A method for three-dimensional reconstruction of a layered geological structure

Abstract

The application discloses a kind of layered geologic structure three-dimensional reconstruction methods, comprising the following steps: step 1: construct the three-dimensional voxel model of study site;Collect borehole data, construct first stage parameter sample set;The input of sample in first stage parameter sample set is voxel space position information, and output is stratum category;Step 2: construct first stage structure identification model, and using first stage parameter sample set is trained and optimized;Step 3: using the first stage structure identification model of training optimization completion, output the prediction result of stratum category of global domain, and output final prediction result by layered sequence constraint;Step 4: construct second stage parameter sample set;The input of sample in second stage parameter sample set is the prediction result of stratum category of global domain of study site and monitoring feature while drilling, and output is soil parameter;Step 5: construct second stage parameter prediction model, and using second stage parameter sample set is trained and optimized.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering technology, and in particular to a method for three-dimensional reconstruction of layered geological structures. Background Technology

[0002] In geotechnical engineering fields such as tunnel engineering, mining engineering, and underground space development, accurately obtaining the underground layered geological structure and its soil parameters is fundamental for engineering design, construction, and numerical simulation analysis. Currently, underground geological structure modeling mainly relies on borehole data, field exposure information, and manual interpretation by engineers, and is completed through two-dimensional profile extrapolation or three-dimensional modeling software. This process is highly dependent on human experience, not only is it labor-intensive and time-consuming, but the modeling results from different personnel also vary significantly, making it difficult to meet the needs of rapid and objective modeling in engineering practice.

[0003] Furthermore, during engineering tunneling and construction, numerical simulation has become an important tool for surrounding rock stability analysis, support design, and risk assessment, and soil parameters are an indispensable key input in numerical simulation. Currently, soil parameters are usually obtained through field sampling and laboratory or in-situ testing. If sampling and testing are carried out again in every area where numerical analysis is required, it is not only time-consuming and costly, but also difficult to obtain in a timely manner under the condition of rapid construction, thus restricting the efficiency of numerical simulation in engineering decision-making.

[0004] In summary, existing technologies are unable to achieve rapid and automatic modeling of layered geological structures, and are also unable to quickly obtain the distribution of soil parameters covering the entire area while ensuring rationality. They cannot meet the dual requirements of rapid modeling and rapid parameter determination in engineering practice.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a three-dimensional reconstruction method for layered geological structures. This method can quickly construct a three-dimensional layered geological structure model based on existing borehole data and predict soil parameters covering the entire area. This provides an intelligent method for numerical simulation with directly usable parameter inputs, thereby reducing the workload of manual modeling, reducing the cost of repeated experiments, and improving the efficiency of engineering analysis and decision-making.

[0007] To achieve the above objectives, the present invention also employs the following technical solution:

[0008] A method for three-dimensional reconstruction of layered geological structures includes the following steps:

[0009] Step 1: Construct a three-dimensional voxel model of the research site; collect borehole data and construct the first-stage parameter sample set; the input of the samples in the first-stage parameter sample set is the spatial location information of voxels, and the output is the stratigraphic category;

[0010] Step 2: Construct the first-stage structure recognition model and train and optimize it using the first-stage parameter sample set;

[0011] Step 3: Using the first-stage structure recognition model that has been trained and optimized, output the prediction results of the stratigraphic categories of the entire domain, and output the final prediction results through stratigraphic sequence constraints;

[0012] Step 4: Construct the second-stage parameter sample set; The input to the samples in the second-stage parameter sample set is the predicted results of the stratigraphic categories of the entire study site and the characteristics of drilling monitoring, and the output is soil parameters;

[0013] Step 5: Construct the second-stage parameter prediction model and train and optimize it using the second-stage parameter sample set;

[0014] Step 6: Use the trained and optimized first-stage structure recognition model to predict the stratigraphic category of the entire target site, and output the final prediction result through stratigraphic sequence constraints; use the trained second-stage parameter prediction model to predict the soil parameters of the target site.

[0015] Furthermore, step 1 includes the following steps:

[0016] Step 1.1: Obtain the spatial extent and boundary information of the research site, divide the research site into regular voxel units, construct a three-dimensional voxel model, and each voxel unit corresponds to a unique spatial location;

[0017] Step 1.2: Obtain borehole data, including borehole planar location, borehole depth, borehole dip angle and formation type, and perform format standardization, anomaly removal and integrity checks on the borehole data;

[0018] Step 1.3: Discretize the borehole data along the borehole depth direction at preset intervals, converting the borehole data into depth-by-depth samples, each sample including voxel spatial location information. and the stratigraphic category corresponding to point z; for vertical boreholes, For oblique holes, , , ;in , , Let be the coordinates of the borehole on the ground, and s be the borehole depth. The drilling angle is... It is the azimuth angle;

[0019] Step 1.4: Based on the borehole depth-by-depth records, construct a first-stage parameter sample set with voxel spatial location information as input and formation category as output.

[0020] Furthermore, step 1.1 also includes the following step: constructing a true-value stratum with upper and lower layers and undulating interfaces:

[0021] Step 1.1.1: Construct two interface functions using random functions. , , respectively representing the undulations of the first and second interfaces at different spatial locations; among which ;

[0022] Step 1.1.2: Generate voxel categories according to a three-layer rule from top to bottom: when When, the stratigraphic type is 0; when When, the stratigraphic type is 1; when At that time, the stratigraphic type was 2.

[0023] Furthermore, step 2 includes the following steps:

[0024] Step 2.1: Voxel spatial location information Scale unification and normalization are performed to eliminate the influence of different units on the model training process, as shown in the following formula:

[0025] ;

[0026] In the formula, x, y, and z are the original spatial coordinates of the voxel unit. These represent the minimum and maximum values ​​of the spatial range in the x-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the y-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the z-direction, respectively. These are the normalized dimensionless eigenvalues, with a range of values ​​of... ;

[0027] Step 2.2: Set the type of the first-stage structure recognition model and set the initial parameters of the first-stage structure recognition model;

[0028] Step 2.3: Use the first-stage parameter sample set to train the first-stage structure recognition model, so that the first-stage structure recognition model learns the correspondence between voxel spatial location information and stratigraphic categories; during the training process of the first-stage structure recognition model, introduce the grid search parameter optimization method, traverse and test the key parameter combinations of the first-stage structure recognition model, and record the model performance under different parameter combinations.

[0029] Step 2.4: Select and adjust the model based on the evaluation results of the model performance; when the classification accuracy of the model reaches the preset requirements, complete the training and form the final first-stage structure recognition model.

[0030] Furthermore, in step 2.2, the first-stage structure recognition model is set as a random forest classification model. The initial parameters of the random forest model are set as follows: n_estimators=200, max_depth=15, min_samples_leaf=3, class_weight=balanced, random_state=0. The model output is the probability vector of the corresponding voxel belonging to each layer category: P=(P0,P1,P2).

[0031] In step 2.4, the parameter search range is: n_estimators ∈ {150, 300}, max_depth ∈ {12,None}, min_samples_leaf ∈ {1, 3}, and KFold is used for cross-validation with a fold number of 3.

[0032] Furthermore, step 3 includes the following steps:

[0033] Step 3.1: Using the first-stage structure recognition model that has been trained and optimized, predict the probability of each stratigraphic category corresponding to the spatial location of all voxels in the research site, except for the spatial location of borehole data voxels.

[0034] Step 3.2: Arrange and filter the combinations of stratigraphic categories that satisfy the stratigraphic sequence constraint at each position on each vertical voxel column; the stratigraphic sequence constraint means that the stratigraphic category at each position on the vertical voxel column from top to bottom satisfies the sequence change of 0→1→2;

[0035] Step 3.3: Based on the results of the sorting and filtering in Step 3.2, calculate the overall cost of each combination of vertical voxel columns, and select the combination with the smallest overall cost as the final result; the formula for calculating the overall cost is as follows:

[0036] ;

[0037] In the formula, voxel spatial position The probability that the stratigraphic category is c.

[0038] Furthermore, step 4 includes the following steps:

[0039] Step 4.1: Construct the second-stage parameter sample set; each sample in the second-stage parameter sample set corresponds to a voxel spatial location, in the following form:

[0040] sample ;

[0041] In the formula, Let c be the spatial location of the voxel, and c be the stratigraphic category corresponding to that location. Features for monitoring while drilling, These are soil properties;

[0042] Step 4.2: Perform unit unification, scale normalization, and missing value handling on the data in the constructed second-stage parameter sample set;

[0043] Step 4.3: Perform spatial consistency, class consistency and parameter rationality checks on the samples, and remove obviously abnormal or conflicting samples.

[0044] Furthermore, step 5 includes the following steps:

[0045] Step 5.1: Set the type of the second-stage parameter prediction model and set the initial parameters of the second-stage parameter prediction model;

[0046] Step 5.2: Train the second-stage parameter prediction model using the second-stage parameter sample set; during the training process, set multiple candidate model parameter combinations, evaluate the prediction performance of the model under different parameter combinations through parameter traversal and cross-validation, select the parameter combination with the smallest prediction error or the best stability as the final model configuration, and complete the training of the second-stage parameter prediction model under this parameter configuration.

[0047] Step 5.3: Evaluate the prediction performance of the trained second-stage parameter prediction model. When the model prediction error meets the preset accuracy requirements, the training is completed, and the final second-stage parameter prediction model is formed.

[0048] Furthermore, in step 5.1, the second-stage structure recognition model is set to the XGBoost model, with the initial parameters using the default parameters; in step 5.2, the XGBoost regression model is configured with the following candidate parameter combinations: n_estimators: {300, 500, 800}, learning_rate: {0.03, 0.05, 0.1}, max_depth: {4, 6, 8}, min_child_weight: {1, 3, 5}, subsample: {0.7, 0.8, 1.0}, colsample_bytree: {0.7, 0.8, 1.0}, reg_lambda: {0.5, 1.0, 2.0}, reg_alpha: {0.0, 0.1, 0.5}, gamma: {0.0, 0.1, 0.3}. Cross-validation is performed for each parameter combination.

[0049] Furthermore, in step 5.2, the second-stage parameter sample set is randomly divided into n non-overlapping subsets, and the... (k=1~n) training sessions are conducted, with the kth subset selected as the validation set and the remaining n-1 subsets used as the training set to complete one model training and evaluation cycle. Finally, the average of the five evaluation results is taken as the overall performance index of the parameter combination.

[0050] Compared with the prior art, the beneficial effects of this invention are as follows:

[0051] 1. Through steps 1 to 3, borehole information can be obtained in real engineering projects. By training with the location information (features) and soil type (labels) at different depths of different boreholes, the stratum classification of other un-drilled parts can be predicted, thus enabling rapid reconstruction of the three-dimensional stratum model.

[0052] 2. Through steps 4 and 5, based on the various features (location, drilling data, etc.) and labels (soil parameters) samples of the existing boreholes and tunnels, the established machine learning model can predict soil parameters at other unknown locations, thus avoiding the need to conduct in-situ experiments to determine soil parameters again. The predicted values ​​can be directly used to enter the numerical simulation process.

[0053] 3. Soil parameters can be directly used as material parameter inputs and element assignments in finite element or numerical simulation software, so that engineers do not need to determine soil parameters through empirical assumptions when conducting numerical simulations, nor do they need to repeatedly conduct a large number of indoor or field tests to obtain parameters, thereby reducing the cost of parameter determination and improving the efficiency and consistency of numerical simulation. Attached Figure Description

[0054] Figure 1 A flowchart of a method for three-dimensional reconstruction of layered geological structures;

[0055] Figure 2 This is a diagram of the actual geological structure.

[0056] Figure 3 Drilling diagram;

[0057] Figure 4 This represents the final prediction result of the first-stage structure recognition model. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0059] Example 1: A method for three-dimensional reconstruction of layered geological structures, such as Figure 1 As shown, it includes the following steps:

[0060] Step 1: Construct a three-dimensional voxel model of the research site; collect borehole data and construct the first-stage parameter sample set; the input of the samples in the first-stage parameter sample set is the spatial location information of voxels, and the output is the stratum category.

[0061] In this embodiment, step 1 includes the following steps:

[0062] Step 1.1: Obtain the spatial extent and boundary information of the research site, divide the research site into regular voxel units, construct a three-dimensional voxel model, and each voxel unit corresponds to a unique spatial location.

[0063] Specifically, let , , ,form Voxel units; each voxel is indexed by its grid. Indicates spatial location, where .

[0064] In this embodiment, step 1.1 further includes the following step: constructing a true-value stratum with upper and lower layers and undulating interfaces:

[0065] Step 1.1.1: Construct two interface functions using random functions. , , respectively representing the undulations of the first and second interfaces at different spatial locations; among which .

[0066] In this embodiment, It involves imposing legal constraints on the interface to prevent inter-layer overlap and ensure that the layer thickness is not less than 2 voxels.

[0067] Step 1.1.2: Generate voxel categories according to a three-layer rule from top to bottom: when When, the stratigraphic type is 0; when When, the stratigraphic type is 1; when At that time, the stratigraphic type was 2.

[0068] In this embodiment, the real-world scenario is simulated through the true-value strata in steps 1.1.1 to 1.1.2. When the real-world data is insufficient or limited, the borehole data required in step 1.2 can be generated through the true-value strata, and the global strata generated by the structure recognition model in the first stage of step 2 can be verified.

[0069] Step 1.2: Obtain borehole data, including borehole planar location, borehole depth, borehole dip angle and formation type, and perform format standardization, anomaly removal and integrity checks on the borehole data.

[0070] In this embodiment, the drilling plane position is the coordinate on the ground. The drilling depth is z.

[0071] In this embodiment, the borehole data can be taken from existing borehole data; when the existing borehole data is insufficient, simulated data can be generated through the true formation in steps 1.1.1 to 1.1.2.

[0072] Step 1.3: Discretize the borehole data along the borehole depth direction at preset intervals, converting the borehole data into depth-by-depth samples, each sample including voxel spatial location information. and the stratigraphic category corresponding to point z; for vertical boreholes, For oblique holes, , , ;in , , Let be the coordinates of the borehole on the ground, and s be the borehole depth. The drilling angle is... It is the azimuth angle.

[0073] In this embodiment, the number of drill holes is set to 20, and the plane position of each drill hole is... The true formation was selected using a uniformly random method. Each borehole penetrated the full depth. For each borehole plane position For depth One record is generated for each meter of voxel layer: the input feature of this record is the spatial location information of the voxel. The output label is the stratum type (0 / 1 / 2) at that depth.

[0074] Step 1.4: Based on the borehole depth-by-depth records, construct a first-stage parameter sample set with voxel spatial location information as input and formation category as output.

[0075] In this embodiment, one borehole generates 50 training samples, and 20 boreholes generate a total of 50 training samples. Training samples.

[0076] Step 2: Construct the first-stage structure recognition model and train and optimize it using the first-stage parameter sample set.

[0077] In this embodiment, step 2 includes the following steps:

[0078] Step 2.1: Voxel spatial location information Scale unification and normalization are performed to eliminate the influence of different units on the model training process, as shown in the following formula:

[0079] ;

[0080] In the formula, x, y, and z are the original spatial coordinates of the voxel unit. These represent the minimum and maximum values ​​of the spatial range in the x-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the y-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the z-direction, respectively. These are the normalized dimensionless eigenvalues, with a range of values ​​of... .

[0081] Step 2.2: Set the type of the first-stage structure recognition model and set the initial parameters of the first-stage structure recognition model.

[0082] In this embodiment, in step 2.2, the first-stage structure recognition model is set as a random forest classification model. The initial parameters of the random forest model are set as follows: n_estimators=200, max_depth=15, min_samples_leaf=3, class_weight=balanced, random_state=0. The model output is the probability vector of the corresponding voxel belonging to each layer category: P=(P0,P1,P2).

[0083] Step 2.3: Use the first-stage parameter sample set to train the first-stage structure recognition model, so that the first-stage structure recognition model learns the correspondence between voxel spatial location information and stratigraphic categories; during the training process of the first-stage structure recognition model, introduce a grid search parameter optimization method to traverse and test the key parameter combinations of the first-stage structure recognition model, and record the model performance under different parameter combinations.

[0084] Step 2.4: Select and adjust the model based on the evaluation results of the model performance; when the classification accuracy of the model reaches the preset requirements, complete the training and form the final first-stage structure recognition model.

[0085] In this embodiment, in step 2.4, the parameter search range is: n_estimators ∈ {150, 300}, max_depth ∈ {12, None}, min_samples_leaf ∈ {1, 3}, and KFold is used for cross-validation with a fold number of 3.

[0086] In this embodiment, in step 2.4, when the model performance is lower than a preset threshold, the model adaptive adjustment mechanism is triggered. The model is optimized by changing the model type, introducing ensemble learning or boosting algorithms, increasing training samples, or adjusting feature combinations. The adjusted model is then retrained and evaluated until the model performance reaches the preset requirements.

[0087] Step 3: Using the first-stage structure recognition model that has been trained and optimized, output the prediction results of the stratigraphic categories of the entire domain, and output the final prediction results through stratigraphic sequence constraints.

[0088] In this embodiment, step 3 includes the following steps:

[0089] Step 3.1: Using the first-stage structure recognition model that has been trained and optimized, predict the probability of each stratigraphic category corresponding to the spatial location of all voxels in the research site, except for the spatial location of borehole data voxels.

[0090] In this embodiment, by combining the data from steps 1 and 3.1, the probabilities of the three stratigraphic categories at each location on each vertical voxel column (fixed x, y, along z) in the three-dimensional voxel model of the research site can be obtained.

[0091] Step 3.2: Arrange and filter the combinations of stratigraphic categories that satisfy the stratigraphic sequence constraint at each position on each vertical voxel column; the stratigraphic sequence constraint means that the stratigraphic categories at each position on the vertical voxel column from top to bottom satisfy the sequence change of 0→1→2.

[0092] Step 3.3: Based on the results of the sorting and filtering in Step 3.2, calculate the overall cost of each combination of vertical voxel columns, and select the combination with the smallest overall cost as the final result; the formula for calculating the overall cost is as follows:

[0093] ;

[0094] In the formula, voxel spatial position The probability that the stratigraphic category is c.

[0095] In this embodiment, the stratigraphic category prediction results at each location on each voxel are obtained through step 3.3, and the stratigraphic category prediction results for the entire study site are obtained by combining them.

[0096] refer to Figures 2-4 In this embodiment, the true formation data is generated through steps 1.1.1 to 1.1.2. Figure 2 As shown, the borehole generated through step 1.3 is as follows Figure 3 As shown, the generated borehole data was used to train and optimize the first-stage structure recognition model. Finally, the stratigraphic category prediction results for the entire research site generated in step 3 are as follows: Figure 4 As shown, accurate predictions were made at both 25m and 41m.

[0097] Step 4: Construct the second-stage parameter sample set; The input of the samples in the second-stage parameter sample set is the predicted results of the stratigraphic categories of the entire study site and the characteristics of drilling monitoring, and the output is soil parameters.

[0098] Specifically, step 4 includes the following steps:

[0099] Step 4.1: Construct the second-stage parameter sample set; each sample in the second-stage parameter sample set corresponds to a voxel spatial location, in the following form:

[0100] sample ;

[0101] In the formula, Let c be the spatial location of the voxel, and c be the stratigraphic category corresponding to that location. Features monitored while drilling, such as drilling pressure, drilling speed, torque, rotational speed, pump pressure, and mud ratio, etc. These are soil parameters, which include initial engineering parameters, field mechanical test results, and laboratory mechanical test results.

[0102] In this embodiment, each sample in the sample set corresponds to a voxel spatial location. That is, each sample includes the voxel spatial location and the formation type, drilling monitoring characteristics and soil parameters corresponding to that location.

[0103] In this embodiment, soil properties specifically include elastic modulus, Poisson's ratio, etc.

[0104] Step 4.2: Perform unit unification, scale normalization, and missing value handling on the data in the constructed second-stage parameter sample set.

[0105] In this embodiment, the units of data from different sources are standardized. For example, the elastic modulus is standardized to MPa, Poisson's ratio to a dimensionless value, drilling pressure to kN, torque to kN·m, and drilling speed to m / min. Spatial coordinates and continuous numerical features are scaled, preferably using a minimum-maximum normalization method.

[0106] In this embodiment, missing values ​​can be handled through interpolation, similar sample inference, or model-based estimation methods to ensure the completeness and consistency of the parameter prediction samples.

[0107] Step 4.3: Perform spatial consistency, class consistency and parameter rationality checks on the samples, and remove obviously abnormal or conflicting samples.

[0108] In this embodiment, a statistical distribution-based outlier detection method is used for continuous features. When a soil parameter satisfies the following formula, it is identified as an outlier and removed:

[0109] ;

[0110] In the formula, For the i-th eigenvalue, The mean of this feature. This represents the standard deviation of the feature.

[0111] Step 5: Build the second-stage parameter prediction model and train and optimize it using the second-stage parameter sample set.

[0112] In this embodiment, the second-stage parameter prediction model predicts and outputs the soil parameters of each voxel in the spatial location of the entire domain based on the input.

[0113] In this embodiment, step 5 includes the following steps:

[0114] Step 5.1: Set the type of the second-stage parameter prediction model and set the initial parameters of the second-stage parameter prediction model.

[0115] In this embodiment, in step 5.1, the second-stage structure recognition model is set to the XGBoost model, and the initial parameters are set to default parameters.

[0116] The XGBoost model can handle nonlinear relationships, has a strong fitting ability for multi-source features, and can suppress overfitting through regularization and subsampling mechanisms, making it suitable for this geotechnical engineering parameter prediction scenario.

[0117] Step 5.2: Train the second-stage parameter prediction model using the second-stage parameter sample set; during the training process, set multiple candidate model parameter combinations, evaluate the prediction performance of the model under different parameter combinations through parameter traversal and cross-validation, select the parameter combination with the smallest prediction error or the best stability as the final model configuration, and complete the training of the second-stage parameter prediction model under this parameter configuration.

[0118] In this embodiment, in step 5.2, an input feature vector is constructed, including the voxel spatial location, the formation category corresponding to the location, and the monitoring features while drilling. The input label is one of the soil properties parameters.

[0119] In this embodiment, in step 5.2, the second-stage parameter sample set is randomly divided into n non-overlapping subsets, and the... (k=1~n) training sessions are conducted, with the kth subset selected as the validation set and the remaining n-1 subsets used as the training set to complete one model training and evaluation cycle. Finally, the average of the five evaluation results is taken as the overall performance index of the parameter combination.

[0120] In this embodiment, in step 5.2, the XGBoost regression model sets the candidate parameter combinations as follows: n_estimators: {300, 500, 800}, learning_rate: {0.03, 0.05, 0.1}, max_depth: {4, 6, 8}, min_child_weight: {1, 3, 5}, subsample: {0.7, 0.8, 1.0}, colsample_bytree: {0.7, 0.8, 1.0}, reg_lambda: {0.5, 1.0, 2.0}, reg_alpha: {0.0, 0.1, 0.5}, gamma: {0.0, 0.1, 0.3}. Cross-validation is performed for each parameter combination.

[0121] Step 5.3: Evaluate the prediction performance of the trained second-stage parameter prediction model. When the model prediction error meets the preset accuracy requirements, the training is completed, and the final second-stage parameter prediction model is formed.

[0122] In this embodiment, step 5.3 uses the root mean square error (RMSE) or mean absolute error (MAE) as the evaluation metric. The minimum RMSE or MAE of the cross-validation is used as the optimal criterion, and the corresponding parameter combination is selected as the final model parameter configuration.

[0123] In this embodiment, if the prediction error is higher than a preset threshold or the prediction result is not stable enough, the dataset is expanded and steps 4 and 5 are re-executed to achieve dynamic updating and continuous optimization of the model.

[0124] Step 6: Use the trained and optimized first-stage structure recognition model to predict the stratigraphic category of the entire target site, and output the final prediction result through stratigraphic sequence constraints; use the trained second-stage parameter prediction model to predict the soil parameters of the target site.

[0125] In this embodiment, soil parameters can be directly used as material parameter inputs and element assignments in finite element or numerical simulation software. This eliminates the need for engineers to determine soil parameters through empirical assumptions when conducting numerical simulations, and also eliminates the need to repeatedly conduct a large number of indoor or field tests to obtain parameters. This reduces the cost of parameter determination and improves the efficiency and consistency of numerical simulation.

[0126] This embodiment of a three-dimensional reconstruction method for layered geological structures, through steps 1 to 3, can obtain borehole information in real engineering projects, and use the location information (features) and soil type (labels) at different depths of different boreholes for training, so as to predict the stratum classification of other un-drilled parts and quickly reconstruct a three-dimensional stratum model.

[0127] This embodiment of a three-dimensional reconstruction method for layered geological structures, through steps 4 and 5, trains a machine learning model based on various features (location, drilling data, etc.) and labels (soil parameters) of the exposed portions of existing boreholes and tunnels. The established model can predict soil parameters at other unknown locations, avoiding the need for in-situ experiments to measure soil parameters again. The predicted values ​​can be directly used to enter the numerical simulation process.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of layered geological structures, characterized in that, Includes the following steps: Step 1: Construct a three-dimensional voxel model of the research site; collect borehole data and construct the first-stage parameter sample set; the input of the samples in the first-stage parameter sample set is the spatial location information of voxels, and the output is the stratigraphic category; Step 2: Construct the first-stage structure recognition model and train and optimize it using the first-stage parameter sample set; Step 3: Using the first-stage structure recognition model that has been trained and optimized, output the prediction results of the stratigraphic categories of the entire domain, and output the final prediction results through stratigraphic sequence constraints; Step 4: Construct the second-stage parameter sample set; The input to the second-stage parameter sample set consists of the predicted stratigraphic categories and drilling monitoring characteristics of the entire study site, and the output consists of soil parameters. Step 5: Construct the second-stage parameter prediction model and train and optimize it using the second-stage parameter sample set; Step 6: Use the trained and optimized first-stage structure recognition model to predict the stratigraphic category of the entire target site, and output the final prediction result through stratigraphic sequence constraints; use the trained second-stage parameter prediction model to predict the soil parameters of the target site. Step 3 includes the following steps: Step 3.1: Using the first-stage structure recognition model that has been trained and optimized, predict the probability of each stratigraphic category corresponding to the spatial location of all voxels in the research site, except for the spatial location of borehole data voxels. Step 3.2: Arrange and filter the combinations of stratigraphic categories that satisfy the stratigraphic sequence constraint at each position on each vertical voxel column; the stratigraphic sequence constraint means that the stratigraphic category at each position on the vertical voxel column from top to bottom satisfies the sequence change of 0→1→2; Step 3.3: Based on the results of the sorting and filtering in Step 3.2, calculate the overall cost of each combination of vertical voxel columns, and select the combination with the smallest overall cost as the final result; The overall cost calculation formula is as follows: ； In the formula, voxel spatial position The probability that the stratigraphic category is c.

2. The method for three-dimensional reconstruction of layered geological structures according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Obtain the spatial extent and boundary information of the research site, divide the research site into regular voxel units, construct a three-dimensional voxel model, and each voxel unit corresponds to a unique spatial location; Step 1.2: Obtain borehole data, including borehole planar location, borehole depth, borehole dip angle and formation type, and perform format standardization, anomaly removal and integrity checks on the borehole data; Step 1.3: Discretize the borehole data along the borehole depth direction at preset intervals, converting the borehole data into depth-by-depth samples, each sample including voxel spatial location information. and the stratigraphic category corresponding to point z; for vertical boreholes, For oblique holes, , , ;where x0, y0, Let be the coordinates of the borehole on the ground. The drilling depth, The drilling angle is... It is the azimuth angle; Step 1.4: Based on the borehole depth-by-depth records, construct a first-stage parameter sample set with voxel spatial location information as input and formation category as output.

3. The method for three-dimensional reconstruction of layered geological structures according to claim 2, characterized in that, Step 1.1 also includes the following step: constructing a true-value stratum with upper and lower layers and undulating interfaces: Step 1.1.1: Construct two interface functions using random functions. , , respectively representing the undulations of the first and second interfaces at different spatial locations; among which ; Step 1.1.2: Generate voxel categories according to a three-layer rule from top to bottom: when When, the stratigraphic type is 0; when When, the stratigraphic type is 1; when At that time, the stratigraphic type was 2.

4. The method for three-dimensional reconstruction of layered geological structures according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Voxel spatial location information Scale unification and normalization are performed to eliminate the influence of different units on the model training process, as shown in the following formula: ； In the formula, x, y, and z are the original spatial coordinates of the voxel unit. These represent the minimum and maximum values ​​of the spatial range in the x-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the y-direction, respectively. These represent the minimum and maximum values ​​of the spatial range in the z-direction, respectively. These are the normalized dimensionless eigenvalues, with a range of values ​​of... ; Step 2.2: Set the type of the first-stage structure recognition model and set the initial parameters of the first-stage structure recognition model; Step 2.3: Use the first-stage parameter sample set to train the first-stage structure recognition model, so that the first-stage structure recognition model learns the correspondence between voxel spatial location information and stratigraphic categories; In the first stage of structural recognition model training, a grid search parameter optimization method is introduced to traverse and test the key parameter combinations of the first stage structural recognition model and record the model performance under different parameter combinations. Step 2.4: Select and adjust the model based on the evaluation results of the model performance; when the classification accuracy of the model reaches the preset requirements, complete the training and form the final first-stage structure recognition model.

5. The method for three-dimensional reconstruction of layered geological structures according to claim 4, characterized in that, In step 2.2, the first-stage structure recognition model is set as a random forest classification model. The initial parameters of the random forest model are set as follows: n_estimators=200, max_depth=15, min_samples_leaf=3, class_weight=balanced, random_state=0. The model output is the probability vector of the corresponding voxel belonging to each layer category: P=(P0,P1,P2). In step 2.4, the parameter search range is: n_estimators ∈ {150, 300}, max_depth ∈ {12,None}, min_samples_leaf ∈ {1, 3}, and KFold is used for cross-validation with a fold number of 3.

6. The method for three-dimensional reconstruction of layered geological structures according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Construct the second-stage parameter sample set; each sample in the second-stage parameter sample set corresponds to a voxel spatial location, in the following form: sample ; In the formula, For voxel spatial location, The stratigraphic category corresponding to this location. Features for monitoring while drilling, These are soil properties; Step 4.2: Perform unit unification, scale normalization, and missing value handling on the data in the constructed second-stage parameter sample set; Step 4.3: Perform spatial consistency, class consistency and parameter rationality checks on the samples, and remove obviously abnormal or conflicting samples.

7. The method for three-dimensional reconstruction of layered geological structures according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: Set the type of the second-stage parameter prediction model and set the initial parameters of the second-stage parameter prediction model; Step 5.2: Train the second-stage parameter prediction model using the second-stage parameter sample set; during the training process, set multiple candidate model parameter combinations, evaluate the prediction performance of the model under different parameter combinations through parameter traversal and cross-validation, select the parameter combination with the smallest prediction error or the best stability as the final model configuration, and complete the training of the second-stage parameter prediction model under this parameter configuration. Step 5.3: Evaluate the prediction performance of the trained second-stage parameter prediction model. When the model prediction error meets the preset accuracy requirements, the training is completed, and the final second-stage parameter prediction model is formed.

8. The method for three-dimensional reconstruction of layered geological structures according to claim 7, characterized in that, In step 5.1, the second-stage structure recognition model is set to the XGBoost model, with the initial parameters using the default parameters. In step 5.2, the XGBoost regression model is configured with the following candidate parameter combinations: n_estimators: {300, 500, 800}, learning_rate: {0.03, 0.05, 0.1}, max_depth: {4, 6, 8}, min_child_weight: {1, 3, 5}, subsample: {0.7, 0.8, 1.0}, colsample_bytree: {0.7, 0.8, 1.0}, reg_lambda: {0.5, 1.0, 2.0}, reg_alpha: {0.0, 0.1, 0.5}, gamma: {0.0, 0.1, 0.3}. Cross-validation is performed for each parameter combination.

9. A method for three-dimensional reconstruction of layered geological structures according to claim 7, characterized in that, In step 5.2, the parameter sample set of the second stage is randomly divided into n non-overlapping subsets. For the k-th training, k=1~n, the k-th subset is selected as the validation set, and the remaining n-1 subsets are selected as the training set to complete one model training and evaluation. Finally, the average of the 5 evaluation results is taken as the overall performance index of the parameter combination.

Images