Three-dimensional geostress field inversion method based on dynamic prior constraints

By employing a three-dimensional geostress field inversion method with dynamic prior constraints and utilizing neural networks and Markov chain Monte Carlo algorithms for global optimization, the problems of insufficient sample coverage and model distortion in the geostress field inversion of deep metal mines are solved, achieving high-precision and reliable inversion results.

CN123116873APending Publication Date: 2026-11-17NORTHEASTERN UNIV CHINA
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202611609966.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for inverting the geostress field in deep metal mines suffer from several problems, including insufficient coverage of high-dimensional parameter space samples, low accuracy of nonlinear mapping, susceptibility to local optima, inversion distortion caused by simplification of physical models, and a lack of scientific quantification and reliable assessment of the non-uniqueness of inversion solutions.

Method used

A three-dimensional geostress field inversion method based on dynamic prior constraints is adopted. By constructing a multi-input multi-output neural network nonlinear surrogate model, combined with multi-chain Markov chain Monte Carlo algorithm and gradient optimization technique, global optimization and uncertainty quantification are performed to ensure the reliability and accuracy of the inversion results.

Benefits of technology

It improves the robustness and reliability of inversion, enhances the ability to represent nonlinear features, and avoids the problems of physical model distortion and local optima caused by insufficient sample space coverage and limited model expressiveness in traditional methods, thus realizing the scientific quantification and reliable evaluation of inversion solutions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116873A_ABST
    Figure CN123116873A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of geostress inversion in deep metal mines, and relates to a three-dimensional geostress field inversion method based on dynamic prior constraints. The method includes: establishing a three-dimensional numerical model and determining the stress components at each measured point; defining the range of prior parameters; constructing an input-output forward modeling dataset; constructing a multi-input multi-output neural network nonlinear surrogate model; obtaining the pre-optimized maximum a posteriori probability point; determining the posterior distribution of the parameters and calculating the chain splitting value to ensure that each chain converges to the same posterior distribution region; when the chain splitting value meets the preset convergence requirements, determining the highest a posteriori density interval and continuous maximum a posteriori probability estimation points, and performing feedback adjustments according to preset verification rules until a preset termination condition is met; and determining the inversion parameters corresponding to the continuous maximum a posteriori probability estimation points that meet the preset termination condition as the optimal inversion parameters. Its beneficial effect is that it ensures that the obtained optimal solution is the globally true optimal solution rather than a pseudo-solution caused by boundary constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of geostress inversion technology for deep metal mines, specifically to a three-dimensional geostress field inversion method based on dynamic prior constraints. Background Technology

[0002] In deep metal mining, the initial geostress field is a key factor in engineering stability analysis, support design, and optimization of stope structural parameters. Currently, due to limited measured data, numerical simulations or intelligent calculations are typically used to invert the geostress field distribution through forward modeling.

[0003] Chinese patents with publication numbers CN120354762A, CN120832833A, CN119885683A, and CN122065225A all involve inverting spatial geostress distribution characteristics based on discrete geostress measurement data, aiming to overcome the shortcomings of traditional linear superposition method and multiple linear regression method in inverting results with large errors under complex geological conditions.

[0004] Chinese patent CN120354762A proposes an intelligent analysis system and method for inverting the three-dimensional geostress field of layered surrounding rock. This patent focuses on optimizing sample generation through sensitivity analysis, nonlinear superposition models, and orthogonal experimental design, and improving fitting performance through self-organizing learning. However, orthogonal experimental design often struggles to comprehensively analyze all high-order interactions when dealing with complex multi-factor problems, leading to distortion in the trained model in sparse sample regions. Although the scheme introduces a closed-loop iterative mechanism to dynamically correct model weights, the model training process does not involve incremental learning; it only iteratively adjusts the weights under preset samples. This limits the model's ability to extrapolate to conditions not covered by the samples, thus restricting its global optimization capabilities.

[0005] Chinese patent CN120832833A discloses a method and apparatus for inverting nonlinear geostress fields based on physical data neural networks. This patent focuses on using physical data neural networks for optimization, effectively characterizing the nonlinear features of the geostress field by constructing a hybrid loss term that integrates physical constraints and data-driven approaches. However, this method also relies on orthogonal experimental design for training sample generation. While this discrete sampling method has balance, it struggles to achieve dense sample coverage in high-dimensional parameter spaces. Furthermore, its optimization mechanism employs an error-triggered sample compensation strategy, where new samples are added to the dataset for retraining when the inversion error does not reach a threshold. While this strategy ensures physical consistency of the data, it is essentially a passive iteration, resulting in high computational costs and a risk of getting trapped in local optima due to the locality of the initial sample space.

[0006] Chinese patent CN119885683A discloses a nonlinear geostress field inversion method, apparatus, equipment, and medium. This patent focuses on constructing nonlinear boundary conditions using a nonlinear variation coefficient matrix. While it has the advantages of low data requirements and fast solution speed, its expressive power is limited compared to neural network methods, making it difficult to capture the complex nonlinear relationships in the geostress field distribution.

[0007] Patent CN122065225A proposes a method for inverting and analyzing geostress fields by integrating physical information neural networks. This patent focuses on introducing equilibrium differential equations and compatibility equations as physical constraints, simplifying the three-dimensional geostress field problem into a two-dimensional planar model, aiming to reduce computational costs while maintaining mechanical equilibrium criteria and geometric continuity. However, this dual simplification of physical constraints and spatial dimension has significant limitations. Specifically, while neglecting the constitutive relations of rock mass (i.e., the physical laws between stress and strain), this method also fails to consider the disturbance effect of geological structures on the geostress field, making it difficult to obtain reliable inversion results when facing complex geological environments with heterogeneous, nonlinear, or strong three-dimensional effects.

[0008] Therefore, it is urgent to study a three-dimensional geostress field inversion method to solve the technical problems of physical model distortion, limited high-dimensional nonlinear feature representation, high cost of passive iterative calculation, and easy getting trapped in local optima in the existing technology. Summary of the Invention

[0009] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a three-dimensional geostress field inversion method based on dynamic prior constraints, which aims to solve the problems of insufficient high-dimensional parameter space sample coverage, low nonlinear mapping accuracy, easy trapping in local optima, inversion distortion caused by simplification of physical models, and lack of scientific quantification and reliable evaluation of the non-uniqueness of inversion solutions in the prior art.

[0010] Technical solution To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0011] This invention provides a three-dimensional geostress field inversion method based on dynamic prior constraints, comprising the following steps:

[0012] Step 1: Establish a three-dimensional numerical model based on geological exploration data, stratigraphic lithology, and measured stress point coordinates, and determine the stress components at each measured point;

[0013] Step 2: Apply linearly varying boundary conditions to the three-dimensional numerical model sequentially, perform univariate forward modeling, establish an input-output dataset, and define the range of prior parameters based on measured geostress components;

[0014] Step 3: Sample multiple initial samples within the range of prior parameters, perform forward modeling calculations, and construct the input-output forward modeling dataset;

[0015] Step 4: Construct a multi-input, multi-output neural network nonlinear surrogate model;

[0016] Step 5: Construct the normalized magnitude likelihood, sample the candidate set within the range of prior parameters, take the points with the smallest negative log-likelihood values ​​as the initial points, and approach the global optimum through multi-starting point gradient optimization to obtain the pre-optimized maximum posterior probability point;

[0017] Step 6: Using the pre-optimized maximum posterior probability point as the initial state, perform posterior sampling using the multi-chain Markov chain Monte Carlo algorithm, traverse the parameter space of the input-output forward modeling dataset, determine the posterior distribution of the parameters, and calculate the splitting of each chain. The value ensures that all chains converge to the same posterior distribution region;

[0018] Step 7: When the chains split When the value meets the preset convergence requirements, the highest posterior density interval and the continuous maximum posterior probability estimation points are determined, and feedback adjustments are made according to the preset test rules until the preset termination conditions are met.

[0019] Step 8: Determine the inversion parameters corresponding to the consecutive maximum posterior probability estimation points that satisfy the preset termination condition as the optimal inversion parameters.

[0020] As a further improvement to the method of the present invention, step 1 specifically includes:

[0021] Based on the geological report, the area covering all the impact range of the project was delineated as the area to be inverted;

[0022] Based on the engineering geological data and stratigraphic lithology data of the area to be inverted, the spatial distribution of rock groups and fault structures is clarified.

[0023] Based on the spatial distribution of rock groups and fault structures, a three-dimensional numerical model reflecting the geological structural characteristics of the area to be inverted is constructed, and the principal stress values ​​of each measured point are collected and expressed as stress components.

[0024] As a further improvement to the method of the present invention, step 2 includes the following steps:

[0025] Step 2.1: Using the controlled variable method, perform univariate gradient forward modeling on the three-dimensional numerical model and collect the calculation results to construct an input-output dataset;

[0026] Step 2.2: Based on the constructed input-output dataset, use multi-output linear fitting mapping and evaluate parameter sensitivity to construct a sensitivity sequence to measure the strength of the effect of each input parameter on different stress components;

[0027] Step 2.3: Under small-amplitude regularization, based on the sensitivity sequence, and combining sequential optimization and weighted least squares method, solve for the combination of input boundary conditions that satisfy the conditions, and obtain the initial boundary parameter vector;

[0028] Step 2.4: Using the initial boundary parameter vector as the center, expand along each parameter dimension to define the range of prior parameters.

[0029] As a further improvement to the method of the present invention, step 2.2 specifically includes:

[0030] Based on the input-output dataset, for the i-th input parameter, least squares fitting is performed on the stress components at n measurement points to obtain multiple linear fitting equations, which serve as a single-parameter linear model:

[0031] ;

[0032] In the formula, y i For the i-th input parameter x i The generated stress component, a i b i These are the slope vector and the intercept vector, respectively;

[0033] The least squares fitting process is repeated for all N input parameters to obtain N sets of single-parameter linear models. Based on the linear superposition assumption of elasticity, the fittings for each stress component are superimposed to construct an overall linear approximation model.

[0034] ;

[0035] In the formula, BS is the stress component vector, and BS is the boundary condition vector. The superposition of intercept terms of each linear model ;

[0036] Using the superposition of the intercept terms of each linear model as the baseline, the deviation between the target stress and the baseline is calculated, and the target required intensity vector Y for each stress component is constructed. * :

[0037] ;

[0038] In the formula, Y is the actual target stress vector;

[0039] Euclidean norm aggregation is performed on the slopes of the same physical quantity in response at different measurement points to construct a sensitivity sequence. :

[0040] ;

[0041] In the formula, n is the total number of measurement points.

[0042] As a further improvement to the method of the present invention, step 3 specifically includes:

[0043] Within the range of the prior parameters, a quasi-Monte Carlo sampling method is used to extract combinations of boundary conditions as initial samples;

[0044] The boundary conditions are combined to perform batch multi-threaded forward modeling calculations, and the calculation results are summarized to obtain the input-output forward modeling dataset.

[0045] As a further improvement to the method of the present invention, step 4 specifically includes:

[0046] A hierarchical sampling strategy is adopted to partition the input-output forward modeling dataset, and a neural network model with residual attention mechanism and multilayer perceptron is integrated to construct a multi-input multi-output neural network nonlinear surrogate model.

[0047] The input-output forward dataset is clustered into multiple clusters using the K-Means algorithm, and the training set and test set are independently divided in each cluster according to a preset ratio to achieve balanced sampling in each parameter range and ensure that the training set and test set have the same coverage in the parameter space.

[0048] The input boundary condition parameters are mapped into a high-dimensional vector sequence. A multi-layer multi-head self-attention mechanism is adopted to achieve global information capture and feature extraction by calculating the correlation between query vector, key vector, and value vector. The coupling relationship between parameters is automatically learned to obtain the encoded global features.

[0049] A multilayer perceptron is used to perform nonlinear regression on the global features encoded by the self-attention mechanism to establish a highly complex nonlinear fitting relationship between the input and output.

[0050] As a further improvement to the method of the present invention, step 5 specifically includes:

[0051] Step 5.1: Construct the normalized magnitude likelihood of the predicted variance by introducing a nonlinear surrogate model, normalize the residual by dividing it by the physical scale of the corresponding stress component, and assign the normal stress component the same penalty weight as the shear stress component.

[0052] Step 5.2: Generate a candidate parameter set using hypercubic Latin sampling within the prior parameter range, and use a nonlinear surrogate model to quickly evaluate the negative log-likelihood of each candidate parameter point, and initially screen out high-likelihood candidate points.

[0053] Step 5.3: Perform the L-BFGS-B quasi-Newton method with boundary constraints on the smooth posterior high probability region, so that the high likelihood candidate points can quickly converge to the local mode of the posterior distribution in the pre-optimization stage, and obtain the pre-optimized maximum posterior probability point.

[0054] As a further improvement to the method of the present invention, step 6 specifically includes:

[0055] Posterior sampling phase: Starting from near the pre-optimized maximum posterior probability point, the multi-chain Markov chain Monte Carlo algorithm is used to initiate parallel sampling using the adaptive random walk Metropolis-Hastings algorithm; the proposal step size of each chain is adaptively adjusted through the Robbins-Monro random approximation strategy. When the acceptance rate is higher than the target value, the step size is automatically increased, and when it is lower than the target value, the step size is decreased, so that the acceptance rate is maintained near the optimal acceptance rate and is subject to independent pruning constraints of the prior range of each parameter.

[0056] Convergence diagnosis phase: Gelman-Rubin based on split chains is used. Diagnostic methods are used based on statistical measures, when all parameters are split. When all values ​​approach 1, it indicates that the sampling of the multi-chain Markov chain Monte Carlo algorithm has converged to the same probability space; otherwise, there is a non-convergence phenomenon, and the current sampling logic needs to be corrected.

[0057] As a further improvement to the method of the present invention, the splitting The calculation of the value includes:

[0058] Divide the m chains of length T equally to form 2m chains of length T. For a semi-chain, define the average intra-chain variance W and the inter-chain variance B as:

[0059] ;

[0060] ;

[0061] In the formula, Let c be the value at step t of the c-th chain; Let be the average value of all t values ​​in the c-th chain; For all The average value;

[0062] Mixed estimators of posterior variance for:

[0063] ;

[0064] pass Calculate splitting with average intra-chain variance W The value is:

[0065] .

[0066] As a further improvement to the method of the present invention, step 7 specifically includes:

[0067] Step 7.1: Define the Bayesian confidence fluctuation range of the inversion parameters using the highest posterior density interval;

[0068] Step 7.2: Select the maximum a posteriori probability estimate point MCMC-MAP from the posterior distribution as the discrete approximate solution;

[0069] Step 7.3: Using the MCMC-MAP point as the initial iteration point, the gradient-driven local refinement is performed using the L-BFGS-B quasi-Newton method to obtain the continuous maximum a posteriori probability estimation point;

[0070] Step 7.4: Examine the spatial relationship between the consecutive maximum a posteriori probability estimation points and the highest a posteriori density interval. If the consecutive maximum a posteriori probability estimation points are located inside the highest a posteriori density interval, it indicates that the optimal solution is located in the high-density core region of the posterior distribution, and proceed to the next step of the examination; if the consecutive maximum a posteriori probability estimation points deviate from the highest a posteriori density interval, it indicates that the consecutive maximum a posteriori probability estimation points are local extrema or that the Markov chain Monte Carlo algorithm has not sampled sufficiently, and it is necessary to increase the sampling number of the Markov chain Monte Carlo algorithm to improve the spatial coverage, and return to step 6 to recalculate;

[0071] Step 7.5: Check whether the consecutive maximum a posteriori probability estimation points are close to the prior boundary. If they are within the range of prior parameters, it indicates that the parameters corresponding to the consecutive maximum a posteriori probability estimation points are optimal, and the test ends. If they are close to the boundary of the prior parameter range, calculate the negative log-likelihood gradient and proceed to the next test.

[0072] Step 7.6: Check whether the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the boundary. If the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the prior parameter range, the check is terminated. If the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the prior parameter range, the prior boundary parameter range needs to be expanded, and the calculation should be repeated in step 3.

[0073] The beneficial effects of this invention are:

[0074] Inversion robustness and reliability assurance: The magnitude difference is eliminated by designing a normalized amplitude likelihood function, and the judgment is made based on the spatial relative position of the refined MAP point with respect to the prior parameters and the direction of the negative log-likelihood gradient. When the obtained solution does not meet the verification requirements, the search range is expanded within the original prior parameter range by a preset expansion coefficient, and samples are supplemented and the nonlinear surrogate model is reconstructed. The process is iterated until the refined MAP point stably converges within the HPD interval corresponding to the preset confidence level, ensuring that the obtained optimal solution is the global true optimum rather than a pseudo solution caused by boundary constraints.

[0075] Enhanced Nonlinear Characterization Capability: By utilizing numerical calculation software incorporating rock mass constitutive relations, forward modeling is driven by quasi-Monte Carlo sampling to generate a physically constrained sample set. A multi-input multi-output (MIMO) nonlinear surrogate model is trained using a hierarchical sampling cross-validation strategy. This achieves synergy in data quality, validation reliability, and model expressive power, overcoming the limitations of traditional methods in representing high-dimensional nonlinear features and physical model distortion caused by insufficient sample space coverage and limited model expressiveness.

[0076] Global optimization and uncertainty quantification: A hybrid inference strategy combining determinism and stochasticity (L-BFGS-B deterministic pre-optimization MAP + MCMC posterior global exploration) is adopted to avoid the problem of traditional passive iterative mechanisms getting stuck in local optima; at the same time, HPD intervals are used to achieve scientific quantification of the non-uniqueness of the inversion solution and definition of the credible fluctuation range.

[0077] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the substantive content of the present invention. Attached Figure Description

[0078] Figure 1 This is a schematic flowchart illustrating a three-dimensional geostress field inversion method based on dynamic prior constraints, according to an exemplary embodiment.

[0079] Figure 2 This is a schematic diagram of a nonlinear proxy model structure according to an exemplary embodiment;

[0080] Figure 3 This is a schematic diagram of layered sampling according to an exemplary embodiment;

[0081] Figure 4 This is a schematic diagram illustrating the training process of a nonlinear proxy model according to an exemplary embodiment. Detailed Implementation

[0082] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0083] Firstly, such as Figure 1 As shown, this embodiment of the invention provides a three-dimensional geostress field inversion method based on dynamic prior constraints, including the following steps:

[0084] Step 1: Establish a three-dimensional numerical model based on geological exploration data, stratigraphic lithology, and measured stress point coordinates, and determine the stress components of each measured point.

[0085] Specifically, based on the geological report, the area covering all engineering impact zones is delineated as the area to be inverted; based on the engineering geological data and stratigraphic lithology data of the area to be inverted, the spatial distribution of rock groups and fault structures is clarified; a three-dimensional numerical model accurately reflecting the geological structural characteristics of the area to be inverted is constructed using modeling software, and the principal stress values ​​(magnitude, azimuth, dip) of each measured point are collected in terms of 6 stress components (σ). xx σ yy σ zz τ xy τ xz τ yz )express.

[0086] Step 2: Apply linearly varying boundary conditions to the three-dimensional numerical model sequentially, perform univariate forward modeling, establish an input-output dataset, and define the range of prior parameters based on measured geostress components.

[0087] To address the weak extrapolation capability of neural networks, it is necessary to define the range of prior parameters to construct a numerical surrogate model with strong interpolation capabilities around the target stress. Under this premise, a robust range of prior parameters is constructed by utilizing limited forward modeling data combined with least squares output fitting and sensitivity evaluation.

[0088] The prior parameter range refers to the range of input parameters that may have solutions, required for forward modeling calculations in numerical computation software. Its determination process includes:

[0089] Step 2.1: Using the controlled variable method, perform univariate gradient forward modeling on the three-dimensional numerical model and collect the calculation results to construct an input-output dataset.

[0090] Specifically, the controlled variable method is used to sequentially apply linearly changing boundary conditions to the three-dimensional numerical model (changing only one input parameter each time while keeping the others constant), and univariate gradient forward modeling is performed using the numerical calculation software FLAC3D to collect forward modeling data and construct an input-output dataset.

[0091] Step 2.2: Based on the constructed input-output dataset, a multi-output linear fitting mapping is used to evaluate parameter sensitivity, and a sensitivity sequence is constructed to measure the strength of the effect of each input parameter on different stress components, including:

[0092] Based on the input-output dataset, for the i-th input parameter, least squares fitting is performed on the six stress components at n measurement points to obtain a total of n×6 linear fitting equations, which serve as a single-parameter linear model:

[0093] ;

[0094] In the formula, y i For the i-th input parameter x i The generated stress component, a i b i These are the slope vector and the intercept vector, respectively.

[0095] Repeat the above least-squares fitting process for all N input parameters to obtain N sets of single-parameter linear models. Based on the linear superposition assumption of elasticity, the fittings for each stress component are superimposed to construct an overall linear approximation model:

[0096] ;

[0097] In the formula, BS is the stress component vector, and BS is the boundary condition vector. The superposition of intercept terms of each linear model .

[0098] Using the superposition of the intercept terms of each linear model as the baseline, the deviation between the target stress and the baseline is calculated, and the target required intensity vector Y for each stress component is constructed. * :

[0099] ;

[0100] In the formula, Y is the actual target stress vector.

[0101] To quantitatively assess the influence of each input parameter on different stress components, a sensitivity sequence was constructed by Euclidean norm aggregation of the slopes of the same physical quantity in the responses at different measurement points. This is used to measure the intensity of the effect of each input parameter on different stress components.

[0102] ;

[0103] In the formula, n is the total number of measurement points.

[0104] Step 2.3: Under small-amplitude regularization, based on the sensitivity sequence, and combining sequential optimization and weighted least squares method, solve for the combination of input boundary conditions that satisfy the conditions, and obtain the initial boundary parameter vector.

[0105] Step 2.4: Using the initial boundary parameter vector as the center, expand it in each parameter dimension to define the range of prior condition parameters.

[0106] Since the overall linear model is constructed based on the linear superposition assumption of elasticity and ignores the coupling effect between parameters, the obtained solution may deviate from the true optimal solution and can only be used as an initial estimate. Therefore, based on this, a priori parameter range containing the potential optimal solution is constructed by expanding the initial boundary parameter vector along a certain range in each parameter dimension, providing a reasonable search space for subsequent global iterative optimization.

[0107] Step 3: Sample multiple initial samples within the range of prior parameters and perform forward modeling to construct the input-output forward modeling dataset.

[0108] Specifically, to ensure the uniform distribution of boundary condition combinations and to fully cover the boundary parameter range space, a certain number of boundary condition combinations are extracted as initial samples within the prior parameter range using a quasi-Monte Carlo sampling method (such as Sobol sequence). The sampled boundary condition combinations are then substituted into numerical calculation software for batch multi-threaded forward modeling calculations, and the calculation results are summarized to obtain the input-output forward modeling dataset.

[0109] Step 4: Construct a nonlinear surrogate model with multiple inputs and multiple outputs in a neural network.

[0110] Specifically, a hierarchical sampling strategy is employed to partition the input-output forward modeling dataset, and a residual attention mechanism and a multilayer perceptron neural network model are integrated to construct a multiple-input multiple-output (MIMO) neural network nonlinear surrogate model. This nonlinear surrogate model is used to replace the forward modeling process in numerical computation software, enabling rapid prediction of possible output results for input parameters.

[0111] Specifically, the input-output forward modeling dataset is clustered into multiple clusters using the K-Means algorithm, and the training set and test set are independently divided within each cluster according to a preset ratio to achieve balanced sampling in each parameter range, ensuring that the training set and test set have the same coverage in the parameter space and eliminating the problem of inconsistent parameter distribution caused by random partitioning.

[0112] The input boundary condition parameters are mapped to a high-dimensional vector sequence. A multi-layer, multi-head self-attention mechanism is employed to capture global information and extract features by calculating the correlation between the query vector Q, key vector K, and value vector V. The coupling relationship between parameters is automatically learned to obtain the encoded global features.

[0113] ;

[0114] In the formula, Q is the query vector; K is the key vector; V is the value vector; d k is the scaling factor; softmax(·) is the normalized exponential function.

[0115] Subsequently, a multilayer perceptron (MLP) was used to perform nonlinear regression on the global features encoded by the self-attention mechanism to establish a highly complex nonlinear fitting relationship between the input and output.

[0116] The constructed nonlinear surrogate model is trained using mini-batch training and constrained by the mean squared error (MSE) loss function. Gradient clipping and an adaptive learning rate decay strategy based on validation loss are combined to ensure model training stability. Furthermore, an early stopping mechanism is introduced to prevent overfitting (training is terminated early when the validation set performance shows no improvement within a few iterations).

[0117] In this process, the model training data propagation in each round of the nonlinear proxy model is as follows: Figure 2 As shown, the specific data flow is as follows:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] In the formula, x represents the actual input data; y represents the model's true label; N p N represents the dimension of the input data. pts The number of geometric sampling points in space; E is the matrix after high-dimensional mapping of the input data x; W e For the embedding weight matrix; b e H is the bias vector of the embedding layer; l X is the output of the attention sublayer of layer l; lis the final output of the l-th residual self-attention block (containing self-attention sub-layers and feedforward network sub-layers); z is the output of fusing all residual attention blocks X. L The resulting global feature vector; For the predicted result; W c b c W represents the parameter of the feature compression layer in an MLP. o b o These are the weights and biases for the output layer.

[0125] The specific calculation method for the loss function is as follows:

[0126] ;

[0127] In the formula, These are the model's predicted values; This represents the actual stress value.

[0128] Based on the above architecture, a multi-input multi-output neural network nonlinear proxy model was constructed. This model first maps the input design parameters into high-dimensional embeddings and establishes the coupling relationship between the input parameters through a residual self-attention module; then, the encoded features are aggregated into a global feature vector and input into the multilayer perceptron regression head to achieve a nonlinear mapping from boundary conditions to six stress components.

[0129] In this embodiment, as Figure 3 The stratified sampling diagram shown illustrates the high overlap in probability density distributions between the training and test sets in the standardized input space, indicating that they share consistent statistical characteristics. This demonstrates that the partitioning is reasonable and the test set is sufficiently representative, providing a reliable basis for evaluating the model's generalization performance. Figure 4 The diagram shown illustrates the training process of the nonlinear surrogate model. The training loss and test loss decrease synchronously with each training batch and converge, with small intervals, indicating no significant overfitting. The coefficient of determination R0 2 The value eventually stabilized around 0.99, indicating that the surrogate model has high fitting accuracy and strong generalization ability, and can effectively replace the high-fidelity model for rapid prediction.

[0130] Step 5: Construct the normalized magnitude likelihood. Sample the candidate set within the range of prior parameters. Take the points with the smallest negative log-likelihood values ​​as the initial points. Approach the global optimum through multi-starting-point gradient optimization to obtain the pre-optimized maximum a posteriori probability point (for ease of description, the pre-optimized maximum a posteriori probability point is simply referred to as the pre-optimized MAP point, and this abbreviation will be used in the following text).

[0131] The established nonlinear surrogate model is used to replace the time-consuming physical forward modeling. By performing Latin hypercube sampling (LHS) within the prior parameter range and combining gradient optimization to screen the maximum a posteriori (MAP) points, the initial state is provided for the subsequent use of the Markov chain Monte Carlo random sampling algorithm.

[0132] Step 5.1: Construct the normalized magnitude likelihood of the predicted variance by introducing a nonlinear surrogate model, normalize the residual by dividing it by the physical scale of the corresponding stress component, and assign the normal stress component the same penalty weight as the shear stress component.

[0133] Specifically, because the stress tensor in measured geostress data exhibits a clear directionality, and the magnitudes of the normal stress and shear stress components differ significantly, directly summing the residuals using traditional least squares or Gaussian likelihood methods would lead to the larger magnitude of normal stress dominating the optimization process, thus masking shear stress information. Therefore, to eliminate the influence of magnitude differences and ensure that the likelihood function objectively reflects physical laws, a normalized amplitude likelihood was constructed. This normalized likelihood normalization normalizes the residuals by dividing them by the physical scale of the corresponding stress components, assigning equivalent penalty weights to the normal stress and shear stress components, thereby effectively eliminating the impact of magnitude differences on the optimization process. Furthermore, the total variance is introduced by fusing relative noise from instrument measurements, relative baseline error from nonlinear surrogate model predictions, and absolute variance from the nonlinear surrogate model estimated by MC Dropout, allowing the likelihood function to automatically relax constraints in regions with large model prediction variance or high observation noise.

[0134] Among them, the normalized magnitude log-likelihood The formula is shown below:

[0135] ;

[0136] ;

[0137] ;

[0138] In the formula, p=1,2…,n are indexed measurement points, and k=1,…,6 are indexed stress components; To normalize the residuals, These are the predicted values ​​from the nonlinear surrogate model. For the systematic bias of the physical model, These are measured values. The physical scale for the k-th stress component; To normalize the total variance, The relative error of the instrument measurement; represents the relative baseline error of the nonlinear surrogate model; VarMC represents the absolute variance of the nonlinear surrogate model estimated by MC Dropout.

[0139] Step 5.2: Generate a candidate parameter set using hypercubic Latin sampling within the prior parameter range, and use a nonlinear surrogate model to quickly evaluate the negative log-likelihood of each candidate point, and initially screen out high-likelihood candidate points.

[0140] Because the Markov chain Monte Carlo algorithm has extremely low convergence efficiency when randomly walking in an unbounded prior space, it is difficult to obtain reliable posterior estimates with limited computational resources. Therefore, a pre-optimization stage is used to locate the point with the maximum posterior probability, and this point is then used as the starting point for subsequent global random exploration to explore the posterior distribution.

[0141] Specifically, in the pre-optimization stage, a candidate parameter set is generated using hypercubic Latin sampling within the range of prior parameters; and the negative log-likelihood of each candidate point is quickly evaluated using a deterministic prediction model (with the nonlinear surrogate model prediction variance set to zero), thus initially screening out high-likelihood candidate points.

[0142] Step 5.3: Perform the L-BFGS-B quasi-Newton method with boundary constraints on the smooth posterior high probability region, so that the high likelihood candidate points can quickly converge to the local mode of the posterior distribution in the pre-optimization stage, and obtain the pre-optimized MAP points, which provide the initial state in the high likelihood density region for the subsequent Markov chain Monte Carlo algorithm multi-chain sampling.

[0143] ;

[0144] In the formula, The pre-optimized MAP points are obtained from the pre-optimization stage; D is the range of prior parameters. Here are the boundary condition parameters to be inverted, and N is the number of inversion parameters; This represents the boundary condition that makes the function reach its minimum value; For the likelihood function under the deterministic prediction mode, For observation data vectors; represents the prior distribution of the inversion parameters.

[0145] Step 6: Using the pre-optimized MAP points as the initial state, perform posterior sampling using the multi-chain Markov chain Monte Carlo algorithm, traverse the parameter space of the input-output forward modeling dataset, determine the posterior distribution of the parameters, and calculate the splitting of each chain. The value ensures that all chains converge to the same posterior distribution region. Here, the posterior distribution refers to the sum of the probabilities of all possible combinations of input parameters occurring within the allowable range of the input parameters.

[0146] In the posterior sampling phase: Multiple Markov Monte Carlo chains are initiated from near the pre-optimized MAP point, and the Metropolis-Hastings (RWMH) algorithm is used for parallel sampling. The proposal step size of each chain is adaptively adjusted using a Robbins-Monro stochastic approximation strategy; the step size is automatically increased when the acceptance rate is higher than the target value (approximately 0.234) and decreased when it is lower, maintaining the acceptance rate near the optimal acceptance rate. This is further constrained by independent pruning of the prior range of each parameter, ensuring efficient traversal of the parameter space.

[0147] During the convergence diagnosis phase, a multi-chain parallel sampling strategy is adopted to obtain the complete posterior distribution: to test whether each chain converges to the same probability space, a split-chain-based Gelman-Rubin algorithm is used. Diagnostic methods are used based on statistical measures, when all parameters are split. When all values ​​approach 1, it indicates that the sampling of the multi-chain Markov chain Monte Carlo algorithm has converged to the same probability space; conversely, if they are far apart, there is a lack of convergence, indicating that the current posterior inference results are unreliable. In this case, it is necessary to diagnose parameter correlation and ambiguity, and to comprehensively adjust the algorithm by increasing iterations, switching samplers, or strengthening observation constraints until all parameters are split. Only when the value satisfies the convergence criterion can the reliability of posterior inference be ensured.

[0148] Wherein, the split The calculation of the value includes:

[0149] Divide the m chains of length T equally to form 2m chains of length T. For a semi-chain, define the average intra-chain variance W and the inter-chain variance B as:

[0150] ;

[0151] ;

[0152] In the formula, Let c be the value at step t of the c-th chain; Let be the average value of all t values ​​in the c-th chain; For all The average value;

[0153] Mixed estimators of posterior variance Defined as:

[0154] ;

[0155] pass Calculate splitting with average intra-chain variance W The value is:

[0156] .

[0157] Step 7: When the chains split When the value meets the preset convergence requirements, the highest posterior density interval and the continuous maximum posterior probability estimation points are determined, and feedback adjustments are made according to the preset test rules until the preset termination conditions are met.

[0158] Specifically, to quantify the reliability of the inversion results, the Bayesian confidence range of the parameters is defined using the highest posterior density interval (hereinafter referred to as the HPD interval for ease of description). Based on this, discrete mode points are located using the sampling results of the Markov chain Monte Carlo algorithm, and are transformed into continuous maximum a posteriori probability estimation points (hereinafter referred to as refined MAP points for ease of description). By analyzing the relative spatial relationship between the refined MAP points and the prior parameter range, the reliability of the prior parameter range is assessed and dynamically corrected to determine the refined MAP points that meet the preset termination conditions.

[0159] Step 7.1: Use HPD intervals to define the Bayesian confidence fluctuation range of the inversion parameters.

[0160] The posterior distribution of the inversion parameters is sampled using the Markov chain Monte Carlo algorithm to characterize the distribution features of high-probability regions in the parameter space. Furthermore, to avoid the problem that traditional symmetric confidence intervals are insufficient in describing skewed posterior distributions, HPD intervals and posterior standard deviations are used to evaluate the degree of posterior shrinkage of the inversion parameters.

[0161] The HPD interval refers to the shortest continuous interval that can cover the sample size given that the cumulative probability of the parameter falling within the interval reaches a set standard (such as 90%, 95%, or 99%).

[0162] The HPD interval definition includes: Let the j-th inversion parameter be... After sorting the S posterior sample sets obtained by merging the chains, the results are as follows: At a level of credibility Below, the sample size is ,in This indicates rounding down to the nearest integer.

[0163] The HPD interval is then defined as: .

[0164] In the formula, Let be the lower and upper bounds of the HPD interval for the j-th inversion parameter.

[0165] Step 7.2: Select the maximum a posteriori probability estimation point (MCMC-MAP) (the maximum a posteriori probability estimation point based on Markov chain Monte Carlo algorithm MCMC sampling location and gradient optimization refinement) from the posterior distribution as a discrete approximate solution.

[0166] Step 7.3: Given the limitations of discrete sampling in the Markov chain Monte Carlo algorithm, MCMC-MAP points often fail to accurately approximate the probability extrema of the continuous posterior surface. Therefore, using MCMC-MAP points as initial iteration points, the L-BFGS-B quasi-Newton method is employed for gradient-driven local refinement to obtain refined MAP points with low error.

[0167] Step 7.4: To verify the global representativeness of the refined solution, examine the spatial relationship between the refined MAP point and the HPD interval. If the refined MAP point is located inside the HPD interval, it indicates that the optimal solution is located in the high-density core region of the posterior distribution, and proceed to the next step of verification. If the refined MAP point deviates from the HPD interval, it indicates that the refined MAP point is a local extremum or that the Markov chain Monte Carlo algorithm has not sampled sufficiently. It is necessary to increase the sampling number of the Markov chain Monte Carlo algorithm to improve the spatial coverage, and return to step 6 to recalculate.

[0168] Step 7.5: Check whether the refined MAP point is close to the prior boundary. If it is within the range of the prior parameters, it means that the parameters corresponding to the refined MAP point are optimal, and the check ends. If it is close to the boundary of the prior parameter range, calculate the negative log-likelihood gradient and proceed to the next check.

[0169] Step 7.6: Check whether the gradient of the likelihood function at the refined MAP point points outside the boundary. If the gradient of the likelihood function at the refined MAP point does not point outside the boundary of the prior parameter range, the check terminates. If the gradient of the likelihood function at the refined MAP point points outside the boundary of the prior parameter range, the prior boundary parameter range needs to be expanded, and the calculation should be repeated in step 3.

[0170] Given that nonlinear surrogate models possess good interpolation capabilities but weak extrapolation capabilities within the training parameter range, the prior parameter range essentially defines the optimization space. Therefore, if the refined MAP point falls within this optimization space, it indicates that it represents the optimal inversion parameters. If the refined MAP point approaches the prior parameter range and its likelihood gradient points out of bounds, it indicates that the current prior parameter range fails to cover the true optimal solution. In other words, the inversion parameters corresponding to the refined MAP point are merely pseudo-optimal solutions constrained by boundaries. Boundary alignment diagnostics are needed to analyze the reliability of the prior parameter range constraints, enabling targeted expansion of the prior range to avoid pseudo-optimal solutions and improve the accuracy and reliability of the inversion results.

[0171] The prior boundary alignment diagnosis includes:

[0172] Based on the defined range of prior parameters: If the refined MAP points satisfy: or This indicates that the solution may be subject to prior constraints from the nonlinear surrogate model. To determine whether it is necessary to expand the parameter range, the negative log-likelihood gradient needs to be further calculated. At this point, when the refined MAP point satisfies and When the parameter is increased further, it indicates that the objective function can still be reduced; when the refined MAP point is close to the lower boundary... and When this occurs, it indicates that further reducing the parameters can still lower the objective function; in this case, it also indicates that the true optimal solution may be located outside the current prior parameter range, and the parameter range can be expanded using the following formula: ;

[0173] In the formula, For boundary expansion coefficient, This represents the expanded range of prior parameters.

[0174] The above process is repeated iteratively until the refined MAP point no longer adheres to the prior boundary.

[0175] In this embodiment, for parameters determined to be boundary-limited, quantitative prior extension suggestions are provided to reduce the risk of pseudo-optimal solutions caused by improper prior range setting.

[0176] Step 8: Determine the inversion parameters corresponding to the refined MAP points that meet the preset termination conditions as the optimal inversion parameters.

[0177] To further illustrate the three-dimensional geostress field inversion method based on dynamic prior constraints provided by this invention, an engineering case is used to discuss the applicability of the method provided in the embodiments of this invention, including the following steps:

[0178] Step 1: Establish a three-dimensional numerical model of the actual engineering area to be inverted, determine the rock mechanics parameters, and collect field-measured in-situ stress data. Simultaneously, transform the measured in-situ stress data from the three principal stresses into six stress components. The field-measured in-situ stress data and the stress components of the transformed measured in-situ stress are shown in Tables 1 and 2, respectively.

[0179] Table 1. Field measured ground stress data

[0180]

[0181] Table 2. Stress components of the measured ground stress after conversion.

[0182]

[0183] Step 2: Apply linearly varying boundary conditions to the 3D numerical model sequentially, and perform univariate gradient forward modeling using the numerical calculation software FLAC3D to determine the range of prior parameters.

[0184] In this embodiment, self-weight stress, as well as horizontal compression, horizontal shear and vertical shear, a total of 7 types of parameters are considered for inversion, and their prior parameter ranges are shown in Table 3.

[0185] Table 3. Defined range of prior parameters

[0186]

[0187] Step 3: Using the Sobol sequence, 300 sets of sample points were generated within the prior parameter range. FLAC3D was then used for batch multi-threaded forward modeling calculations to construct the input-output forward modeling dataset. Some of the sample points generated by Sobol are shown in Table 4. Some of the forward modeling data are shown in Table 5.

[0188] Table 4. Sample points generated by some Sobol

[0189]

[0190] Table 5. Partial Forward Modeling Data

[0191]

[0192] Step 4: A nonlinear surrogate model is constructed by using a hierarchical sampling strategy and a neural network that integrates residual self-attention mechanism and multilayer perceptron.

[0193] Step 5: Within the prior parameter range, a candidate parameter set is roughly generated using hypercubic Latin sampling. The prediction variance of the nonlinear surrogate model in the likelihood function is set to zero (in the normalized amplitude likelihood, the measurement noise error is set to 2%, and the prediction error of the nonlinear surrogate model is set to 8%). The negative log-likelihood value of each candidate parameter is quickly evaluated, and the 8 points with the smallest negative log-likelihood are selected. These, along with the center point of the prior parameters, result in a total of 9 starting points. L-BFGS-B gradient optimization is performed on these 9 starting points to find the global optimum, denoted as the pre-optimized MAP points. The parameters of the pre-optimized MAP points are shown in Table 6.

[0194] Table 6 Pre-optimized MAP point parameters

[0195]

[0196] Step 6: Centering on the pre-optimized MAP point, perform a truncated Gaussian perturbation based on the standard deviation of each parameter's prior range by 5% of its width. This provides different starting points for the four Markov chain Monte Carlo algorithm chains. Posterior sampling is then performed, traversing the parameter space of the input-output forward modeling dataset to determine the posterior distribution region of the parameters. Each chain is then evaluated. The value is used to ensure that all four Markov chain Monte Carlo algorithm chains converge to the same region. The starting points of the four Markov chain Monte Carlo algorithm chains are shown in Table 7.

[0197] Table 7. Starting Points of Each Chain

[0198]

[0199] Step 7: When the chains split When the value meets the preset convergence requirements, the HPD interval and refined MAP point are determined, and feedback adjustments are made according to the preset verification rules until the preset termination condition is met.

[0200] Step 7.1: Define the Bayesian confidence range of the parameters using the 95% HPD interval. The HPD interval is shown in Table 8.

[0201] Table 8 HPD Intervals

[0202]

[0203] Step 7.2: Select MCMC-MAP points as discrete approximate solutions from the posterior space evaluated by the Markov chain Monte Carlo algorithm.

[0204] Step 7.3: Perform L-BFGS-B gradient optimization on the discrete approximate solution to obtain the refined MAP point. The parameters of the refined MAP point are shown in Table 9.

[0205] Table 9 Refined MAP Point Parameters

[0206]

[0207] Step 7.4: Determine the spatial relationship between the refined MAP point and the HPD interval. If the refined MAP point is located inside the HPD interval, it means that the optimal solution is located in the high-density core region of the posterior distribution, and the next step of verification is performed. If the refined MAP point deviates from the HPD interval, it means that the point is a local extremum or that the Markov chain Monte Carlo algorithm has not sampled sufficiently, and resampling should be performed to improve the spatial coverage.

[0208] Step 7.5: Determine whether the refined MAP points are close to the upper and lower bounds by calculating their relative positions within the prior boundary. Perform a prior parameter range boundary (parameter interval for training the nonlinear surrogate model) fit test on the refined MAP points. When the relative position is less than or equal to 0.03, it is considered close to the lower bound; when it is greater than or equal to 0.97, it is considered close to the upper bound. The relationship between the prior range and the relative positions of the refined MAP points is shown in Table 10.

[0209] Table 10 Relationship between the range of prior parameters and the relative positions of refined MAP points

[0210]

[0211] Step 7.6: When the refined MAP point is close to the prior boundary, the likelihood gradient direction is determined. The likelihood gradient direction of the parameters close to the prior boundary is determined as shown in Table 11.

[0212] Table 11 shows the likelihood gradient direction determination for parameters adhering to the prior boundary.

[0213]

[0214] Parameters that are close to the prior boundary and whose likelihood gradient direction points outward are expanded by 30% of the width of the prior boundary parameter interval. Based on the expanded interval, 100 supplementary samples are obtained using Sobol sampling, and FLAC3D forward modeling is performed to obtain a supplementary training set. The nonlinear surrogate model is then retrained using the original training set. The above steps are repeated until the refined MAP point is no longer close to the prior boundary. The inversion parameters corresponding to the refined MAP point that meets the preset termination condition are determined as the optimal inversion parameters. The final HPD range and refined MAP point are shown in Table 12. The comparison between the stress component inversion values ​​of the final refined MAP point and the measured values ​​is shown in Table 13.

[0215] Table 12 Final HPD Range and Refined MAP Points

[0216]

[0217] Table 13 Comparison of the stress component inversion values ​​and measured values ​​at the final refined MAP points

[0218]

[0219] Referring to Table 13 above, in terms of absolute error, the maximum absolute error between the 24 inversion values ​​and the measured values ​​in this embodiment is 1.050 MPa (σ at measuring point 2). yy The minimum error is 0.009 MPa (σ at measuring point 2). xx The mean absolute error is 0.433 MPa, and the coefficient of determination R0 is 0.433 MPa. 2The value of 0.9879 indicates that the geostress field inversion results obtained in this embodiment are in high agreement with the actual stress field distribution.

[0220] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

[0222] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A three-dimensional geostress field inversion method based on dynamic prior constraints, characterized in that, Includes the following steps: Step 1: Establish a three-dimensional numerical model based on geological exploration data, stratigraphic lithology, and measured stress point coordinates, and determine the stress components at each measured point; Step 2: Apply linearly varying boundary conditions to the three-dimensional numerical model sequentially, perform univariate forward modeling, establish an input-output dataset, and define the range of prior parameters based on measured geostress components; Step 3: Sample multiple initial samples within the range of prior parameters, perform forward modeling calculations, and construct the input-output forward modeling dataset; Step 4: Construct a multi-input, multi-output neural network nonlinear surrogate model; Step 5: Construct the normalized magnitude likelihood, sample the candidate set within the range of prior parameters, take the points with the smallest negative log-likelihood values ​​as the initial points, and approach the global optimum through multi-starting point gradient optimization to obtain the pre-optimized maximum posterior probability point; Step 6: Using the pre-optimized maximum posterior probability point as the initial state, perform posterior sampling using the multi-chain Markov chain Monte Carlo algorithm, traverse the parameter space of the input-output forward modeling dataset, determine the posterior distribution of the parameters, and calculate the splitting of each chain. The value ensures that all chains converge to the same posterior distribution region; Step 7: When the chains split When the value meets the preset convergence requirements, the highest posterior density interval and the continuous maximum posterior probability estimation points are determined, and feedback adjustments are made according to the preset test rules until the preset termination conditions are met. Step 8: Determine the inversion parameters corresponding to the consecutive maximum posterior probability estimation points that satisfy the preset termination condition as the optimal inversion parameters.

2. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 1 specifically includes: Based on the geological report, the area covering all the impact range of the project was delineated as the area to be inverted; Based on the engineering geological data and stratigraphic lithology data of the area to be inverted, the spatial distribution of rock groups and fault structures is clarified. Based on the spatial distribution of rock groups and fault structures, a three-dimensional numerical model reflecting the geological structural characteristics of the area to be inverted is constructed, and the principal stress values ​​of each measured point are collected and expressed as stress components.

3. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Using the controlled variable method, perform univariate gradient forward modeling on the three-dimensional numerical model and collect the calculation results to construct an input-output dataset; Step 2.2: Based on the constructed input-output dataset, use multi-output linear fitting mapping and evaluate parameter sensitivity to construct a sensitivity sequence to measure the strength of the effect of each input parameter on different stress components; Step 2.3: Under small-amplitude regularization, based on the sensitivity sequence, and combining sequential optimization and weighted least squares method, solve for the combination of input boundary conditions that satisfy the conditions, and obtain the initial boundary parameter vector; Step 2.4: Using the initial boundary parameter vector as the center, expand along each parameter dimension to define the range of prior parameters.

4. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 3, characterized in that, Step 2.2 specifically includes: Based on the input-output dataset, for the i-th input parameter, least squares fitting is performed on the stress components at n measurement points to obtain multiple linear fitting equations, which serve as a single-parameter linear model: ; In the formula, y i For the i-th input parameter x i The generated stress component, a i b i These are the slope vector and the intercept vector, respectively; The least squares fitting process is repeated for all N input parameters to obtain N sets of single-parameter linear models. Based on the linear superposition assumption of elasticity, the fittings for each stress component are superimposed to construct an overall linear approximation model. ; In the formula, BS is the stress component vector, and BS is the boundary condition vector. The superposition of intercept terms of each linear model ; Using the superposition of the intercept terms of each linear model as the baseline, the deviation between the target stress and the baseline is calculated, and the target required intensity vector Y of each stress component is constructed. * : ; In the formula, Y is the actual target stress vector; Euclidean norm aggregation is performed on the slopes of the same physical quantity in response at different measurement points to construct a sensitivity sequence. : ; In the formula, n is the total number of measurement points.

5. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 3 specifically includes: Within the range of the prior parameters, a quasi-Monte Carlo sampling method is used to extract combinations of boundary conditions as initial samples; The boundary conditions are combined to perform batch multi-threaded forward modeling calculations, and the calculation results are summarized to obtain the input-output forward modeling dataset.

6. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 4 specifically includes: A hierarchical sampling strategy is adopted to partition the input-output forward modeling dataset, and a neural network model with residual attention mechanism and multilayer perceptron is integrated to construct a multi-input multi-output neural network nonlinear surrogate model. The input-output forward dataset is clustered into multiple clusters using the K-Means algorithm, and the training set and test set are independently divided in each cluster according to a preset ratio to achieve balanced sampling in each parameter range and ensure that the training set and test set have the same coverage in the parameter space. The input boundary condition parameters are mapped into a high-dimensional vector sequence. A multi-layer multi-head self-attention mechanism is adopted to achieve global information capture and feature extraction by calculating the correlation between query vector, key vector, and value vector. The coupling relationship between parameters is automatically learned to obtain the encoded global features. A multilayer perceptron is used to perform nonlinear regression on the global features encoded by the self-attention mechanism to establish a highly complex nonlinear fitting relationship between the input and output.

7. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1: Construct the normalized magnitude likelihood of the predicted variance by introducing a nonlinear surrogate model, normalize the residual by dividing it by the physical scale of the corresponding stress component, and assign the normal stress component the same penalty weight as the shear stress component. Step 5.2: Generate a candidate parameter set using hypercubic Latin sampling within the prior parameter range, and use a nonlinear surrogate model to quickly evaluate the negative log-likelihood of each candidate parameter point, and initially screen out high-likelihood candidate points. Step 5.3: Perform the L-BFGS-B quasi-Newton method with boundary constraints on the smooth posterior high probability region, so that the high likelihood candidate points can quickly converge to the local mode of the posterior distribution in the pre-optimization stage, and obtain the pre-optimized maximum posterior probability point.

8. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 6 specifically includes: Posterior sampling phase: Starting from near the pre-optimized maximum posterior probability point, the multi-chain Markov chain Monte Carlo algorithm is used to initiate parallel sampling using the adaptive random walk Metropolis-Hastings algorithm; the proposal step size of each chain is adaptively adjusted through the Robbins-Monro random approximation strategy. When the acceptance rate is higher than the target value, the step size is automatically increased, and when it is lower than the target value, the step size is decreased, so that the acceptance rate is maintained near the optimal acceptance rate and is subject to independent pruning constraints of the prior range of each parameter. Convergence diagnosis phase: Gelman-Rubin based on split chains is used. Diagnostic methods are used based on statistical measures, when all parameters are split. When all values ​​approach 1, it indicates that the sampling of the multi-chain Markov chain Monte Carlo algorithm has converged to the same probability space; otherwise, there is a non-convergence phenomenon, and the current sampling logic needs to be corrected.

9. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 8, characterized in that, The split The calculation of the value includes: Divide the m chains of length T equally to form 2m chains of length T. For a semi-chain, define the average intra-chain variance W and the inter-chain variance B as: ; ; In the formula, Let c be the value at step t of the c-th chain; Let be the average value of all t values ​​in the c-th chain; For all The average value; Mixed estimators of posterior variance for: ; pass Calculate splitting with average intra-chain variance W The value is: 。 10. The three-dimensional geostress field inversion method based on dynamic prior constraints according to claim 1, characterized in that, Step 7 specifically includes: Step 7.1: Define the Bayesian confidence fluctuation range of the inversion parameters using the highest posterior density interval; Step 7.2: Select the maximum a posteriori probability estimate point MCMC-MAP from the posterior distribution as the discrete approximate solution; Step 7.3: Using the MCMC-MAP point as the initial iteration point, the gradient-driven local refinement is performed using the L-BFGS-B quasi-Newton method to obtain the continuous maximum a posteriori probability estimation point; Step 7.4: Examine the spatial relationship between the consecutive maximum a posteriori probability estimation points and the highest a posteriori density interval. If the consecutive maximum a posteriori probability estimation points are located inside the highest a posteriori density interval, it indicates that the optimal solution is located in the high-density core region of the posterior distribution, and proceed to the next step of the examination; if the consecutive maximum a posteriori probability estimation points deviate from the highest a posteriori density interval, it indicates that the consecutive maximum a posteriori probability estimation points are local extrema or that the Markov chain Monte Carlo algorithm has not sampled sufficiently, and it is necessary to increase the sampling number of the Markov chain Monte Carlo algorithm to improve the spatial coverage, and return to step 6 to recalculate; Step 7.5: Check whether the consecutive maximum a posteriori probability estimation points are close to the prior boundary. If they are within the range of prior parameters, it indicates that the parameters corresponding to the consecutive maximum a posteriori probability estimation points are optimal, and the test ends. If they are close to the boundary of the prior parameter range, calculate the negative log-likelihood gradient and proceed to the next test. Step 7.6: Check whether the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the boundary. If the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the prior parameter range, the check is terminated. If the gradient of the likelihood function at the consecutive maximum a posteriori probability estimation points to the outside of the prior parameter range, the prior boundary parameter range needs to be expanded, and the calculation should be repeated in step 3.

Citation Information

Patent Citations

  • Nonlinear crustal stress field inversion method, device, equipment and medium

    CN119885683A

  • Layered surrounding rock three-dimensional crustal stress field inversion intelligent analysis system and method

    CN120354762A

  • Nonlinear crustal stress field inversion method and device based on physical data neural network

    CN120832833A

  • Ground stress field inversion analysis method fusing physical information neural network

    CN122065225A