A multi-source data stratum layering prediction method based on random field and deep learning

Abstract

The application discloses a kind of based on random field and deep learning's multi-source data stratum layering prediction method, comprising: extracting drilling data, static data and layering information, and through HGB capture depth trend, combine the complex nonlinear residual between Gaussian process fitting HGB model prediction value and true value, finally generate two kinds of continuous random field;Based on the generated continuous random field, construct multi-task deep learning model, utilize Transformer to extract deep semantic features, combine LSTM modeling deep direction long-distance dependence, multi-head attention mechanism strengthens local and global feature learning, CRF layer accurately optimizes layering boundary prediction, to realize the unity of stratum layering and boundary accurate prediction;Through Monte Carlo random sampling, introduce random field uncertainty, realize the confidence calculation and risk quantification of stratum layering.The application has the advantages that: accurately capture stratum characteristics and boundary, provide robust and reliable stratum layering solution.

Description

Technical Field

[0001] This invention relates to the technical field of stratigraphic stratification in geotechnical engineering, and in particular to a multi-source data stratigraphic stratification prediction method based on random fields and deep learning. Background Technology

[0002] Stratigraphic delineation plays a crucial foundational role in geotechnical engineering construction. It is not only key to designing foundations and support structures but also provides essential information for analyzing soil behavior and deformation. However, the stratigraphic delineation process is often accompanied by significant uncertainty, especially when field survey data is limited. Improper handling of these uncertainties in stratigraphic delineation and classification can lead to unforeseen problems during construction, impacting project progress and the performance of geotechnical structures. A survey of 28 construction projects in the UK revealed that 22% of geotechnical engineering problems were related to uncertainties in stratigraphic delineation, highlighting the necessity for the proper assessment and quantification of stratification uncertainty.

[0003] CN113945975A discloses a method for jointly inverting stratigraphic structure based on Love waves and Rayleigh waves, belonging to the field of seismic exploration. The technical solution involves converting the obtained surface wave seismic records to obtain Love wave and Rayleigh wave dispersion curves respectively, and then substituting these dispersion curves into the target equation for joint inversion. Compared with existing technologies, this invention has the following obvious advantages: It comprehensively utilizes Love wave and Rayleigh wave signals, greatly improving the accuracy of the inversion and effectively curbing the multiple solutions of the inversion problem. The inversion results of this invention match well with the actual stratigraphic model. However, the data and methods used in this application are completely different, and it only uses one type of data for stratigraphic layering, failing to integrate more comprehensive stratigraphic information.

[0004] CN 118131344 A discloses a method for stratigraphic stratification in an exploration area, comprising the following steps: S10, determining the resistivity of the strata in the exploration area based on electromagnetic observation data collected from the exploration area, and determining the preliminary stratification of the strata and the upper and lower interfaces of each layer based on resistivity logging data or geological drilling lithology logging data of the exploration area; S20, dividing the strata into multiple unit blocks, and filling the unit blocks in the strata with different colors according to the different resistivity values ​​based on the resistivity values ​​determined in S10, forming a resistivity mosaic profile; S30, determining the resistivity values ​​of the upper and lower interfaces of the resistivity mosaic profile based on the preliminary stratification and the resistivity mosaic profile; S40, determining the stratification of the resistivity mosaic profile based on the resistivity values ​​of the upper and lower interfaces, which constitutes the stratigraphic stratification of the exploration area. The method of this invention can improve the accuracy and reliability of stratigraphic stratification. This method is completely different from the method used in this application, and the data source used is only resistivity, without using multi-source fusion data to comprehensively characterize the soil layer.

[0005] CN114880950A discloses a method, apparatus, equipment, and medium for soil quantification and stratification based on XGBoost pore pressure static cone penetration test data. The method includes the following steps: S1, collection and organization of pore pressure static cone penetration test data and soil classification information; S2, conversion of raw static cone penetration test data; S3, establishment of a pore pressure static cone penetration test soil quantification and stratification XGBoost prediction model using a set of calibrated optimal hyperparameters; S4, training the soil quantification and stratification XGBoost prediction model; S5, using the trained soil quantification and stratification XGBoost prediction model to predict the soil type of another site, using a softprob objective function to output the probability of each soil type corresponding to each soil layer; S6, determining the stratification accuracy, merging the stratification results, and finally obtaining the soil quantification and stratification results. This method only performs very coarse stratification and does not reach the sub-layer level. It relies on only one data source for prediction and does not use multi-source data for stratification. It does not calculate the uncertainty of the stratigraphic random field, nor does it calculate the confidence level; therefore, the reliability of the prediction is questionable. This method only discloses a small part of the technical features of the project solution, and the main technical principles and details are different.

[0006] Previous methods of soil layer division mostly used single data and did not use deep learning to explore the correlation between features and soil layers. Therefore, previous methods of soil layer division lacked regional adaptability. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a multi-source data-based stratigraphic stratification prediction method based on random fields and deep learning. First, drilling data, static sounding data, and stratification information are extracted from a geotechnical engineering exploration database to construct two continuous random fields: one uses Hidden Gaussian Grounding (HGB) to capture the overall trend of drilling data along depth and combines Gaussian processes to model nonlinear residuals, generating a continuous random field of drilling physical and mechanical parameters; the other uses optimized HGB and Gaussian processes to generate a continuous random field of static sounding PS values ​​from static sounding data, compensating for the lack of static sounding data at drilling locations. Then, based on the data input from these two continuous random fields, a Transformer-LSTM-CRF hybrid deep learning model is constructed. The Transformer captures the contextual semantic features of drilling and static sounding data, the LSTM further models long-distance dependencies along the depth direction, a multi-head attention mechanism enhances the model's ability to learn local and global features, and the CRF layer optimizes stratification boundary prediction, thereby achieving a unified approach to stratigraphic stratification and accurate boundary prediction. Meanwhile, this method incorporates Monte Carlo random sampling, introducing the uncertainty of continuous random fields into the model prediction process, providing confidence quantification for stratification results, and improving the ability to identify high-risk areas. Compared with traditional Roberson and Hidden Markov Models (HMMs) based on static probe data, this method significantly improves the applicability of stratigraphic stratification prediction, its ability to capture nonlinear relationships, and its ability to quantify uncertainty through multi-source data fusion and deep modeling, providing a more robust and efficient solution for soil layer division, which is crucial in geotechnical engineering design and construction.

[0008] The objective of this invention is achieved through the following technical solutions:

[0009] A multi-source data stratigraphic prediction method based on random fields and deep learning, the method comprising the following steps:

[0010] S1: Obtain drilling data, static exploration data and stratification data from the geotechnical engineering investigation database, and generate drilling data table, static exploration data table and stratification data table;

[0011] S2: Based on the drilling data, a continuous random field of drilling physical and mechanical parameters is generated using the HGB model and Gaussian process;

[0012] S3: Based on the static probe data, a continuous random field of static probe PS values ​​is generated using the HGB model and Gaussian process;

[0013] S4: Based on the continuous random field of the drilling physical and mechanical parameters and the continuous random field of the static exploration PS value, the Transformer model is used to capture the contextual semantic features of the drilling data and the static exploration data, the bidirectional LSTM layer is used to model the long-distance dependency in the depth direction, the multi-head attention mechanism is used to enhance the model's learning ability for local and global features, and the CRF layer is used to optimize the prediction of the strata boundary, thereby achieving the unification of accurate prediction of strata stratification and boundary.

[0014] S5: Monte Carlo random sampling is used to introduce the uncertainty of the continuous random field into the model prediction process, providing a confidence quantification for the hierarchical results;

[0015] S6: Evaluate the model training and prediction results using evaluation metrics for soil layer division and visualization plots.

[0016] In step S1, the StandardScaler function is used to preprocess the drilling data, the static exploration data, and the stratified data to ensure that the feature scale is consistent.

[0017] In step S2, the specific steps are as follows:

[0018] S2.1: Use a hyperparameter-optimized HGB model to capture the trend of the drilling physical and mechanical parameters along the depth;

[0019] S2.2: Use a Gaussian process to capture the residuals between the trend of the drilling physical and mechanical parameters along the depth and the true values.

[0020] In step S3, the specific steps are as follows:

[0021] S3.1: Use the hyperparameter-optimized HGB model to capture the trend of the static probe PS value along the depth;

[0022] S3.2: For each depth point of the static exploration data, select the nearest borehole point to construct a Kriging model, and use a Gaussian process to compensate for the residuals at each depth to capture the geological characteristics at different depth levels, and combine them to obtain the overall prediction results.

[0023] In step S4, the specific steps are as follows:

[0024] S4.1: Align the continuous random field of the drilling physical and mechanical parameters generated in step S2 with the continuous random field of the static PS value generated in step S3. Based on the borehole location and depth information, match the drilling physical and mechanical parameters, the static PS value with the layered data to ensure data consistency and integrity.

[0025] S4.2: Standardize the numerical features of the drilling data and the static exploration data and convert them into strings; perform two-step processing on the layered data: first, perform preliminary mapping of the stratigraphic markers, and then further standardize individual geological layer labels to reduce the diversity and redundancy of labels;

[0026] S4.3: Encode the drilling data and the static exploration data using a pre-trained Transformer model to obtain the contextual embedding representation of the input sequence, thereby extracting deep semantic information and capturing the contextual relationships of geological parameters.

[0027] S4.4: Add a bidirectional LSTM layer to further extract the contextual information of the sequence and capture long-distance dependencies in depth;

[0028] S4.5: Introduced a multi-head attention mechanism, which focuses on different features at different attention heads, enabling the model to better capture the diversity and complexity in geological data;

[0029] S4.6: Add layer normalization and batch normalization to ensure the stability of data distribution during feature extraction at different levels and reduce the gradient vanishing problem;

[0030] S4.7: Add a boundary prediction module, use a linear layer to map the output of the bidirectional LSTM layer to the emission score of the boundary label, use a CRF layer to perform sequence labeling on the boundary label, and capture the dependency relationship between the boundary and the drilling data and the static exploration data;

[0031] S4.8: Introducing L2 regularization and Dropout mechanisms to prevent model overfitting and ensure the model's generalization ability on different geological datasets;

[0032] S4.9: Combine classification loss, CRF loss, and L2 regularization to obtain the loss function;

[0033] S4.10: The dataset is partitioned using borehole numbers to improve the model's generalization ability and avoid overfitting;

[0034] S4.11: Gradient Clipping is used to control the gradient norm and avoid gradient explosion; the ReduceLROnPlateau learning rate scheduler is introduced to automatically reduce the learning rate when the validation loss does not decrease significantly, so as to achieve more refined model optimization and improve the convergence ability of the model in the later training stage.

[0035] S4.12: During model training, the model state and validation loss for the last few training rounds are recorded. By comparing these, the optimal model state is selected and saved, ensuring the final performance of the model.

[0036] In step S5, the specific steps are as follows:

[0037] Each test sample is sampled multiple times to simulate different possibilities of the input features, estimate the model's prediction distribution, and calculate the prediction probability for each category, thereby providing a measure of uncertainty for different tiers.

[0038] In step S6, the specific steps are as follows:

[0039] The model's prediction performance was evaluated using the F1 score, accuracy, and recall. The model's prediction results were further evaluated using the loss function's variation with the number of training epochs, the ROC curve, the confusion matrix, and the distribution of actual predictions with depth.

[0040] The advantages of this invention are:

[0041] 1) Advantages of multi-source data fusion: The Roberson method and HMM usually only rely on static drilling data for stratigraphic stratification, making it difficult to comprehensively consider information from other data sources. This may cause the model to ignore physical and mechanical parameters (such as density, void ratio, shear strength, etc.) contained in the drilling data, thus limiting the comprehensive utilization of soil stratification information. This method combines drilling data with static drilling data and constructs two continuous random fields (random field of drilling physical and mechanical parameters and random field of static drilling PS values) to achieve alignment and complementarity of multi-source data, improve the ability to characterize soil properties, and comprehensively utilize effective information for soil stratification.

[0042] 2) Separating the trend from the residuals allows the model to learn both macro trends and micro local variations simultaneously, resulting in better fitting ability and interpretability.

[0043] 3) Parameter search is performed using Bayesian optimization, which improves the efficiency of hyperparameter selection and avoids the blindness of manual parameter tuning, especially suitable for models with large parameter spaces;

[0044] 4) By using the standard deviation in the Gaussian process, the uncertainty range of each prediction point can be given, which can clearly characterize the confidence level of the model at different depths and help to understand the reliability of the prediction. This uncertainty estimation based on the prediction standard deviation provides an important reference for geological engineering decision-making, enabling engineers to better weigh risks.

[0045] 5) The supervised training-obtained multi-source data multi-index prediction model GEO-MULTI for underground space was used to verify the original data. The statistical box plot method was used to identify potential outliers by comparing the model's prediction results with the measured values.

[0046] 6) The results show that the overall accuracy for classification prediction is 1.00, achieving a very good prediction effect; the regression prediction R for different target values ​​is also very good. 2 A relatively ideal prediction accuracy is achieved when the value is between 0.91 and 0.92. Attached Figure Description

[0047] Figure 1 This is a flowchart of the multi-source data stratigraphic prediction method based on random fields and deep learning according to the present invention;

[0048] Figure 2 This is a schematic diagram of the overall structure of the deep learning model of the present invention;

[0049] Figure 3 This is a distribution diagram of the boreholes and static boreholes of the present invention;

[0050] Figure 4 This is a schematic diagram of the continuous random field of compression modulus of the present invention;

[0051] Figure 5 This is a schematic diagram of the heavily continuous random field of the present invention;

[0052] Figure 6 This is a schematic diagram of the continuous random field of cohesion in this invention;

[0053] Figure 7 This is a schematic diagram of the continuous random field of internal friction angle of the present invention;

[0054] Figure 8 This is a schematic diagram of the continuous random field of water content according to the present invention;

[0055] Figure 9 This is a schematic diagram of the continuous random field for static PS value detection according to the present invention;

[0056] Figure 10 This is a schematic diagram of the confusion matrix between the predicted classification and the measured classification of the present invention;

[0057] Figure 11 This is a schematic diagram of the ROC curves of the soil layer classification prediction results of this invention.

[0058] Figure 12 This is a schematic diagram showing the comparison between borehole-predicted soil layers and artificial soil layers, as well as the prediction confidence level, according to the present invention.

[0059] Figure 13This is a table showing the calculation results of the prediction and evaluation indicators for each stratum category in this invention. Detailed Implementation

[0060] The features and other related features of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments, so as to facilitate understanding by those skilled in the art:

[0061] Example: Figure 1 As shown, this embodiment relates to a multi-source data stratigraphic prediction method based on random fields and deep learning. The method mainly includes the following steps:

[0062] S1: As Figure 3 As shown, the original drilling data (soil layer parameters), static test data (soil layer parameters), and stratification data of each exploration project are obtained from the geotechnical engineering exploration database. Drilling data table, static test data table, and stratification data table are generated. The StandardScaler function is used to preprocess the drilling data, static test data, and stratification data to ensure that the characteristic scale is consistent.

[0063] The drilling data includes compression modulus, unit weight, cohesion, internal friction angle, water content, and borehole coordinates, while the static exploration data includes static exploration PS value and borehole coordinates.

[0064] S2: As Figure 2 as well as Figures 4-8 As shown, based on drilling data, a continuous random field of drilling physical and mechanical parameters is generated through the HGB (HistGradientBoosting) model and Gaussian Process (GP). That is, a continuous random field of drilling physical and mechanical parameters with continuous distribution in the depth direction is generated using drilling data that is discontinuous in the depth direction.

[0065] In step S2, the specific steps are as follows:

[0066] S2.1: Use the hyperparameter-optimized HGB model to capture the trend of drilling physical and mechanical parameters along depth.

[0067] The HGB model learns complex functions through incremental tree construction, making it suitable for capturing the main trend of data. For drilling data, the Optuna hyperparameter optimization framework is used to optimize the hyperparameters of the HGB model. This optimization method determines the optimal hyperparameters, ensuring that the trend model achieves optimal performance on drilling data.

[0068] S2.2: Use Gaussian processes to capture the residuals between the trend of drilling physical and mechanical parameters along depth and the true values.

[0069] Among these methods, Bayesian optimization is used for parameter search, which improves the efficiency of hyperparameter selection and avoids the blindness of manual parameter tuning, making it particularly suitable for models with large parameter spaces. To address the nonlinear characteristics of drilling data, different kernel functions are used to capture more complex nonlinear local features, ensuring that the Gaussian process can capture complex local residuals that the trend model fails to explain.

[0070] Specifically, the HGB model learns complex functions through incremental tree construction:

[0071]

[0072] In the formula, M represents the total number of weak learners (regression trees), η is the learning rate, which controls the contribution of each tree, and is optimized using the Optuna hyperparameter optimization framework. m (x) represents the output of the m-th tree, i.e., the main trend of soil parameters.

[0073] The HGB model is optimized with the following hyperparameters: learning_rate, with a value range of [0.3, 0.5]; max_depth, which controls the maximum depth of the tree and is used to limit the model complexity, with a value range of [15, 20]; min_samples_leaf, which is the minimum number of samples per leaf node, with a value range of [20, 25]; max_iter, which is the number of iterations (i.e., the maximum number of trees), with a value range of [500, 5000]; and l2_regularization, which is the L2 regularization coefficient used to prevent overfitting, with a value range of [0.0, 1.0].

[0074] To address the nonlinear characteristics of drilling data, different kernel functions are used to capture more complex nonlinear local features, ensuring that the Gaussian process can capture complex local residuals that the trend model fails to explain.

[0075]

[0076] In the formula, K is the kernel matrix, which is composed of kernel functions.

[0077] This embodiment uses the following combination of kernel functions:

[0078] k(x,x')=C·RBF(x,x')+WhiteKernel(x,x');

[0079] ConstantKernel is used to define the global magnitude of the predicted values, and RBF radial basis function kernel is used to capture the smooth correlation of input features.

[0080]

[0081] In the formula, ||x-x'|| is the Euclidean distance of the input features, and length_scale is the length scale, which determines the decay rate of the correlation.

[0082] The WhiteKernel is used to represent observation noise. Maximum likelihood estimation (MLE) is used to optimize the parameters of the Gaussian process.

[0083] S3: As Figure 2 and Figure 9 As shown, based on static exploration data, a continuous random field of static exploration PS values ​​is generated through the HGB model and Gaussian process. That is, a continuous random field of static exploration PS values ​​continuously distributed along the horizontal direction at each depth is generated using horizontally discontinuous static exploration data to make up for the problem of missing static exploration data at drilling locations, so as to better achieve the fusion and alignment of multi-source data.

[0084] In step S3, the specific steps are as follows:

[0085] S3.1: Use the hyperparameter-optimized HGB model to capture the trend of static probe PS values ​​along depth.

[0086] Specifically, the Optuna hyperparameter optimization framework was used to optimize the hyperparameters of HGB based on static probe data. During the optimization of the HGB model, KFold cross-validation combined with parallel computation (via joblib's Parallel) was used, taking into account the characteristics of static probe data, to ensure the robustness and generalization ability of the hyperparameters while significantly accelerating model training.

[0087] S3.2: For each depth point in the static exploration data, select the nearest borehole point to construct a Kriging model, and use a Gaussian process to compensate for the residuals at each depth to capture the geological characteristics at different depth levels, and combine them to obtain the overall prediction results.

[0088] Specifically, the Gaussian process model was trained using the Adam optimizer and the ExactMarginalLogLikelihood (used to calculate the marginal log-likelihood of the Gaussian process) as the loss function. This approach enables the model to better fit the nonlinear residuals of static exploration data and exhibits strong adaptability in formations at different depths.

[0089] S4: As Figure 2As shown, based on the continuous random field of drilling physical and mechanical parameters and the continuous random field of static exploration PS values, the Transformer model is used to capture the contextual semantic features of drilling and static exploration data. The bidirectional LSTM (Long Short-Term Memory) layer is used to model the long-distance dependency in the depth direction. The multi-head attention mechanism is used to enhance the model's learning ability for local and global features. The CRF (Conditional Random Field Layer) layer is used to optimize the prediction of the stratification boundary, thereby achieving the unification of strata stratification and boundary accurate prediction.

[0090] In step S4, the specific steps are as follows:

[0091] S4.1: Align the continuous random field of drilling physical and mechanical parameters generated in step S2 with the continuous random field of static PS values ​​generated in step S3. Based on the borehole location and depth information, match the drilling physical and mechanical parameters, static PS values ​​with the layered data to ensure data consistency and integrity.

[0092] S4.2: Standardize the numerical characteristics of drilling and static exploration data and convert them into strings; perform two-step processing on the strata data: first, perform preliminary mapping of strata markers, and then further standardize individual geological stratum labels to reduce label diversity and redundancy.

[0093] S4.3: The pre-trained Transformer model is used to encode the drilling and static exploration data to obtain the contextual embedding representation of the input sequence, thereby extracting deep semantic information and capturing the contextual relationships of geological parameters. The output of the Transformer model is used as the input for subsequent layers.

[0094] S4.4: A bidirectional LSTM layer is added to further extract contextual information from the sequence and capture long-range dependencies at depth. The bidirectional LSTM layer can effectively handle the relationships between adjacent time points in the sequence, enhancing the understanding of the strata.

[0095] S4.5: A multi-head attention mechanism is introduced, which focuses on different features at different attention heads, enabling the model to better capture the diversity and complexity in geological data. Combining the multi-head attention mechanism with a bidirectional LSTM layer enhances the model's balance between capturing global information and local features.

[0096] S4.6: Add layer normalization and batch normalization to ensure the stability of data distribution during feature extraction at different levels and reduce the gradient vanishing problem.

[0097] S4.7: Add a boundary prediction module, using a linear layer to map the output of the bidirectional LSTM layer to the emission score of the boundary label (to reflect the model's confidence in each label), and use a CRF layer to perform sequence labeling on the boundary labels to capture the dependency between the boundary and drilling and static exploration data.

[0098] S4.8: Introduces L2 regularization and Dropout mechanisms to prevent model overfitting and ensure the model's generalization ability on different geological datasets.

[0099] S4.9: Combine classification loss, CRF loss, and L2 regularization to obtain the loss function.

[0100] S4.10: The dataset is partitioned using borehole numbering. This ensures that the model does not see the same borehole data on the training and test sets, thereby improving the model's generalization ability and avoiding overfitting.

[0101] S4.11: Gradient Clipping is used to control the gradient norm and avoid gradient explosion, especially when training deep models, to ensure the stability of the model training process; the ReduceLROnPlateau learning rate scheduler is introduced to automatically reduce the learning rate when the validation loss does not decrease significantly, so as to achieve more refined model optimization and improve the convergence ability of the model in the later training stage.

[0102] S4.12: During model training, the model state and validation loss are recorded in the last few training epochs. By comparing these, the optimal model state is selected and saved, ensuring the final performance of the model.

[0103] S5: Monte Carlo random sampling is used to introduce the uncertainty of continuous random fields into the model prediction process, providing a confidence quantification for the stratified results.

[0104] In step S5, the specific steps are as follows:

[0105] By sampling each test sample multiple times to simulate different probabilities of the input features, estimating the model's prediction distribution, and calculating the prediction probability for each category, an uncertainty measure for different stratigraphic layers is provided. This probability distribution not only identifies the most likely layer but also reflects the degree of uncertainty in the model's stratigraphic classification, helping engineers conduct risk assessments and optimize decisions.

[0106] S6: Evaluate the model training and prediction results using evaluation metrics for soil layer division and visualization plots.

[0107] In step S6, the specific steps are as follows:

[0108] The model's prediction performance was evaluated using F1 score, accuracy, and recall. The results were further assessed using the loss function versus training epochs curve, ROC curve, confusion matrix, and the distribution of predicted values ​​with depth. The confusion matrix of the predicted names is shown below. Figure 10 As shown in the figure. The ROC curves of the prediction results for each soil layer classification are as follows. Figure 11 As shown in the bar chart, the classification prediction accuracy evaluation scores for each soil layer are as follows. Figure 12 As shown.

[0109] like Figure 13 As shown, the weighted average of the prediction evaluation index calculation results for each stratum category is: accuracy 0.91, precision 0.82, recall 0.85, and f1 value 0.83.

[0110] Specifically, Kriging interpolation, as a traditional spatial statistical method, is often used for spatial interpolation of data from limited exploration points (such as boreholes and static penetration test points), but it has significant limitations. The Kriging method primarily assumes the stationarity and linear correlation of variables, thus performing poorly when dealing with areas with complex geological conditions or significant nonlinear variations. Furthermore, Kriging interpolation focuses more on generating "optimal estimates," and is relatively weak in quantifying and predicting interpolation uncertainties, making it difficult to meet the needs of probabilistic analysis. More importantly, Kriging interpolation relies solely on spatial location correlation, failing to effectively combine regional historical exploration experience and prior knowledge, thereby limiting its predictive accuracy.

[0111] To overcome these shortcomings, the Kriging method, which combines machine learning techniques with Gaussian processes, has emerged as a more promising alternative. This approach leverages machine learning's ability to capture nonlinear relationships and Gaussian processes' ability to model local features, enabling the generation of more refined random fields and more accurate quantification and prediction of uncertainties in stratigraphic division.

[0112] In stratified prediction, relying solely on static borehole data (such as PS values) has certain limitations. While static borehole data is densely distributed, it primarily reflects the resistance characteristics of soil layers and cannot directly characterize their physical and mechanical properties. Drilling data, on the other hand, contains key parameters such as density, void ratio, and shear strength, providing a more comprehensive characterization of soil layer properties. Furthermore, empirical classification charts based on static borehole data (such as SBT charts) are developed based on global data and may perform poorly in specific sites, while drilling data can provide regionalized information to optimize stratification results. Therefore, combining drilling and static borehole data to construct a continuous random field and aligning the two can significantly improve the model's accuracy in characterizing soil properties.

[0113] By combining machine learning and Kriging Gaussian processes, two continuous random fields can be constructed: one is a physical-mechanical parameter random field based on drilling data, which captures the overall trend in the depth direction through HistGB and compensates for nonlinear residuals using Gaussian processes, thus aligning with the stratified data; the other is a PS value random field based on static drilling data, which predicts the static drilling values ​​at the drilling location using static drilling data, thereby compensating for the lack of static drilling data at drilling points. Ultimately, the combination of the two random fields can generate a complete drilling-static drilling dataset, which, when combined with the stratified data, forms a comprehensive characterization of the formation.

[0114] To achieve more accurate stratigraphic prediction, a deep learning-based Transformer-LSTM-CRF model was introduced. The Transformer model captures the complex relationships hidden in drilling and static exploration data through contextual embedding, extracting deep semantic information. The LSTM model further models long-distance dependencies along the depth direction, making it particularly suitable for reflecting continuous changes in strata along depth. The CRF model, through global optimization of sequence labeling, effectively addresses the problem of ambiguous stratigraphic boundaries, significantly improving the prediction accuracy of stratigraphic boundaries. The accuracy of stratigraphic boundaries is crucial for stratigraphic delineation, as only accurate boundaries can ensure correct stratigraphic results.

[0115] Furthermore, the Transformer-LSTM-CRF hybrid model demonstrates superior capabilities in prediction processes based on Monte Carlo random sampling. This model captures the residuals of the input continuous random field, incorporating the uncertainty of the continuous random field into the stratification prediction, providing prediction probabilities for each category, and calculating the confidence level of the final stratification results as a function of depth. This not only helps identify the most probable strata number but also quantifies the uncertainty of the model in stratigraphic classification, providing engineers with a basis for decision optimization.

[0116] Capturing uncertainty is crucial for soil layer delineation. On one hand, it reduces construction risks and avoids unexpected problems caused by incorrect soil layer delineation; on the other hand, it improves the robustness of the model under different site conditions, making the prediction results more reliable. Furthermore, quantifying uncertainty provides key input for probabilistic analysis, facilitating further reliability analysis in slope stability assessments or foundation design.

[0117] In summary, by combining drilling and static exploration data, generating continuous random fields through machine learning and Kriging Gaussian processes, and using a Transformer-LSTM-CRF model for stratification prediction, this approach not only significantly improves the accuracy and reliability of soil stratification but also provides a powerful tool for quantifying uncertainties in stratification. This method can comprehensively optimize the decision-making process in geotechnical engineering design and construction, laying a solid foundation for improving engineering quality and reducing risks.

[0118] The beneficial technical effects of this embodiment are as follows:

[0119] 1) Advantages of multi-source data fusion: The Roberson method and HMM usually only rely on static drilling data for stratigraphic stratification, making it difficult to comprehensively consider information from other data sources. This may cause the model to ignore physical and mechanical parameters (such as density, void ratio, shear strength, etc.) contained in the drilling data, thus limiting the comprehensive utilization of soil stratification information. This method combines drilling data with static drilling data and constructs two continuous random fields (random field of drilling physical and mechanical parameters and random field of static drilling PS values) to achieve alignment and complementarity of multi-source data, improve the ability to characterize soil properties, and comprehensively utilize effective information for soil stratification.

[0120] 2) Separating the trend from the residuals allows the model to learn both macro trends and micro local variations simultaneously, resulting in better fitting ability and interpretability.

[0121] 3) Parameter search is performed using Bayesian optimization, which improves the efficiency of hyperparameter selection and avoids the blindness of manual parameter tuning, especially suitable for models with large parameter spaces;

[0122] 4) By using the standard deviation in the Gaussian process, the uncertainty range of each prediction point can be given, which can clearly characterize the confidence level of the model at different depths and help to understand the reliability of the prediction. This uncertainty estimation based on the prediction standard deviation provides an important reference for geological engineering decision-making, enabling engineers to better weigh risks.

[0123] 5) The supervised training-obtained multi-source data multi-index prediction model GEO-MULTI for underground space was used to verify the original data. The statistical box plot method was used to identify potential outliers by comparing the model's prediction results with the measured values.

[0124] 6) The results show that the overall accuracy for classification prediction is 1.00, achieving a very good prediction effect; the regression prediction R for different target values ​​is also very good. 2 A relatively ideal prediction accuracy is achieved when the value is between 0.91 and 0.92.

[0125] Although the above embodiments have described the concept and embodiments of the present invention in detail with reference to the accompanying drawings, those skilled in the art will recognize that various improvements and modifications can still be made to the present invention without departing from the scope of the claims, and therefore will not be elaborated here.

Claims

1. A method for multi-source data stratigraphic layering prediction based on random fields and deep learning, characterized in that The method comprises the following steps: S1: obtaining drilling data, static exploration data and stratification data from a geotechnical engineering investigation database to generate drilling data table, static exploration data table and stratification data table; S2: based on the drilling data, a continuous random field of drilling physical and mechanical parameters is generated by an HGB model and a Gaussian process; S3: based on the static exploration data, a continuous random field of static exploration PS values is generated by an HGB model and a Gaussian process; S4: a Transformer-LSTM-CRF hybrid deep learning model is constructed; based on the continuous random field of drilling physical and mechanical parameters and the continuous random field of static exploration PS values, the context semantic features of the drilling data and the static exploration data are captured by a Transformer model, the long-distance dependence in the depth direction is modeled by a bidirectional LSTM layer, the learning ability of the model for local and global features is enhanced by a multi-head attention mechanism, and the stratification boundary prediction is optimized by a CRF layer, so that the unification of stratification and boundary accurate prediction is realized; In step S4, the specific steps are as follows: S4.1: align the continuous random field of drilling physical and mechanical parameters and the continuous random field of static exploration PS values, and match the drilling physical and mechanical parameters, the static exploration PS values and the stratification data according to the drilling position and depth information; S4.2: using a pre-trained Transformer model to encode the drilling data and the static exploration data, for obtaining the context embedding representation of the input sequence, to extract deep semantic information and capture the context relationship of geological parameters; S4.3: adding a bidirectional LSTM layer to further extract the context information of the sequence and capture the long-distance dependence in the depth direction; S4.4: introducing a multi-head attention mechanism, which can better capture the diversity and complexity of geological data by focusing on different features in different attention heads; S4.5: adding layer normalization and batch normalization to ensure the stability of data distribution during feature extraction at different levels and reduce the gradient vanishing problem; S4.6: adding a boundary prediction module, using a linear layer to map the output of the bidirectional LSTM layer to the boundary label, using a CRF layer to perform sequence labeling on the boundary label, and capturing the dependence between the boundary and the drilling data and the static exploration data; S5: using Monte Carlo random sampling to introduce the uncertainty of the continuous random field into the prediction process of the hybrid deep learning model, and providing confidence quantification for the stratification result; S6: using evaluation indexes for soil layer division and visualization to evaluate the training and prediction results of the hybrid deep learning model.

2. The multi-source data stratigraphic layer prediction method based on random fields and deep learning according to claim 1, characterized in that In step S1, the StandardScaler function is used to preprocess the drilling data, the static exploration data and the stratification data to ensure consistent feature scales.

3. The method of claim 1, wherein the method further comprises: determining a plurality of initial layer boundaries based on the plurality of initial layer boundaries and the plurality of initial layer boundaries determined by the deep learning model; and determining a plurality of final layer boundaries based on the plurality of initial layer boundaries and the plurality of initial layer boundaries determined by the deep learning model. In step S2, the specific steps are as follows: S2.1: using an HGB model optimized by hyperparameters to capture the trend of the drilling physical and mechanical parameters along the depth; S2.2: Capture the residual between the trend of the drilling physical and mechanical parameters along the depth and the true value using Gaussian process.

4. The method of claim 3, wherein the method further comprises: determining a plurality of initial layer boundaries based on the plurality of initial layer boundaries and the plurality of initial layer thicknesses; and determining a plurality of final layer boundaries based on the plurality of initial layer boundaries and the plurality of initial layer thicknesses. In step S3, the specific steps are as follows: S3.1: Capture the trend of the static exploration PS value along the depth using the HGB model optimized by hyperparameters; S3.2: For each depth point of the static exploration data, select the nearest drilling point to construct the Kriging model, compensate the residual of each depth by Gaussian process to capture the geological characteristics of different depth levels, and combine to obtain the overall prediction result.

5. The method of claim 4, wherein the method further comprises: In step S5, the specific steps are as follows: Multiple sampling is performed on each test sample to simulate different possibilities of input features, estimate the model prediction distribution, calculate the prediction probability of each category, and thus provide uncertainty measurement of different layer numbers.

6. The method of claim 5, wherein the method further comprises: In step S6, the specific steps are as follows: The prediction effect of the model is evaluated using f1 value, accuracy, recall rate, and the prediction result of the model is evaluated using loss function curve changing with training round number, roc curve, confusion matrix, and prediction measured distribution along depth.

Images