Digital mine three-dimensional twinborn model construction method, equipment and medium

Abstract

The invention discloses a digital mine three-dimensional twinborn model construction method and device and a medium, and relates to the technical field of digital twines.The digital mine three-dimensional twinborn model construction device comprises a voxel fusion module, a geological intelligent deduction module and a model construction module, and the geological intelligent deduction module is used for conducting spatio-temporal evolution on a mine area structure to obtain twinborn model constraint parameters; the geological intelligent deduction module comprises a topology analysis unit, a dynamic deduction unit and a constraint quantization unit; the topology analysis unit is provided with an FPGA board card, isosurface extraction is carried out on the initial snapshot of the underground structure, confidence estimation is carried out, and a mine three-dimensional structural body is obtained. According to the method, topological relation analysis, finite element analysis and multi-physics field coupling deduction are executed in the FPGA board card, so that the mechanical stability of the twin model in a complex stress field and a seepage field is improved. Through a PLAXIS simulation engine and a constraint quantization unit, constraint and correction of the twin model are realized, and the intelligent evolution ability of the twin model is improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method, equipment and medium for constructing a three-dimensional twin model of a digital mine. Background Technology

[0002] With the continuous development of mining engineering towards digitalization and intelligence, 3D twin models of mines are of great significance in geological exploration, mine planning, and safety monitoring. Traditional 3D modeling methods mainly rely on static geological mapping and other means to reconstruct and restore the geometry and geological structure of mining areas in both spatial and temporal dimensions. In recent years, geological simulation methods such as finite element analysis (FEM) and physical modeling-based digital twin technology have been widely used, improving the accuracy and reliability of geological evolution prediction to a certain extent.

[0003] However, traditional methods for constructing 3D twin models of mines still have some problems that urgently need to be addressed. First, existing static geometric reconstruction strategies neglect the topological constraints and physical coupling laws within geological structures, leading to evolutionary deviations in the twin model under varying ground pressure conditions and posing a risk of mechanical distortion. Second, the lack of a dynamic constraint mechanism based on multiphysics feedback means that the boundary conditions of the twin model cannot be corrected in real time according to geological structure feedback. This results in the twin model's accuracy relying on manual intervention and empirical calibration, affecting its intelligent evolution capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method, equipment, and medium for constructing a three-dimensional twin model of a digital mine to solve the problems of easy mechanical distortion and insufficient intelligent evolution capabilities.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a digital mine three-dimensional twin model construction device, which includes a voxel fusion module, a geological intelligent inference module, and a model construction module. The geological intelligent inference module is used to perform spatiotemporal evolution of the mining area structure to obtain twin model constraint parameters. The geological intelligent inference module includes a topology analysis unit, a dynamic inference unit, and a constraint quantization unit; The topology parsing unit is equipped with an FPGA board, which can extract isosurfaces from the initial snapshot of the underground structure, perform confidence estimation, obtain the three-dimensional structure of the mine, and perform topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine, outputting a connected topology skeleton map of the mining area and a multi-category semantic tag set. The dynamic simulation unit integrates the PLAXIS simulation engine, which can perform the spatiotemporal evolution of geological structures and multi-physics coupling based on the connected topology skeleton map of the mining area and multi-category semantic tag set, and output a multi-physics coupling dataset. The constraint quantization unit, which embeds a finite element analysis algorithm, identifies key parameters and quantizes constraint boundaries in a multiphysics coupled dataset, and outputs constraint parameters for the twin model.

[0007] As a preferred embodiment of the digital mine three-dimensional twin model construction device of the present invention, the voxel fusion module and the geological intelligent inference module are connected through an industrial Ethernet bus to collect multi-source heterogeneous data of the mine and perform data processing and voxel fusion reconstruction. The voxel fusion module includes a data acquisition unit, a data regularization unit, and a voxel fusion reconstruction unit; The data acquisition unit is equipped with lidar and depth camera to collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area. The data normalization unit binds device identifiers to multi-source heterogeneous data from the mine and unifies the timestamps, outputting a standard raw dataset. The voxel fusion reconstruction unit performs multi-frame point cloud registration and voxel fusion on the standard original dataset to generate an initial snapshot of the underground structure.

[0008] As a preferred embodiment of the digital mine three-dimensional twin model construction device of the present invention, the model construction module and the geological intelligent inference module are connected through an EtherCAT interface, which is used to receive twin model constraint parameters and perform parameter initialization and model construction; The model building module includes an initialization unit, a hierarchical building unit, and a hierarchical stacking unit; The initialization unit has a built-in PyTorch framework and can call generative adversarial architectures and graph convolutional neural networks; it encodes the constraint parameters of the Siamese model into constraint embedding vectors, which are used to limit the parameter search space of the generative adversarial architecture and graph convolutional neural network, and obtain the constraint-optimized generative adversarial architecture and graph convolutional neural network. The hierarchical building unit is configured with a GPU accelerator to deploy the constraint-optimized generative adversarial architecture and graph convolutional neural network in parallel, and to build a geometric reconstruction layer and a semantic reasoning layer. The hierarchical stacking unit integrates cross-level interactive routing, performs hierarchical stacking on the geometric reconstruction layer and the semantic reasoning layer, and outputs a three-dimensional digital twin model of the mine.

[0009] As a preferred embodiment of the digital mine 3D twin model construction device of the present invention, the output of standard raw dataset specifically refers to reading the frame source field of multi-source heterogeneous data of the mine, binding the device identifier to the multi-source heterogeneous data of the mine according to the frame source field, and using the PTP time synchronization protocol to unify the timestamps, and outputting standard raw dataset.

[0010] As a preferred embodiment of the digital mine 3D twin model construction device of the present invention, the generation of the initial snapshot of the underground structure specifically refers to performing spatiotemporal alignment and multi-frame point cloud registration on the standard original dataset to obtain a unified coordinate point cloud, and performing voxel fusion and attribute field reconstruction to generate the initial snapshot of the underground structure.

[0011] As a preferred embodiment of the digital mine 3D twin model construction device of the present invention, the output of the mine area connectivity topology skeleton map and multi-category semantic tag set specifically includes the following steps. The Marching Cubes algorithm is embedded in the FPGA board to extract isosurfaces from the initial snapshot of the underground structure, forming a voxel isosurface mesh, and confidence estimation is performed to obtain the three-dimensional structure of the mine. Perform skeleton extraction and topology analysis on the three-dimensional structure of the mine to generate a connected topology skeleton map of the mining area; Based on the semantic rule base, semantic segmentation is performed on the three-dimensional structure of the mine, and a multi-category semantic tag set is output.

[0012] In a preferred embodiment of the digital mine 3D twin model construction device of the present invention, the output of twin model constraint parameters specifically includes the following steps. The topological skeleton map of the mining area and the multi-category semantic tag set are input into the PLAXIS simulation engine to perform the spatiotemporal evolution of geological structure and form a spatiotemporal prior field of ground pressure seepage. Multiphysics coupling and boundary condition mapping are performed on the spatiotemporal prior field of geopressure seepage, and a multiphysics coupling dataset is output. The finite element analysis algorithm is applied to identify key parameters in a multiphysics coupled dataset and generate safety control parameters. Based on the interconnected topology of the mining area, the constraint boundaries of the safety control parameters are quantified, and the constraint parameters of the twin model are output.

[0013] As a preferred embodiment of the digital mine 3D digital twin model construction device of the present invention, the output of the mine 3D digital twin model specifically includes the following steps. Using a gated attention mechanism, feature channel weights are adjusted and information flow paths are controlled for constrained optimization generative adversarial architectures and graph convolutional neural networks to build initial reconstruction layers and inference layers. Parallel deployment of the initial reconstruction layer and inference layer is performed using a GPU accelerator to obtain the geometric reconstruction layer and semantic inference layer; Construct cross-level interactive routing through the DDS message event bus and multi-head attention mechanism; Based on cross-level interaction routing, feature interaction and hierarchical stacking are performed on the geometric reconstruction layer and semantic reasoning layer to output a 3D digital twin model of the mine.

[0014] Secondly, this invention provides a method for constructing a three-dimensional twin model of a digital mine, including, Collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area; bind equipment identifiers to the multi-source heterogeneous data of the mine and unify the timestamps to output a standard raw dataset; perform multi-frame point cloud registration and voxel fusion on the standard raw dataset to generate an initial snapshot of the underground structure. Configure an FPGA board to extract isosurfaces from the initial snapshot of the underground structure and perform confidence estimation to obtain the three-dimensional structure of the mine. Simultaneously, perform topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine, and output a connected topological skeleton map of the mining area and a multi-category semantic label set. Based on the connected topological skeleton map of the mining area and the multi-category semantic label set, use the PLAXIS simulation engine to perform spatiotemporal evolution of geological structures and multi-physics coupling, output a multi-physics coupling dataset, and use the finite element analysis algorithm to identify key parameters and quantize constraint boundaries of the multi-physics coupling dataset, and output twin model constraint parameters. In the PyTorch framework, generative adversarial architecture (GAP) and graph convolutional neural network (GCNN) are invoked; the constraint parameters of the twin model are encoded into constraint embedding vectors, and the parameter search space of the GAP and GCNN is limited to obtain constraint-optimized GAP and GCNN; a GPU accelerator is configured to deploy the constraint-optimized GAP and GCNN in parallel, build a geometric reconstruction layer and a semantic inference layer, and integrate cross-level interaction routing to perform hierarchical stacking of the geometric reconstruction layer and the semantic inference layer, outputting a 3D digital twin model of the mine.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for constructing a three-dimensional twin model of a digital mine as described in the second aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By performing topological relationship analysis, finite element analysis, and multiphysics coupling deduction on an FPGA board, the topological constraints and physical coupling laws of geological structures can be incorporated into the twin model, improving the mechanical stability of the twin model under complex stress and seepage fields and avoiding mechanical distortion. Through the PLAXIS simulation engine and constraint quantization unit, the constraints and corrections of the twin model are realized, enabling the twin model to update boundary conditions and control parameters according to environmental changes such as ground pressure and seepage, eliminating reliance on manual experience calibration and improving the intelligent evolution capability of the twin model. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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 these drawings without creative effort.

[0018] Figure 1 A schematic diagram of equipment for building 3D twin models of digital mines.

[0019] Figure 2 A flowchart for generating an initial snapshot of the underground structure.

[0020] Figure 3 A flowchart for generating the topology skeleton diagram of the mining area connectivity.

[0021] Figure 4 A flowchart for generating constraint parameters for a twin model. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4This is one embodiment of the present invention, which provides a method, equipment, and medium for constructing a three-dimensional twin model of a digital mine, including the following steps: The voxel fusion module includes a data acquisition unit, a data warping unit, and a voxel fusion reconstruction unit.

[0026] The data acquisition unit is equipped with lidar and depth cameras to collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area.

[0027] Specific operations: Multi-source heterogeneous data in mines includes laser point cloud data, surface image data, and operational environment parameters. LiDAR is installed in the internal roadways of the mine structure to collect laser point cloud data; depth cameras are placed on the working face of the mining area to collect surface image data; and industrial environmental monitoring software (such as VentSim) is used to monitor the temperature, humidity, gas concentration, and wind speed of the mining area to obtain operational environment parameters.

[0028] The data normalization unit reads the frame source field of the multi-source heterogeneous data in the mine, binds the device identifier to the multi-source heterogeneous data in the mine according to the frame source field, and uses the PTP time synchronization protocol to unify the timestamps and outputs the standard raw dataset.

[0029] Specific operations: The asynchronous IoSession session of the Apache Mina library is used to perform frame-level unpacking of multi-source heterogeneous data from the mine, obtain the data frame header information, and extract the frame source fields in the data frame header information, such as device IP address, MAC address, etc. It should be noted that asynchronous IoSession sessions achieve frame-level unpacking by performing streaming buffer parsing and byte sequence reassembly on multi-source heterogeneous data from the mine.

[0030] A consistent hashing algorithm is used to perform hash matching between the frame source field and the device registration mapping table to obtain the corresponding device number and acquisition channel information; a JSON serializer is used to perform key-value mapping between the device number and the acquisition channel information to obtain a device identification data object, and a uniqueness verification is performed on the device identification data object to generate a unique device identifier; the unique device identifier is bound to the multi-source heterogeneous data of the mine through the Kafka data bus pipeline to form unique identifier multi-source data; It should be noted that the device registration mapping table is directly loaded through a distributed device management service (such as Redis Cluster); the consistent hashing algorithm achieves hash matching by performing hash ring mapping and virtual node positioning on the frame source field and the device registration mapping table; uniqueness verification refers to the process of conflict detection and redundancy verification of device identification data objects based on the device registration mapping table.

[0031] Based on the Transparent Clock service of the PTP time synchronization protocol, the local PTP time that uniquely identifies the arrival time of multi-source data at the data acquisition gateway (such as the OPC UA gateway) is recorded and compared with the acquisition timestamp of multi-source heterogeneous data in the mine. When the time deviation exceeds the error threshold, it is judged as an abnormal timestamp. Linear interpolation is used to resample and smooth the time series of abnormal timestamps to achieve timestamp unification and obtain a standard original dataset. It should be noted that the error threshold is defined based on the statistical distribution characteristics of historical time deviations. For example, local PTP time is recorded within a rolling 24-hour window, the mean (μ) and standard deviation (σ) of the time deviation are obtained, and the error threshold is set to μ ± 3σ to cover the normal fluctuation range of the time deviation.

[0032] Ideally, by binding device identifiers and unifying timestamps for multi-source heterogeneous data, asynchronous errors caused by data source ambiguity and time jitter can be eliminated, ensuring that stable processing throughput can still be maintained when high-concurrency multi-source heterogeneous data floods in.

[0033] The voxel fusion reconstruction unit performs spatiotemporal alignment and multi-frame point cloud registration on the standard original dataset to obtain a unified coordinate point cloud, and performs voxel fusion and attribute field reconstruction to generate an initial snapshot of the underground structure.

[0034] Specific operations: By sorting time frames of the standard original dataset using a sliding window, temporal alignment is achieved, and spatial interpolation is performed on the standard original dataset to eliminate geometric distortion and obtain spatiotemporally aligned data packets. Point cloud reconstruction is performed on the spatiotemporally aligned data packets to generate initial point cloud data. The initial point cloud data is then denoised and outlier removed to eliminate noise and artifact interference caused by equipment jitter, forming a clean point cloud set. The clean point cloud set is then converted to a unified coordinate system through affine transformation, and each clean point cloud is batch-transformed to a unified coordinate system. Statistical filtering and voxel downsampling are performed simultaneously to improve the distribution uniformity of the clean point cloud and reduce redundancy, resulting in a unified coordinate point cloud. It should be noted that point cloud reconstruction refers to the process of performing depth reprojection and coordinate registration on spatiotemporally aligned data packets using the NDT matching algorithm.

[0035] Based on the octree spatial index, the unified coordinate point cloud is projected onto a 3D voxel grid and sparsification is performed to reduce storage burden while preserving structural details, resulting in a sparse voxel set. The Marching Cubes fusion pipeline is then applied to perform voxel-by-voxel aggregation on the sparse voxel set to form voxel fused data. It should be noted that the octree spatial index is a hierarchical partitioned data structure, implemented through the Octree interface in the PCL library.

[0036] The attribute field is reconstructed based on the voxel fusion data. Further, attribute mapping and unit normalization are performed on the voxel fusion data to obtain a multi-source attribute volume. Axial slicing sampling is performed on the multi-source attribute volume to generate attribute volume slices. The attribute volume slices are then divided and indexed to obtain a three-dimensional attribute field. The spatial range and version information of the three-dimensional attribute field are extracted, and a packaging registration is performed to output an initial snapshot of the underground structure. It should be noted that axial slice sampling refers to the process of performing axial step traversal and equidistant truncation projection on a multi-source attribute volume.

[0037] The geological intelligent simulation module includes a topological analysis unit, a dynamic simulation unit, and a constraint quantization unit.

[0038] The topology parsing unit, equipped with an FPGA board, can extract isosurfaces from the initial snapshot of the underground structure and perform confidence estimation to obtain the three-dimensional structure of the mine. At the same time, it performs topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine, and outputs a connected topology skeleton map of the mining area and a multi-category semantic label set.

[0039] Specific operations: The Marching Cubes algorithm is embedded in an FPGA board to extract isosurfaces and perform confidence estimation on the initial snapshot of the underground structure. Furthermore, the Marching Cubes algorithm is written into the FPGA board using a synthesis toolchain (such as Vitis HLS), enabling the FPGA board to have hardware acceleration capabilities for parallel processing of voxel data. The FPGA board is used to perform voxel block traversal and edge table lookup on the initial snapshot of the underground structure to obtain the set of edge intersections. Linear interpolation and Delaunay triangulation are performed on the edge intersection set to output triangular isosurfaces. Vertex clustering and neighborhood search are performed on the triangular isosurfaces to obtain the voxel isosurface mesh. It should be noted that Delaunay triangulation is an optimal triangular mesh generation method that satisfies the empty circle property. It refers to the process of topologically connecting the set of edge intersections and verifying the empty circumcircle. Vertex clustering refers to the process of merging vertices and simplifying the mesh by using the Euclidean distance metric to measure the isosurfaces of triangular pieces.

[0040] Confidence estimation is performed on the voxel isosurface mesh to obtain the surface-level confidence score. The specific mathematical formula is as follows. ; in, Indicates the face-level confidence level. This represents the total number of vertices contained in the voxel isosurface mesh. Represents the vertex index. Indicates the first The three-dimensional coordinates of each vertex. Indicates the first The distance gradient value corresponding to the three-dimensional coordinates of each vertex. Represents the gradient sensitivity coefficient. Represents the natural exponential function; It should be noted that the three-dimensional coordinates of the vertex are obtained by performing vertex extraction and coordinate normalization on the voxel isosurface mesh; the distance gradient is obtained by performing spatial difference differentiation on the three-dimensional coordinates of the vertex; the gradient sensitivity coefficient is used to adjust the influence intensity of the distance gradient value, and is defined based on the distribution dispersion rate of historical distance gradient values, with an exemplary value range of 0.5-2.0.

[0041] When the face-level confidence of the voxel isosurface mesh is lower than the quality threshold, it indicates that there is reconstruction uncertainty and is marked as a region to be optimized; when the face-level confidence is lower than the quality threshold, it indicates that the geometry is reliable and is marked as a region to be directly retained; surface interpolation is performed on the region to be optimized to fill geometric gaps and smooth surface continuity to obtain the repaired region; vertex remapping and edge fusion are performed on the repaired region and the directly retained region to generate the 3D structure of the mine; It should be noted that the quality threshold is defined based on the patch curvature distribution of the historical voxel isosurface mesh. For example, by statistically analyzing the patch curvature samples of the historical voxel isosurface mesh and performing Gaussian fitting, the patch curvature distribution is obtained. The patch curvature distribution is then truncated to percentiles to obtain the quality threshold. An exemplary value range is 0.15 to 0.35. Vertex remapping is achieved by aligning the coordinate system and matching neighboring vertices between the repaired region and the directly retained region. Edge blending refers to the process of using the Laplacian smoothing algorithm to smooth the seams and optimize the normal consistency of the repaired region and the directly retained region after vertex remapping.

[0042] Skeleton extraction and topological relationship analysis are performed on the three-dimensional structure of the mine. Furthermore, the boundary structure voxels of the three-dimensional structure of the mine are extracted and morphological erosion and hole filling are performed to obtain a continuous internal space. Under the premise of preserving the geometric shape, constraint contraction and central axis plane approximation are performed on the continuous internal space to generate a candidate set of central domains. It should be noted that the constraint contraction is achieved by performing boundary anchoring and net radius constraint operations on the continuous internal space; the approach of the central axis plane is achieved by performing equidistant layer stripping and symmetrical projection operations on the continuous internal space. Dijkstra's algorithm is used to perform a global path traversal on the candidate set of central regions to obtain the main centerline. Based on the spatial constraints of the mine roadways, the main centerline is merged and bifurcated to obtain a directed topological skeleton. The directed topological skeleton is then weighted and fused to form a weighted connected skeleton graph. Kalman filtering is applied to remove noise loops from the weighted connected skeleton graph and output the connected topological skeleton graph of the mine area. It should be noted that Dijkstra's algorithm achieves global path traversal by expanding nodes and updating paths in the candidate set of the central domain; the spatial constraint relationship of the mine roadway is based on the safety distance definition of the roadway design and is used for topology simplification and path optimization of the main trunk centerline.

[0043] Based on the semantic rule base, the three-dimensional structure of the mine is semantically segmented, and a multi-category semantic tag set is output. Furthermore, based on the geological structure of the mine and the lithological distribution law of the rock strata (i.e., the spatial distribution characteristics and combination patterns of different rock types in the mine geological body), a semantic rule base is defined to constrain the generation logic and classification boundary of the semantic tags. Local geometric features (such as surface curvature, cross-sectional area, dip angle, roughness, etc.) of the three-dimensional structure of the mine are extracted and semantically segmented with semantic tags in the semantic rule base. For example, when the local geometric features of a certain area in the three-dimensional structure of the mine simultaneously satisfy "cross-sectional area > 10㎡" and "surface curvature < 0.05", the semantic tag "main transport roadway" is assigned. When the local geometric features of a certain area satisfy "dip angle > 30°" and "roughness > 0.8", the semantic tag "sharply dipping rock fault zone" is assigned. All assigned semantic tags are integrated to generate a multi-category semantic tag set.

[0044] The dynamic simulation unit integrates the PLAXIS simulation engine, which can perform spatiotemporal evolution of geological structures and multi-physics coupling based on the connected topology skeleton map of the mining area and multi-category semantic label set, and output a multi-physics coupling dataset.

[0045] Specific operations: The connected topological skeleton map of the mining area and the multi-category semantic label set are input into the PLAXIS simulation engine to perform the spatiotemporal evolution of geological structure. Furthermore, the connected topological skeleton map of the mining area and the multi-category semantic label set are dimensionally aligned, and semantic fusion is performed using the graph attention mechanism to obtain the topological skeleton with semantic annotation. The rock mass constitutive parameters in the topological skeleton with semantic annotation are extracted as the simulation condition input. It should be noted that the PLAXIS simulation engine is a geotechnical engineering simulation platform based on the finite element method, capable of simulating multi-physics interaction processes such as rock mass deformation and seepage coupling under complex geological conditions.

[0046] The simulation conditions are imported into the PLAXIS simulation engine. During the simulation, multiple rounds of cyclic perturbation loads are applied to the simulation conditions (the number of rounds is defined based on the actual needs of the engineering safety assessment, such as 12 rounds) to simulate the process of rock deformation and pore water pressure diffusion, and to obtain the simulated gravitational field. After the simulation is completed, the nodal stress values ​​and pore water pressure data of the simulated stress field are collected and integrated to obtain the spatiotemporal evolution data of the stress field.

[0047] Equal time interval reconstruction is performed on the spatiotemporal evolution data of stress field: Principal component analysis is applied to perform feature dimensionality reduction and modal decomposition on the spatiotemporal evolution data of stress field to obtain stress modal components. Amplitude normalization and phase synchronization are performed on the stress modal components to generate a standard stress field distribution. Then, radial basis functions are used to perform spatiotemporal surface fitting on the standard stress field distribution to form a spatiotemporal prior field of ground pressure seepage. It should be noted that spatiotemporal surface fitting refers to the process of kernel basis expansion and nodal interpolation of the standard stress field distribution.

[0048] Multiphysics coupling is performed on the spatiotemporal prior field of ground pressure seepage. Furthermore, the ADMM coupling optimization algorithm is used to decouple and reconstruct the spatiotemporal prior field of ground pressure seepage, generating intermediate coupled physical field data. Relaxation iteration is performed on the intermediate coupled physical field data to eliminate local non-equilibrium oscillations and obtain equilibrium multiphysics data. The equilibrium multiphysics data is then dimension-unified to generate a multiphysics coupling vector. It should be noted that the ADMM coupled optimization algorithm achieves decoupling and reconstruction by splitting variables and projecting consistent constraints onto the spatiotemporal prior field of ground pressure seepage; relaxation iteration refers to the process of local linearization approximation and global constraint relaxation of the intermediate data of the coupled physical field according to the residual convergence criterion (based on the stress relative residual norm definition of the intermediate data of the coupled physical field).

[0049] According to the working condition configuration specification, the multiphysics coupling vector is mapped to the boundary conditions to obtain the initial boundary conditions; a Gaussian filter is used to perform bilateral filtering on the initial boundary conditions to enhance the boundary continuity and generate optimized boundary conditions; the location index, boundary type and constraint parameters of the optimized boundary conditions are extracted and weighted and fused to generate a multiphysics coupling dataset. It should be noted that the working condition configuration specification is based on the mechanical characteristics definition of actual mining conditions, and achieves boundary condition mapping by performing type matching and parameter assignment on multi-physics coupling vectors.

[0050] The constraint quantization unit, which embeds a finite element analysis algorithm, identifies key parameters and quantizes constraint boundaries in a multiphysics coupled dataset, and outputs constraint parameters for the twin model.

[0051] Specific operations: The finite element method (FEM) algorithm is applied to identify key parameters in a multiphysics coupled dataset. Furthermore, singular value decomposition (SVD) is used to perform tensor decomposition on the dataset, generating a low-dimensional feature space. Principal component analysis (PCA) is then applied to perform variance rotation and orthogonalization on this low-dimensional feature space to obtain the finite element feature basis. Eigenvalue decomposition is performed on the finite element feature basis, and the top u eigenvalues ​​are extracted as modal key parameters. Based on the working condition safety assessment criteria, safety criterion constraints are embedded into the modal key parameters to generate safety control parameters. It should be noted that singular value decomposition achieves tensor decomposition by extracting eigenvalues ​​and constructing orthogonal bases from multi-physics coupled datasets; the value of u is defined based on the eigenvalue contribution rate of the finite element eigenbase; and the working condition safety assessment criteria are defined based on the stability grading standard of mine rock masses.

[0052] Based on the connected topology skeleton diagram of the mining area, the safety control parameters are constrained and quantified. Furthermore, a bidirectional index relationship is established between the nodes (intersections, endpoints) and edges (roadway segments, connecting segments, mining passages) of the connected topology skeleton diagram of the mining area and the three-dimensional structure of the mine. For example, spatial coordinate mapping is performed on the nodes, mapping the nodes to the nearest neighbor voxels of the three-dimensional structure of the mine. At the same time, the edges are topologically linked with the three-dimensional structure of the mine through the ICP iterative nearest point algorithm, and the skeleton-entity association index table is output. It should be noted that the ICP iterative nearest neighbor algorithm achieves topological links by performing nearest neighbor matching and rigid body transformation on edges and the 3D structure of the mine.

[0053] Spatial domain resampling of safety control parameters yields a discrete safety control parameter field. Constraint parameters are matched to the discrete safety control parameter field based on the skeleton-entity association index table. For example, when the location index of the discrete safety control parameter field is located at the intersection of roadways, structural bearing constraint parameters are matched in the skeleton-entity association index table; when the location index of the discrete safety control parameter field is located in the edge area of ​​the mining area, surrounding rock deformation constraint parameters are matched in the skeleton-entity association index table. It should be noted that spatial domain resampling refers to the process of spatial domain discretization and resampling of safety control parameters; the position index is obtained by performing coordinate system alignment and position encoding on the discrete safety control parameter field.

[0054] The matched constraint parameters are time-series aligned and smoothed using a sliding window to obtain a time-varying constraint parameter sequence. Min-Max standardization is applied to normalize the time-varying constraint parameter sequence to obtain the constraint parameters of the twin model.

[0055] The model building module includes an initialization unit, a hierarchical building unit, and a hierarchical stacking unit.

[0056] The initialization unit, with a built-in PyTorch framework, can call generative adversarial architectures and graph convolutional neural networks, and encodes the constraint parameters of the Siamese model into constraint embedding vectors to limit the parameter search space of the generative adversarial architecture and graph convolutional neural network, thereby obtaining the constraint-optimized generative adversarial architecture and graph convolutional neural network.

[0057] Specific operations: In the PyTorch framework, the adversarial architecture is generated by calling the nn.Module parameter and the graph convolutional neural network is called by calling the nn.GCNConv parameter. The constraint parameters of the twin model are structured into binary encoding. For example, the hard constraint flag of the twin model constraint parameters is encoded as bit 1, and the soft constraint flag of the twin model constraint parameters is encoded as bit 0, forming a binary field stream. Then, LDA linear discriminant analysis is applied to compress the dimension of the binary field stream to obtain the constraint embedding vector. It should be noted that the constraint flags of the twin model constraint parameters are obtained by performing a logical state determination on the twin model constraint parameters; LDA linear discriminant analysis achieves dimensionality compression by maximizing inter-class divergence and minimizing intra-class divergence on the binary field stream.

[0058] Based on the constraint embedding vector, the parameter search space of the generative adversarial architecture and the graph convolutional neural network is limited. Furthermore, a multilayer perceptron is used to decode and map the constraint embedding vector to obtain multidimensional control parameters. Based on the parameter control parameters, the learning rate of the generative adversarial architecture is decayed through the StepLR learning rate scheduler in the PyTorch framework, and gradient truncation is performed by applying norm-constrained gradient clipping (Clip-Norm) to achieve search space limitation and form a constraint-optimized generative adversarial architecture. It should be noted that the multilayer perceptron achieves decoding mapping by performing linear transformations and nonlinear activations on the constraint embedding vectors. For example, by combining fully connected layers with the ReLU activation function, the constraint embedding vectors are mapped to multidimensional control parameters. Norm-constrained gradient clipping achieves gradient truncation by scaling the gradient projection of the generative adversarial architecture. For example, the threshold for norm-constrained gradient clipping is set to 2, and the gradients of the generative adversarial architecture that exceed the threshold of norm-constrained gradient clipping are scaled until the L2 norm falls within the threshold range of norm-constrained gradient clipping.

[0059] Orthogonal projection transformation is applied to perform spatial projection on the constraint embedding vector to generate a constraint subspace. Based on the constraint subspace, adjacency weight constraints and propagation path regulation are applied to the graph convolutional neural network to obtain a spatially constrained graph convolutional neural network. Sparse regularization is applied to optimize the stability of the spatially constrained graph convolutional neural network to obtain a constraint-optimized graph convolutional neural network. It should be noted that adjacency weight constraint refers to the process of spectral norm normalization and feature space projection on the graph convolutional neural network; propagation path regulation refers to the process of using gating mechanisms to prune redundant paths and enhance effective paths on the graph convolutional neural network.

[0060] The hierarchical building blocks are configured with GPU accelerators to deploy constrained optimization generative adversarial architectures and graph convolutional neural networks in parallel, building geometric reconstruction layers and semantic reasoning layers.

[0061] Specific operations: Using a gated attention mechanism, feature channel weights and information flow paths are adjusted for a constrained optimization generative adversarial architecture and a graph convolutional neural network. Furthermore, feature tensor projection is performed on the constrained optimization generative adversarial architecture to obtain query vectors and key vectors, and scaled dot products are performed on the query vectors and key vectors to generate channel attention weights. Based on the channel attention weights, redundant information is suppressed for the constrained optimization generative adversarial architecture, thereby adjusting the feature channel weights and generating the initial reconstruction layer. It should be noted that tensor projection refers to the process of applying a gated attention mechanism to perform linear transformation and feature space mapping on a constrained optimization generative adversarial architecture; redundancy information suppression refers to the process of assigning channel weights and attenuating redundant components on a constrained optimization generative adversarial architecture.

[0062] The node features of the constrained optimized graph convolutional neural network are extracted by graph convolution and linearly transformed to obtain the node response vector. The node response vector is nonlinearly mapped and numerically normalized using the sigmoid activation function to generate propagation path control coefficients. Based on the propagation path control coefficients, the constrained optimized graph convolutional neural network is gated and truncated to achieve information flow path control and obtain the initial inference layer. It should be noted that gating truncation refers to the process of modulating path intensity and guiding information flow in a constrained optimization graph convolutional neural network through a gating attention mechanism.

[0063] Parallel deployment of the initial reconstruction layer and inference layer is performed through GPU accelerator. Furthermore, the CUDA stream management API is called and shared memory is allocated to build a dual-channel CUDA thread. The dual-channel CUDA thread is embedded into the GPU accelerator through a unified virtual address space (such as NVIDIA UVA) to obtain a parallel processing pipeline. It should be noted that shared memory allocation refers to the process of using the CUDA memory management interface to lock video memory pages and configure memory space using the CUDA stream management API.

[0064] Through a parallel processing pipeline, the initial reconstruction layer and inference layer are decomposed into tasks and data blocks. Based on the complexity balance principle, the initial reconstruction layer and inference layer are divided into several processing task sub-blocks and allocated to different GPU accelerators for parallel processing. This achieves load balancing and reduces memory bandwidth consumption, thus obtaining a distributed processing layer. The distributed processing layer is then stream-scheduled through a cross-GPU stream scheduler to achieve efficient parallel deployment and generate the geometric reconstruction layer and semantic inference layer. It should be noted that streaming scheduling refers to the process of orchestrating task pipelines and dynamically allocating resources across distributed processing layers using a cross-GPU streaming scheduler.

[0065] The hierarchical stacking unit integrates cross-level interactive routing, performs hierarchical stacking on the geometric reconstruction layer and semantic reasoning layer, and outputs a 3D digital twin model of the mine.

[0066] Specific operations: Cross-level interactive routing is constructed using the DDS message event bus and multi-head attention mechanism. Further, the DDS standard API is used to call the DDS message event bus and perform topic registration to obtain reliable publish-subscribe channels. Feature-channel association modeling is then performed on the multi-head attention mechanism and reliable publish-subscribe channels: the CCA canonical correlation alignment algorithm is used to align the multi-head attention mechanism and reliable publish-subscribe channels across modalities, obtaining aligned feature representations. These aligned feature representations are then associated and mapped according to channel matching rules to form a feature-channel interaction tensor. The feature-channel interaction tensor is then spatiotemporally encoded to generate a dynamically aware routing table. Finally, the routing controller optimizes the path and shapes the traffic in the dynamically aware routing table to form cross-level interactive routing. It should be noted that the CCA canonical correlation alignment algorithm achieves cross-modal alignment by mapping the feature space of the multi-head attention mechanism and the reliable publish-subscribe channel; the channel matching rule is based on the correlation strength definition of the multi-head attention mechanism and the reliable publish-subscribe channel; the spatiotemporal relationship encoding refers to the process of embedding positional information and establishing temporal relationships in the feature-channel interaction tensor through a position encoder.

[0067] It should also be noted that in the process of building cross-level interactive routing, the application of the DDS message event bus ensures the timeliness of cross-level data interaction, while the application of the multi-head attention mechanism realizes the capture of multi-granularity feature relationships. Compared with the existing centralized polling scheduling scheme, it has the advantages of low latency and high throughput.

[0068] Based on the cross-level interactive routing, feature interaction and hierarchical stacking are performed on the geometric reconstruction layer and the semantic inference layer. Further, in the cross-level interactive routing, feature splicing and residual connections are performed on the geometric reconstruction layer and the semantic inference layer to achieve feature interaction and obtain cross-modal feature tensors. A multi-head attention mechanism is used to perform channel weighting on the cross-modal feature tensors to obtain weighted cross-modal features. The Softmax normalization function is used to transform the probability distribution of the weighted cross-modal features to generate hierarchical fusion weights. Based on the hierarchical fusion weights, the geometric reconstruction layer and the semantic inference layer are stacked hierarchically to complete the construction of the 3D digital twin model of the mine. It should be noted that the three-dimensional digital twin model of a mine can realize the dynamic coupling and deduction of geological structure and mining activities. Compared with the existing static geometric modeling methods, it has the ability to perceive and model the state of the entire life cycle of the mine.

[0069] This embodiment also provides a method for constructing a three-dimensional twin model of a digital mine, including: Collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area; bind equipment identifiers to the multi-source heterogeneous data of the mine and unify the timestamps to output a standard raw dataset; perform multi-frame point cloud registration and voxel fusion on the standard raw dataset to generate an initial snapshot of the underground structure. Configure an FPGA board to extract isosurfaces from the initial snapshot of the underground structure and perform confidence estimation to obtain the three-dimensional structure of the mine. Simultaneously, perform topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine, and output a connected topological skeleton map of the mining area and a multi-category semantic label set. Based on the connected topological skeleton map of the mining area and the multi-category semantic label set, use the PLAXIS simulation engine to perform spatiotemporal evolution of geological structures and multi-physics coupling, output a multi-physics coupling dataset, and use the finite element analysis algorithm to identify key parameters and quantize constraint boundaries of the multi-physics coupling dataset, and output twin model constraint parameters. In the PyTorch framework, generative adversarial architecture (GAP) and graph convolutional neural network (GCNN) are invoked; the constraint parameters of the twin model are encoded into constraint embedding vectors, and the parameter search space of the GAP and GCNN is limited to obtain constraint-optimized GAP and GCNN; a GPU accelerator is configured to deploy the constraint-optimized GAP and GCNN in parallel, build a geometric reconstruction layer and a semantic inference layer, and integrate cross-level interaction routing to perform hierarchical stacking of the geometric reconstruction layer and the semantic inference layer, outputting a 3D digital twin model of the mine.

[0070] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for constructing a three-dimensional twin model of a digital mine as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0071] In summary, this invention, by performing topological relationship analysis, finite element analysis, and multiphysics coupling derivation on an FPGA board, integrates the topological constraints and physical coupling laws of geological structures into the twin model, improving its mechanical stability under complex stress and seepage fields and avoiding mechanical distortion. Through the PLAXIS simulation engine and constraint quantization unit, the twin model is constrained and corrected, enabling it to update boundary conditions and control parameters based on environmental changes such as ground pressure and seepage, eliminating reliance on manual experience calibration and enhancing the twin model's intelligent evolution capabilities.

[0072] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device for constructing a three-dimensional twin model of a digital mine, characterized in that: It includes a voxel fusion module, a geological intelligent inference module, and a model building module. The geological intelligent inference module is used to obtain twin model constraint parameters by spatiotemporally transforming the structure of the mining area. The geological intelligent inference module includes a topology analysis unit, a dynamic inference unit, and a constraint quantization unit; The topology parsing unit is equipped with an FPGA board. It extracts isosurfaces from the initial snapshot of the underground structure and performs confidence estimation to obtain the three-dimensional structure of the mine. At the same time, it performs topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine and outputs a connected topology skeleton map of the mining area and a multi-category semantic tag set. The dynamic simulation unit integrates the PLAXIS simulation engine, which performs the spatiotemporal evolution of geological structures and multi-physics coupling based on the mining area connectivity topology skeleton map and multi-category semantic tag set, and outputs a multi-physics coupling dataset. The constraint quantization unit, which embeds a finite element analysis algorithm, identifies key parameters and quantizes constraint boundaries in a multiphysics coupled dataset, and outputs constraint parameters for the twin model.

2. The digital mine three-dimensional twin model construction equipment as described in claim 1, characterized in that: The voxel fusion module and the geological intelligent inference module are connected via an industrial Ethernet bus to collect multi-source heterogeneous data from the mine and perform data processing and voxel fusion reconstruction. The voxel fusion module includes a data acquisition unit, a data regularization unit, and a voxel fusion reconstruction unit; The data acquisition unit is equipped with lidar and depth camera to collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area. The data normalization unit binds device identifiers to multi-source heterogeneous data from the mine and unifies the timestamps, outputting a standard raw dataset. The voxel fusion reconstruction unit performs multi-frame point cloud registration and voxel fusion on the standard original dataset to generate an initial snapshot of the underground structure.

3. The digital mine three-dimensional twin model construction equipment as described in claim 1, characterized in that: The model building module and the geological intelligent inference module are connected through the EtherCAT interface, which is used to receive twin model constraint parameters and perform parameter initialization and model building; The model building module includes an initialization unit, a hierarchical building unit, and a hierarchical stacking unit; The initialization unit has a built-in PyTorch framework and calls the generative adversarial architecture and graph convolutional neural network; it encodes the constraint parameters of the Siamese model into constraint embedding vectors, which are used to limit the parameter search space of the generative adversarial architecture and graph convolutional neural network, and obtain the constraint-optimized generative adversarial architecture and graph convolutional neural network. The hierarchical building unit is configured with a GPU accelerator to deploy the constraint-optimized generative adversarial architecture and graph convolutional neural network in parallel, and to build a geometric reconstruction layer and a semantic reasoning layer. The hierarchical stacking unit integrates cross-level interactive routing, performs hierarchical stacking on the geometric reconstruction layer and the semantic reasoning layer, and outputs a three-dimensional digital twin model of the mine.

4. The digital mine three-dimensional twin model construction equipment as described in claim 2, characterized in that: The aforementioned standard raw dataset specifically refers to reading the frame source field of multi-source heterogeneous data from the mine, binding device identifiers to the multi-source heterogeneous data from the mine based on the frame source field, and simultaneously using the PTP time synchronization protocol to unify timestamps, thereby outputting a standard raw dataset.

5. The digital mine three-dimensional twin model construction equipment as described in claim 2, characterized in that: The process of generating an initial snapshot of the underground structure specifically involves performing spatiotemporal alignment and multi-frame point cloud registration on the standard original dataset to obtain a unified coordinate point cloud, and then performing voxel fusion and attribute field reconstruction to generate an initial snapshot of the underground structure.

6. The digital mine three-dimensional twin model construction equipment as described in claim 1, characterized in that: The output of the connected topology skeleton map of the mining area and the multi-category semantic tag set specifically includes the following steps. The Marching Cubes algorithm is embedded in the FPGA board to extract isosurfaces from the initial snapshot of the underground structure, forming a voxel isosurface mesh, and confidence estimation is performed to obtain the three-dimensional structure of the mine. Perform skeleton extraction and topology analysis on the three-dimensional structure of the mine to generate a connected topology skeleton map of the mining area; Based on the semantic rule base, semantic segmentation is performed on the three-dimensional structure of the mine, and a multi-category semantic tag set is output.

7. The digital mine three-dimensional twin model construction equipment as described in claim 1, characterized in that: The output twin model constraint parameters specifically include the following steps. The topological skeleton map of the mining area and the multi-category semantic tag set are input into the PLAXIS simulation engine to perform the spatiotemporal evolution of geological structure and form a spatiotemporal prior field of ground pressure seepage. Multiphysics coupling and boundary condition mapping are performed on the spatiotemporal prior field of geopressure seepage, and a multiphysics coupling dataset is output. The finite element analysis algorithm is applied to identify key parameters in a multiphysics coupled dataset and generate safety control parameters. Based on the interconnected topology of the mining area, the constraint boundaries of the safety control parameters are quantified, and the constraint parameters of the twin model are output.

8. The digital mine three-dimensional twin model construction equipment as described in claim 3, characterized in that: The output of the three-dimensional digital twin model of the mine specifically includes the following steps. Using a gated attention mechanism, feature channel weights are adjusted and information flow paths are controlled for constrained optimization generative adversarial architectures and graph convolutional neural networks to build initial reconstruction layers and inference layers. Parallel deployment of the initial reconstruction layer and inference layer is performed using a GPU accelerator to obtain the geometric reconstruction layer and semantic inference layer; Construct cross-level interactive routing through the DDS message event bus and multi-head attention mechanism; Based on cross-level interaction routing, feature interaction and hierarchical stacking are performed on the geometric reconstruction layer and semantic reasoning layer to output a 3D digital twin model of the mine.

9. A method for constructing a digital mine three-dimensional twin model, based on the digital mine three-dimensional twin model construction equipment according to any one of claims 1 to 8, characterized in that: include, Collect multi-source heterogeneous data of the mine in the internal roadways and working faces of the mining area; bind equipment identifiers to the multi-source heterogeneous data of the mine and unify the timestamps to output a standard raw dataset; perform multi-frame point cloud registration and voxel fusion on the standard raw dataset to generate an initial snapshot of the underground structure. Configure an FPGA board to extract isosurfaces from the initial snapshot of the underground structure and perform confidence estimation to obtain the three-dimensional structure of the mine. Simultaneously, perform topological relationship analysis and semantic segmentation on the three-dimensional structure of the mine, and output a connected topological skeleton map of the mining area and a multi-category semantic label set. Based on the connected topological skeleton map of the mining area and the multi-category semantic label set, use the PLAXIS simulation engine to perform spatiotemporal evolution of geological structures and multi-physics coupling, output a multi-physics coupling dataset, and use the finite element analysis algorithm to identify key parameters and quantize constraint boundaries of the multi-physics coupling dataset, and output twin model constraint parameters. In the PyTorch framework, generative adversarial architecture (GAP) and graph convolutional neural network (GCNN) are invoked; the constraint parameters of the twin model are encoded into constraint embedding vectors, and the parameter search space of the GAP and GCNN is limited to obtain constraint-optimized GAP and GCNN; a GPU accelerator is configured to deploy the constraint-optimized GAP and GCNN in parallel, build a geometric reconstruction layer and a semantic inference layer, and integrate cross-level interaction routing to perform hierarchical stacking of the geometric reconstruction layer and the semantic inference layer, outputting a 3D digital twin model of the mine.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for constructing a three-dimensional twin model of a digital mine as described in any of claims 9.

Images