A geological modeling method based on slope digital twinning

Abstract

The application provides a geological modeling method based on slope digital twinning, comprising: acquiring and fusing multi-source point cloud data to construct a surface three-dimensional model of the slope; integrating real drilling data and geophysical interpretation profile data to construct a geological attribute graph topology structure containing spatial nodes and inter-node correlation; constructing a stratum attribute deduction model based on a graph neural network, taking the geological attribute topology structure as input, training the model by introducing a geophysical constraint loss function, deducing the stratum attribute of a virtual drilling position, and generating encrypted geological data; fusing the real drilling data and the encrypted geological data to reconstruct a stratum interface, and integrating with the surface three-dimensional model to generate a three-dimensional geological entity model of the slope. The geological modeling method provided by the application realizes twinning mapping of slope full-factor geometry and geological attributes, and effectively improves the modeling accuracy.

Description

Technical Field

[0001] This invention belongs to the field of slope modeling technology, specifically relating to a geological modeling method based on slope digital twins. Background Technology

[0002] In the field of geotechnical engineering, slope stability monitoring and safety assessment are of paramount importance. With the development of digital technology, constructing three-dimensional digital twin models that can accurately reflect geological structure and engineering conditions has become an important trend for realizing intelligent slope monitoring and management.

[0003] However, existing slope geological modeling methods still have significant shortcomings: on the one hand, modeling usually relies on a single data source or simple data splicing, making it difficult to achieve a high-precision unified expression of surface topography and underground geological structure; on the other hand, underground geological modeling is highly dependent on limited borehole data, resulting in low inference accuracy in data-sparse areas, and the models built are mostly static results, making it impossible to quickly and locally update them when new monitoring or exposure data is obtained, thus making it difficult to support dynamic safety analysis and early warning throughout the entire life cycle of slopes. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, the present invention provides a geological modeling method based on slope digital twins, comprising: Acquire and fuse multi-source point cloud data to construct a three-dimensional surface model of the slope; By integrating real borehole data with geophysical interpretation profile data, a geological attribute map topology structure containing spatial nodes and inter-node relationships is constructed. A stratigraphic attribute inference model based on graph neural network is constructed. The geological attribute topology is used as input. The model is trained by introducing a loss function with geophysical constraints to infer the stratigraphic attributes of the virtual borehole location and generate encrypted geological data. By integrating real borehole data with the encrypted geological data, the stratigraphic interface is reconstructed and integrated with the three-dimensional surface model to generate a three-dimensional geological entity model of the slope. An incremental learning and update mechanism is established. When new slope monitoring data or geological exposure information is acquired, the local parameters of the stratigraphic attribute inference model are fine-tuned and the local mesh of the three-dimensional geological entity model is reconstructed to achieve dynamic updating of the three-dimensional geological entity model.

[0005] According to the technical solution provided by the present invention, the step of acquiring and fusing multi-source point cloud data to construct a high-precision three-dimensional surface model of the slope includes: Macroscopic texture point clouds of slopes are obtained by oblique photography of drones, and high-density point clouds of slope facades and hidden areas are obtained by ground laser scanning. By using image control points deployed in the slope area, the macro-texture point cloud and the high-density point cloud are unified to the absolute geographic coordinate system; A probability-weighted algorithm based on the incident angle is used to fuse multi-source point clouds after unifying the coordinate system to construct a three-dimensional surface model of the slope.

[0006] According to the technical solution provided by the present invention, the image control point is a dual-mode target with a high-contrast pattern printed on its surface and a high-reflectivity sphere installed at its center, and the absolute geographic coordinates of its geometric center are determined by a measuring instrument.

[0007] According to the technical solution provided by the present invention, the step of using a probability weighting algorithm based on the incident angle to fuse multi-source point clouds after unifying the coordinate system includes: For points in the overlapping region, calculate their incident angle relative to each scanning source. The closer the incident angle is to the vertical incident angle, the higher the fusion confidence is assigned. The coordinate information of corresponding points from different scanning sources is weighted and averaged based on the fusion confidence level to obtain the fused point cloud.

[0008] According to the technical solution provided by the present invention, the step of integrating real borehole data and geophysical interpretation profile data to construct a geological attribute map topology structure containing spatial nodes and inter-node relationships includes: Real borehole data and geophysical interpretation profile data are discretized into spatial nodes; Assign node characteristics containing spatial coordinates and geophysical properties to each node; Based on the Euclidean distance between nodes and their geological relevance, the edges connecting the nodes and their features are constructed to form the initial geological attribute map topology.

[0009] According to the technical solution provided by this invention, a low-level attribute inference model based on a graph neural network is constructed. Using the geological attribute topology as input, the model is trained by introducing a loss function with geophysical constraints to infer the stratigraphic attributes of the virtual borehole location and generate encrypted geological data, including: Construct a graph neural network model incorporating a geological attention mechanism; Using the geological attribute map topology as input, the geological attention mechanism is used to automatically learn the feature weights of neighboring nodes in different geological deposition directions, thereby realizing the transfer and aggregation of geological attribute information from known nodes to unknown nodes. During model training, a composite loss function is collected, which includes a data fitting term, a physical constraint term, and a topology smoothing term. The physical constraint term is constructed based on the underlying order law and is used to penalize prediction results that violate the deposition law. The trained model is used to deduce the underlying lithology and physical and mechanical parameters of the virtual borehole location, generating encrypted geological data containing virtual borehole information.

[0010] According to the technical solution provided by the present invention, the geological attention mechanism introduces a relative elevation coding function based on the vertical distance between two nodes for dynamic weight correction when calculating the attention coefficient between nodes.

[0011] According to the technical solution provided by the present invention, real borehole data and the encrypted geological data are integrated to reconstruct the stratigraphic interface, and then integrated with the three-dimensional surface model to generate a three-dimensional geological entity model of the slope, including: The virtual borehole data and the real borehole data are fused according to a preset confidence weight to form a hybrid geological control point set; Based on the aforementioned set of mixed geological control points, the radial basis function implicit potential field method is used to reconstruct the interfaces of each stratigraphy. The reconstructed stratigraphic interface meshes are subjected to Boolean operations and spatial clipping with the surface 3D model to generate a slope 3D geological entity model with continuous internal strata and geometrically consistent surface.

[0012] According to the technical solution provided by the present invention, the method of reconstructing various strata interfaces using the radial basis function implicit potential field includes: For each formation interface, construct a scalar implicit function such that the function value is zero at the interface contact point, positive above the interface, and negative below the interface; The weighting coefficients of the implicit function are solved using the aforementioned set of mixed geological control points; The moving cube algorithm is used to extract negative isosurfaces from the implicit function to generate a continuous and smooth triangular mesh of the stratigraphic interface.

[0013] According to the technical solution provided by the present invention, the incremental learning and updating mechanism, when acquiring new slope monitoring data or geological exposure information, performs local parameter fine-tuning on the stratigraphic attribute inference model and local mesh reconstruction on the three-dimensional geological entity model to achieve dynamic updating of the three-dimensional geological entity model, includes: Establish a replay cache to store representative historical training samples; When monitoring data is updated or new geological information is revealed, the backbone network parameters of the stratigraphic property inference model are frozen, and a local training subgraph is constructed using the new data and historical training samples in the replay cache. Fine-tune the weights of the model's output layer and the network weights adjacent to the newly added data space, and introduce a parameter importance regularization term into the loss function; Based on the fine-tuned model deduction results, the local mesh of the affected area in the three-dimensional geological entity model is reconstructed and seamlessly stitched together.

[0014] Compared with existing technologies, the advantages of this invention are as follows: By fusing multi-source point cloud data to construct a high-precision surface model and integrating real borehole and geophysical data to form a geological attribute map topology, a unified and rich data foundation is provided for subsequent modeling. On this basis, based on graph neural networks, geophysical constraints are introduced to extrapolate stratigraphic attributes, which can efficiently generate spatially continuous encrypted geological data based on sparse known points, significantly improving the accuracy and rationality of geological modeling. Furthermore, by fusing real and virtual data, stratigraphic interface reconstruction and surface model integration are completed to form a consistent three-dimensional geological entity model. Finally, by using an incremental learning mechanism to perform local fine-tuning and mesh updates on the model, efficient dynamic maintenance of the digital twin model is achieved, enabling it to respond promptly to changes in slope condition and improving the timeliness and engineering applicability of slope modeling. Attached Figure Description

[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of a geological modeling method based on slope digital twins provided in this embodiment of the invention; Figure 2 This is a schematic diagram of control points provided in an embodiment of the present invention; Figure 3 A schematic diagram of a slope model based on UAV oblique photography 3D reconstruction provided by the present invention. Detailed Implementation

[0016] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] As mentioned in the background section, there are technical issues such as... Figure 1 As shown, this embodiment proposes a geological modeling method based on slope digital twins, including steps S100-S500: S100: Acquire and fuse multi-source point cloud data to construct a three-dimensional surface model of the slope.

[0019] Specifically, in step S100, firstly, various technical means are used to acquire surface point cloud data of the target slope area. These data come from different acquisition platforms and sensors, collectively forming a multi-source point cloud dataset. Next, through spatial registration and fusion algorithms, these point cloud data from different sources and perspectives are unified under the same spatial reference system, eliminating misalignments and gaps between data, ultimately generating a complete, continuous, and high-precision three-dimensional model of the slope surface. This model accurately reflects the geometric morphology of the slope surface.

[0020] Furthermore, step S100 specifically includes the following steps S110-S130: S110: Obtain macroscopic texture point clouds of slopes through oblique photography by UAVs, and obtain high-density point clouds of slope facades and hidden areas through ground laser scanning.

[0021] Specifically, in step S110, a complementary air-ground data acquisition strategy is adopted. For the top of the slope, gentle slope areas, and large-scale macroscopic surface morphology, an unmanned aerial vehicle (UAV) equipped with an oblique photogrammetry system is used for aerial surveying. The UAV flies along a preset flight path and simultaneously acquires image data of the slope from multiple angles in the air using its onboard five-lens camera group (one vertical lens and four oblique lenses), such as... Figure 3 As shown, after photogrammetric processing including aerial triangulation (AT) and dense matching, high-resolution point cloud data with accurate 3D coordinates and realistic RGB texture information is generated; this is the macroscopic texture point cloud. This point cloud can accurately reflect the geometric and textural features of the slope top surface and gentle areas.

[0022] Meanwhile, for areas where drones have limited field of view or where high-quality data is difficult to obtain, such as steep slopes, reverse slopes, grooves, slope toes, and concealed areas partially obscured by vegetation, ground-based 3D laser scanning technology is used as a supplement. Ground-based laser scanners are set up at multiple stations around the slope to perform detailed scans of these areas. By emitting laser beams and receiving echoes, ground-based laser scanners can directly acquire a large number of high-precision 3D coordinate points on the target surface, forming a high-density point cloud. The point cloud density and the ability to restore the geometric details of the facade are generally superior to aerial photogrammetry.

[0023] S120: Using image control points deployed in the slope area, the macroscopic texture point cloud and the high-density point cloud are unified to the absolute geographic coordinate system.

[0024] Specifically, in step S120, within the slope mapping area, representative locations such as the slope toe, slope platform, and slope crest are selected, and at least three non-collinear image control points are evenly distributed. These image control points must be clearly identifiable and accurately measured in subsequent UAV aerial imagery and ground-based laser scanning point clouds. In this embodiment, the image control points are dual-modal targets with high-contrast red and white patterns printed on their surfaces (e.g., ...). Figure 2 As shown in the image, the "texture center" is easily and accurately identified. A high-reflectivity sphere is installed at the geometric center of the pattern; this "geometric center" can be extracted with high precision from the laser point cloud using intensity information and a sphere fitting algorithm. Using high-precision surveying instruments such as RTK or a total station, the absolute geographic coordinates of the geometric center of the high-reflectivity sphere on each target are determined in the field. .

[0025] First, during the UAV aerial survey data processing, the pixel coordinates of these control points on the aerial image are correlated with their known absolute geographic coordinates. Through this step, the macroscopic texture point cloud generated by the UAV photogrammetry solution is directly corrected and output to a specified absolute geographic coordinate system (such as WGS-84 or the National Geodetic Coordinate System).

[0026] Secondly, when processing ground-based laser scanning point clouds, the coordinates of each target center in the radar local coordinate system are extracted through intensity channel segmentation and spherical fitting. Then, using the two sets of coordinates (local coordinates and absolute geographic coordinates) of all control points, the optimal rigid body transformation parameters (rotation matrix) are solved through algorithms such as Singular Value Decomposition (SVD). Translation vector Align the entire high-density point cloud of the ground to an absolute geographic coordinate system that is consistent with the macro-texture point cloud.

[0027] S130: A probability-weighted algorithm based on the incident angle is used to fuse multi-source point clouds after unifying the coordinate system to construct a three-dimensional surface model of the slope.

[0028] Specifically, in step S130, after the precise registration of the point clouds is completed, the multi-source point clouds in the overlapping areas are intelligently fused to generate a single surface model of optimal quality. This step implements a probabilistic weighted fusion algorithm, the core of which is to assign a confidence weight reflecting the observation quality to each data point.

[0029] Furthermore, in step S130, a probability-weighted algorithm based on the incident angle is used to fuse the multi-source point clouds after unifying the coordinate system, specifically including the following steps S131-S132: S131: For points in the overlapping region, calculate their incident angle relative to each scanning source. The closer the incident angle is to the vertical incident angle, the higher the fusion confidence level is assigned.

[0030] Specifically, in step S131, for the registered point cloud, the algorithm first identifies overlapping regions. For each point in the overlapping region, its line-of-sight vector relative to the UAV camera center and the ground-based radar transmission center is calculated; ray casting is used to monitor for occlusion and remove pseudo-data points falling into the "visual shadow area". Then, for points of the same name, the angle between the line-of-sight beams of each scanning source and the local surface normal vector of that point (i.e., the angle of incidence) is calculated; the fusion confidence level is defined, and the closer the angle of incidence is to 0 degrees (perpendicular incidence), the higher the confidence level.

[0031] S132: Based on the fusion confidence level, the coordinate information of the same points from different scanning sources is weighted and averaged to obtain the fused point cloud.

[0032] Specifically, in step S132, the coordinates of the two source point clouds are fused by weighted average based on the confidence level:

[0033] In the formula, For the merged point cloud data, and Macro-texture point cloud data and high-density point cloud data, respectively. and These represent the confidence weights of macroscopic texture point clouds and high-density point clouds, respectively, thus preserving the advantages of the UAV's top-view perspective while maximizing the ground-based radar's ability to geometrically reconstruct steep slopes.

[0034] S200: Integrates real borehole data with geophysical interpretation profile data to construct a geological attribute map topology that includes spatial nodes and the relationships between nodes.

[0035] Specifically, in step S200, the results data of the exploratory boreholes drilled in the slope area (i.e., actual borehole data) and the interpretation profile data obtained through geophysical exploration methods are collected. The geological information points (such as stratigraphic boundaries) contained in these data are abstracted into a series of discrete points with spatial location information, each defined as a spatial node. Each node is assigned characteristic information, and connections (i.e., "edges") are established between nodes based on their spatial and geological relationships, thereby constructing a graph structure composed of nodes and edges—the initial geological attribute graph topology. This graph structure transforms the geological spatial problem into a topological representation that can be processed by a graph neural network.

[0036] Furthermore, step S200 specifically includes the following steps S210-S230: S210: Discretize the actual borehole data and geophysical interpretation profile data into spatial nodes.

[0037] Specifically, in step S210, the collected raw geological data is first spatially discretized to generate the basic units of the graph structure—nodes. These data mainly include two types: first, real borehole data obtained through field drilling, which records the borehole's latitude and longitude coordinates, elevation, and the burial depth, lithological description, and corresponding physical and mechanical parameter test values ​​of different stratigraphic interfaces along the depth direction; second, geophysical interpretation profile data obtained through geophysical exploration methods (such as high-density electrical resistivity tomography and seismic refraction methods), which typically provide the spatial distribution trend of stratigraphic interfaces or lithological boundaries in the form of two-dimensional profiles. During processing, the borehole opening point of each borehole, as well as the top and bottom plate locations of different lithological strata within the borehole, are extracted as independent spatial nodes. Simultaneously, a series of stratigraphic control points are extracted from the geophysical interpretation profile at certain intervals or based on characteristic change points, and these are also used as spatial nodes. This process essentially digitizes continuous, simulated geological information into a series of discrete control points with clearly defined spatial locations (three-dimensional coordinates).

[0038] S220: Assign node characteristics to each node, including spatial coordinates and geophysical properties.

[0039] Specifically, in step S220, each spatial node generated above is assigned its characteristic attributes, i.e., node features. Each node's feature vector contains at least two main categories of information: the first is spatial coordinate features, typically obtained by normalizing the node's absolute geographic coordinates as the basic feature. The second is the physical and mechanical properties of the soil and rock mass. For nodes from actual boreholes, attribute values ​​obtained experimentally or empirically at the corresponding depth can be directly assigned, such as lithology type coding, density, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. For nodes extracted from geophysical profiles, their attributes may be incomplete or inferred; therefore, they can be assigned a predicted lithology type and an estimated range of attribute values. Through this step, each node is no longer a simple geometric point but becomes a feature vector carrying geographic location and geological attribute information.

[0040] S230: Based on the Euclidean distance between nodes and the geological correlation, construct the edges connecting the nodes and their features to form the initial geological attribute map topology.

[0041] Specifically, in step S230, the connections (edges) between these spatial nodes are constructed, and the characteristics of the edges are defined, thus forming a complete graph topology. The construction of connections follows the principle of combining spatial proximity and geological correlation. First, using a spatial indexing algorithm (such as KD-Tree), the nearest neighbor node or neighboring nodes within a certain radius in three-dimensional space are found for each node, establishing preliminary connections. This connection based on Euclidean distance constitutes the skeleton of the graph, ensuring the propagation of local information. More importantly, geological correlation needs to be introduced as an important basis for constructing or weighting edges. For example, based on regional geological data or geophysical interpretation results, it can be determined which nodes may belong to the same strata or the same tectonic unit. Even if their spatial distance is slightly far, connections can be established between them or the weight of existing connections can be enhanced. The characteristics of each "edge" can include basic Euclidean distance and a metric representing the strength of geological correlation. The metric can be derived based on prior knowledge or the statistical properties of the data itself. Finally, all nodes and edges together constitute an initial geological attribute graph topology that can characterize the geological spatial structure and attribute correlation of the slope. This graph structure serves as the direct input to the subsequent graph neural network model, transforming the geological modeling problem into a node classification and regression problem on the graph.

[0042] S300: Construct a geological attribute inference model based on graph neural network. Using the geological attribute topology as input, train the model by introducing a loss function with geophysical constraints, infer the geological attributes of the virtual borehole location, and generate encrypted geological data.

[0043] Specifically, in step S300, a graph neural network (GNN) model is constructed as the core algorithm for stratigraphic attribute inference. The topology of the geological attribute map constructed in step S200 is used as the model input. During the model training phase, a composite loss function is designed to guide model learning. This loss function not only includes a data fitting term that measures the accuracy of model predictions, but also introduces a "geophysical constraint term" based on fundamental geological laws to penalize predictions that clearly violate geological laws. The model is trained using existing real borehole data, enabling it to learn the distribution patterns of stratigraphic attributes in three-dimensional space from sparse known point information. After training, this model is applied to the "virtual borehole location" nodes in the geological attribute map that do not have real boreholes, inferring the stratigraphic lithology, physical and mechanical parameters, and other attributes of these locations, thereby generating a set of more spatially denser, encrypted geological data.

[0044] Furthermore, step S300 specifically includes the following steps S310-S340: S310: Construct a graph neural network model that incorporates a geological attention mechanism.

[0045] Specifically, in step S310, a graph neural network model specifically designed for geological attribute extrapolation is first constructed. The input to this model is the topological structure of the geological attribute map generated in step S200 and its corresponding node feature matrix. The node feature matrix... It gathers the attribute information of all nodes. Indicates the total number of nodes. This represents the initial feature dimension, specifically including the normalized spatial coordinates of each node and the known geophysical parameters of the soil and rock mass at that point.

[0046] Considering that the heterogeneity of the strata is much greater in the vertical direction than in the horizontal direction, a relative elevation coding function is introduced to correct the attention weights. For the center node... and neighboring nodes Its attention coefficient The calculation formula is:

[0047] In the formula, , Represents the node feature vector; || represents a learnable linear transformation matrix; || represents a vector concatenation operation. Represents the attention vector parameters; This represents the relative elevation coding function. , and They are nodes and nodes The vertical distance. This factor forces the grid to dynamically adjust its weights based on the vertical distance between two points. If the vertical distance is too large (spanning multiple layers), the weight is reduced; if the vertical distance is small and the horizontal distance is close, a high weight is assigned.

[0048] The attention coefficients are normalized using the Softmax function and then weighted and aggregated.

[0049] In the formula, Indicates the central node Updated feature vector; To represent a non-linear activation function, functions such as ReLU or ELU are typically used. Indicates the central node A set of neighboring nodes; This represents the normalized attention coefficient.

[0050] By stacking multiple layers of such networks, the model is able to capture high-order features ranging from local lithological variations to global geological structures such as folds and fault dips.

[0051] S320: Using the geological attribute map topology as input, the geological attention mechanism is used to automatically learn the feature weights of neighboring nodes in different geological deposition directions, thereby realizing the transfer and aggregation of geological attribute information from known nodes to unknown nodes.

[0052] Specifically, in step S320, the aforementioned node feature matrix and the adjacency relationship of the graph are input into the constructed model. Driven by a geological attention mechanism, the model, through its stacked multi-layer network, automatically learns and aggregates feature information from each node's neighboring nodes. During this process, the model can adaptively determine the weights of geological feature propagation in different directions (especially the vertical and horizontal directions), thereby achieving efficient and reasonable transmission and aggregation of geological attribute information from known real borehole nodes to unknown nodes to be deduced (i.e., virtual borehole locations), gradually generating updated feature representations containing high-order geological context information for all nodes in the graph.

[0053] S330: During model training, a composite loss function is collected, which includes a data fitting term, a physical constraint term, and a topology smoothing term. The physical constraint term is constructed based on the underlying order law and is used to penalize prediction results that violate the deposition law.

[0054] In step S330, a "first principle of soil layers (sequence rate)" constraint term is introduced into the loss function of the model training, constructing a hybrid loss function that includes prediction error, topological consistency, and geophysical constraints. This function imposes a high penalty on anomalous projections that violate the sedimentary rule of "older strata below, newer strata above." The "first principle of soil layers" specifically includes: The strata within the study area were coded from youngest to oldest according to their depositional age. .For example, It is miscellaneous fill soil. It is a silty clay. It is a strongly weathered rock. It is moderately weathered rock. In a normal slope without overturning structures, for any two points on the same vertical axis, the stratigraphic sequence number of the deep node must be greater than or equal to the stratigraphic sequence number of the shallow node.

[0055] To guide the model to learn geological distributions that conform to physical laws under conditions of limited labeled samples, this step employs a composite loss function consisting of three parts. : ; In the formula, This represents the data fitting loss; Indicates physical constraint loss; Represents topological smoothing loss; and These are the weighting coefficients for physical constraint loss and topology smoothing loss, respectively.

[0056] Among them, data fitting loss Used to supervise the model's prediction accuracy at known borehole nodes. Weighted cross-entropy loss is employed.

[0057] In the formula, Represents the actual set of borehole nodes; Indicates the true label; Indicates the predicted probability; This represents the category weight, used to address the category imbalance problem in soil and rock samples where there are fewer interlayer samples and more bedrock samples.

[0058] Physical constraint loss Used to penalize predictions that violate stratigraphic order laws. A ReLU-based one-sided penalty term is constructed:

[0059] In the formula, This represents the set of vertically adjacent edges, which is the set of all pairs of nodes that are "vertically adjacent" in space. Indicates the total number of vertical sides; and Indicates model prediction nodes and Belongs to the The probability value at the class level is a number between 0 and 1, and , The mathematical expectation is a formula for calculating the expected stratigraphic sequence number of the predicted strata.

[0060] Among them, if deep nodes The expected ordinal number is greater than that of the shallow nodes. (Right now ), conforms to the laws of physics, the difference is negative, the ReLU output is 0, and there is no loss; if (When the formation is reversed), the difference is positive, ReLU produces a positive loss, forcing the network to correct its parameters through backpropagation.

[0061] Topological smoothing loss Based on Laplace regularization, drastic jumps in prediction results between adjacent nodes are prevented:

[0062] In the formula, Represents the set of edges in the entire graph, including all connected pairs of nodes in the graph (not only vertical, but also horizontal). and Represents a node and The complete predicted probability vector.

[0063] S340: Using the trained model, the underlying lithology and physical and mechanical parameters of the virtual borehole location are deduced, generating encrypted geological data containing virtual borehole information.

[0064] Specifically, in step S340, the graph neural network model, which has been fully trained in step S330 and has learned the spatial distribution patterns of geological attributes, is used to perform forward inference on all unknown nodes marked as "virtual borehole locations" in the geological attribute map. The model outputs the probability distribution of stratigraphic lithology categories and the predicted values ​​of various physical and mechanical parameters at each virtual node. This newly generated, spatially continuous, and high-density geological attribute data constitutes the encrypted geological data. Combined with the original sparse real borehole data, this data significantly improves the abundance and spatial coverage of constraint information used to construct the three-dimensional geological model, laying a reliable data foundation for the next step of high-precision geological interface reconstruction.

[0065] S400: Integrate real borehole data with the encrypted geological data, reconstruct the stratigraphic interface, and integrate it with the surface 3D model to generate a 3D geological entity model of the slope.

[0066] Specifically, in step S400, the actual borehole data from step S200 is fused with the encrypted geological data generated in step S300 to form a mixed geological control point set that is more spatially evenly distributed and has a richer data volume. Based on this mixed control point set, a 3D geological modeling algorithm is used to reconstruct the interfaces of each stratigraphic layer, forming a continuous 3D stratigraphic interface model. Finally, the generated stratigraphic interface model is spatially integrated and geometrically calculated with the high-precision 3D surface model constructed in step S100, so that the top surface of the stratigraphic model accurately matches the actual surface model, thereby obtaining a 3D geological entity model whose interior is filled with various stratigraphic entities and whose external geometry is consistent with the actual slope. This completes the construction of the static geometry and attribute base of the slope digital twin.

[0067] Furthermore, based on the encrypted geological data generated in step S300, this method executes step S400 to fuse multi-source constraints and construct the final three-dimensional geological entity model. Step S400 specifically includes S410-S430: S410: The virtual borehole data and the real borehole data are fused according to a preset confidence weight to form a hybrid geological control point set.

[0068] Specifically, in step S410, different confidence weights are assigned to the data sources and then fused. Real borehole data from field drilling and testing has the highest confidence level, and its weight... It is usually set to the maximum value (e.g., 1.0). The confidence level of the virtual borehole data output in step S300 is... This is related to the uncertainty of model predictions. A feasible quantification method is based on the variance of the predicted probability in the model output: the smaller the variance, the more certain the model's prediction for that point is, and the higher the weight assigned. During spatial fusion, all real borehole control points are merged with virtual borehole control points to form a more spatially evenly distributed and densely packed set of mixed geological control points. Before merging, a stratigraphic sequence logic check can be performed, for example, to check for outliers in the vertical direction that violate the stratigraphic sequence law (the age code of deep strata should be greater than that of shallow strata), in order to ensure the inherent consistency of the input data.

[0069] S420: Based on the aforementioned set of mixed geological control points, the interfaces of each stratigraphic unit are reconstructed using the radial basis function implicit potential field method.

[0070] Specifically, step S420 includes the following steps S421-S423.

[0071] S421: Construct a scalar implicit function for each formation interface, such that the function value is zero at the interface contact point, positive above the interface, and negative below the interface.

[0072] Specifically, in step S421, for each stratigraphic interface to be constructed (e.g., the bottom interface of the silty clay layer, the top interface of the strongly weathered rock, etc.), a continuous surface reconstruction is performed using the radial basis function (RBF) implicit potential field method. Specifically, a three-dimensional scalar implicit function is constructed for each target interface. ,in Let be the coordinates of any point in space. The definition of this function must satisfy a clear geometric meaning: it applies to all control points traversed by the interface. At that point, the function value is zero ( ); On the upper part of the interface (i.e., the side with the relatively newer strata), the function value is positive ( ); at the bottom of the interface (i.e., on the side of the relatively older strata), the function value is negative ( The specific form of an implicit function is usually expressed as:

[0073] In the formula, Represents the radial basis kernel function; This indicates that the coordinates of the point to be calculated are... These are the coordinates of the control points; This represents the weight coefficient to be determined; This represents a low-order polynomial used to simulate overall geological trends.

[0074] S422: Solve for the weight coefficients of the implicit function using the mixed geological control point set.

[0075] Specifically, in step S422, by all interface control points Must meet Substituting this into the equation, we can solve the system of linear equations to obtain the weights. With polynomial coefficients.

[0076] S423: The moving cube algorithm is used to extract negative isosurfaces from the implicit function to generate a continuous and smooth triangular mesh of the stratigraphic interface.

[0077] Specifically, in step S423, the implicit function is obtained. Then, the moving cube algorithm is used to extract the data from the 3D spatial mesh. isosurfaces (i.e.) (the boundary between negative and positive values), which is equivalent to extracting The boundary surfaces are then used to generate a continuous, smooth triangular mesh model of the stratigraphic interface. This process is repeated for all stratigraphic interfaces to obtain a series of surface meshes representing the top and bottom plates of different strata.

[0078] S430: Perform Boolean operations and spatial clipping on the reconstructed strata interface meshes and the surface 3D model to generate a slope 3D geological entity model with continuous internal strata and geometrically consistent surface.

[0079] Specifically, in step S430, all the generated stratigraphic interface meshes are subjected to Boolean operations and spatial trimming with the surface 3D model (triangular mesh) obtained in step S100. The process is as follows: First, adjacent stratigraphic interfaces are used to close the sides and bottom, forming several preliminary, independent stratigraphic solid blocks (for example, the solid between the surface and the first interface represents the first layer of soil). Then, the surface model is introduced as a "tool," and a Construction Solid Geometry (CSG) Boolean intersection operation is performed on all stratigraphic solid blocks, i.e.: In the formula Represents a solid block of strata; This represents the three-dimensional surface model obtained in step S100.

[0080] This operation precisely removes all invalid mesh portions that protrude above the actual ground surface, ensuring that the final model's geometric top surface perfectly matches the real terrain. Ultimately, a high-precision 3D geological solid model is generated, characterized by continuous internal stratigraphic attitude, correct sequence logic, and accurate surface geometry. This solid model can be further meshed using tetrahedrons or hexahedrons, and the physical and mechanical parameters derived in step S300 (such as elastic modulus, Poisson's ratio, cohesion, and internal friction angle) can be mapped to each mesh cell or node, thus completing the static construction of a digital twin base with both accurate geometry and rich attributes. This embodiment constructs a hybrid control point set by fusing multi-source geological data and reconstructs continuous stratigraphic interfaces using the radial basis function implicit potential field method, strictly adhering to the stratigraphic sequence law. Boolean operations are then used to achieve seamless integration with a high-precision surface model, thereby generating a three-dimensional geological entity model with continuous internal strata and accurately matched surface geometry. This method effectively solves the problems of inaccurate interface inference, stratigraphic sequence contradictions, and mismatch with terrain caused by data sparsity in traditional modeling, providing a high-fidelity static geometric foundation for slope digital twins.

[0081] S500: Establish an incremental learning and update mechanism. When new slope monitoring data or geological exposure information is acquired, the local parameters of the stratigraphic attribute inference model are fine-tuned and the local mesh of the three-dimensional geological entity model is reconstructed to realize the dynamic update of the three-dimensional geological entity model.

[0082] Specifically, in step S500, to address changes in geological conditions or updates in monitoring data during slope construction or operation, this method establishes a dynamic model update mechanism. This mechanism is based on incremental learning. When the system receives new data, it does not require rerunning the entire modeling process from scratch. Instead, it triggers an update process: First, the graph neural network stratigraphic attribute inference model trained in step S300 undergoes "local fine-tuning." While retaining most of the knowledge already learned by the model, it primarily utilizes new data and some historical data related to the newly added area to quickly adjust the model's local weights, enabling the model to adapt to new information. Then, based on the updated geological attribute data inferred from the fine-tuned model, only the geological interfaces and entity meshes of the affected local areas in the 3D geological entity model generated in step S400 are recalculated and reconstructed. Finally, the updated local meshes are seamlessly stitched with the remaining unchanged parts of the model, thereby efficiently achieving dynamic and real-time updates of the entire digital twin model, ensuring it remains synchronized with the actual geological state of the slope.

[0083] Furthermore, step S500 specifically includes the following steps S510-S540: S510: Establish a replay cache to store representative historical training samples.

[0084] Specifically, in step S510, the system first establishes and maintains a replay buffer. This buffer is built immediately after the initial model training is completed, aiming to retain a portion of key historical samples that can represent the distribution of the original training data, called "backbone samples." The principle for selecting these samples balances representativeness and discriminative power: on the one hand, samples closest to the cluster centers of each category in the feature space are selected to retain the core features of that category; on the other hand, samples close to the classification decision boundary are also selected to retain the complex discriminative information already learned by the model. By storing these samples, the buffer can effectively represent "old knowledge" in subsequent incremental learning, which is the basis for preventing the model from catastrophically forgetting new data.

[0085] S520: When monitoring data is updated or new geological information is revealed, the backbone network parameters of the stratigraphic property inference model are frozen, and a local training subgraph is constructed using the new data and historical training samples in the replay cache.

[0086] Specifically, in step S520, when new data arrives, the system does not retrain the entire complex model. First, it freezes the weight parameters of the backbone network layers responsible for general feature extraction in the already trained graph neural network model, maintaining its stable feature extraction capability. Then, using the newly added data points as the core, and combining them with related historical backbone samples from the replay cache, the system dynamically constructs a local training subgraph within the original large-scale geological attribute map topology, containing only the affected nodes and their close neighbors. This significantly reduces the scale and scope of data required for each update.

[0087] S530: Fine-tune the output layer of the model and the network weights adjacent to the newly added data space, and introduce a parameter importance regularization term into the loss function.

[0088] Specifically, in step S530, the system mainly unfreezes and updates the weights of the model's output layer (classification / regression head), as well as the network parameters (such as attention weights) of nodes within the local subgraph directly related to the spatial location of the newly added data points. In the loss function for fine-tuning training, in addition to the original data fitting and physical constraints, the core innovation lies in introducing a regularization term based on an "elastic weight consolidation" mechanism, which takes the form:

[0089] In the formula, This represents the total loss value during incremental training; This represents the base loss on the mixed dataset in step S330; This represents the set of newly collected data and cached old backbone data B; Represents the regularization coefficient; express The diagonal elements of the information matrix are used to measure the first... The importance of each parameter to the old task; This indicates the current model (currently being fine-tuned and updated) at the [number]th [number]. The values ​​of the parameters; This indicates the first (frozen) of the old model (frozen before the update). The values ​​of the parameters.

[0090] This regularization term penalizes significant changes to important parameters, thereby effectively learning new knowledge while firmly consolidating the memory of old knowledge, achieving stable and efficient incremental learning.

[0091] S540: Based on the fine-tuned model deduction results, the local mesh of the affected area in the three-dimensional geological entity model is reconstructed and seamlessly stitched together.

[0092] Specifically, in step S540, the system first delineates the local area requiring updating by setting a radius R based on the geological influence range, centered on the newly added or significantly parameter-changed virtual borehole points. Then, utilizing the local characteristics of the radial basis function (RBF) implicit potential field method from step S420, the system recalculates the control point weights of the implicit function only within this area based on the updated virtual borehole data. Next, the system deletes the old triangular mesh within this local area and regenerates the stratigraphic interface mesh based on the updated implicit equations. Finally, at the boundary of the local area, techniques such as constrained triangulation are used to smoothly and seamlessly stitch the newly generated local mesh with the externally retained, unchanged old mesh, ensuring the geometric and topological continuity and closure of the entire 3D geological entity model. Through this process, the entire slope digital twin model can achieve rapid and accurate dynamic synchronization and updating without the high cost of global reconstruction.

[0093] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A geological modeling method based on slope digital twins, characterized in that, include: Acquire and fuse multi-source point cloud data to construct a three-dimensional surface model of the slope; By integrating real borehole data with geophysical interpretation profile data, a geological attribute map topology structure containing spatial nodes and inter-node relationships is constructed. A stratigraphic attribute inference model based on graph neural network is constructed. The geological attribute topology is used as input. The model is trained by introducing a loss function with geophysical constraints to infer the stratigraphic attributes of the virtual borehole location and generate encrypted geological data. By integrating real borehole data with the encrypted geological data, the stratigraphic interface is reconstructed and integrated with the three-dimensional surface model to generate a three-dimensional geological entity model of the slope. An incremental learning and update mechanism is established. When new slope monitoring data or geological exposure information is acquired, the local parameters of the stratigraphic attribute inference model are fine-tuned and the local mesh of the three-dimensional geological entity model is reconstructed to achieve dynamic updating of the three-dimensional geological entity model.

2. The geological modeling method based on slope digital twins according to claim 1, characterized in that, The acquisition and fusion of multi-source point cloud data to construct a high-precision 3D surface model of the slope includes: Macroscopic texture point clouds of slopes are obtained by oblique photography of drones, and high-density point clouds of slope facades and hidden areas are obtained by ground laser scanning. By using image control points deployed in the slope area, the macro-texture point cloud and the high-density point cloud are unified to the absolute geographic coordinate system; A probability-weighted algorithm based on the incident angle is used to fuse multi-source point clouds after unifying the coordinate system to construct a three-dimensional surface model of the slope.

3. The geological modeling method based on slope digital twins according to claim 2, characterized in that, The image control point is a dual-mode target with a high-contrast pattern printed on its surface and a high-reflectivity sphere installed at its center. The absolute geographic coordinates of its geometric center are determined by measuring instruments.

4. The geological modeling method based on slope digital twins according to claim 2, characterized in that, The method employs a probability-weighted algorithm based on the incident angle to fuse multi-source point clouds after unifying the coordinate system, including: For points in the overlapping region, calculate their incident angle relative to each scanning source. The closer the incident angle is to the vertical incident angle, the higher the fusion confidence is assigned. The coordinate information of corresponding points from different scanning sources is weighted and averaged based on the fusion confidence level to obtain the fused point cloud.

5. The geological modeling method based on slope digital twins according to claim 1, characterized in that, The integration of real borehole data and geophysical interpretation profile data to construct a geological attribute map topology structure containing spatial nodes and inter-node relationships includes: Real borehole data and geophysical interpretation profile data are discretized into spatial nodes; Assign node characteristics containing spatial coordinates and geophysical properties to each node; Based on the Euclidean distance between nodes and their geological relevance, the edges connecting the nodes and their features are constructed to form the initial geological attribute map topology.

6. The geological modeling method based on slope digital twins according to claim 1, characterized in that, A low-level attribute inference model based on a graph neural network is constructed. Using the geological attribute topology as input, the model is trained by introducing a loss function with geophysical constraints to infer the stratigraphic attributes of the virtual borehole location, generating encrypted geological data, including: Construct a graph neural network model incorporating a geological attention mechanism; Using the geological attribute map topology as input, the geological attention mechanism is used to automatically learn the feature weights of neighboring nodes in different geological deposition directions, thereby realizing the transfer and aggregation of geological attribute information from known nodes to unknown nodes. During model training, a composite loss function is collected, which includes a data fitting term, a physical constraint term, and a topology smoothing term. The physical constraint term is constructed based on the underlying order law and is used to penalize prediction results that violate the deposition law. The trained model is used to deduce the underlying lithology and physical and mechanical parameters of the virtual borehole location, generating encrypted geological data containing virtual borehole information.

7. The geological modeling method based on slope digital twins according to claim 6, characterized in that, The geological attention mechanism introduces a relative elevation coding function based on the vertical distance between two nodes for dynamic weight correction when calculating the attention coefficient between nodes.

8. The geological modeling method based on slope digital twins according to claim 6, characterized in that, By integrating real borehole data with the encrypted geological data, stratigraphic interfaces are reconstructed and integrated with the surface 3D model to generate a 3D geological entity model of the slope, including: The virtual borehole data and the real borehole data are fused according to a preset confidence weight to form a hybrid geological control point set; Based on the aforementioned set of mixed geological control points, the radial basis function implicit potential field method is used to reconstruct the interfaces of each stratigraphy. The reconstructed stratigraphic interface meshes are subjected to Boolean operations and spatial clipping with the surface 3D model to generate a slope 3D geological entity model with continuous internal strata and geometrically consistent surface.

9. The geological modeling method based on slope digital twins according to claim 8, characterized in that, The method of reconstructing various stratigraphic interfaces using the radial basis function implicit potential field includes: For each formation interface, construct a scalar implicit function such that the function value is zero at the interface contact point, positive above the interface, and negative below the interface; The weighting coefficients of the implicit function are solved using the aforementioned set of mixed geological control points; The moving cube algorithm is used to extract negative isosurfaces from the implicit function to generate a continuous and smooth triangular mesh of the stratigraphic interface.

10. The geological modeling method based on slope digital twins according to claim 1, characterized in that, The incremental learning and updating mechanism, when acquiring new slope monitoring data or geological exposure information, involves fine-tuning local parameters of the stratigraphic attribute extrapolation model and reconstructing the local mesh of the three-dimensional geological entity model to achieve dynamic updating of the three-dimensional geological entity model, including: Establish a replay cache to store representative historical training samples; When monitoring data is updated or new geological information is revealed, the backbone network parameters of the stratigraphic property inference model are frozen, and a local training subgraph is constructed using the new data and historical training samples in the replay cache. Fine-tune the weights of the model's output layer and the network weights adjacent to the newly added data space, and introduce a parameter importance regularization term into the loss function; Based on the fine-tuned model deduction results, the local mesh of the affected area in the three-dimensional geological entity model is reconstructed and seamlessly stitched together.

Images