A method for constructing a multi-source data fusion stratum parameter field by incremental dynamic updating

Abstract

The application provides a kind of incremental dynamic updating multi-source data fusion stratum parameter field construction method, the construction method includes: integration of seismic, adjacent well logging geological priori data arrangement work area induced point, each point is configured local covariance matrix, realize the anisotropic transformation of kernel function coordinate;Sparse Gaussian process variation target function is constructed, which is composed of data lower bound, geological gradient constraint, wavelet fidelity, uncertainty covariance regularization term weighting;Regular term adjusts covariance correlation scale according to the stratum variance of last round, expands in high uncertainty area and shrinks in low uncertainty area;First, use priori data to obtain initial stratum parameter field;After adding new while-drilling logging data, determine whether to add induced point according to uncertainty adaptive sampling rule, update model parameter and covariance matrix by joint optimization;The grid of work area is predicted by using the optimized model, and the three-dimensional stratum parameter field with mean value and uncertainty variance is output.

Description

Technical Field

[0001] This application belongs to the field of geology, and in particular relates to a method for constructing a stratigraphic parameter field by fusion of multi-source data with incremental dynamic updates. Background Technology

[0002] Static geological modeling methods, based on pre-drilling seismic data and adjacent well logging data, establish an initial formation parameter field through deterministic interpolation or geostatistical methods. However, subsurface geological conditions typically exhibit high spatial heterogeneity, non-stationarity, and anisotropy. Low-resolution pre-drilling seismic data and sparse adjacent well logging data alone cannot represent the spatial distribution patterns of formation parameters, and the initial model contains significant uncertainty. Logging while drilling (LoWL) technology can acquire high-precision formation information around the well trajectory in real time. In actual LoWL updates, with the continuous influx of data, global recalculation for model updates would impose a computational burden; while simple local data assimilation could easily disrupt the continuity and multi-scale characteristics of the overall geological structure, leading to morphological distortion or abrupt changes in local parameters in the updated model. The anisotropy of the subsurface medium varies drastically with spatial location. Without macroscopic constraints, the model is prone to overfitting or failing to adjust spatial correlation scales. In sparsely covered, high-uncertainty regions and densely covered, low-uncertainty regions, it is impossible to detect local geological features. Furthermore, existing methods lack mechanisms for utilizing and transmitting geological uncertainties when fusing multi-source heterogeneous data. This results in the output formation parameter field providing only deterministic predictions, failing to offer reliable risk assessments for drilling decisions. Therefore, there is an urgent need for a formation parameter field construction method that can handle multi-source data, adaptively represent spatial anisotropy, and update and quantify prediction uncertainties in real time during drilling. Summary of the Invention

[0003] This invention proposes an incrementally dynamically updated multi-source data fusion method for constructing a stratigraphic parameter field, addressing the problem that existing technologies lack a method for constructing a stratigraphic parameter field that can handle multi-source data, adaptively represent spatial anisotropy, and update and quantify prediction uncertainties in real-time during drilling. The method includes:

[0004] Prior geological data including seismic and adjacent well logging data are acquired. Initial induction points are set up in the target work area. A local covariance matrix is ​​configured for each induction point. The inverse of the local covariance matrix is ​​used to perform anisotropic linear transformation on the input coordinates during kernel function calculation. A variational inference objective function for a sparse Gaussian process model is constructed. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation.

[0005] Based on the prior geological data, the variational inference objective function is initially optimized to obtain the initial formation parameter field; new logging-while-drilling data is acquired and added to the training set; based on the adaptive sampling criterion of model uncertainty, it is determined whether to add induced points in the neighborhood of the new logging-while-drilling data; the variational inference objective function is jointly optimized to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points;

[0006] The uncertainty-driven covariance regularization term uses the uncertainty variance of the previous formation parameter field to constrain the local covariance matrix of the current round; the optimized sparse Gaussian process model is used to predict the grid points of the target work area, and outputs a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

[0007] Furthermore, this invention also relates to an incrementally dynamically updated multi-source data fusion stratigraphic parameter field construction system, comprising:

[0008] The module is used to acquire prior geological data including seismic and adjacent well logging data, set up initial induction points in the target work area, and configure a local covariance matrix for each induction point. The inverse of the local covariance matrix is ​​used to perform anisotropic linear transformation on the input coordinates during kernel function calculation. The variational inference objective function of the sparse Gaussian process model is constructed. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation.

[0009] The update module is used to perform initial optimization of the variational inference objective function based on the prior geological data to obtain an initial formation parameter field; acquire new logging-while-drilling data and add the new logging-while-drilling data to the training set; determine whether to add induced points in the neighborhood of the new logging-while-drilling data based on the adaptive sampling criterion of model uncertainty; and perform joint optimization of the variational inference objective function to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points.

[0010] The output module is used to constrain the local covariance matrix of the current round using the uncertainty-driven covariance regularization term of the previous round formation parameter field; and to predict the grid points of the target work area using the optimized sparse Gaussian process model, and output a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

[0011] This invention, by setting up induced points and configuring a local covariance matrix in the target work area, and combining a variational inference objective function composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, can fuse prior geological data and detect the anisotropy and multi-scale characteristics of the formation. During the drilling phase, by fusing new logging data and adding induced points based on model uncertainties, the variational parameters and covariance matrix are jointly updated. Utilizing the uncertainty-driven covariance regularization term, the correlation scale is expanded in high-uncertainty regions and contracted in low-uncertainty regions. This method improves the accuracy and resolution of three-dimensional formation parameter field prediction and provides reliable prediction uncertainty variance, laying a solid foundation for drilling trajectory optimization and construction risk assessment. Attached Figure Description

[0012] Figure 1 A flowchart of a method for constructing a multi-source data fusion stratigraphic parameter field for incremental dynamic updates;

[0013] Figure 2 This is a spatial distribution map of the data and the induced points;

[0014] Figure 3 A diagram showing the relationship between the receptive field scale and uncertainty. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] This invention proposes a method for constructing a multi-source data fusion stratigraphic parameter field with incremental dynamic updates, such as... Figure 1 As shown, it includes:

[0017] S1. Acquire prior geological data including seismic and adjacent well logging data. Deploy initial induction points in the target work area. Configure a local covariance matrix for each induction point. When calculating the kernel function, use the inverse of the local covariance matrix to perform anisotropic linear transformation on the input coordinates. Construct a variational inference objective function for a sparse Gaussian process model. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation.

[0018] Seismic impedance volume data containing 3D coordinate information and sonic transit time logging curves from adjacent wells are read as prior geological data. The coordinates and geological attributes are normalized to the range of 0 to 1. Unsupervised clustering is performed on the 3D coordinate grid of the target work area, and the center coordinates of each cluster are extracted as the initial induction point locations.

[0019] A local covariance matrix is ​​initialized for each induced point. This matrix is ​​obtained by multiplying a lower triangular matrix with its transpose to ensure positive definiteness. When calculating the squared exponential kernel function, the inverse of this local covariance matrix is ​​used to perform Mahalanobis distance operation on the input three-dimensional coordinate vector to achieve anisotropic linear transformation.

[0020] In one embodiment, the model's predicted distribution parameters and observed data are input into the function, and the computational graph performs an expected log-likelihood evaluation operation to obtain a specific scalar value of the goodness of fit between the model's predicted distribution and the actual observed data. The function extracts the statistical parameters of the variational posterior probability distribution and the preset prior distribution, calculates the relative entropy deviation between the two, and obtains the specific value of the divergence as the penalty function. The expected log-likelihood value and the divergence penalty value obtained above are algebraically summed, and the output value is used as the final lower bound of the data evidence.

[0021] On the three-dimensional grid for predicting the stratigraphic parameter field, a three-dimensional Sobel convolution kernel, composed of two-dimensional difference and one-dimensional Gaussian smoothing, is applied to perform discrete convolution filtering to calculate and output the three-dimensional spatial gradient vector of each grid point. The azimuth and apparent dip values ​​are read from the seismic attribute volume grid by grid through complex seismic trace analysis, and converted into unit structure normal vectors perpendicular to the strata using trigonometric function projection. At each grid point, the sine value of the angle between the spatial gradient vector and the unit normal vector is calculated using vector dot product operation, and its square is calculated. Then, the coherence coefficient or envelope amplitude is extracted from the seismic attributes and normalized to serve as the spatial weighting coefficient for that point. The squares of the sine values ​​at all grid points in the entire area are multiplied by their corresponding weighting coefficients, and all the product results are summed to obtain the geological structure gradient constraint term.

[0022] Using inverse distance weighted or kriging interpolation algorithms, the scattered prior logging data is interpolated and aggregated onto a three-dimensional coarse-resolution grid magnified with a fixed step size to generate a prior reference three-dimensional matrix. The prediction parameter matrix of the current model is then output at the same grid points. Using specified compactly supported orthogonal wavelet basis functions, a three-level progressive standard three-dimensional discrete wavelet transform operation is performed on these two three-dimensional matrices. After the transform, the first-level detail coefficient matrix containing high-frequency fluctuations is discarded, and only the third-level low-frequency approximation coefficient matrix representing the macroscopic structure, as well as the second and third-level mid-frequency structural detail coefficient matrices, are extracted. The absolute values ​​of the squared L2 norm errors of the two matrices on the extracted low-frequency and mid-frequency blocks are calculated respectively. The calculated mid-frequency error is multiplied by a preset adjustment coefficient and then directly added to the low-frequency error to output the total multi-scale wavelet coefficient fidelity term.

[0023] During each iteration of forward propagation, the predicted uncertainty variance value output by the model at each induction point is extracted, and the sum of the diagonal elements of the local covariance matrix associated with each induction point, which is the trace of the matrix, is calculated simultaneously. For induction points whose predicted variance is greater than a set variance threshold, the trace of its matrix is ​​subtracted from a preset lower bound constant. If the result is greater than zero, the result is squared and multiplied by a specified adjustment coefficient as a penalty value. For induction points whose predicted variance is less than or equal to the threshold, the trace of its matrix is ​​multiplied by another smaller suppression coefficient as a penalty value. After calculating the piecewise penalty values ​​for all individual induction points, all penalty values ​​in the entire field are arithmetically summed, and the sum obtained is the specific value of the uncertainty-driven covariance regularization term.

[0024] An uncertainty-driven covariance regularization term is constructed using the trace of the uncertainty variance and local covariance matrix predicted in the previous iterations. The four terms are then divided by their absolute values ​​from the initial iterations to perform dimensionless dimensionality reduction. By using hyperparameters as weighting coefficients, tensor addition is applied to linearly weight the four terms to form a variational inference objective function.

[0025] As one possible implementation, the step of setting up initial guidance points in the target work area and configuring a local covariance matrix for each guidance point includes:

[0026] The target work area is divided into three-dimensional grids. Based on the spatial sampling density of prior geological data and the variance distribution of stratigraphic parameters, the K-means clustering algorithm is used to generate the coordinates of initial induced points in three-dimensional space.

[0027] For each initial induction point, the spatial distribution covariance of the prior geological data in the neighborhood is calculated. The spatial distribution covariance is used as the initial local covariance matrix of the induction point, which serves as an initial approximate representation of the spatial geometric distribution and density of the prior data in the local area.

[0028] For the target work area, such as a three-dimensional space with a physical range of 10km×10km×2km, the three-dimensional mesh is divided according to the preset mesh resolution, such as a mesh step size of 50m in the x and y directions and a depth step size of 5m in the z direction.

[0029] Prior geological data, such as sonic transit time, resistivity logging curves, and spatial coordinates of 3D seismic impedance volumes, are extracted from each point within the grid. These coordinates are then combined with the sampling density weights of the prior data and the local variance distribution characteristics of formation parameters. For example, regions with local variance greater than 1.5 times the global variance are assigned significantly increased clustering weights. A weighted K-means clustering algorithm is then applied to extract the coordinates of the induced points. The number of cluster centers, K, is determined based on the total size of the observed data and a preset compression rate. For instance, when the training sample size is 100,000, K=2000 is set, and the coordinates of the 2000 cluster centers obtained after iterative convergence are the initial coordinates of the induced points. Figure 2 As shown in the figure, the spatial distribution characteristics of prior data points and initial induced points in the target work area are presented in cross-sectional form, and the spatial coverage and relative density of induced points can be observed.

[0030] For the k-th initial induction point, the local neighborhood is delineated by spatially truncating the search radius, for example, by setting a spherical search radius R=300m or selecting the N observation points with the closest spatial distance, preferably N=150.

[0031] Within this neighborhood, extract the three-dimensional coordinate sequence of all data points containing prior geological data. Calculate the three-dimensional empirical distribution covariance matrix of this set of spatial coordinates, i.e. ,in Let be the geometric centroid of the neighborhood coordinate distribution.

[0032] Add a diagonal regularization term ϵI to the main diagonal of the calculated initial covariance matrix, where I is the identity matrix and the parameter ϵ is fixed. This three-dimensional empirical covariance matrix characterizes the geometric distribution structure of local data clusters. It serves as the initial parameter for the non-stationary kernel function of the sparse Gaussian process, establishing the initial local receptive field search space degree for the induced point. For example, in continuous sedimentary sections, the horizontal correlation scale reaches 500m, while the vertical depth scale is only 10m. This provides the covariance matrix for the non-stationary kernel function to express the local spatial anisotropy during the drilling process.

[0033] As one possible implementation, the variational inference objective function for constructing the sparse Gaussian process model includes:

[0034] The variational inference objective function is set as maximizing the lower bound of the marginal likelihood and the negative penalty weighted sum of the regularization constraints after dimensionless dimensionality reduction transformation;

[0035] The lower bound term of the data evidence is used to fit the probability distribution between the observed data and the model predictions; the geological structure gradient constraint term is used to constrain the dip angle of the stratigraphic interface predicted by the model to be consistent with the dip angle extracted from the seismic data; the multi-scale wavelet coefficient fidelity term is used to maintain the spatial structure consistency between the predicted parameter field and the prior geological data at low and medium frequency scales; and the uncertainty-driven covariance regularization term is used to adjust the local receptive field of the kernel function.

[0036] The network structure of the sparse Gaussian process model consists of an input layer, a latent feature mapping layer, and an inference output layer. The input layer receives multi-dimensional coordinate vectors, the latent feature mapping layer performs spatial correlation mapping based on induced points containing local covariance matrices and variational distribution parameters through a kernel function, and the inference output layer generates target posterior parameters based on probability distribution calculations.

[0037] The input to the sparse Gaussian process model is the three-dimensional spatial coordinate vector of each data point or grid point in the target formation. The output of the sparse Gaussian process model is the posterior predicted mean of the predicted formation parameters at the corresponding coordinate points. And the variance of prediction uncertainty, which expresses the confidence level of prediction. The computational structure of the variational inference objective function is constructed as a maximization of a multinomial joint weighted function. Among them, the lower bound of data evidence. It includes the expected log-likelihood and the KL divergence penalty function between the variational posterior probability distribution and the prior distribution.

[0038] The penalty constraint term is mapped using a dimensionless dimensionality reduction transformation mechanism. Specifically, during the first round of forward evaluation of the objective function initialization, the initial absolute scalar benchmarks of each loss term are extracted. For example, the initial value of the structure gradient constraint term is... The wavelet fidelity term is In subsequent iterative optimizations, division operations are performed to normalize and scale each term by its initial absolute value or the exponential smoothing mean of the last ten rounds, thus mapping and controlling its operating range within the [0,1] interval.

[0039] When determining the linear weight allocation for each regularization constraint, a hyperparameter value space with clear geophysical boundary significance is adopted: weight parameters controlling the conformity of the geological structure. The preferred range is set to [0.5, 2.0], and the exemplary calculation can be fixed at this value. This ensures that the internal phase axes of the reconstructed prediction parameter field conform to the layer geometry constraints extracted from the 3D seismic data. The weight parameters responsible for macroscopic multi-scale low-frequency lattice matching... The preferred range is [0.1, 0.8], and an example value of 0.4 can be used to prevent unreasonable high-frequency mathematical oscillations in the far-well blank region where there are no logging point constraints. The covariance regularization weight parameter is used to adjust the local receptive field of model uncertainty. The preferred range is [0.01, 0.1], and the example value is 0.05.

[0040] After fusing the dimensionless scalars through the above hyperparameters, the analytical gradient is fed back to the model, and the Adam optimizer based on momentum decay is used. The initial learning rate is set to 0.01, and the model is iterated and updated at least 500 times in a single round. Thus, the joint inference of variational distribution parameters and local covariance matrix is ​​completed in a unified high-dimensional cost space.

[0041] As one possible implementation, the geological structure gradient constraint term includes:

[0042] The spatial gradient vector of the predicted stratigraphic parameter field is calculated using the three-dimensional Sobel operator, and the three-dimensional structure dip angle extracted from the seismic data is converted into a unit structure normal vector.

[0043] The weighted sum of the squares of the sine values ​​of the angles between the spatial gradient vector and the unit structure normal vector at each spatial grid point is calculated and used as a geological structure gradient constraint term. This term drives the spatial gradient vector of the predicted formation parameter field to approach and parallel the unit structure normal vector, making the gradient variation of formation parameters along the bedding plane approach zero.

[0044] On the uniform three-dimensional discrete evaluation grid where the predicted parameter field of the model output is located, three-dimensional Sobel edge detection convolution kernels with a size configuration of 3×3×3 are deployed along the three orthogonal directions of space x, y, and z respectively; for example, the three-dimensional Sobel operator along the horizontal x direction can be generated by the outer product tensor expansion of the standard two-dimensional Sobel difference operator [[1,0,-1],[2,0,-2],[1,0,-1]] and the one-dimensional Gaussian smoothing kernel [1,2,1] along the vertical z direction.

[0045] By applying this operator combination to the predicted three-dimensional continuous formation parameter field scalar matrix f(x, y, z) with zero-padding discrete convolution filtering, the approximate values ​​of the three partial derivatives at each spatial grid point are simultaneously output, thereby combining them to obtain the three-dimensional spatial gradient vector representing the direction of formation property variation. Simultaneously, complex seismic trace analysis is used to extract structural geometric parameters grid-by-grid from 3D seismic attribute data, such as the local seismic phase axis envelope or body dip coherence attributes, covering azimuth angles. and viewing angle Then, through a triangular projection transformation from a spherical coordinate system to a Cartesian coordinate system, a unit structural normal vector v=[sinϕcosθ,sinϕsinθ,cosϕ] is generated perpendicular to the strike plane of the local geological structure. T .

[0046] For M constraint evaluation mesh nodes within the entire target work area, the node size in engineering application examples is typically taken as... On the order of magnitude, using the dot product formula for vector space and the algebraic relationship of the 2-norm distance, the gradient vector of the above spatial parameters is calculated. With unit structure normal vector The square of the sine of the included angle at each specific grid point i. The specific discretization calculation formula is constructed as follows: Among them, the protection constant Set as Avoid floating-point overflow exceptions caused by dividing the denominator by zero during calculation.

[0047] In this operator, the seismic continuity envelope amplitude or local coherence coefficient value is used as a grid-by-grid spatially adaptive weighting coefficient. The range is smoothly normalized to [0,1], such as in the fractured zone where fractures and faults are developed. Weighting in sequence-complete continuous sedimentary regions By penalizing the squared term of the sinusoidal value to drive loss gradient descent, the spatial gradient vector is made to approach and parallel the structural normal vector. This allows for the penalty of the physical angle offset between the gradient normal of the predicted parameter field and the actual seismic reflection bedding plane in continuous stratigraphic profiles with high geological signal-to-noise ratio.

[0048] As one possible implementation, the multi-scale wavelet coefficient fidelity term includes:

[0049] Discrete prior geological data are interpolated and mapped to a three-dimensional downsampled coarse grid to obtain a prior reference parameter field;

[0050] A predicted formation parameter field is generated on the three-dimensional downsampling coarse grid. Three-dimensional discrete wavelet transform is applied to the predicted formation parameter field and the prior reference parameter field respectively to extract the wavelet decomposition coefficients at preset low-frequency and mid-frequency scales.

[0051] The L2 norm difference between wavelet decomposition coefficients at corresponding scales of the two parameter fields is calculated and used as a fidelity term for multi-scale wavelet coefficients to penalize the deviation of the model prediction results at the macroscopic geological structure scale.

[0052] By loading pre-configured geological anisotropy search ellipsoid parameters, and employing either a spatial inverse distance weighted algorithm or ordinary kriging interpolation from geostatistics, the irregularly distributed prior data is re-aggregated, sampled, and mapped onto a specific group of three-dimensional downsampled coarse-resolution grid nodes. Typically, the three-dimensional grid step size is set to be on the order of four times the high-resolution original output grid parameters, for example, downscaling from 50m×50m×5m to 200m×200m×20m. This process is used to construct a three-dimensional matrix of prior reference parameter fields with a stable macroscopic low-frequency structural base. .

[0053] Meanwhile, on this 3D downsampled coarse grid with a completely consistent topological spatial structure and the same coordinate node indices, the sparse Gaussian process inference model is invoked to calculate the expected mean of its current iteration state, thereby generating a predicted stratigraphic parameter field aligned with the aforementioned physical spatial scale registration. After these two data sets are prepared, a dedicated wavelet basis function with tight support and orthogonal smoothness is used, preferably the Daubechies wavelet family, with the db4 wavelet type assigned by default. and Two equal-dimensional three-dimensional data volumes are simultaneously executed using a standard three-dimensional discrete wavelet transform algorithm that includes three progressive decomposition levels.

[0054] After the wavelet multi-scale cascade decomposition process is completed, the high-frequency scale coefficient matrix representing local micro-fluctuations is truncated and discarded. For example, the highest frequency detail coefficient tensor of the first decomposition level is cleared, and the frequency band wavelet decomposition coefficient matrix group representing the macroscopic framework of the deep strata in the target work area is extracted: that is, the set of low-frequency approximate general profile coefficients containing the output of the third-level positive filter is extracted. And the set of intermediate frequency construction detail coefficients distributed in the bandpass outputs of layers 2 and 3. and .

[0055] Construct the formula for the multi-scale wavelet coefficient fidelity penalty loss term in algebraic residual form: The multi-scale control weighting coefficients used to adjust spectral band preferences The preset parameter setting is 0.5 to ensure that the weight of the macroscopic low-frequency main structure fitting is higher than that of the mid-frequency secondary fine-tuning structure. The third-layer low-frequency approximate profile coefficient matrix is ​​extracted from the stratigraphic parameter field predicted by the model after undergoing a three-layer three-dimensional discrete wavelet transform. The third-level low-frequency approximate profile coefficient matrix is ​​extracted from the prior reference parameter field generated by interpolation of known data after undergoing a three-level three-dimensional discrete wavelet transform.

[0056] By driving the optimizer to sum and minimize the difference in L2 norm error between the corresponding scale frequency band coefficient blocks of the two in this feature transformation domain, a macroscopic correction thrust can be generated in the macroscopic large-scale frequency band range of more than 400m, which punishes and reduces the frequency domain divergence and spatial distortion bias that occur in the variational inference model in cross-grid large-scale spatial geological sedimentary sequence prediction.

[0057] S2, based on the prior geological data, perform initial optimization of the variational inference objective function to obtain the initial formation parameter field; acquire new logging-while-drilling data and add the new logging-while-drilling data to the training set; determine whether to add induced points in the neighborhood of the new logging-while-drilling data based on the adaptive sampling criterion of model uncertainty; perform joint optimization of the variational inference objective function to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points.

[0058] The constructed variational inference objective function is initially optimized by maximizing the initial learning rate and iterating 500 times. The kernel function hyperparameters and variational distribution parameters of the sparse Gaussian process model are updated by backpropagation algorithm, and the predicted mean of the grid across the entire region is calculated as the initial stratigraphic parameter field.

[0059] The system receives the latest acoustic transit time and gamma ray logging curves, along with the corresponding 3D coordinates of the drill bit trajectory, from the logging-while-drilling (WWL) tool via a real-time data interface. This new WWL data is then appended to the existing training set tensor. The model prediction uncertainty variance at the coordinates of the new WWL data is calculated. An adaptive sampling criterion based on model uncertainty is used to set a variance threshold. If the calculated variance exceeds this threshold, the model is deemed to have insufficient expressive power for that region. The system then identifies the three closest existing induced points to the new WWL data. If the nearest distance is still greater than a set spatial distance threshold, a new induced point is added at the coordinates of the new WWL data, and its corresponding variational parameters and local covariance matrix are initialized.

[0060] The updated training set and the set of induced points are input into the model to jointly optimize the variational inference objective function. During backpropagation, the gradient of the loss function with respect to the newly added variational parameters and the lower triangular elements of the local covariance matrix of the affected induced points in the neighborhood of the drilling trajectory is calculated. The gradient descent method is used to update the above parameters and fit the new logging-while-drilling data.

[0061] As one possible implementation, the adaptive sampling criterion based on model uncertainty for determining whether to add an induction point in the neighborhood of new logging-while-drilling data includes:

[0062] When a new data point is input during logging, the uncertainty variance predicted by the sparse Gaussian process model at the location of the new data point during logging is calculated.

[0063] If the uncertainty variance is higher than the set first variance threshold, and the minimum Euclidean distance between the new logging-while-drilling data point and all existing induced points is greater than the set distance threshold, then it is determined that a new induced point will be added at the location of the new logging-while-drilling data point.

[0064] Otherwise, no additional induction points will be added; instead, new logging-while-drilling data points will be added to the training set to participate in subsequent updates of variational parameters.

[0065] As drilling progresses, downhole logging-while-drilling (LMWD) instruments continuously perform in-situ detection along the drill bit trajectory and upload new digital acquisition sequences in real time. For example, a data point set containing natural gamma logging values ​​and spatial rectangular coordinates is transmitted at a spatial frequency of 0.5m per drilling depth and is set as a new LMWD data point. The corresponding well logging observation true value is The inference module inputs the three-dimensional spatial coordinates into the baseline sparse Gaussian process computation map that has completed the previous round of model training and optimization. Based on covariance analysis, it extracts the diagonal scalar element of the predicted distribution matrix generated at that point. This value is the posterior prediction uncertainty variance value at that location. .

[0066] The adaptive topological incremental sampling operator based on confidence boundaries is invoked for logical verification and judgment. The first step is to extract the first variance evaluation threshold boundary. This boundary is preferably set to the 85th to 90th percentile of the previous full-field prediction variance spatial distribution array, sorted from smallest to largest. Alternatively, in manual mode, it can be set to a standardized variance constant value, such as... The vectorized Euclidean distance solution matrix kernel function is activated to calculate the spatial distance between the 3D coordinate array of the newly entered data points and the coordinate sequences of all K surviving induced points in the current network architecture set, and to find the global minimum Euclidean distance index. .

[0067] In the constraint condition logic judgment stage, if the following two judgment criteria are met: firstly, the prediction uncertainty variance at that coordinate is greater than the set first variance threshold, that is... This phenomenon indicates that the current local space is within the blind zone of the baseline model's predicted field of view; secondly, the calculated minimum Euclidean distance is greater than the preset spatial isolation distance threshold parameter. The engineering demonstration example sets the anti-overlap safety distance limit as follows: This characteristic indicates that the local topological space lacks nuclear radiation control from an existing induced core. At this point, executing the extension procedure instruction will change the spatial three-dimensional coordinates of the newly measured data point. As a newly added global induction node At the same time, the well logging distribution features within the nearest distance cutoff are adjusted to complete a new positive definite local covariance matrix assignment, allowing the model to fully induce the number of core nodes to complete the spatial scale auto-increment operation from K to K+1.

[0068] Conversely, if the above threshold conditions are not met simultaneously—for example, if the distance between the point and a certain existing induced point is only 30m, triggering the anti-overlap mechanism; or if the measured uncertainty variance is only 0.15—then no new induced point will be generated, and the point attribute pair set containing the true values ​​of drilling measurements will be generated according to the normal operating procedure. Add it to the observation point queue so that it can participate as a label target and guide the parameter fine-tuning update operation in the mini-batch parameter gradient iteration of the next mini-batch loop.

[0069] S3, wherein the uncertainty-driven covariance regularization term uses the uncertainty variance of the previous stage formation parameter field to constrain the local covariance matrix of the current stage; the optimized sparse Gaussian process model is used to predict the grid points of the target work area, and outputs a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

[0070] When calculating the uncertainty-driven covariance regularization term, the predicted uncertainty variance of the formation parameter field output from the previous iteration is extracted at each induced point. Based on this uncertainty variance, a piecewise penalty function is constructed to constrain the trace of the local covariance matrix. During backpropagation optimization, if the uncertainty variance of the region in the previous round is high, a lower bound constraint is triggered, penalizing the reduction of the trace of the local covariance matrix. The optimizer tends to increase the eigenvalues ​​of the local covariance matrix, prompting the local covariance matrix to expand the correlation scale in space to cover more sample points. If the region has low uncertainty, an upper bound suppression is triggered, penalizing the increase of the trace of the local covariance matrix, forcing the eigenvalues ​​of the local covariance matrix to decrease, shrinking the correlation scale to detect local details.

[0071] After optimization, a three-dimensional regular grid point coordinate array covering the entire target work area is generated. This grid point coordinate array is then input into the optimized sparse Gaussian process model in batches for forward propagation inference. The multivariate normal distribution object from the GPyTorch library is used to extract the prediction mean and prediction uncertainty variance for each grid point. The one-dimensional prediction results are then inversely transformed back into a three-dimensional spatial matrix shape and exported as a standard grid format file containing coordinates, mean, and variance data, thus outputting a three-dimensional stratigraphic parameter field.

[0072] As one possible implementation, the uncertainty-driven covariance regularization term includes:

[0073] Extract the uncertainty variance at each induced point position of the previous round of optimization output;

[0074] A penalty function based on the uncertainty variance is constructed. When the uncertainty variance is higher than the second variance threshold, the penalty function penalizes the decrease of the trace of the local covariance matrix, forcing the correlation scale to expand to smooth the parameter field. When the uncertainty variance is lower than or equal to the second variance threshold, the penalty function penalizes the increase of the trace of the local covariance matrix, forcing the correlation scale to shrink to fit the high-frequency local data changes.

[0075] Before each round of joint optimization iteration, forward propagation is performed to calculate the inferred state from the previous round of the sparse Gaussian process at all K induced points. Marginal prediction uncertainty variance sequence .

[0076] Based on this uncertainty, a continuously differentiable piecewise covariance regularization penalty function is established: the trace of the local covariance matrix associated with each induced point is extracted. The trace of this matrix maps the associated scale of the receptive field range of the corresponding local Gaussian kernel function. Simultaneously, a second variance control threshold reflecting the confidence level of data control is set; for example, this threshold is set based on the standardized empirical variance of the target logging-while-drilling parameters. .like Figure 3 As shown in the figure, the positive correlation between the variance of prediction uncertainty and the receptive field scale of the kernel function is illustrated, and the dashed line marks the boundary of the second variance threshold.

[0077] Based on the control logic of this piecewise composite penalty function, the cluster of induced points that fall into the control blind zone without data constraints or the high uncertainty zone is identified, i.e., when the judgment condition satisfies the uncertainty variance. When this is done, a lower bound constraint regularization term is added to the overall objective loss function. The penalty coefficient The value is 0.1, and the lower bound parameter is... Set as The reduction behavior of the local covariance matrix trace is penalized, thereby driving the three axial correlation extension scales of its corresponding kernel function to increase, thus expanding its local search range. For example, the lateral extension scale is expanded to a region of 500m, thereby enhancing the ability of this local region to borrow remote observation data and obtain a macroscopic low-frequency smooth and coherent distribution.

[0078] Conversely, for strong confidence regions induced by data sources within dense logging-while-drilling data coverage, i.e., those meeting the criteria... When adding a linear proportional suppression regularization penalty constraint term. The penalty coefficient Setting the value to 0.05 penalizes the increase in the trace of the covariance matrix at this point, causing the trace to decrease and reducing the correlation perception scale. For example, it compresses the kernel function search radius from 500m to within 50m, giving the neighborhood of this induced point a greater ability to attenuate spatial gradient partial derivatives, thus achieving high-resolution fitting of characteristic parameters of high-frequency local fluctuations such as thin sand-slurry interactions during drilling. The penalty terms calculated for all K induced points in the field are accumulated to form the covariance regularization term in the overall objective function, denoted as . .

[0079] As one possible implementation, the step of using an optimized sparse Gaussian process model to predict grid points in the target work area and outputting a three-dimensional formation parameter field with a prediction mean and a prediction uncertainty variance includes:

[0080] Extract the optimized coordinates of all induced points, the variational mean vector, the variational covariance matrix, and the local covariance matrix corresponding to each induced point;

[0081] For any grid point within the target work area, the cross-covariance vector between the grid point and all induced points is calculated using a non-stationary anisotropic kernel function;

[0082] Based on the cross-variance vector and variational parameters, the posterior predicted mean and posterior predicted variance of the stratigraphic parameters at the grid points are calculated, and the three-dimensional stratigraphic parameter field of the entire work area is generated by summarizing them.

[0083] After the variational inference process reaches the maximum iteration constraint, such as setting a global iteration of 1000 times, or after the objective loss function converges, the optimized set of three-dimensional coordinate matrices of the global K induced points is extracted. The K×1 variational distribution mean vector m, the K×K global variational covariance matrix S, and the 3×3 local covariance matrix assigned to each induced point. .

[0084] For any spatial grid point coordinate to be solved within the 3D target work area Assign a prior background covariance matrix to the predicted location. Let them be isotropic square matrices, such as , Given the set prior background signal variance constant, we simultaneously establish the dedicated matrices for each induction point. Substitute each value into the Gibbs nonstationary anisotropic kernel function calculation formula: , To control the prior signal variance parameter of the kernel function output amplitude, the distance probability scores between the grid point and K induced points are calculated sequentially, forming a cross-correlation characteristic row vector reflecting the weight of the surrounding non-uniform radiation. .

[0085] Based on the posterior analytical rule of the normal marginal probability Gaussian distribution, the generated cross-covariance eigenvectors are incorporated into the prediction inference framework: the inference formula is derived through the mean expectation equation. Solve for the posterior predicted mean of formation parameters at the target grid point location, such as acoustic transit time or resistivity. And simultaneously based on the covariance evaluation rule equation ,in The cross-covariance matrix among the internal induced point groups. To predict the static prior constant variance of grid point calibration, such as 1.0, calculate the variance representing the posterior prediction uncertainty of single-point extrapolation. .

[0086] Parallel computation architecture is used to perform parallel calculations on the grid distribution point group. The predicted mean obtained from each test point and the evaluated variance scalar are woven into a three-dimensional raster field with a geographic topology. After adding geodetic origin and grid spacing header description files, the control layer persistently dumps and generates three-dimensional solid model data that conforms to industry standards, such as the standard seismic data volume format SEG-Y or the grid VTK format, presenting a three-dimensional stratigraphic parameter data volume with an uncertain early warning layer.

[0087] This invention also proposes an incrementally dynamically updated multi-source data fusion stratigraphic parameter field construction system, comprising the following modules:

[0088] The module is used to acquire prior geological data including seismic and adjacent well logging data, set up initial induction points in the target work area, and configure a local covariance matrix for each induction point. The inverse of the local covariance matrix is ​​used to perform anisotropic linear transformation on the input coordinates during kernel function calculation. The variational inference objective function of the sparse Gaussian process model is constructed. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation.

[0089] The update module is used to perform initial optimization of the variational inference objective function based on the prior geological data to obtain an initial formation parameter field; acquire new logging-while-drilling data and add the new logging-while-drilling data to the training set; determine whether to add induced points in the neighborhood of the new logging-while-drilling data based on the adaptive sampling criterion of model uncertainty; and perform joint optimization of the variational inference objective function to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points.

[0090] The output module is used to constrain the local covariance matrix of the current round using the uncertainty-driven covariance regularization term of the previous round formation parameter field; and to predict the grid points of the target work area using the optimized sparse Gaussian process model, and output a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

[0091] As one possible implementation, the step of setting up initial guidance points in the target work area and configuring a local covariance matrix for each guidance point includes:

[0092] The target work area is divided into three-dimensional grids. Based on the spatial sampling density of prior geological data and the variance distribution of stratigraphic parameters, the K-means clustering algorithm is used to generate the coordinates of initial induced points in three-dimensional space.

[0093] For each initial induction point, the spatial distribution covariance of the prior geological data in the neighborhood is calculated. The spatial distribution covariance is used as the initial local covariance matrix of the induction point, which serves as an initial approximate representation of the spatial geometric distribution and density of the prior data in the local area.

[0094] As one possible implementation, the variational inference objective function for constructing the sparse Gaussian process model includes:

[0095] The variational inference objective function is set as maximizing the lower bound of the marginal likelihood and the negative penalty weighted sum of the regularization constraints after dimensionless dimensionality reduction transformation;

[0096] The lower bound term of the data evidence is used to fit the probability distribution between the observed data and the model predictions; the geological structure gradient constraint term is used to constrain the dip angle of the stratigraphic interface predicted by the model to be consistent with the dip angle extracted from the seismic data; the multi-scale wavelet coefficient fidelity term is used to maintain the spatial structure consistency between the predicted parameter field and the prior geological data at low and medium frequency scales; and the uncertainty-driven covariance regularization term is used to adjust the local receptive field of the kernel function.

[0097] As one possible implementation, the geological structure gradient constraint term includes:

[0098] The spatial gradient vector of the predicted stratigraphic parameter field is calculated using the three-dimensional Sobel operator, and the three-dimensional structure dip angle extracted from the seismic data is converted into a unit structure normal vector.

[0099] The weighted sum of the squares of the sine values ​​of the angles between the spatial gradient vector and the unit structure normal vector at each spatial grid point is calculated and used as a geological structure gradient constraint term. This term drives the spatial gradient vector of the predicted formation parameter field to approach and parallel the unit structure normal vector, making the gradient variation of formation parameters along the bedding plane approach zero.

[0100] As one possible implementation, the multi-scale wavelet coefficient fidelity term includes:

[0101] Discrete prior geological data are interpolated and mapped to a three-dimensional downsampled coarse grid to obtain a prior reference parameter field;

[0102] A predicted formation parameter field is generated on the three-dimensional downsampling coarse grid. Three-dimensional discrete wavelet transform is applied to the predicted formation parameter field and the prior reference parameter field respectively to extract the wavelet decomposition coefficients at preset low-frequency and mid-frequency scales.

[0103] The L2 norm difference between wavelet decomposition coefficients at corresponding scales of the two parameter fields is calculated and used as a fidelity term for multi-scale wavelet coefficients to penalize the deviation of the model prediction results at the macroscopic geological structure scale.

[0104] As one possible implementation, the adaptive sampling criterion based on model uncertainty for determining whether to add an induction point in the neighborhood of new logging-while-drilling data includes:

[0105] When a new data point is input during logging, the uncertainty variance predicted by the sparse Gaussian process model at the location of the new data point during logging is calculated.

[0106] If the uncertainty variance is higher than the set first variance threshold, and the minimum Euclidean distance between the new logging-while-drilling data point and all existing induced points is greater than the set distance threshold, then it is determined that a new induced point will be added at the location of the new logging-while-drilling data point.

[0107] Otherwise, no additional induction points will be added; instead, new logging-while-drilling data points will be added to the training set to participate in subsequent updates of variational parameters.

[0108] As one possible implementation, the uncertainty-driven covariance regularization term includes:

[0109] Extract the uncertainty variance at each induced point position of the previous round of optimization output;

[0110] A penalty function based on the uncertainty variance is constructed. When the uncertainty variance is higher than the second variance threshold, the penalty function penalizes the decrease of the trace of the local covariance matrix, forcing the correlation scale to expand to smooth the parameter field. When the uncertainty variance is lower than or equal to the second variance threshold, the penalty function penalizes the increase of the trace of the local covariance matrix, forcing the correlation scale to shrink to fit the high-frequency local data changes.

[0111] As one possible implementation, the step of using an optimized sparse Gaussian process model to predict grid points in the target work area and outputting a three-dimensional formation parameter field with a prediction mean and a prediction uncertainty variance includes:

[0112] Extract the optimized coordinates of all induced points, the variational mean vector, the variational covariance matrix, and the local covariance matrix corresponding to each induced point;

[0113] For any grid point within the target work area, the cross-covariance vector between the grid point and all induced points is calculated using a non-stationary anisotropic kernel function;

[0114] Based on the cross-variance vector and variational parameters, the posterior predicted mean and posterior predicted variance of the stratigraphic parameters at the grid points are calculated, and the three-dimensional stratigraphic parameter field of the entire work area is generated by summarizing them.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0116] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a stratigraphic parameter field by incrementally and dynamically updating multi-source data fusion, characterized in that, include: Prior geological data including seismic and adjacent well logging data are acquired. Initial induction points are set up in the target work area. A local covariance matrix is ​​configured for each induction point. The inverse of the local covariance matrix is ​​used to perform anisotropic linear transformation on the input coordinates during kernel function calculation. A variational inference objective function for a sparse Gaussian process model is constructed. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation. Based on the prior geological data, the variational inference objective function is initially optimized to obtain the initial formation parameter field; new logging-while-drilling data is acquired and added to the training set; based on the adaptive sampling criterion of model uncertainty, it is determined whether to add induced points in the neighborhood of the new logging-while-drilling data; the variational inference objective function is jointly optimized to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points; The uncertainty-driven covariance regularization term uses the uncertainty variance of the previous formation parameter field to constrain the local covariance matrix of the current round; the optimized sparse Gaussian process model is used to predict the grid points of the target work area, and outputs a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

2. The method according to claim 1, characterized in that, The step of setting up initial guidance points in the target work area and configuring a local covariance matrix for each guidance point includes: The target work area is divided into three-dimensional grids. Based on the spatial sampling density of prior geological data and the variance distribution of stratigraphic parameters, the K-means clustering algorithm is used to generate the coordinates of initial induced points in three-dimensional space. For each initial induction point, the spatial distribution covariance of the prior geological data in the neighborhood is calculated. The spatial distribution covariance is used as the initial local covariance matrix of the induction point, which serves as an initial approximate representation of the spatial geometric distribution and density of the prior data in the local area.

3. The method according to claim 1, characterized in that, The variational inference objective function for constructing the sparse Gaussian process model includes: The variational inference objective function is set as maximizing the lower bound of the marginal likelihood and the negative penalty weighted sum of the regularization constraints after dimensionless dimensionality reduction transformation; The lower bound term of the data evidence is used to fit the probability distribution between the observed data and the model predictions; the geological structure gradient constraint term is used to constrain the dip angle of the stratigraphic interface predicted by the model to be consistent with the dip angle extracted from the seismic data; the multi-scale wavelet coefficient fidelity term is used to maintain the spatial structure consistency between the predicted parameter field and the prior geological data at low and medium frequency scales; and the uncertainty-driven covariance regularization term is used to adjust the local receptive field of the kernel function.

4. The method according to claim 1, characterized in that, The geological structure gradient constraint term includes: The spatial gradient vector of the predicted formation parameter field is calculated using the three-dimensional Sobel operator, and the three-dimensional structure dip angle extracted from the seismic data is converted into a unit structure normal vector. The weighted sum of the squares of the sine values ​​of the angles between the spatial gradient vector and the unit structure normal vector at each spatial grid point is calculated and used as a geological structure gradient constraint term. This term drives the spatial gradient vector of the predicted formation parameter field to approach and parallel the unit structure normal vector, making the gradient variation of formation parameters along the bedding plane approach zero.

5. The method according to claim 1, characterized in that, The multi-scale wavelet coefficient fidelity term includes: Discrete prior geological data are interpolated and mapped to a three-dimensional downsampled coarse grid to obtain a prior reference parameter field; A predicted formation parameter field is generated on the three-dimensional downsampling coarse grid. Three-dimensional discrete wavelet transform is applied to the predicted formation parameter field and the prior reference parameter field respectively to extract the wavelet decomposition coefficients at preset low-frequency and mid-frequency scales. The L2 norm difference between wavelet decomposition coefficients at corresponding scales of the two parameter fields is calculated and used as a fidelity term for multi-scale wavelet coefficients to penalize the deviation of the model prediction results at the macroscopic geological structure scale.

6. The method according to claim 1, characterized in that, The adaptive sampling criterion based on model uncertainty for determining whether to add an induction point in the neighborhood of new logging-while-drilling data includes: When a new data point is input during logging, the uncertainty variance predicted by the sparse Gaussian process model at the location of the new data point during logging is calculated. If the uncertainty variance is higher than the set first variance threshold, and the minimum Euclidean distance between the new logging-while-drilling data point and all existing induced points is greater than the set distance threshold, then it is determined that a new induced point will be added at the location of the new logging-while-drilling data point. Otherwise, no additional induction points will be added; instead, new logging-while-drilling data points will be added to the training set to participate in subsequent updates of variational parameters.

7. The method according to claim 1, characterized in that, The uncertainty-driven covariance regularization term includes: Extract the uncertainty variance at each induced point position of the previous round of optimization output; A penalty function based on the uncertainty variance is constructed. When the uncertainty variance is higher than the second variance threshold, the penalty function penalizes the decrease of the trace of the local covariance matrix, forcing the correlation scale to expand to smooth the parameter field. When the uncertainty variance is lower than or equal to the second variance threshold, the penalty function penalizes the increase of the trace of the local covariance matrix, forcing the correlation scale to shrink to fit the high-frequency local data changes.

8. The method according to claim 1, characterized in that, The method of using an optimized sparse Gaussian process model to predict grid points in the target work area and outputting a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance includes: Extract the optimized coordinates of all induced points, the variational mean vector, the variational covariance matrix, and the local covariance matrix corresponding to each induced point; For any grid point within the target work area, the cross-covariance vector between the grid point and all induced points is calculated using a non-stationary anisotropic kernel function; Based on the cross-variance vector and variational parameters, the posterior predicted mean and posterior predicted variance of the stratigraphic parameters at the grid points are calculated, and the three-dimensional stratigraphic parameter field of the entire work area is generated by summarizing them.

9. A system for constructing a stratigraphic parameter field by incrementally and dynamically updating multi-source data fusion, characterized in that, include: The module is used to acquire prior geological data including seismic and adjacent well logging data, set up initial induction points in the target work area, and configure a local covariance matrix for each induction point. The inverse of the local covariance matrix is ​​used to perform anisotropic linear transformation on the input coordinates during kernel function calculation. The variational inference objective function of the sparse Gaussian process model is constructed. The variational inference objective function is composed of a data evidence lower bound term, a geological structure gradient constraint term, a multi-scale wavelet coefficient fidelity term, and an uncertainty-driven covariance regularization term, which are linearly weighted after dimensionless dimensionality reduction transformation. The update module is used to perform initial optimization of the variational inference objective function based on the prior geological data to obtain an initial formation parameter field; acquire new logging-while-drilling data and add the new logging-while-drilling data to the training set; determine whether to add induced points in the neighborhood of the new logging-while-drilling data based on the adaptive sampling criterion of model uncertainty; and perform joint optimization of the variational inference objective function to update the variational parameters of the sparse Gaussian process model and the local covariance matrix of the affected induced points. The output module is used to constrain the local covariance matrix of the current round using the uncertainty-driven covariance regularization term of the previous round formation parameter field; and to predict the grid points of the target work area using the optimized sparse Gaussian process model, and output a three-dimensional formation parameter field with prediction mean and prediction uncertainty variance.

10. The system according to claim 9, characterized in that, The step of setting up initial guidance points in the target work area and configuring a local covariance matrix for each guidance point includes: The target work area is divided into three-dimensional grids. Based on the spatial sampling density of prior geological data and the variance distribution of stratigraphic parameters, the K-means clustering algorithm is used to generate the coordinates of initial induced points in three-dimensional space. For each initial induction point, the spatial distribution covariance of the prior geological data in the neighborhood is calculated. The spatial distribution covariance is used as the initial local covariance matrix of the induction point, which serves as an initial approximate representation of the spatial geometric distribution and density of the prior data in the local area.

Images