Tunnel stratum modeling method and system based on multi-source feature driving and spatial constraint

Abstract

The application provides a tunnel stratum modeling method and system based on multi-source feature driving and space constraint, which comprises the following steps: obtaining the position, stratum category, water content, plasticity index and liquidity index of a drilling sample point in a tunnel engineering area, discretizing the area into a horizontal-depth two-dimensional grid and defining a stratum marker field; training a stratum classification model based on standardized multi-source rock-soil features, establishing a mapping of features and stratum category probability; reconstructing a continuous feature field according to an anisotropic distance weight of a horizontal correlation scale greater than a vertical correlation scale, and inputting the classification model to generate a stratum category probability field; further constructing a Markov random field data item from the probability field, combining transverse adaptive smoothing, vertical sequence constraint and drilling hard constraint for discrete optimization to obtain a tunnel stratum model; and simultaneously calculating an uncertainty index and associating the output; the application can better improve the continuity, sequence rationality and prediction reliability of stratum modeling under the condition of sparse drilling.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering and underground engineering geological modeling technology, and in particular to a tunnel stratum modeling method and system based on multi-source feature driving and spatial constraints. Background Technology

[0002] Before tunnel construction, a geological model is typically built based on borehole data to support route design, selection of support parameters, and assessment of construction risks. In actual projects, the number of boreholes is often sparsely distributed due to limitations in site, cost, and schedule, making it difficult to fully reflect the geological variations between boreholes. Especially in soft soil, reclaimed areas, or areas with significant lateral geological changes, geological interfaces may exhibit undulations, lenses, or local abrupt changes. Relying solely on manual profile extension or conventional interpolation methods can easily result in overly smooth interfaces or omission of local structures.

[0003] While existing stratigraphic modeling methods can use spatial interpolation or machine learning models to predict inter-bore areas, they still have shortcomings: On the one hand, conventional interpolation methods often fail to take into account both the horizontal continuity of strata and the anisotropic characteristics of rapid changes in the depth direction, which may lead to unreasonable mixing of different strata in the vertical direction; on the other hand, simple data-driven classification models mainly rely on local geotechnical parameters to make category judgments, lacking constraints on the horizontal continuity of strata, vertical sequence, and consistency of borehole measured labels, and the prediction results are prone to discrete patches, interface fractures, or layer skipping phenomena.

[0004] In addition, existing methods mostly output deterministic stratigraphic category distributions, but they are insufficient in expressing the uncertainty of predictions for stratigraphic boundaries, sparsely drilled areas, and areas with multiple competing categories, making it difficult to provide a reliable basis for subsequent supplementary exploration and construction risk control.

[0005] Therefore, there is an urgent need for a tunnel stratum modeling method and system based on multi-source feature-driven and spatial constraints, which can integrate multi-source soil and rock features and spatial constraints under sparse borehole conditions to generate a continuous and reliable tunnel stratum model and quantify uncertainty. Summary of the Invention

[0006] To address the aforementioned shortcomings of the existing technology, the purpose of this invention is to provide a tunnel stratum modeling method and system based on multi-source feature-driven and spatial constraints, aiming to solve the technical problems of poor continuity, unstable sequence, and difficulty in quantifying prediction uncertainty in tunnel stratum models under sparse borehole conditions.

[0007] To achieve the above objectives, this invention provides a tunnel stratum modeling method based on multi-source feature-driven and spatial constraints, comprising the following steps: Obtain borehole sample data within the tunnel engineering area. The borehole sample data includes the spatial location of the sample, the stratum category label, and the water content, plasticity index, and liquidity index. The tunnel engineering area is discretized into a horizontal-depth two-dimensional grid domain, and a stratigraphic label field to be solved is defined within the two-dimensional grid domain. Multi-source soil and rock characteristics are constructed using the water content, plasticity index, and liquidity index of borehole samples. The multi-source soil and rock characteristics are standardized and combined with the corresponding stratigraphic category labels to train a stratigraphic classification model (Data-driven Virtual Borehole, DVB) to establish a mapping relationship between multi-source soil and rock characteristics and stratigraphic category probabilities. Based on the spatial location of borehole samples and the standardized multi-source soil and rock characteristics, the features of each grid point to be estimated in the two-dimensional grid domain are reconstructed according to the anisotropic distance weight to obtain a continuous feature field. The horizontal correlation scale of the anisotropic distance weight is greater than the vertical correlation scale. Input the continuous feature field into the trained stratigraphic classification model to obtain the stratigraphic category probability field of each grid point; The data terms of the energy function of the Markov Random Field (MRF) are constructed using the stratigraphic category probability field, and a spatial constraint optimization model is formed by combining the lateral adaptive smoothing term and the vertical sequence constraint term. Hard constraints consistent with the measured stratigraphic category labels are applied to the grid points corresponding to the borehole sample points, and the spatial constraint optimization model is discretized to obtain the tunnel stratigraphic model. The uncertainty index of each grid point is calculated based on the probability field of the stratum category, and the uncertainty index is output in association with the tunnel stratum model.

[0008] As a further improvement to the above technical solution, the two-dimensional mesh domain is represented as: ; Where Ω represents the two-dimensional grid domain corresponding to the tunnel engineering area, i represents the grid number in the depth direction, j represents the grid number in the horizontal direction, m represents the number of discrete elements in the depth direction, and n represents the number of discrete elements in the horizontal direction. The stratigraphic marker field is represented as follows: ; Where L represents the stratigraphic marker field, L(i,j) represents the stratigraphic category at grid point (i,j), and K represents the total number of stratigraphic categories. Different stratigraphic categories are numbered according to a preset stratigraphic sequence.

[0009] As a further improvement to the above technical solution, the multi-source soil and rock characteristic vector of the borehole sample points is represented as follows: ; Where x represents the multi-source soil and rock characteristic vector, w represents the water content, PI represents the plasticity index, and LI represents the liquidity index; To eliminate the influence of different characteristic dimensions and numerical ranges, the d-th soil and rock characteristic is standardized according to the following formula: ; in, and The first The mean and standard deviation of each feature. The d-th standardized soil and rock characteristic value is used; after processing, the standardized feature vector is obtained. , as input to the stratigraphic classification model.

[0010] As a further improvement to the above technical solution, the stratigraphic classification model is a multilayer perceptron (MLP) model, and outputs category scores according to the following formula: ; in It is a nonlinear mapping function composed of a multi-layer fully connected network; These are model parameters; Represents the category score vector, This represents the score corresponding to stratigraphic category c; This represents the standardized multi-source soil and rock feature vector; The category scores are converted into stratum category probabilities using the Softmax function as follows: ; in, This indicates that the sample points are within a given standardized multi-source soil and rock characteristics. L represents the probability that a given stratigraphic class belongs to stratigraphic class c, where L represents the stratigraphic class variable and c represents the candidate stratigraphic class. This represents the score corresponding to category c. This represents the score corresponding to category k, where K represents the total number of stratigraphic categories; This is a temperature parameter used to adjust the smoothness of the probability distribution. When When the probability distribution is small, it tends to be sharp, which is beneficial for strengthening classification decisions; when... When the value is large, the probability distribution is smoother, which is beneficial for expressing uncertainty. This model can be used to obtain the probability distribution of stratigraphic categories for each location.

[0011] As a further improvement to the above technical solution, a weighted cross-entropy loss function is used when training the stratigraphic classification model: ; in, Represents the loss function. This represents the category weight corresponding to stratigraphic category c. This represents the one-hot encoded value of the true label in category c; the category weight is determined according to the following formula: ;in, This represents the number of samples of formation category c in the borehole sampling points; by minimizing the loss function and optimizing the model parameters θ, the optimal classification model is obtained.

[0012] As a further improvement to the above technical solution, the set of pore points is represented as: ; in, This represents the set of borehole sampling points. For the first Spatial location of each borehole sample point; is the multi-source soil and rock feature vector of the kth borehole sample point; This indicates the measured stratigraphic category label for the k-th borehole sample point; Indicates the total number of borehole samples; For any grid point (i,j) to be estimated, its multi-source soil and rock characteristics are reconstructed according to the following formula: ; in, Let represent the estimated multi-source soil and rock characteristic vector at the grid point (i,j) to be estimated. This represents the normalized spatial weight of the k-th borehole sample point for the estimated grid point (i,j). This represents the multi-source soil and rock feature vector of the k-th borehole sample point.

[0013] As a further improvement to the above technical solution, considering that the strata have obvious directional characteristics in space: that is, strong horizontal continuity and more significant vertical variation; the normalized spatial weight is determined by the following anisotropic distance weight function: ; in It is a horizontally correlated scale; For vertically related scales; The distance decay exponent; > To enhance horizontal continuity and suppress excessive diffusion in the depth direction; To ensure the weights are normalized, the anisotropic distance weight function is normalized: ; in, This represents the normalized spatial weights.

[0014] As a further improvement to the above technical solution, the continuous feature field is represented as: ; Represents a continuous characteristic field within a two-dimensional grid domain. Represents grid points The estimated multi-source soil and rock feature vectors are obtained at the location; the continuous feature field is input into the stratigraphic classification model to obtain the following probability field: ; in, Let (i,j) represent the probability distribution of the stratigraphic category at grid point (i,j). This represents the probability that grid point (i,j) belongs to stratigraphic category c; The mapping relationship between multi-source soil and rock characteristics and stratigraphic category probabilities is shown below: .

[0015] As a further improvement to the above technical solution, the optimization objective of the Markov random field energy function is: ; in, This represents the optimized stratigraphic marker field. This represents the stratigraphic marker field to be optimized. The total energy function representing the stratigraphic marker field; The total energy function is expressed as: ; in, Represents grid points Data items, This represents the lateral smoothing term between horizontally adjacent grid points p and q. Represents grid points Vertical formation constraint terms, It represents the set of neighborhoods in the horizontal direction.

[0016] As a further improvement to the above technical solution, the data item is constructed according to the following formula: ; in, Represents grid points Stratigraphic Class Energy of data items at time This indicates the estimation of multi-source soil and rock eigenvectors. Below, grid points Stratigraphic category The probability. The negative logarithmic form is used to convert probability into energy. These values ​​embody data-driven information: the higher the probability, the lower the energy, and the more inclined one is to choose that category; the lower the probability, the higher the energy, and the less inclined one is to choose that category. The horizontal smoothing term is constructed according to the following formula: ; in, Indicates horizontally adjacent grid points and Energy of smoothing terms between them Represents grid points The adaptive smoothing coefficient at the location, Indicates an indicator function, Represents grid points Stratigraphic categories, Represents grid points Stratigraphic categories; Considering that type variation should be allowed at formation interfaces while homogeneous regions should remain continuous, the adaptive smoothing coefficient is determined according to the following formula: ; in, Represents grid points The adaptive smoothing coefficient at the location, Represents the basic smoothing coefficient. Represents grid points The probability of the maximum stratigraphic class at that location. Represents grid points Estimate the spatial gradient of multi-source soil and rock eigenvectors. The norm of the spatial gradient is represented. Indicates the adjustment parameter; The vertical sequence constraint term is constructed according to the following formula: ; in, Represents grid points Energy of the vertical sequence constraint term at the location, Indicates the stratigraphic type of the current depth grid point. Indicates the stratigraphic category of adjacent grid points at the previous depth. This represents the penalty coefficient applied when the preset vertical sequence relationship is violated; To avoid discontinuous skipping of layers between adjacent grid points in the depth direction, a difference penalty is introduced in the vertical sequence constraint term: ;in, This represents the difference penalty function. This represents the difference in stratigraphic category number between the current depth grid point and the adjacent grid point at the previous depth.

[0017] As a further improvement to the above technical solution, the hard constraint on the location of the borehole sample point is expressed as follows: ; in, Indicates the first The location of each borehole sample point This indicates the optimal stratigraphic category for that location. Indicates the first The measured formation category labels for each borehole sample point; the hard constraints remain valid throughout the energy function optimization process. The hard constraints remain valid throughout the optimization process, ensuring consistency between the results and measured data; providing spatial anchors to improve overall stability; and suppressing error propagation.

[0018] As a further improvement to the above technical solution, when performing discrete optimization on the spatially constrained optimization model, an iterative conditional mode algorithm is adopted, and the stratigraphic category of the grid points is updated according to the following formula: ; in, Indicates the first After the second iteration, the grid points Stratigraphic categories, Indicates the number of iterations. Indicates the candidate stratigraphic category, Represents grid points Select candidate stratigraphic categories Local energy at time; The initialization of the iterative conditional pattern algorithm is performed according to the following formula: ; in, Represents grid points The initial stratigraphic category at the location, Represents grid points Stratigraphic category The probability of [the probability]; iteration stops when the following condition is met or the maximum number of iterations is reached: ; in, Indicates the first Stratigraphic marker field after the next iteration This represents the convergence threshold.

[0019] As a further improvement to the above technical solution, the uncertainty index includes information entropy, which is calculated according to the following formula: ; in, Represents grid points The stratigraphic category prediction information entropy at that location Represents grid points Stratigraphic category The probability, This represents the total number of stratigraphic categories; the higher the information entropy, the higher the uncertainty in predicting the stratigraphic category at that grid point. The uncertainty index also includes classification margin, which is calculated according to the following formula: ; in, Represents grid points Classification margin at the location, Represents grid points The probability of the maximum stratigraphic class at that location. Represents grid points The probability of the second largest stratigraphic category at a given grid point; the smaller the classification margin, the lower the stability of the stratigraphic category prediction at that grid point.

[0020] Secondly, the present invention also provides a tunnel stratum modeling system based on multi-source feature-driven and spatial constraints, comprising: The data acquisition module is used to acquire borehole sample data within the tunnel engineering area. The borehole sample data includes the spatial location of the sample, the stratum category label, and the water content, plasticity index, and liquidity index. The mesh modeling module is used to determine the horizontal-depth range of the tunnel engineering area based on the spatial location of the sample points, discretize the range into a two-dimensional mesh domain, and define the stratigraphic marker field to be solved within the two-dimensional mesh domain; The classification modeling module is used to construct multi-source soil and rock characteristics based on the water content, plasticity index, and liquidity index of borehole samples, standardize the multi-source soil and rock characteristics, and train the stratigraphic classification model in combination with the corresponding stratigraphic category labels. The spatial reconstruction module is used to reconstruct the features of each grid point to be estimated in the two-dimensional grid domain based on the spatial location of the borehole sample points and the standardized multi-source soil and rock characteristics, according to the anisotropic distance weight, to obtain a continuous feature field. The horizontal correlation scale of the anisotropic distance weight is greater than the vertical correlation scale. The probability field generation module is used to input the continuous feature field into the trained stratigraphic classification model to obtain the stratigraphic category probability field of each grid point in the two-dimensional grid domain. The spatial constraint optimization module is used to construct the data terms of the Markov random field energy function with the stratigraphic category probability field, and combine the lateral adaptive smoothing term and the vertical sequence constraint term to form a spatial constraint optimization model. After applying hard constraints consistent with the measured stratigraphic category labels to the grid points corresponding to the borehole sample points, the spatial constraint optimization model is discretized to obtain the tunnel stratigraphic model. The uncertainty quantification module is used to calculate the uncertainty index of each grid point based on the probability field of the stratum category, and output the uncertainty index in association with the tunnel stratum model.

[0021] As a further improvement to the above technical solution, the spatial reconstruction module includes a weight calculation unit and a feature overlay unit; the weight calculation unit is used to calculate and normalize the spatial weight of each borehole sample point based on the distance between the grid point to be estimated and the borehole sample point in the horizontal and depth directions, the horizontal correlation scale, the vertical correlation scale, and the distance decay index; the feature overlay unit is used to perform weighted overlay of the multi-source soil and rock features of the borehole sample points based on the normalized spatial weights.

[0022] As a further improvement to the above technical solution, the spatial constraint optimization module includes a data item construction unit, a lateral smoothing item construction unit, a vertical constraint item construction unit, a hard constraint unit, and an iterative optimization unit. The data item construction unit is used to construct data items based on the negative logarithm of the stratigraphic category probability. The lateral smoothing item construction unit is used to construct lateral smoothing items based on the category differences of laterally adjacent grid points and an adaptive smoothing coefficient. The vertical constraint item construction unit is used to construct vertical sequence constraint items based on preset stratigraphic sequence relationships and difference penalties. The hard constraint unit is used to fix the borehole sample point positions to the measured stratigraphic categories. The iterative optimization unit is used to solve the stratigraphic marker field using an iterative conditional mode algorithm.

[0023] As a further improvement to the above technical solution, the uncertainty quantification module includes an information entropy calculation unit and a classification margin calculation unit; the information entropy calculation unit is used to calculate the information entropy based on the probability distribution of the stratigraphic category of each grid point; the classification margin calculation unit is used to calculate the difference between the probability of the largest stratigraphic category and the probability of the second largest stratigraphic category, and output the information entropy and classification margin in association with the tunnel stratigraphic model.

[0024] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the tunnel stratum modeling method based on multi-source feature driving and spatial constraints as described in any one of the first aspects.

[0025] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows: This invention provides a tunnel stratum modeling method based on multi-source feature-driven and spatial constraints. By combining multi-source soil and rock features, anisotropic spatial reconstruction, stratum category probability prediction, Markov random field spatial constraints, and uncertainty quantification, it can continuously model the stratum structure of tunnel engineering areas under the condition of limited borehole sampling points. Compared with methods that rely solely on manual profile extension, conventional interpolation, or single-point classification prediction, this invention can improve the spatial continuity and sequence rationality of the stratum model and express the reliability of the stratum prediction results.

[0026] Specifically, the present invention has the following beneficial effects:

[0027] 1. This invention uses water content, plasticity index, and liquidity index to construct multi-source soil and rock characteristics, and then trains a DVB stratigraphic classification model after standardization. Since these parameters respectively reflect the soil's water content, plasticity, and hardness, their combined use can comprehensively characterize the differences in soil and rock properties across different strata. Standardization reduces the impact of variations in the dimensions and numerical ranges of different parameters on model training, thus facilitating the establishment of a more stable mapping relationship between multi-source soil and rock characteristics and stratigraphic category probabilities.

[0028] 2. This invention discretizes the tunnel engineering area into a horizontal-depth two-dimensional grid domain, and defines a stratigraphic marker field within the two-dimensional grid domain, so that borehole samples, continuous feature fields, stratigraphic category probability fields, and the final stratigraphic model are all within the same spatial representation framework. Therefore, the measured information of borehole locations can correspond to the areas to be estimated between boreholes, facilitating subsequent global stratigraphic inference and spatial constraint optimization.

[0029] 3. This invention spatially reconstructs the soil and rock characteristics of borehole samples according to anisotropic distance weights, and makes the horizontal correlation scale larger than the vertical correlation scale. Since tunnel strata typically have strong horizontal extensibility and rapid changes in depth, this setting can enhance the continuous expression of soil and rock characteristics in the horizontal direction, while suppressing their excessive diffusion in the vertical direction, thereby reducing unreasonable mixing of different strata in the depth direction.

[0030] 4. This invention inputs a continuous feature field into the trained DVB stratigraphic classification model to obtain the stratigraphic category probability field for each grid point. This probability field not only reflects the relative probability of each grid point belonging to different stratigraphic categories, but also serves as the source of data terms in the subsequent Markov random field energy function, ensuring that the spatial optimization results are still constrained by the multi-source soil and rock feature classification results.

[0031] 5. This invention constructs a data term for the energy function of a Markov random field using a stratigraphic category probability field, and combines it with a lateral adaptive smoothing term and a vertical sequence constraint term to form a spatially constrained optimization model. The data term is used to maintain consistency between the optimization results and the probabilistic prediction results; the lateral adaptive smoothing term is used to suppress discrete patches within the stratigraphic layers and local category jumps; the vertical sequence constraint term is used to limit anomalous layer jumps in the depth direction that do not conform to the stratigraphic sequence relationship. The combined effect of these constraints helps improve the lateral continuity and vertical sequence rationality of the stratigraphic model.

[0032] 6. This invention applies hard constraints to the grid points corresponding to borehole sample points, ensuring that the optimized stratigraphic model remains consistent with the measured stratigraphic category labels at the borehole locations. These hard constraints provide stable constraints for stratigraphic inference in areas between boreholes, reducing deviations from measured points caused by probabilistic errors or over-smoothing, thereby improving the consistency of the modeling results with existing exploration data.

[0033] 7. This invention calculates the uncertainty index of each grid point based on the probability field of the stratigraphic category and outputs the uncertainty index in association with the tunnel stratigraphic model. Therefore, it can not only obtain deterministic stratigraphic distribution results, but also identify the predicted unstable locations of stratigraphic boundaries, sparse borehole areas, and multi-category competing areas, providing a reference for supplementary exploration, construction risk assessment, and design parameter adjustment. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram showing the true formation values ​​and borehole sample distribution in an engineering case according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatial distribution of three soil and rock parameters at borehole sampling points in an embodiment of the present invention, wherein, Figure 2 (a) in the diagram is a schematic diagram of water content distribution. Figure 2 (b) in the diagram is a schematic diagram of the plasticity index distribution. Figure 2 (c) in the diagram is a schematic diagram of the liquidity index distribution; Figure 3 This is a histogram of three soil and rock parameters at the borehole sample points in this embodiment of the invention, used to represent the numerical distribution of water content, plasticity index, and liquidity index. Figure 4 This is a schematic diagram of the interpolation results of borehole sample point parameters in an embodiment of the present invention, wherein, Figure 4 (a) in the diagram is a schematic diagram of the water content interpolation results. Figure 4 (b) in the diagram is a schematic diagram of the plasticity index interpolation results. Figure 4 (c) in the figure is a schematic diagram of the interpolation results of the liquidity index; Figure 5 This is a schematic diagram of the predicted stratigraphic results obtained by using multi-source feature driving and spatial constraints in an embodiment of the present invention; Figure 6 This is a schematic diagram of the misalignment distribution of the predicted stratigraphic results in an embodiment of the present invention; Figure 7 This is a statistical chart showing the identification performance of different stratigraphic categories in an embodiment of the present invention; Figure 8 This is a spatial uncertainty entropy map of the stratigraphic category prediction results in an embodiment of the present invention; Figure 9 This is a schematic diagram of the superposition of predicted formation results and entropy values ​​in an embodiment of the present invention; Figure 10 This is a classification margin diagram of the stratigraphic category prediction results in an embodiment of the present invention; Figure 11 This is a probability stability analysis diagram extracted along the centerline in an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the iterative convergence of the MRF spatial constraint optimization process in an embodiment of the present invention; Figure 13 This is a comparison chart of ablation experiments under different model configurations in the embodiments of the present invention; Figure 14 This is a flowchart illustrating the tunnel stratum modeling method based on multi-source feature driving and spatial constraints disclosed in this invention.

[0036] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0039] Example 1

[0040] See Figure 14This invention provides a tunnel stratum modeling method based on multi-source feature-driven and spatial constraints. The method uses the geotechnical parameters and stratum category labels of borehole samples as a foundation. First, it establishes a mapping relationship between multi-source geotechnical features and stratum category probabilities. Then, it extends discrete borehole information to the entire modeling area and optimizes the initial probability results using spatial constraints, thereby obtaining a tunnel stratum model that considers data features, spatial continuity, and sequence relationships. The method includes the following steps: First, borehole sampling data is acquired within the tunnel engineering area. Each borehole sampling data point includes its spatial location, stratum category label, and soil and rock parameters. These parameters include water content, plasticity index, and liquidity index. Water content reflects the soil's water content, plasticity index reflects its plasticity, and liquidity index reflects its hardness. The combination of these three parameters distinguishes different soil and rock layers, providing input for subsequent classification models.

[0041] The tunnel engineering area is discretized into a horizontal-depth two-dimensional grid domain, which is represented as follows: ; in, This represents a two-dimensional grid domain corresponding to the tunnel engineering area; The grid number indicating the depth direction; Indicates the grid number in the horizontal direction; Indicates the number of discrete units in the depth direction; This represents the number of discrete units in the horizontal direction.

[0042] In a two-dimensional grid domain Define the stratigraphic marker field to be solved above: ; in, Indicates stratigraphic marker field; Represents grid points Stratigraphic category at the location; This represents the total number of stratigraphic categories. Different stratigraphic categories are numbered according to a preset stratigraphic sequence to facilitate the subsequent establishment of vertical sequence constraints. Through the above discretization process, the tunnel stratigraphic modeling problem can be transformed into a problem of solving stratigraphic categories at grid points, which is convenient for unified processing by combining probabilistic models and spatial optimization models.

[0043] Then, the multi-source soil and rock characteristics of the borehole samples are constructed. For any borehole sample, its multi-source soil and rock characteristic vector is represented as: ; in, Represents the characteristic vector of multi-source soil and rock; Indicates water content; Indicates the plasticity index; Represents liquid index; This represents the transpose of a vector.

[0044] To reduce the impact of differences in the dimensions and numerical ranges of different geotechnical parameters on model training, the first... Standardize the individual soil and rock characteristics: ; in, Indicates the first One original soil and rock characteristic value; Indicates the first The mean value of each soil and rock characteristic at the borehole sampling points; Indicates the first Standard deviation of individual soil and rock characteristics at borehole sampling points; Represents the standardized first Each soil and rock characteristic value. The standardized eigenvector is denoted as... This approach allows water content, plasticity index, and liquidity index to participate in training on a uniform scale, preventing any single parameter from having an unreasonable dominance over the classification results due to its large numerical range.

[0045] Based on this, a multi-source feature-driven stratigraphic classification model, DVB, is constructed. This stratigraphic classification model can employ a multilayer perceptron model to establish a nonlinear mapping relationship between multi-source soil and rock features and stratigraphic categories. The model outputs category scores as shown below: ; in, Represents the category score vector, ; Indicates stratigraphic category The corresponding score; This represents a nonlinear mapping function composed of multiple fully connected networks; Indicates model parameters; This represents the standardized multi-source soil and rock feature vector.

[0046] To convert category scores into probabilistic representations, the Softmax function is introduced: ; in, This indicates that the sample points are within a given standardized multi-source soil and rock characteristics. It belongs to the stratigraphic category The probability of; Represents the stratigraphic category variable; Indicates the candidate stratigraphic category; Indicate category The corresponding score; Indicate category The corresponding score; Indicates the total number of stratigraphic categories; This indicates a temperature parameter. Used to adjust the smoothness of the probability distribution. When the value is smaller, the probability distribution is more concentrated. A larger probability distribution results in a smoother probability distribution. The probability output not only provides an initial stratigraphic classification but also lays the foundation for MRF data item construction and uncertainty quantification.

[0047] During model training, a weighted cross-entropy loss function is used: ; in, Represents the loss function; Indicates stratigraphic category Corresponding category weights; Indicates the true label in the category One-hot encoded value; This indicates that the sample point belongs to a stratigraphic category. The probability of class division. Class weights can be determined as follows: ; in, Indicates stratigraphic category The number of samples in the borehole sampling points. By adjusting the class weights, the impact of the imbalance in the number of samples from different strata on model training can be mitigated, allowing the strata with fewer samples to still be effectively learned by the model.

[0048] Since the training samples for the DVB model come from borehole locations, while actual modeling requires obtaining the geological distribution of the entire tunnel area, it is necessary to reconstruct the multi-source soil and rock characteristics of the borehole samples into a two-dimensional grid domain. Let the set of borehole samples be: ; in, Represents the set of borehole samples; Indicates the first Spatial location of each borehole sample point; Indicates the first Multi-source soil and rock feature vectors of each borehole sample point; Indicates the first Measured stratigraphic category labels for each borehole sample point; This indicates the total number of borehole samples.

[0049] For any grid point to be estimated within the two-dimensional grid domain The estimated characteristics of multi-source soil and rock were obtained by weighted superposition of borehole samples: ; in, Indicates the grid points to be estimated Estimated multi-source soil and rock characteristic vectors at the location; Indicates the first Each borehole sample point is to be estimated grid point Normalized space weights; Indicates the first Multi-source soil and rock feature vectors of borehole samples.

[0050] Considering that tunnel strata typically exhibit strong horizontal extensibility while changing rapidly in the depth direction, this embodiment employs an anisotropic distance weighting function. The unnormalized weights corresponding to each borehole sample point are represented as follows: ; And normalize it: ; in, Indicates the first Each borehole sample point is to be estimated as a grid point. Unnormalized weights; Represents the normalized spatial weights; Indicates the first The location of each borehole sample point; Indicates the horizontal correlation scale; Indicates the vertically related scale; Indicates the distance decay index; Indicates the borehole sample point number; This indicates the total number of borehole sampling points. During implementation, [the following data is used]. This makes the horizontal influence range greater than the depth influence range. This enhances the continuous expression of strata in the horizontal direction, while suppressing the excessive diffusion of soil and rock characteristics in the vertical direction and reducing unreasonable mixing between different strata.

[0051] After reconstructing the multi-source soil and rock characteristics of each grid point to be estimated, the full-field continuous characteristic field within the two-dimensional grid domain is obtained: ; in, Represents a continuous characteristic field; Represents grid points Estimated multi-source soil and rock characteristic vectors at the location; Represents a two-dimensional grid domain. Continuous feature fields Input the trained DVB model and predict the stratigraphic class probability for each grid point in the two-dimensional grid domain to obtain the stratigraphic class probability field for each grid point: ; in, Represents grid points The probability distribution of stratigraphic categories at the location; This indicates that, given the reconstructed feature vector Under the condition of grid points Stratigraphic category The probability of; Indicates the total number of stratigraphic categories; Represents grid points Stratigraphic category at the location; This indicates the candidate stratigraphic category. This step realizes the transformation from discrete borehole samples to a global probability field, providing a data foundation for spatially constrained optimization.

[0052] It should be noted that during the training phase, the DVB model uses the multi-source soil and rock features and corresponding stratigraphic category labels of borehole samples as supervision information to learn the mapping relationship between multi-source soil and rock features and stratigraphic category probabilities; during the prediction phase, it uses the continuous feature field obtained from spatial reconstruction. As input, a global prediction is performed on the area to be estimated between boreholes. Thus, the discrete observation information at the borehole locations can be expanded into a stratigraphic category probability field covering the entire two-dimensional grid domain. This step realizes the transformation from discrete borehole samples to a global probability field. On the one hand, the continuous feature field retains the geotechnical parameter information provided by the borehole samples and reflects the spatial characteristics of lateral extension and vertical variation of strata through anisotropic spatial reconstruction. On the other hand, the probability field output by the DVB model can provide a data basis for the subsequent MRF spatial constraint model, ensuring that the subsequent optimization process is constrained by the multi-source geotechnical feature classification results and can continue to be spatially corrected by combining lateral smoothing constraints, vertical sequence constraints, and borehole hard constraints.

[0053] Furthermore, to avoid the problems of discrete patches, interface fractures, or layer skipping caused by relying solely on local probability classification, this embodiment introduces a Markov random field spatial constraint model. The optimization objective of the stratigraphic marker field is expressed as: ; in, This represents the optimized stratigraphic marker field; This represents the stratigraphic marker field to be optimized. This represents the total energy function of the stratigraphic marker field.

[0054] The total energy function consists of a data term, a lateral smoothing term, and a vertical sequence constraint term: ; in, Represents grid points The data items reflect the consistency between the classification results and feature information; Indicates horizontally adjacent grid points and The lateral smoothing term between them is used to constrain formation continuity; Represents grid points The vertical sequence constraint term is used to demonstrate the vertical sequence structure; Represents the set of neighborhoods in the horizontal direction; This represents a two-dimensional grid domain. This energy function allows the local classification probabilities, lateral stratigraphic continuity, and vertical sequence relations given by the DVB model to be unified within a single optimization framework.

[0055] The data items are determined by the probability field output by the DVB model: ; in, Represents grid points Stratigraphic Class Energy of data items at time; This indicates the estimation of multi-source soil and rock eigenvectors. Under the condition, grid points Stratigraphic category The probability of the data item is determined by the negative logarithmic form. The higher the probability of a category, the lower the energy of the data item corresponding to it, and vice versa. This ensures that the optimization results remain consistent with the multi-source soil and rock feature classification results.

[0056] To enhance the horizontal continuity of the formation, a lateral smoothing term is introduced: ; in, Indicates horizontally adjacent grid points and Energy of smoothing terms between; Represents grid points Adaptive smoothing coefficient at the location; This indicates an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Represents grid points Stratigraphic categories; Represents grid points The stratigraphic category. This item is used to suppress unnecessary category jumps within the same stratigraphic unit, making the stratigraphic results more continuous laterally.

[0057] The adaptive smoothing coefficient can be expressed as: ; in, Represents grid points Adaptive smoothing coefficient at the location; Indicates the basic smoothing coefficient; Represents grid points The probability of the maximum stratigraphic class at that location; Represents grid points Estimate the spatial gradient of multi-source soil and rock eigenvectors; The norm of the spatial gradient is represented; This represents the adjustment parameter. By incorporating the maximum class probability and feature gradient into the smoothing coefficient, neighborhood coordination can be strengthened in areas of classification uncertainty, while reducing the risk of over-smoothing near interfaces where soil and rock characteristics change significantly, thus helping to maintain the true stratigraphic interface.

[0058] Considering that strata typically exhibit a sequence relationship along the depth direction, usually showing a soft-over-hard or young-to-old sequence, this embodiment further includes a vertical sequence constraint term. When the stratum category numbers are arranged according to the preset sequence direction, the vertical sequence constraint term can be expressed as: ; in, Represents grid points Energy of the vertical sequence constraint term at the location; Indicates the stratigraphic category of the current depth grid point; Indicates the stratigraphic category of adjacent grid points at the previous depth; This represents the penalty coefficient applied when a predefined vertical sequence relation is violated. This term is used to limit reverse jumps in the depth direction that are inconsistent with the predefined sequence, thereby improving the sequence rationality of the stratigraphic model.

[0059] To further avoid cross-level jumps, a difference penalty can be introduced into the vertical sequence constraint term: ; in, This represents the difference penalty function; This represents the difference in stratigraphic category number between the current depth grid point and the adjacent grid point at the previous depth. The larger the difference in category number, the more pronounced the skipping of layers, and the higher the penalty can be set to suppress discontinuous skipping of layers.

[0060] To ensure that the modeling results do not deviate from the measured borehole data, hard constraints are applied to the mesh points corresponding to the borehole sample points: ; in, Indicates the first The location of each borehole sample point; Indicates the optimal stratigraphic category at this location; Indicates the first The measured stratigraphic category labels for each borehole sample point. This hard constraint remains valid throughout the optimization process, ensuring that the model results are consistent with the measured results at the borehole locations, while also providing stable anchor points for the surrounding area and reducing the propagation of errors across the entire region.

[0061] After constructing the MRF spatial constraint model, the tunnel stratigraphic inversion problem can be transformed into a discrete marker optimization problem, i.e., finding the stratigraphic marker field that minimizes the global energy under a given energy function. Since each grid point in the two-dimensional grid domain needs to select from multiple candidate stratigraphic categories, directly solving for the global optimum would be computationally intensive. To balance solution efficiency and modeling stability, this embodiment employs an iterative conditional mode algorithm to approximate the solution of the spatial constraint optimization model.

[0062] The specific solution steps are as follows: During the initialization phase, the initial stratigraphic label field is determined based on the stratigraphic category probability field output by the DVB model: ; in, Represents grid points The initial stratigraphic category at the location; Represents grid points Stratigraphic category The probability of; This indicates the candidate stratigraphic category. This initialization method ensures that the initial stratigraphic label field remains consistent with the multi-source soil and rock feature classification results, providing a more reasonable initial solution for subsequent spatial optimization. Simultaneously, the stratigraphic category of the grid points corresponding to the borehole samples is set as the measured stratigraphic category label, ensuring that the borehole locations always serve as reliable control points throughout the iteration process.

[0063] Subsequently, the local energy corresponding to the candidate category is calculated for each grid point in the two-dimensional grid domain, and updated according to the following rules: ; in, Indicates the first After the second iteration, the grid points Stratigraphic categories; Indicates the number of iterations; Indicates the candidate stratigraphic category; Represents grid points Select candidate stratigraphic categories The local energy at time, which is determined by the aforementioned data terms, the lateral adaptive smoothing term, the vertical sequence constraint term, and the borehole hard constraint.

[0064] During the iterative update phase, each grid point within the two-dimensional grid domain is traversed in a preset order. For the current grid point, its local energy is calculated for different candidate stratigraphic categories, and the grid point is updated to the stratigraphic category with the minimum local energy. After each round of traversal, the measured stratigraphic category label is restored for the grid points corresponding to the borehole samples to avoid deviations in measured points caused by spatial smoothing or local probability errors. This process not only improves the spatial continuity between borehole regions using neighborhood constraints but also ensures that the model results are consistent with the measured data at the borehole locations.

[0065] After each iteration, the stratigraphic marker field can be locally smoothed and stabilized to reduce the impact of isolated grid points or small noise patches on the overall stratigraphic structure. This process does not change the measured labels at the borehole sampling points and is mainly used to improve the continuity and stability of the results within the stratigraphic region.

[0066] Iteration stops when the following conditions are met or the preset maximum number of iterations is reached: ; in, Indicates the first Stratigraphic marker field after the next iteration Indicates the first Stratigraphic marker field after the next iteration; Indicates the convergence threshold; This measures the change in the stratigraphic marker field. A convergence threshold is reached when the change in the stratigraphic marker field between two consecutive iterations is less than the convergence threshold, or when the number of iterations reaches the preset maximum number of iterations. At that time, it is assumed that the stratigraphic marker field has reached a stable state, and the optimized tunnel stratigraphic model is output.

[0067] Through iterative optimization, a relatively stable stratigraphic marker field can be obtained under controllable computational load, improving modeling efficiency. Through the above solution process, the class probabilities output by the DVB model can serve as the initial discrimination criterion, the MRF spatial constraints can gradually correct local discrete patches, interface fractures, and unreasonable layer skipping phenomena during the iteration process, and the borehole hard constraints ensure that the optimization results do not deviate from the measured borehole data. Therefore, this solution method can obtain a relatively stable stratigraphic marker field under controllable computational load conditions, making it suitable for tunnel stratigraphic modeling under limited borehole sampling conditions.

[0068] The uncertainty index of each grid point is calculated based on the probability field of the stratum category, and the uncertainty index is output in association with the tunnel stratum model.

[0069] Using the above method, under the condition of limited borehole sampling points, multi-source soil and rock parameters such as water content, plasticity index, and liquidity index can be transformed into global stratigraphic category probabilities. Then, through anisotropic spatial reconstruction, MRF spatial constraints, borehole hard constraints, and uncertainty quantification, a tunnel stratigraphic model is obtained. Compared with methods that rely solely on manual profile extension, conventional interpolation, or simple data-driven classification, this method can better maintain the lateral continuity of the stratigraphy, reduce unreasonable vertical mixing and skipping of layers, and express the predictive reliability of stratigraphic boundaries and sparse borehole areas.

[0070] As a preferred embodiment, after obtaining the optimized tunnel geological model, to further reflect the reliability of the prediction results, this implementation method quantifies the uncertainty of each grid point within the two-dimensional grid domain based on the geological category probability distribution output by the DVB model. It should be noted that the output of the MRF spatial constraint optimization is a deterministic geological category label, but its prediction reliability can still be characterized by the geological category probability field before optimization. Therefore, while outputting the geological model, the corresponding uncertainty distribution can be given, providing a reference for geological interpretation, supplementary exploration, and construction risk assessment.

[0071] For any grid point within the two-dimensional grid domain The probability distribution of its stratigraphic categories is expressed as follows: ; in, Represents grid points The probability distribution of stratigraphic categories at the location; Represents grid points The probability of belonging to stratigraphic category c; c represents the candidate stratigraphic category; K represents the total number of stratigraphic categories.

[0072] For any grid point Its information entropy is expressed as: ; in, Represents grid points The stratigraphic category prediction information entropy at that location; Represents grid points Stratigraphic category The probability of; This indicates the total number of stratigraphic categories.

[0073] Information entropy is used to reflect the degree of uncertainty in classification results. When the probability of a certain stratigraphic category is close to 1, the probabilities of other categories are lower. A probability close to 0 indicates that the classification result of this grid point is relatively certain; when the probabilities of multiple stratigraphic categories are relatively close... An increase in the value indicates strong category competition at the grid point, resulting in higher prediction uncertainty. This indicator can be used to identify locations with lower reliability in stratigraphic boundaries or sparsely drilled areas.

[0074] Besides information entropy, classification margin can also be used to characterize classification stability. Classification margin is calculated using the following formula: ;

[0075] in, Represents grid points Classification margin at the location; Represents grid points The probability of the maximum stratigraphic class at that location; Represents grid points The probability of the second largest stratigraphic category at a given location. Classification margin is used to measure the degree of distinction between the best and second-best categories. The larger the classification margin, the more obvious the distinction between the best and second-best categories, and the more stable the prediction results; the smaller the classification margin, the stronger the competition among multiple categories, the lower the classification stability, and this area can be regarded as a stratigraphic boundary or a key area of ​​focus for supplementary exploration.

[0076] In the engineering implementation, information entropy and classification margin can be overlaid with the predicted stratigraphic model for display. For intra-stratum regions, the maximum class probability is typically high, information entropy is low, and classification margin is large, indicating relatively stable prediction results. For areas near stratigraphic interfaces or with weak borehole constraints, multiple stratigraphic class probabilities are similar, information entropy increases, and classification margin decreases, indicating high uncertainty in these areas. Therefore, this implementation not only outputs a deterministic tunnel stratigraphic model but also simultaneously provides the spatial distribution of prediction reliability, making the modeling results more suitable for stratigraphic interpretation and engineering decision-making.

[0077] To further illustrate the inventive concept of this invention, an open-cut section of a subway line station is selected as the modeling object. This section is located in a coastal soft soil development area with significant strata heterogeneity, exhibiting local lateral variations and lenticular structures. Based on engineering geological data, the study area is divided into a two-dimensional modeling region with a horizontal length of 100m and a depth of 60m, and a horizontal-depth two-dimensional grid is established. See [link / reference] Figure 1 The study area mainly consists of three types of soil and rock layers: cohesive soil, silty clay, and silty mud, denoted as C, SC, and M, respectively. Figure 1 The distribution of borehole samples within the study area is also shown.

[0078] To simulate the sparse borehole data encountered in actual engineering projects, five boreholes were laid out within the modeling area, and ten sampling points were selected from each borehole along the depth direction, resulting in a total of 50 borehole sampling points. For each sampling point, its spatial location, corresponding stratigraphic type, and three geotechnical parameters—water content, plasticity index, and liquidity index—were recorded. See [link / reference]. Figure 2 , Figure 2 (a) shows the spatial distribution of water content at the borehole sampling points. Figure 2 (b) shows the spatial distribution of the plasticity index at the borehole sampling points. Figure 2 (c) shows the spatial distribution of the liquid index at the borehole sampling points. The distribution range of soil and rock parameters differs for different strata types, and a certain disturbance is superimposed on the parameters to reflect the dispersion of engineering soil and rock parameters.

[0079] Statistical analysis was performed on the three types of soil and rock parameters at the borehole sampling points, and the results are as follows: Figure 3 As shown. By Figure 3 As can be seen, the water content distribution ranges widely, from approximately 20% to 45%, reflecting the differences in water content among different soil layers; the plasticity index mainly ranges from 10 to 20, and has a certain distinguishing effect on categories such as cohesive soil and silty clay; the liquidity index has a relatively small range, but exhibits stable fluctuations with changes in strata. Therefore, a single parameter is insufficient to reliably distinguish all strata categories, while the combined use of water content, plasticity index, and liquidity index provides a more complete representation of soil and rock characteristics, which is beneficial for the DVB strata classification model to learn the correspondence between soil and rock parameters and strata categories.

[0080] In the spatial feature reconstruction process, the aforementioned borehole samples were used as input, and feature reconstruction was performed on the grid points to be estimated within the two-dimensional grid according to anisotropic distance weights. Because the strata in this engineering area have strong lateral extension and rapid changes in the depth direction, the horizontal correlation scale was set larger than the vertical correlation scale. This ensures that the borehole samples have a greater influence on adjacent areas in the horizontal direction, while their diffusion in the depth direction is limited. See also... Figure 4 , Figure 4 The interpolation results of water content, plasticity index and liquidity index in the data show strong lateral continuity and retain local features near the borehole. In the sparse area of ​​the borehole, the parameter field shows a certain smooth diffusion, indicating that there is incomplete information in this area due to insufficient sampling points, which needs to be further corrected by combining probability models and spatial constraints.

[0081] Will Figure 4 The DVB stratigraphic classification model, trained by inputting the continuous feature field shown, obtains the stratigraphic category probability field for each grid point within the two-dimensional grid. Then, a data term for the Markov random field energy function is constructed using this stratigraphic category probability field. Spatial optimization is then performed by combining a lateral adaptive smoothing term, a vertical sequence constraint term, and a hard constraint on borehole location to obtain the predicted stratigraphic results. See also... Figure 5 The optimized predicted strata can reproduce the layered structure characteristics of the reference strata, with good lateral continuity. The location of large-scale layered interfaces is basically consistent with the actual interfaces, and the overall classification accuracy reaches 90.97%. Figure 5 The superimposed true boundary line further indicates that the predicted interface and the true interface have a high degree of consistency in most areas, but there is still a certain offset in areas with large local fluctuations.

[0082] See Figure 6 The misclassified areas are mainly concentrated in stratigraphic boundaries and locally complex structural areas, exhibiting a banded or blocky distribution, while misclassification within the stratigraphic layers is less frequent. This indicates that the method of the present invention can reliably identify homogeneous stratigraphic regions even with a limited number of borehole samples, and limits the main errors to stratigraphic transitions and locally complex structural areas.

[0083] Further, see Figure 7 The identification performance of different stratigraphic categories was statistically analyzed. Figure 7 The results show that the cohesive soil category (Category C) has a high precision but a relatively low recall, the silty soil category (Category M) has a high recall but some overidentification, and the silty clay category (Category SC) performs relatively evenly. Overall, the F1 values ​​of each category remain at a high level, indicating that the combined effect of multi-source soil and rock characteristics and spatial constraints can improve the comprehensive stability of stratigraphic classification results.

[0084] To evaluate the reliability of the prediction results, this embodiment further calculates information entropy and classification margin. See also Figure 8 In the spatial uncertainty entropy map, high-entropy regions are mainly distributed along stratigraphic boundaries, and are associated with... Figure 6 The misclassified regions in the model have a good spatial correspondence, indicating that the prediction uncertainty mainly comes from class aliasing and parameter transition near the interface.

[0085] See Figure 9 By overlaying the predicted stratigraphic results with the entropy values, it can be seen that areas of high uncertainty are concentrated in the transition zones between different stratigraphic layers, while the entropy values ​​within the layers are close to zero, indicating that the stratigraphic categories within the layers are relatively clear. This result demonstrates that information entropy can effectively indicate the reliability of predictions near stratigraphic interfaces and in areas with weak borehole constraints.

[0086] See Figure 10 The classification margin plot shows that the margin is close to 1 in most intra-layer regions, indicating that the probability of the largest category is significantly higher than that of the second largest category, and the classification decision is relatively clear; while the margin decreases significantly near the stratigraphic interface, compared to... Figure 8 The high-entropy regions are largely consistent, indicating that classification margin can reflect the degree of competition between stratigraphic categories. (See also...) Figure 11Further analysis of the maximum class probability, classification margin, and normalized entropy extracted along the centerline reveals that the maximum class probability and classification margin are close to 1, while the entropy is close to 0 within the layer. Near the interface, these indicators show significant fluctuations, indicating strong class competition at that location. Therefore, information entropy and classification margin can serve as auxiliary indicators for identifying uncertain areas at stratigraphic interfaces, and can be used for subsequent supplementary exploration or construction risk assessment.

[0087] For iterative solutions, see [link to relevant documentation]. Figure 12 The energy value in the MRF spatial optimization process quickly stabilizes within the first two iterations, and the label change ratio decreases synchronously to near zero, indicating that spatial constraint optimization can achieve stable results in a relatively small number of iterations, and has good computational efficiency and numerical stability.

[0088] To analyze the impact of various technical features on the modeling results, this embodiment also includes ablation comparisons. See [link / reference] Figure 13 After removing prior labels, the model accuracy and mean crossover ratio both decreased significantly, indicating that borehole measured labels, as hard constraints or prior information, play an important role in maintaining the reliability of stratigraphic results. After weakening the spatial constraints of MRF, the local classification accuracy may be slightly improved, but the spatial continuity decreases, indicating that pursuing only single-point classification results can easily damage the rationality of stratigraphic structure. When only the DVB model is used, the overall performance further decreases, indicating that relying solely on local probability classification is difficult to guarantee the spatial continuity and stratigraphic rationality of the stratigraphic model.

[0089] pass Figures 1-13 The results show that the method of this invention can combine multi-source soil and rock parameters, anisotropic spatial feature reconstruction, DVB probability classification, MRF spatial constraints, and uncertainty quantification to obtain a tunnel stratum model with good continuity, stable sequence, and reflecting the reliability of prediction, even under conditions of sparse borehole data. The implementation results demonstrate that this invention is applicable to tunnel stratum modeling under complex geological conditions, especially suitable for engineering scenarios with a limited number of boreholes, significant undulations in stratum interfaces, or the need to identify uncertain regions.

[0090] Example 2

[0091] This invention also provides a tunnel stratum modeling system based on multi-source feature-driven and spatial constraints. The system can be deployed on an engineering geological modeling platform, server, or computing device with data processing capabilities to execute the aforementioned tunnel stratum modeling method based on multi-source feature-driven and spatial constraints. The system includes a data acquisition module, a mesh modeling module, a classification modeling module, a spatial reconstruction module, a probability field generation module, a spatial constraint optimization module, and an uncertainty quantification module. Each module works collaboratively in the order of borehole data input, feature modeling, probability prediction, spatial optimization, and result output, enabling discrete borehole samples to be transformed into a continuous tunnel stratum model.

[0092] The data acquisition module is used to acquire borehole sampling data within the tunnel engineering area. This data includes the spatial location of the sampling points, stratigraphic category labels, and water content, plasticity index, and liquidity index. Simultaneously acquiring spatial location, category labels, and soil parameters provides a unified data foundation for subsequent model training, spatial feature reconstruction, and borehole hard constraints. Specifically, water content characterizes the soil's water content state, the plasticity index characterizes the soil's plasticity, and the liquidity index characterizes the soil's hardness; these three parameters together constitute the fundamental features for stratigraphic category identification.

[0093] The mesh modeling module is used to determine the horizontal-depth range of the tunnel engineering area based on the spatial location of sample points, and discretizes this range into a two-dimensional mesh domain. Within this two-dimensional mesh domain, the stratigraphic marker field to be solved is defined. Through this module, borehole sample points, the area to be estimated, the continuous feature field, the stratigraphic category probability field, and the final stratigraphic model can be unified under the same spatial coordinate framework, which facilitates subsequent stratigraphic category prediction and spatial constraint optimization at the mesh scale.

[0094] The classification modeling module is used to construct multi-source soil and rock features based on the water content, plasticity index, and liquidity index of borehole samples. These features are then standardized, and a stratigraphic classification model is trained using corresponding stratigraphic category labels. This stratigraphic classification model can be a multi-source feature-driven model, i.e., a data-driven Virtual Borehole model. Standardization reduces the impact of differences in the dimensions and numerical ranges of different soil and rock parameters on model training. After supervised training with stratigraphic category labels, the classification modeling module establishes a mapping relationship between multi-source soil and rock features and stratigraphic category probabilities, providing a model foundation for subsequent global stratigraphic category probability prediction.

[0095] The spatial reconstruction module is used to reconstruct the features of each grid point to be estimated within a two-dimensional grid domain based on the spatial location of borehole samples and standardized multi-source soil and rock characteristics, according to anisotropic distance weights, to obtain a continuous feature field. The horizontal correlation scale of the anisotropic distance weights is larger than the vertical correlation scale. Since tunnel strata typically exhibit strong horizontal extensibility and rapid changes in depth, this setting enhances the continuous expression of soil and rock characteristics in the horizontal direction while limiting excessive diffusion in the depth direction, thereby reducing unreasonable mixing of different strata in the vertical direction.

[0096] In one embodiment, the spatial reconstruction module includes a weight calculation unit and a feature overlay unit. The weight calculation unit calculates the spatial weight of each borehole sample point for the grid point to be estimated based on the distances between the grid point to be estimated and the borehole sample points in the horizontal and depth directions, the horizontal correlation scale, the vertical correlation scale, and the distance decay index, and then normalizes the spatial weights. The feature overlay unit performs weighted overlay of the multi-source soil and rock features of the borehole sample points based on the normalized spatial weights to obtain the estimated multi-source soil and rock features at the grid point to be estimated. Through the cooperation of these units, the discrete soil and rock parameters at the borehole location can be expanded into a continuous feature field covering the entire two-dimensional grid domain, providing input for stratigraphic inference in the region between boreholes.

[0097] The probability field generation module is used to input the continuous feature field into the trained stratigraphic classification model to obtain the stratigraphic category probability field for each grid point within the two-dimensional grid domain. This probability field represents the relative likelihood of each grid point belonging to different stratigraphic categories. Compared with directly outputting deterministic stratigraphic categories, the probability field can preserve category competition relationships and classification confidence information, and provide a basis for constructing data items in subsequent MRF spatial constraint optimization, ensuring that the spatial optimization process is still constrained by the multi-source soil and rock feature classification results.

[0098] The spatial constraint optimization module constructs a Markov random field energy function data term using the stratigraphic category probability field. This data term, combined with a lateral adaptive smoothing term and a vertical sequence constraint term, forms a spatial constraint optimization model. After applying hard constraints consistent with the measured stratigraphic category labels to the grid points corresponding to the borehole samples, the spatial constraint optimization model is discretized to obtain the tunnel stratigraphic model. The data term ensures consistency between the optimization results and the probability prediction results; the lateral adaptive smoothing term suppresses discrete patches and local category jumps within the stratigraphic region; the vertical sequence constraint term restricts abnormal layer jumps in the depth direction that do not conform to the preset stratigraphic sequence relationship; and the hard constraints ensure that the optimization results at the borehole sample locations are consistent with the measured stratigraphic category labels. These constraints work together to improve the lateral continuity and vertical sequence rationality of the stratigraphic model.

[0099] In one embodiment, the spatial constraint optimization module includes a data item construction unit, a lateral smoothing item construction unit, a vertical constraint item construction unit, a hard constraint unit, and an iterative optimization unit. The data item construction unit constructs data items based on the negative logarithm of the stratigraphic category probability, ensuring that categories with higher probabilities correspond to smaller data item energies. The lateral smoothing item construction unit constructs lateral smoothing items based on the category differences between laterally adjacent grid points and an adaptive smoothing coefficient, reducing unnecessary category jumps within the same stratigraphic layer. The vertical constraint item construction unit constructs vertical sequence constraint items based on preset stratigraphic sequence relationships and difference penalties, suppressing unreasonable reverse sequence or cross-layer jumps. The hard constraint unit fixes the borehole sample point locations to the measured stratigraphic categories, ensuring that the modeling results are consistent with the exploration data at the borehole locations. The iterative optimization unit uses an iterative conditional mode algorithm to solve the stratigraphic marker field, obtaining a stable tunnel stratigraphic model under controllable computational load.

[0100] An uncertainty quantification module is used to calculate the uncertainty index of each grid point based on the stratigraphic category probability field and output the uncertainty index in association with the tunnel stratigraphic model. In one embodiment, the uncertainty quantification module includes an information entropy calculation unit and a classification margin calculation unit. The information entropy calculation unit is used to calculate the information entropy based on the stratigraphic category probability distribution of each grid point to reflect the degree of uncertainty in the category prediction at that location; the classification margin calculation unit is used to calculate the difference between the probability of the largest stratigraphic category and the probability of the second largest stratigraphic category to reflect the degree of distinction between the best and second-best categories. By associating the information entropy and classification margin with the tunnel stratigraphic model, stratigraphic boundary areas, sparse borehole areas, and areas with significant multi-category competition can be identified, providing a reference for supplementary exploration, geological interpretation, and construction risk assessment.

[0101] In the above system, each module can be implemented by a processor executing program instructions from memory, or by functional units within the engineering geological modeling software. The data transfer relationships between modules are as follows: the data acquisition module outputs borehole sample data to the mesh modeling module and the classification modeling module; the mesh modeling module outputs the two-dimensional mesh domain and stratigraphic marker field definitions; the classification modeling module outputs the trained stratigraphic classification model; the spatial reconstruction module outputs a continuous feature field; the probability field generation module outputs a stratigraphic category probability field; the spatial constraint optimization module outputs the tunnel stratigraphic model; and the uncertainty quantification module outputs uncertainty indices associated with the tunnel stratigraphic model. This system architecture enables the generation of tunnel stratigraphic modeling results with spatial continuity, sequence constraints, and reliable expression, even with a limited number of borehole samples.

[0102] Example 3

[0103] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein, when the program is executed, it controls the device on which the storage medium is located to perform some or all of the steps in Embodiment 1.

[0104] The storage medium may include high-speed RAM memory, and may also include nonvolatile memory, such as at least one disk storage device. It is understood that the storage medium can be any machine-readable medium capable of storing program code, such as random access memory (RAM), magnetic disk, hard disk, solid state disk (SSD), or nonvolatile memory.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or storage media. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of a computer program product implemented 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.

[0106] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct or indirect applications in other related technical fields, are within the patent protection scope of the present invention.

Claims

1. A tunnel stratum modeling method based on multi-source feature-driven and spatially constrained approaches, characterized in that, Includes the following steps: Obtain borehole sample data within the tunnel engineering area. The borehole sample data includes the spatial location of the sample, the stratum category label, and the water content, plasticity index, and liquidity index. The tunnel engineering area is discretized into a horizontal-depth two-dimensional grid domain, and a stratigraphic label field to be solved is defined within the two-dimensional grid domain. Multi-source soil and rock characteristics are constructed using the water content, plasticity index, and liquidity index of borehole samples. The multi-source soil and rock characteristics are standardized and combined with the corresponding stratigraphic category labels to train a stratigraphic classification model to establish a mapping relationship between multi-source soil and rock characteristics and stratigraphic category probabilities. Based on the spatial location of borehole samples and the standardized multi-source soil and rock characteristics, the features of each grid point to be estimated in the two-dimensional grid domain are reconstructed according to the anisotropic distance weight to obtain a continuous feature field. The horizontal correlation scale of the anisotropic distance weight is greater than the vertical correlation scale. Input the continuous feature field into the trained stratigraphic classification model to obtain the stratigraphic category probability field of each grid point; The data terms of the Markov random field energy function are constructed using the stratigraphic category probability field, and a spatial constraint optimization model is formed by combining the lateral adaptive smoothing term and the vertical sequence constraint term. Hard constraints consistent with the measured stratigraphic category labels are applied to the grid points corresponding to the borehole sample points, and the spatial constraint optimization model is discretized to obtain the tunnel stratigraphic model. The uncertainty index of each grid point is calculated based on the stratigraphic category probability field, and the uncertainty index is correlated with the tunnel stratigraphic model for output.

2. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The two-dimensional grid domain is represented as follows: ； Where Ω represents the two-dimensional grid domain corresponding to the tunnel engineering area, i represents the grid number in the depth direction, j represents the grid number in the horizontal direction, m represents the number of discrete elements in the depth direction, and n represents the number of discrete elements in the horizontal direction. The stratigraphic marker field is represented as follows: ； Where L represents the stratigraphic marker field, L(i,j) represents the stratigraphic category at grid point (i,j), and K represents the total number of stratigraphic categories. Different stratigraphic categories are numbered according to a preset stratigraphic sequence.

3. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The multi-source soil and rock characteristic vectors of borehole samples are represented as follows: ； in, represents the characteristic vector of multi-source soil and rock, w represents water content, PI represents plasticity index, and LI represents liquidity index; The d-th soil and rock feature is standardized according to the following formula: ； in, and The first The mean and standard deviation of each feature. The d-th standardized soil and rock characteristic value is used; after processing, the standardized feature vector is obtained. , as input to the stratigraphic classification model.

4. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The stratigraphic classification model is a multilayer perceptron model, and it outputs category scores according to the following formula: ； in It is a nonlinear mapping function composed of a multi-layer fully connected network; These are model parameters; Represents the category score vector, This represents the score corresponding to stratigraphic category c; This represents the standardized multi-source soil and rock feature vector; The category scores are converted into stratum category probabilities using the Softmax function as follows: ； in, This indicates that the sample points are within a given standardized multi-source soil and rock characteristics. L represents the probability that a given stratigraphic class belongs to stratigraphic class c, where L represents the stratigraphic class variable and c represents the candidate stratigraphic class. This represents the score corresponding to category c. This represents the score corresponding to category k, where K represents the total number of stratigraphic categories; For temperature parameters; The weighted cross-entropy loss function is used when training the stratigraphic classification model: ； in, Represents the loss function. This represents the category weight corresponding to stratigraphic category c. This represents the one-hot encoded value of the true label in category c; the category weight is determined according to the following formula: ;in, This indicates the number of samples of formation type c in the borehole sampling points.

5. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, Let the set of borehole points be represented as: ； in, This represents the set of borehole sampling points. For the first Spatial location of each borehole sample point; is the multi-source soil and rock feature vector of the kth borehole sample point; This indicates the measured stratigraphic category label for the k-th borehole sample point; Indicates the total number of borehole samples. For any grid point (i,j) to be estimated, its multi-source soil and rock characteristics are reconstructed according to the following formula: ； in, Let represent the estimated multi-source soil and rock characteristic vector at the grid point (i,j) to be estimated. This represents the normalized spatial weight of the k-th borehole sample point for the estimated grid point (i,j). This represents the multi-source soil and rock feature vector of the k-th borehole sample point.

6. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 5, characterized in that, The normalized spatial weights are determined by the following anisotropic distance weighting function: ； in It is a horizontally correlated scale; For vertically related scales; The distance decay exponent; > To enhance horizontal continuity and suppress excessive diffusion in the depth direction; To ensure the weights are normalized, the anisotropic distance weight function is normalized: ； in, This represents the normalized spatial weights.

7. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The continuous feature field is represented as: ； Represents a continuous characteristic field within a two-dimensional grid domain. Represents grid points The estimated multi-source soil and rock feature vectors are obtained at the location; the continuous feature field is input into the stratigraphic classification model to obtain the following probability field: ； in, Let (i,j) represent the probability distribution of the stratigraphic category at grid point (i,j). This represents the probability that grid point (i,j) belongs to stratigraphic category c; The mapping relationship between multi-source soil and rock characteristics and stratigraphic category probabilities is shown below: 。 8. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The optimization objective of the Markov random field energy function is: ； in, This represents the optimized stratigraphic marker field. This represents the stratigraphic marker field to be optimized. The total energy function representing the stratigraphic marker field; The total energy function is expressed as: ； in, Represents grid points Data items, This represents the lateral smoothing term between horizontally adjacent grid points p and q. Represents grid points Vertical formation constraint terms, It represents the set of neighborhoods in the horizontal direction.

9. The tunnel strata modeling method based on multi-source feature-driven and spatial constraints as described in claim 1, characterized in that, The hard constraint on the location of the borehole sample point is expressed as: ； in, Indicates the first The location of each borehole sample point This indicates the optimal stratigraphic category for that location. Indicates the first The measured formation category labels of each borehole sample point; the hard constraints remain valid during the energy function optimization process; The uncertainty index includes information entropy, which is calculated according to the following formula: ； in, Represents grid points The stratigraphic category prediction information entropy at that location Represents grid points Stratigraphic category The probability, This represents the total number of stratigraphic categories; the higher the information entropy, the higher the uncertainty in predicting the stratigraphic category at that grid point. The uncertainty index also includes classification margin, which is calculated according to the following formula: ； in, Represents grid points Classification margin at the location, Represents grid points The probability of the maximum stratigraphic class at that location. Represents grid points The probability of the second largest stratigraphic category at a given grid point; the smaller the classification margin, the lower the stability of the stratigraphic category prediction at that grid point.

10. A tunnel stratum modeling system based on multi-source feature-driven and spatial constraints, characterized in that, include: The data acquisition module is used to acquire borehole sample data within the tunnel engineering area. The borehole sample data includes the spatial location of the sample, the stratum category label, and the water content, plasticity index, and liquidity index. The mesh modeling module is used to determine the horizontal-depth range of the tunnel engineering area based on the spatial location of the sample points, discretize the range into a two-dimensional mesh domain, and define the stratigraphic marker field to be solved within the two-dimensional mesh domain; The classification modeling module is used to construct multi-source soil and rock characteristics based on the water content, plasticity index, and liquidity index of borehole samples, standardize the multi-source soil and rock characteristics, and train the stratigraphic classification model in combination with the corresponding stratigraphic category labels. The spatial reconstruction module is used to reconstruct the features of each grid point to be estimated in the two-dimensional grid domain based on the spatial location of the borehole sample points and the standardized multi-source soil and rock characteristics, according to the anisotropic distance weight, to obtain a continuous feature field. The horizontal correlation scale of the anisotropic distance weight is greater than the vertical correlation scale. The probability field generation module is used to input the continuous feature field into the trained stratigraphic classification model to obtain the stratigraphic category probability field of each grid point in the two-dimensional grid domain. The spatial constraint optimization module is used to construct the data terms of the Markov random field energy function with the stratigraphic category probability field, and combine the lateral adaptive smoothing term and the vertical sequence constraint term to form a spatial constraint optimization model. After applying hard constraints consistent with the measured stratigraphic category labels to the grid points corresponding to the borehole sample points, the spatial constraint optimization model is discretized to obtain the tunnel stratigraphic model. The uncertainty quantification module is used to calculate the uncertainty index of each grid point based on the probability field of the stratum category, and output the uncertainty index in association with the tunnel stratum model.

Images