis method; The value of the sample confidence weighting factor is determined based on the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging. The correlation degree between the characteristic data of each type and the P-wave velocity is determined based on the improved grey relational analysis method; further, the calculation formula of the improved grey relational analysis method is as follows: (3) In the formula, γ(X) i V) represents the i-th type of feature data X i The correlation with the P-wave velocity parameter V; ρ is the resolution coefficient, ranging from 0.5 to 0.6; n is the number of characteristic data sampling points; Δ i (k) represents the absolute difference between the feature data and the measured velocity value at the k-th sampling point; Δ min The minimum absolute difference of the entire sequence, Δ max ω is the maximum absolute difference of the entire sequence. k is the sample confidence weighting factor for the kth sampling point, and its value is determined by the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging. A correlation threshold is determined, and the feature data with a correlation greater than the correlation threshold are identified as the target feature data; The target feature data is clustered using a clustering algorithm. Based on the clustering results, the target feature data is grouped to obtain target feature data groups. The target feature dataset is then constructed based on the target feature data groups.

[0014] According to the above scheme, in step S3, an adaptive grid tomography inversion with well constraints is performed based on the target feature dataset to obtain an initial high-precision velocity model, including: Based on the target work area's geological body scale, structural complexity, and well density, adaptive subdivision and inversion grids are used. In areas with developed faults, complex structures, and dense wells, the grids are densified, while in areas with gentle strata and no wells, the grids are sparsed. Using the superimposed velocity spectrum as the initial constraint, hard constraints of well point measured P-wave velocity, geological layer interface constraints, and soft constraints of seismic travel time residuals are added to carry out grid tomography iterative inversion; The inversion is terminated by using the well-seismic matching error of the inversion results. When the relative error of the well-seismic velocity is less than a preset threshold, the inversion is terminated, and an initial high-precision velocity model is obtained.

[0015] According to the above scheme, the method for standardizing and extracting features from the multi-source dataset to construct a multi-source feature dataset in step S1 includes: The seismic data is denoised, and seismic attribute data is extracted from the denoised seismic data. Identify the marker well and the marker layer within the marker well, and standardize the logging data based on the logging data of the marker layer to obtain standard logging data. Based on the rock physical property data, the static Young's modulus and static Poisson's ratio are determined. Based on the static Young's modulus and static Poisson's ratio, combined with wellbore collapse analysis, the maximum horizontal principal stress is determined. The seismic attribute data, the standard well logging data, and the maximum horizontal principal stress are mapped to a unified feature space and fused to obtain a fused feature data volume, which is then determined as the multi-source feature dataset.

[0016] The method for training a velocity optimization model based on well-seismic fusion and deep learning, as described in this invention, has the following beneficial effects: 1. This invention proposes for the first time a three-dimensional spatial velocity prediction quality control factor suitable for seismic velocity modeling, realizing a quantitative characterization of the reliability of velocity prediction results and filling the technical gap in the industry where deep learning velocity modeling lacks a built-in quality control mechanism. 2. This invention is the first to embed the quality control factor into the entire process of training, encoding-decoding, and inference of a deep learning model. It solves the problem of poor model generalization from the root of training, breaks the limitations of black box models from the structural perspective, and achieves the simultaneous completion of prediction and quality control from the output perspective, forming a complete technical closed loop. 3. This invention solves the core bottleneck restricting the industrialization of deep learning speed modeling technology. While improving the accuracy of speed modeling, it achieves full-space quantitative quality control, significantly reducing exploration risks. It has strong promotion and application value in complex structural areas and deep and ultra-deep oil and gas exploration. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the speed optimization model training method with quality control factor constraints provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the practical application method for speed optimization provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the deep learning network structure for the entire process of quality control factors embedded in an embodiment of the present invention. Figure 4 This is a comparison diagram of the velocity field before and after optimization and the corresponding offset imaging profile provided in this embodiment of the invention. Detailed Implementation

[0018] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] This invention relates to a high-precision grid tomography and deep learning-based velocity optimization model training method based on well-seismic fusion, comprising the following steps: S1. Construct a multi-source dataset based on the acquired seismic data, well logging data, rock physical property data and geological stratification data of the target work area. Perform standardization processing and feature extraction on the multi-source dataset to construct a multi-source feature dataset. S2. Determine the correlation between each type of feature data and P-wave velocity in the multi-source feature dataset, and select target feature data from the feature data based on the correlation to construct a target feature dataset; S3. Based on the target feature dataset, perform adaptive grid tomography inversion with well constraints to obtain an initial high-precision velocity model. With the measured velocity at well points and the tomography inversion velocity as the core labels, construct a velocity-feature association label dataset. S4. Based on the target feature dataset and the velocity-feature association label dataset, train a well-seismic fusion velocity optimization deep learning model to obtain the trained velocity optimization model. The deep learning model has a built-in three-dimensional spatial velocity prediction quality control factor throughout the training, encoding-decoding, and inference processes. The quality control factor is composed of a weighted sub-factor of well point constraint confidence, a sub-factor of physical mechanism fit, and a sub-factor of feature matching, which is used for loss constraints in the training process, feature weighting in the encoding-decoding process, and reliability quantification in the inference process. The velocity optimization model includes an encoder module with quality control constraints, a multi-scale attention feature fusion module, and a decoder module with well-seismic dual constraints.

[0020] The innovative three-dimensional spatial velocity prediction quality control factor proposed in this invention is designed for the three-dimensional spatial characteristics of seismic velocity modeling, covering the entire three-dimensional mesh of the work area. The calculation formula is as follows: (1) Q(x,y,z): The velocity prediction quality control factor of the grid at coordinates (x,y,z) in three-dimensional space. The value ranges from 0 to 1. The closer the value is to 1, the higher the reliability of the velocity prediction result of the grid. Q well (x,y,z): Well point constraint confidence sub-factor, calculated based on the spatial distance between the prediction grid and the nearest well point. The closer the distance, the stronger the well point constraint, and the higher the value. Specifically: 0.9~1 when the distance is <500m, 0.7~0.9 when the distance is 500m≤distance<1000m, and 0.5~0.7 when the distance is ≥1000m. Q phy (x,y,z): Physical mechanism fit factor, calculated based on the seismic travel time residuals at this grid. The smaller the travel time residuals, the higher the degree of fit with the physical mechanism of seismic wave propagation, and the higher the value. Specifically: 0.9~1 when residual <1ms, 0.7~0.9 when 1ms≤residual<3ms, and 0.5~0.7 when residual ≥3ms. Q feat (x,y,z): Feature matching degree sub-factor, calculated based on the KL divergence between the input feature and the training set feature distribution at this grid point. The higher the matching degree, the higher the reliability of the model's generalization prediction, and the higher the value. Specifically: when KL divergence < 0.1, the value is 0.9~1; when 0.1 ≤ KL divergence < 0.3, the value is 0.7~0.9; when KL divergence ≥ 0.3, the value is 0.5~0.7. w1, w2, and w3 are weighting coefficients, and w1+w2+w3=1. The preferred values ​​are 0.4, 0.35, and 0.25, respectively, which can be dynamically adjusted according to the well point density and structural complexity of the work area.

[0021] The role of quality control factors in the deep learning training process includes: Based on the quality control factors, a quality control weighted joint loss function is constructed to constrain the model training process. The calculation formula for the quality control weighted joint loss function is as follows: (2) In the formula, N is the number of training samples, and Q i Loss is the quality control factor for the grid corresponding to the i-th sample. fit Loss is the mean squared error loss for velocity fitting. well Loss is the average absolute error loss of the wellpoint velocity. phy The value represents the earthquake travel time residual loss, and α, β, and γ are weighting coefficients, with preferred values ​​of 0.5, 0.3, and 0.2, respectively.

[0022] By dynamically weighting the training samples using quality control factors, the learning weights of high-quality control value samples near well points and conforming to physical mechanisms are strengthened, while noise interference from low-quality control value samples with large residuals in areas without wells is suppressed. This addresses the model overfitting problem at its root and significantly improves the model's generalization ability in areas with few wells and complex structures.

[0023] In the encoder module with quality control constraints, the channel and spatial dimensions of the input features are dynamically weighted by quality control factors to enhance feature extraction in high quality control value regions and suppress invalid features and noise in low quality control value regions. In the decoder module with well and seismic dual constraints, the deconvolution decoding process is constrained by quality control factors to enhance the accuracy of velocity prediction results in high quality control value regions and supplement well point and physical mechanism constraints in low quality control value regions to avoid outliers in the prediction results that do not conform to geological laws, thus breaking the limitations of traditional black box models.

[0024] During model inference, the optimized 3D velocity field and corresponding 3D quality control factor field are output simultaneously to quantify the reliability of the velocity prediction results for each grid in the full 3D space of the work area. Based on the quality control factor field, hierarchical quality control of velocity modeling results can be achieved: the prediction results are directly adopted for high reliability areas with quality control factor ≥ 0.8, tomographic inversion constraints are added for medium reliability areas with quality control factor 0.5~0.8, and manual quality control and parameter correction are performed for low reliability areas with quality control factor < 0.5. For the first time, the entire process of velocity modeling from "prediction" to "quality control" has been automated, which greatly improves the efficiency of industrial applications.

[0025] Furthermore, in step S2, the correlation between each type of feature data and the P-wave velocity in the multi-source feature dataset is determined, and target feature data is selected from the feature data based on the correlation to construct a target feature dataset, including: An improved grey relational analysis method is obtained by setting a sample confidence weight factor term in the grey relational analysis method; The value of the sample confidence weighting factor is determined based on the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging. The correlation degree between the characteristic data of each type and the P-wave velocity is determined based on the improved grey relational analysis method; further, the calculation formula of the improved grey relational analysis method is as follows: (3) In the formula, γ(X) i V) represents the i-th type of feature data X. i The correlation with the P-wave velocity parameter V; ρ is the resolution coefficient, ranging from 0.5 to 0.6; n is the number of characteristic data sampling points; Δ i (k) represents the absolute difference between the feature data and the measured velocity value at the k-th sampling point; Δ min The minimum absolute difference of the entire sequence, Δ max ω is the maximum absolute difference of the entire sequence. k The sample confidence weighting factor for the k-th sampling point is determined by the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging.

[0026] Furthermore, the sample confidence weighting factor ω k The value selection rule is as follows: when the logging curve integrity is ≥90%, the layer matching error is ≤2ms, and the well point imaging coverage is ≥80%, ω k Take 1.2; otherwise ω k Take 0.8.

[0027] A correlation threshold is determined, and the feature data with a correlation greater than the correlation threshold are identified as the target feature data; The target feature data is clustered using a clustering algorithm. Based on the clustering results, the target feature data is grouped to obtain target feature data groups. The target feature dataset is then constructed based on the target feature data groups.

[0028] Furthermore, in step S3, an adaptive grid tomography inversion with well constraints is performed based on the target feature dataset to obtain an initial high-precision velocity model, including: Based on the target work area's geological body scale, structural complexity, and well density, adaptive subdivision and inversion grids are used. In areas with developed faults, complex structures, and dense wells, the grids are densified, while in areas with gentle strata and no wells, the grids are sparsed. Using the superimposed velocity spectrum as the initial constraint, hard constraints of well point measured P-wave velocity, geological layer interface constraints, and soft constraints of seismic travel time residuals are added to carry out grid tomography iterative inversion; The inversion is terminated by using the well-seismic matching error of the inversion results. When the relative error of the well-seismic velocity is less than a preset threshold, the inversion is terminated, and an initial high-precision velocity model is obtained.

[0029] Furthermore, in step S1, the multi-source dataset is standardized and features are extracted to construct a multi-source feature dataset, including: The seismic data is denoised, and seismic attribute data is extracted from the denoised seismic data. Identify the marker well and the marker layer within the marker well, and standardize the logging data based on the logging data of the marker layer to obtain standard logging data. Based on the rock physical property data, the static Young's modulus and static Poisson's ratio are determined. Based on the static Young's modulus and static Poisson's ratio, combined with wellbore collapse analysis, the maximum horizontal principal stress is determined. The seismic attribute data, the standard well logging data, and the maximum horizontal principal stress are mapped to a unified feature space and fused to obtain a fused feature data volume, which is then determined as the multi-source feature dataset.

[0030] This invention also provides a speed optimization method for high-precision grid tomography and deep learning based on well-seismic fusion, comprising: A target data set to be detected is constructed based on the seismic data, well logging data, rock physical property data, and geological stratification data to be detected. After preprocessing the target data set, it is input into the trained velocity optimization model to obtain the optimized high-precision three-dimensional velocity field and the corresponding three-dimensional quality control factor field output by the model. The speed optimization model is a model trained according to the speed optimization model training method.

[0031] This invention also provides a training device for a high-precision grid tomography and deep learning velocity optimization model based on well-seismic fusion, comprising: The first module is used to construct a multi-source dataset based on the acquired seismic data, well logging data, rock physical property data and geological stratification data, and to perform standardization processing and feature extraction on the multi-source dataset to construct a multi-source feature dataset. The second module is used to determine the correlation between each type of feature data and the P-wave velocity in the multi-source feature dataset, and to select target feature data from the feature data based on the correlation to construct a target feature dataset; The third module is used to perform well-constrained adaptive grid tomography inversion based on the target feature dataset to obtain an initial high-precision velocity model and construct a velocity-feature association label dataset. The fourth module is used to train a well-seismic fusion velocity optimization deep learning model based on the target feature dataset and the velocity-feature association label dataset, to obtain the trained velocity optimization model; the deep learning model has a built-in three-dimensional spatial velocity prediction quality control factor in the entire process of training, encoding-decoding, and inference; the velocity optimization model includes an encoder module with quality control constraints, a multi-scale attention feature fusion module, and a decoder module with well-seismic dual constraints.

[0032] The present invention also provides an electronic device, comprising: Memory, used to store computer programs; A processor for implementing the speed optimization model training method, or the speed optimization method, when executing the computer program.

[0033] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the speed optimization model training method or the speed optimization method.

[0034] This embodiment focuses on the Middle Permian thin-layered shale gas reservoirs in the central Sichuan Basin, covering an area of ​​1200 km². The area has 60 full-coverage 3D seismic data runs at a dominant frequency of 35 Hz. There are 12 exploration wells, including 8 imaging logging wells and 6 core wells. The target layer is buried at a depth of 3000–5000 m, characterized by well-developed faults and strong reservoir heterogeneity. Existing technologies in this area have velocity prediction errors exceeding 15% in well-free areas, and the reliability of these predictions cannot be quantified, resulting in extremely high exploration risks. This invention's method is used to optimize velocity and verify the innovative effect of the invention's quality control factors.

[0035] The following combination Figure 1 , Figure 1 This is a flowchart of a speed optimization model training method with quality control factor constraints provided in an embodiment of the present invention. The method specifically includes the following steps: S101. Based on the acquired seismic data, well logging data, rock physical property data and geological stratification data of the target work area, construct a multi-source dataset, and perform standardization processing and feature extraction on the multi-source dataset to construct a multi-source feature dataset.

[0036] In this embodiment, the three-dimensional pre-stack seismic data is first preprocessed by bandpass filtering for noise reduction, inverse Q filtering for frequency enhancement, offset repositioning, and spherical diffusion compensation to extract 16 seismic features, including stacking velocity spectrum, coherence volume, curvature volume, amplitude attribute, and frequency attribute. A 28m thick stable mudstone section in well A of the work area is selected as a marker layer, and the logging data of 12 wells in the entire work area are standardized and corrected to obtain standard logging data. Logging features such as sonic transit time, P-wave velocity, and density are extracted. Based on core experimental data, a rock physical quantity plate is calibrated, and rock physical features such as P-wave / S-wave velocity ratio and static Young's modulus are calculated. Using 6 sets of geological stratification data as a benchmark, all features are mapped to a unified three-dimensional grid space, and a multi-source feature dataset with dimensions of 200×300×150×8 is obtained through principal component analysis fusion.

[0037] S102. Determine the correlation between each type of feature data and the P-wave velocity in the multi-source feature dataset, and construct a target feature dataset by filtering target feature data from the feature data based on the correlation.

[0038] Table 1. Calculation results of the correlation between characteristic data and P-wave velocity

[0039] This embodiment uses an improved grey relational analysis method to calculate the correlation between each feature and the P-wave velocity. The correlation threshold is set to 0.8, and eight strongly correlated features, such as superposition velocity spectrum, acoustic transit time, coherence volume, and P-wave / S-wave velocity ratio, are selected. The target feature dataset is constructed by grouping by stratigraphic unit through K-Means clustering.

[0040] S103. Based on the target feature dataset, perform adaptive grid tomography inversion with well constraints to obtain an initial high-precision velocity model. With the measured velocity at well points and the tomography inversion velocity as the core labels, construct a velocity-feature association label dataset.

[0041] This embodiment uses an adaptive meshing method based on the structural complexity of the work area. The mesh is refined to 12.5m×12.5m×5m in the fractured area and sparsed to 50m×50m×20m in the flat area. Using the stacked velocity spectrum as the initial model, the velocity hard constraints, stratigraphic constraints, and travel time residual constraints of 12 wells are added. Gauss-Newton tomographic inversion is carried out. After 12 iterations, the average relative error of well-seismic analysis is reduced to 2.8%, resulting in an initial high-precision velocity model. The measured velocity at well points is used as the high-precision label, and the tomographic inversion velocity is used as the spatial constraint label to construct a velocity-feature association label dataset.

[0042] S104. Based on the target feature dataset and the velocity-feature association label dataset, train the well-seismic fusion velocity optimization deep learning model to obtain the trained velocity optimization model; the deep learning model has a built-in three-dimensional spatial velocity prediction quality control factor in the entire process of training, encoding-decoding, and inference.

[0043] This step is the core innovative implementation step of the present invention, and the specific implementation process is as follows: 1. Calculate the full three-dimensional spatial quality control factor field of the work area. According to the quality control factor calculation formula, the quality control factor for each grid in the work area was calculated, where w1=0.4, w2=0.35, and w3=0.25, resulting in a three-dimensional quality control factor field covering the entire work area. Statistical analysis showed that the average quality control factor within 500m of the well point was 0.92, and the average quality control factor in the no-well area was 0.68, achieving a quantitative characterization of full-space reliability.

[0044] 2. Construct a quality control weighted joint loss function and conduct model training. In this embodiment, the velocity optimization model adopts a 3D-UNet backbone network, consisting of an encoder module with quality control constraints, a multi-scale attention feature fusion module, and a decoder module with well-seismic dual constraints. The target feature dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer is used with an initial learning rate of 0.001, a batch size of 4, and a maximum of 100 iterations. A quality control-weighted joint loss function is used for parameter optimization.

[0045] During training, samples are dynamically weighted using quality control factors. Well-point samples with high quality control values ​​have a learning weight 1.5 to 2 times that of low-quality control samples without wells, effectively strengthening the learning of reliable samples and suppressing noise interference. Training is terminated early when the validation set loss does not decrease for 10 consecutive rounds, and training is completed after 76 iterations.

[0046] 3. Verification of Model Accuracy and Innovation Effect Testing showed that the model trained by this invention had an average relative error of ≤2.1% in velocity prediction on the test set and a wellpoint velocity matching rate of ≥97%. In the high reliability region with a quality control factor ≥0.8, the average relative error in prediction was ≤1.5%, which is 72% more accurate than traditional deep learning methods. In the medium reliability region with a quality control factor of 0.5~0.8, the average relative error in prediction was ≤3.2%. In the low reliability region with a quality control factor <0.5, the average relative error in prediction was ≤5.8%, which fully meets the needs of industrial exploration.

[0047] More importantly, the model of this invention can simultaneously output a quality control factor field, realizing full-space quantification of the reliability of prediction results, solving the core pain point that traditional black box models cannot control quality, and providing reliable risk warnings for subsequent well location deployment.

[0048] The following combination Figure 2 , Figure 2 The flowchart below shows a practical application method for speed optimization provided in this embodiment of the invention. The method includes: A dataset to be processed is constructed based on seismic, well logging, rock physics, and geological stratification data of the work area to be inspected; The dataset to be processed is standardized, hierarchically matched, and features are extracted to obtain the feature dataset to be optimized. Input the feature dataset to be optimized into the trained velocity optimization model to obtain the optimized high-precision three-dimensional velocity field and the corresponding three-dimensional quality control factor field.

[0049] In this embodiment, the optimized velocity field is used for pre-stack depth migration imaging, and the results are as follows: Figure 4 As shown. Figure 4 In the diagram, the horizontal axis represents the CDP number (1168-1220), and the vertical axis represents the depth (m, 2800~3900). Compared with traditional methods, the optimized velocity field of this invention is more uniform and has stronger stratification. The corresponding imaging profile shows clearer micro-amplitude structures, more accurate layer positioning, and richer imaging details. The well-seismic composite record matching error is reduced from 8.7% to 2.1%. The quality control factor field grading evaluation clearly indicates the high reliability area around the well point and the low reliability area at the edge of the work area, providing a comprehensive quality control basis for subsequent exploration deployment.

[0050] The following combination Figure 3 , Figure 3 This is a schematic diagram of the deep learning network structure embedded in the entire process of quality control factors provided in an embodiment of the present invention. The device includes: Module 1: Input, used to construct multi-source datasets and multi-source feature datasets; The second module: the encoder, used for feature correlation calculation and target feature dataset construction; The third module: Feature fusion, used for adaptive grid tomography inversion and label dataset construction; The fourth module is the decoder, used for training deep learning models with built-in quality control factors to obtain speed-optimized models.

[0051] Based on the above embodiments, the core innovation and inventive advantage of the present invention are as follows: For the first time, a three-dimensional spatial velocity prediction quality control factor suitable for seismic velocity modeling was proposed, realizing a quantitative characterization of the reliability of velocity prediction results and filling the technical gap in the industry where deep learning velocity modeling lacks a built-in quality control mechanism. For the first time, the quality control factor is embedded into the entire process of training, encoding-decoding, and inference of a deep learning model. This solves the problem of poor model generalization from the root of training, breaks the limitations of black box models from the structural perspective, and achieves simultaneous completion of prediction and quality control from the output perspective, forming a complete technical closed loop. It solves the core bottleneck restricting the industrialization of deep learning speed modeling technology. While improving the accuracy of speed modeling, it achieves full-space quantitative quality control, significantly reducing exploration risks. It has strong promotion and application value in complex structural areas and deep and ultra-deep oil and gas exploration.

[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for training a velocity optimization model based on well-seismic fusion using grid tomography and deep learning, characterized in that, Includes the following steps: S1. Construct a multi-source dataset based on the acquired seismic data, well logging data, rock physical property data and geological stratification data of the target work area. Perform standardization processing and feature extraction on the multi-source dataset to construct a multi-source feature dataset. S2. Determine the correlation between each type of feature data and P-wave velocity in the multi-source feature dataset, and select target feature data from the feature data based on the correlation to construct a target feature dataset; S3. Based on the target feature dataset, perform adaptive grid tomography inversion with well constraints to obtain an initial high-precision velocity model. With the measured velocity at well points and the tomography inversion velocity as the core labels, construct a velocity-feature association label dataset. S4. Based on the target feature dataset and the velocity-feature association label dataset, train a well-seismic fusion velocity optimization deep learning model to obtain the trained velocity optimization model. The deep learning model has a built-in three-dimensional spatial velocity prediction quality control factor throughout the entire process of training, encoding-decoding, and inference.

2. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 1, characterized in that, The quality control factors include a weighted average of well point constraint confidence factor, physical mechanism fit factor, and feature matching factor, which are used for loss constraints in the training process, feature weighting in the encoding-decoding process, and reliability quantification in the inference process.

3. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 2, characterized in that, The velocity optimization model includes an encoder module with quality control constraints, a multi-scale attention feature fusion module, and a decoder module with well-seismic dual constraints.

4. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 3, characterized in that, The formula for calculating the three-dimensional spatial velocity prediction quality control factor is as follows: （1） In the formula, Q(x,y,z) is the velocity prediction quality control factor of the grid at coordinates (x,y,z) in three-dimensional space. well (x,y,z) is the wellpoint constraint confidence factor, Q phy (x,y,z) is the physical mechanism fit factor, Q feat (x,y,z) are feature matching degree sub-factors; w1, w2, and w3 are weighting coefficients, and w1 + w2 + w3 = 1.

5. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 4, characterized in that, Based on the quality control factors, a quality control weighted joint loss function is constructed to constrain the model training process. The calculation formula for the quality control weighted joint loss function is as follows: （2） In the formula, N is the number of training samples, and Q i Loss is the quality control factor for the grid corresponding to the i-th sample. fit Loss is the velocity fitting loss. well Loss is the wellpoint constraint loss. phy The loss is determined by the physical mechanism constraint, and α, β, and γ are weighting coefficients.

6. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 3, characterized in that, In the encoder module with quality control constraints, the channel and spatial dimensions of the input features are dynamically weighted by the quality control factor to enhance feature extraction in high quality control value regions and suppress invalid features and noise in low quality control value regions. In the decoder module with well-seismic dual constraints, the deconvolution decoding process is constrained by a quality control factor. The accuracy of velocity prediction results in high quality control value areas is enhanced, and well point and physical mechanism constraints are added to low quality control value areas to avoid geological anomalies in the prediction results.

7. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 3, characterized in that, During model inference, the optimized three-dimensional velocity field and the corresponding three-dimensional quality control factor field are output simultaneously to quantify the reliability of the velocity prediction results for each grid. Based on the quality control factor field, the velocity prediction results are subjected to hierarchical quality control. The prediction results are directly adopted in the high reliability region with a quality control factor ≥ 0.8, the tomographic inversion constraint is added to the medium reliability region with a quality control factor of 0.5~0.8, and the low reliability region with a quality control factor < 0.5 is subject to manual quality control and parameter correction, so as to achieve full-process quality control of velocity modeling results.

8. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 1, characterized in that, Step S2 includes: An improved grey relational analysis method is obtained by setting a sample confidence weight factor term in the grey relational analysis method; The value of the sample confidence weighting factor is determined based on the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging. The correlation degree between the characteristic data of each type and the P-wave velocity is determined based on the improved grey relational analysis method; further, the calculation formula of the improved grey relational analysis method is as follows: （3） In the formula, γ(X) i V) represents the i-th type of feature data X i The correlation with the P-wave velocity parameter V; ρ is the resolution coefficient, ranging from 0.5 to 0.6; n is the number of characteristic data sampling points; Δ i (k) represents the absolute difference between the feature data and the measured velocity value at the k-th sampling point; Δ min The minimum absolute difference of the entire sequence, Δ max ω is the maximum absolute difference of the entire sequence. k is the sample confidence weighting factor for the kth sampling point, and its value is determined by the integrity of the logging curve, the accuracy of the layer matching, and the coverage of the well point imaging. A correlation threshold is determined, and the feature data with a correlation greater than the correlation threshold are identified as the target feature data; The target feature data is clustered using a clustering algorithm. Based on the clustering results, the target feature data is grouped to obtain target feature data groups. The target feature dataset is then constructed based on the target feature data groups.

9. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 1, characterized in that, In step S3, an adaptive grid tomography inversion with well constraints is performed based on the target feature dataset to obtain an initial high-precision velocity model, including: Based on the target work area's geological body scale, structural complexity, and well density, adaptive subdivision and inversion grids are used. In areas with developed faults, complex structures, and dense wells, the grids are densified, while in areas with gentle strata and no wells, the grids are sparsed. Using the superimposed velocity spectrum as the initial constraint, hard constraints of well point measured P-wave velocity, geological layer interface constraints, and soft constraints of seismic travel time residuals are added to carry out grid tomography iterative inversion; The inversion is terminated by using the well-seismic matching error of the inversion results. When the relative error of the well-seismic velocity is less than a preset threshold, the inversion is terminated, and an initial high-precision velocity model is obtained.

10. The method for training a velocity optimization model based on well-seismic fusion and deep learning according to claim 1, characterized in that, In step S1, the method for standardizing and extracting features from the multi-source dataset to construct a multi-source feature dataset includes: The seismic data is denoised, and seismic attribute data is extracted from the denoised seismic data. Identify the marker well and the marker layer within the marker well, and standardize the logging data based on the logging data of the marker layer to obtain standard logging data. Based on the rock physical property data, the static Young's modulus and static Poisson's ratio are determined. Based on the static Young's modulus and static Poisson's ratio, combined with wellbore collapse analysis, the maximum horizontal principal stress is determined. The seismic attribute data, the standard well logging data, and the maximum horizontal principal stress are mapped to a unified feature space and fused to obtain a fused feature data volume, which is then determined as the multi-source feature dataset.

Images