Tunnel overbreak space boundary three-dimensional image construction system

Abstract

The present application relates to the technical field of tunnel engineering monitoring and three-dimensional visualization, in particular to a tunnel over-under excavation body space boundary three-dimensional image construction system; the system comprises a multi-modal data acquisition module, a feature mapping module, a voxelization extraction module, a risk field generation module and a visualization rendering module; used for acquiring spatial geometric point cloud data, two-dimensional surface texture images and tunnel theoretical design contour models of a target region of a tunnel to be measured, extracting and mapping surface semantic features, generating a spatial boundary grid with semantic features, extracting the spatial boundary of the over-under excavation body and performing voxelization processing, calculating voxelized geometric features, combining a physical information neural network model to generate physical response risk field data, and outputting a multi-channel three-dimensional visualization image; the present application realizes integrated processing of three-dimensional expression, risk assessment and visualization output of over-under excavation boundaries, and reduces misjudgments caused by simply relying on volume deviation.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring and three-dimensional visualization technology, specifically a three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation. Background Technology

[0002] Quality control of tunnel excavation is a critical aspect of underground engineering construction. To ensure that the excavation profile meets design requirements and to reduce the risk of surrounding rock disturbance, it is usually necessary to identify and display the over-excavation and under-excavation results at the tunnel face or excavated areas. In existing technologies, spatial point cloud data, surface image data, or design cross-section models of the tunnel target area are typically acquired. By comparing the measured profile with the theoretical design profile, over-excavation and under-excavation areas are determined, thus providing a basis for construction adjustments, backfilling, or support adjustments.

[0003] However, when analyzing over- or under-excavation of tunnels, existing methods mostly focus on calculating geometric deviations, cross-sectional area differences, or volume differences. They typically use a single deviation index or statistical quantity to characterize excavation quality. While this approach can reflect whether the excavation depth exceeds the design range, it is difficult to simultaneously consider the instability risks corresponding to surface crack textures, structural surface distribution, and local boundary abrupt changes. Especially in areas with fractured surrounding rock, developed fissures, or locally sharp under-excavation, judging solely based on the magnitude of geometric deviations can easily lead to insufficient risk identification or distorted treatment priorities. This results in a focus on geometric representation rather than engineering risk assessment capabilities, thus affecting rapid and accurate comprehensive decision-making on-site. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a three-dimensional image reconstruction system for the spatial boundary of tunnel over- or under-excavation. Specifically, the technical solution of this invention includes:

[0005] The multimodal data acquisition module is used to acquire spatial geometric point cloud data and two-dimensional surface texture images of the target area of ​​the tunnel under test, and to acquire the theoretical design contour model of the tunnel in the target area.

[0006] The feature mapping module is used to extract surface semantic features from two-dimensional surface texture images, obtain the spatial coordinate registration relationship between two-dimensional surface texture images and spatial geometric point cloud data, and map the surface semantic features to spatial geometric point cloud data based on the spatial coordinate registration relationship to generate a spatial boundary mesh with semantic features.

[0007] The voxelization extraction module is used to perform spatial Boolean operations on the spatial boundary mesh with semantic features and the tunnel theoretical design contour model to extract the spatial boundary of the over-excavation and under-excavation body, perform voxelization processing on the spatial boundary of the over-excavation and under-excavation body, extract multiple discrete voxel blocks, and calculate voxelized geometric features based on multiple discrete voxel blocks.

[0008] The risk field generation module is used to take surface semantic features and voxelized geometric features as input to a preset physical information neural network model, and calculate and generate physical response risk field data based on the physical mechanical parameters output by the physical information neural network model.

[0009] The visualization rendering module is used to generate multi-channel 3D visualization images based on physical response risk field data and spatial geometric point cloud data, and adjusts rendering parameters according to the physical response risk field data to output the multi-channel 3D visualization images.

[0010] Preferably, the feature mapping module includes a semantic extraction unit, a coordinate registration unit, and a mesh generation unit;

[0011] The semantic extraction unit is used to extract features from two-dimensional surface texture images based on computer vision algorithms to obtain surface semantic features, which include texture orientation data and crack distribution network.

[0012] The coordinate registration unit is used to register the two-dimensional pixel coordinates of the two-dimensional surface texture image with the three-dimensional spatial coordinates of the spatial geometric point cloud data to obtain the spatial coordinate registration relationship.

[0013] Mesh generation units are used to assign surface semantic features to spatial geometric point cloud data based on spatial coordinate registration relationships, in order to construct spatial boundary meshes with semantic features.

[0014] Preferably, the voxelization extraction module includes a spatial discretization unit, a curvature calculation unit, a gradient calculation unit, and a feature aggregation unit;

[0015] Spatial discrete elements are used to divide the spatial boundary of the under-excavated volume into multiple discrete voxel blocks according to a preset resolution.

[0016] The curvature calculation unit is used to calculate the local Gaussian curvature for each discrete voxel; the gradient calculation unit is used to calculate the gradient in the normal space for each discrete voxel.

[0017] The feature aggregation unit is used to aggregate local Gaussian curvature and normal vector space gradient to generate voxelized geometric features.

[0018] Preferably, the risk field generation module includes a data splicing unit, a model inference unit, and a risk assessment unit;

[0019] The data stitching unit is used to stitch surface semantic features and voxelized geometric features into a multidimensional feature vector;

[0020] The model inference unit is used to input multi-dimensional feature vectors into the physical information neural network model for forward inference, and output the local stress concentration coefficient and the probability of structural instability as physical and mechanical parameters.

[0021] The risk assessment unit is used to calculate a comprehensive risk index by weighting and summing the normalized local stress concentration coefficient and the structural instability probability based on preset weight coefficients. If the comprehensive risk index is greater than or equal to a preset risk threshold, the corresponding area is marked as high-risk data, and if the comprehensive risk index is less than the preset risk threshold, the corresponding area is marked as low-risk data, so as to generate physical response risk field data.

[0022] Preferably, the visualization rendering module includes a geometric channel rendering unit, a physical channel rendering unit, and an image fusion unit;

[0023] The geometric channel rendering unit is used to calculate the spatial depth deviation based on spatial geometric point cloud data and tunnel theory design contour model, and to calculate the ratio of the absolute value of the spatial depth deviation to the preset deviation limit through a preset normalization function, thus mapping the spatial depth deviation to transparency channel data.

[0024] The Physical Channel Rendering Unit is used to map physical response risk field data to color texture channel data;

[0025] The image fusion unit is used to fuse transparency channel data and color texture channel data to generate a multi-channel 3D visualization image.

[0026] Preferably, the visualization rendering module further includes an adaptive adjustment unit, used to adjust the color saturation of the corresponding high-risk image region in the multi-channel 3D visualization image to a first preset saturation threshold and add pulsating texture data when the physical response risk field data is characterized as high-risk state data, and to adjust the color saturation of the corresponding low-risk image region to a second preset saturation threshold and maintain static texture data when the physical response risk field data is characterized as low-risk state data, wherein the first preset saturation threshold is greater than the second preset saturation threshold.

[0027] Preferably, the target area of ​​the tunnel to be measured is the tunnel excavation face, and the multimodal data acquisition module includes a three-dimensional scanning unit, an image acquisition unit, and a model acquisition unit;

[0028] The 3D scanning unit is used to acquire high-density 3D point cloud data of the tunnel excavation face through 3D laser scanning equipment, as spatial geometric point cloud data;

[0029] The image acquisition unit is used to acquire high-resolution images of the tunnel excavation face through high-definition camera equipment, which are used as two-dimensional surface texture images; the model acquisition unit is used to acquire the theoretical design contour model corresponding to the tunnel excavation face, which is used as the theoretical design contour model of the tunnel.

[0030] Preferably, the physical information neural network model is a deep learning model that integrates prior knowledge of rock mechanics. The loss function of the physical information neural network model includes a data-driven loss term and a physical equation constraint loss term, wherein the physical equation constraint loss term is constructed based on the rock yield criterion.

[0031] Preferably, the physical response risk field data includes a stress concentration risk index heatmap.

[0032] Preferably, the system also includes a data compression module, which is used to compress the multi-channel 3D visualization image based on a preset encoding algorithm, and transmit the compressed multi-channel 3D visualization image to the remote monitoring terminal based on a block transmission strategy, so as to display an interactive 3D digital twin on the remote monitoring terminal.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention acquires spatial geometric point cloud data, two-dimensional surface texture images, and a theoretical design contour model of the tunnel target area. It then maps surface semantic features such as texture direction data and crack distribution network in the two-dimensional surface texture image to the spatial geometric point cloud data based on spatial coordinate registration. This transforms the spatial boundary grid from a simple geometric surface to an engineering surface with geological semantic attributes. In addition, by performing spatial Boolean operations on the spatial boundary grid and the theoretical design contour model of the tunnel, and by performing voxelization processing on the spatial boundary of the over-excavation and under-excavation body, voxelized geometric features such as local Gaussian curvature and spatial gradient of normal vector can be stably extracted, thereby effectively distinguishing between sharp under-excavation and gentle over-excavation.

[0035] 2. This invention combines surface semantic features and voxelized geometric features into a physical information neural network model, which can output local stress concentration coefficient, structural instability probability and comprehensive risk index, significantly improving the risk identification ability of surrounding rock fracture, fissure development and local boundary abrupt change areas, and avoiding the distortion of treatment priority caused by relying solely on the size of geometric deviation.

[0036] 3. This invention maps spatial depth deviation to transparency channel data and physical response risk field data to color texture channel data and performs fusion rendering. It can also simultaneously display over-excavation and under-excavation geometric deviations and engineering hazards in the same 3D view. It enhances on-site judgment efficiency by using saturation enhancement and pulsating texture in high-risk areas. At the same time, it uses compression and block transmission strategies to ensure that key risk information can be stably transmitted to remote monitoring terminals for interactive 3D digital twin display. Thus, it realizes the integrated processing of 3D expression, risk assessment and visualization output of tunnel over-excavation and under-excavation boundaries. Attached Figure Description

[0037] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0038] Figure 1 This is a schematic diagram of the modules of the three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation provided in the embodiments of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0040] A 3D image reconstruction system for the spatial boundary of tunnel over- and under-excavation bodies, comprising:

[0041] The multimodal data acquisition module is used to acquire spatial geometric point cloud data and two-dimensional surface texture images of the target area of ​​the tunnel under test, and to acquire the theoretical design contour model of the tunnel in the target area.

[0042] The feature mapping module is connected to the multimodal data acquisition module. It is used to extract surface semantic features from the two-dimensional surface texture image, obtain the spatial coordinate registration relationship between the two-dimensional surface texture image and the spatial geometric point cloud data, and map the surface semantic features to the spatial geometric point cloud data based on the spatial coordinate registration relationship to generate a spatial boundary mesh with semantic features.

[0043] The voxel extraction module is connected to the feature mapping module. It is used to perform spatial Boolean operations on the spatial boundary mesh with semantic features and the tunnel theoretical design contour model to extract the spatial boundary of the over-excavation and under-excavation body. The spatial boundary of the over-excavation and under-excavation body is voxelized to extract multiple discrete voxel blocks. Based on the multiple discrete voxel blocks, the voxelized geometric features are calculated.

[0044] The risk field generation module, connected to the voxelization extraction module, is used to take surface semantic features and voxelized geometric features as input to a preset physical information neural network model, and calculate and generate physical response risk field data based on the physical mechanical parameters output by the physical information neural network model. The visualization rendering module, connected to the risk field generation module, is used to generate multi-channel 3D visualization images based on physical response risk field data and spatial geometric point cloud data, and adjust rendering parameters according to the physical response risk field data to output multi-channel 3D visualization images.

[0045] This embodiment provides a mechanism for constructing a three-dimensional image of the spatial boundary of a tunnel under-excavation or under-excavation body, such as... Figure 1As shown; specifically, the scenario is set as the daily excavation quality review process after the left tunnel face of a mountain tunnel enters the Class IV surrounding rock section; after the completion of one cycle of blasting and the initial risk mitigation, the site needs to provide a joint conclusion within 30 minutes on whether the geometric deviation needs to be treated and whether there are any high-risk areas of local instability; to this end, the system runs continuously in the order of acquisition—mapping—extraction—inference—rendering, so that the output result is not a simple over- or under-excavation volume map, but a three-dimensional image with engineering risk implications;

[0046] The multimodal data acquisition module includes a time synchronization trigger controller, which controls the 3D laser scanning equipment and the high-definition camera equipment to perform joint calibration and synchronous acquisition of the target area within the same time window, so as to ensure that the 2D surface texture image and the spatial geometric point cloud data have a unified initial spatial pose reference.

[0047] Specifically, the multimodal data acquisition module first obtains the spatial geometric point cloud data of the tunnel face, the two-dimensional surface texture image, and the theoretical design contour model. The point cloud data can be obtained by laser scanning equipment, the two-dimensional image can be obtained by a fixed focal length industrial camera, and the theoretical design contour model can be directly read from the cross-sectional parameter file or three-dimensional linear model issued by the design institute.

[0048] Specifically, the feature mapping module extracts surface semantic features from the two-dimensional image, such as joint orientation, fracture density, and weathering zone boundary texture. Then, using the registration relationship between image pixel coordinates and three-dimensional point cloud coordinates, these semantic labels are assigned to the three-dimensional surface to form a spatial boundary grid with geometric and geological semantics.

[0049] Based on this, the voxel extraction module performs spatial Boolean operations on the spatial boundary mesh and the theoretical design contour model. If the measured mesh extends beyond the outer surface of the design contour at a certain point, the corresponding area is marked as over-excavation; if the measured mesh is inside the design contour, the corresponding area is marked as under-excavation. After the obtained boundary is divided into discrete voxels, geometric features such as local Gaussian curvature and spatial gradient of normal vector are calculated in each voxel. Then, the risk field generation module feeds the surface semantic features and voxelized geometric features into the physical information neural network model, outputs physical and mechanical parameters such as local stress concentration coefficient and structural instability probability, and forms a physical response risk field accordingly.

[0050] Specifically, the visualization rendering module generates multi-channel 3D images based on risk field data and raw point cloud data. One channel represents geometric deviation, and the other channel represents the risk level, which makes it easy for the construction manager to identify the amount of geometric deviation and the degree of deviation danger in the same view at the same time.

[0051] To facilitate understanding, a specific calculation example is given; assume that only three surface sampling areas P1, P2, and P3 are taken at the working face; the deviation of the measured point cloud of area P1 from the design contour is +0.18 meters, indicating over-excavation; the deviation of P2 is -0.05 meters, indicating under-excavation; and the deviation of P3 is -0.04 meters, also indicating under-excavation.

[0052] Two-dimensional image analysis revealed that: P1 has few cracks and continuous texture; P2 has a set of cracks nearly perpendicular to the hole axis; P3 has two sets of intersecting cracks; after voxelization, the local curvature of P1 is 0.12 and the normal gradient is 0.08; the local curvature of P2 is 0.41 and the normal gradient is 0.35; the local curvature of P3 is 0.46 and the normal gradient is 0.52.

[0053] The model outputs the following values ​​after inference: P1 stress concentration factor 1.1, instability probability 0.08; P2 stress concentration factor 1.8, instability probability 0.42; P3 stress concentration factor 2.3, instability probability 0.71. Therefore, the system will not simply prioritize over-excavation over under-excavation, but will instead mark P3 and P2 as areas requiring review and treatment, while P1 will be treated as a regular filling area but not a high-risk structural area.

[0054] In some special scenarios, on-site data collection is not always ideal. If there are occlusions or gaps in the point cloud, local patching can be performed based on the continuity of the normal of neighboring points. If the missing area still exceeds the preset proportion after patching, such as exceeding 15% of the target area, then this local area will not participate in high-confidence inference and will only output a marker to be patched.

[0055] If the clarity of a 2D image is lower than the threshold due to dust or water mist, for example, if the effective texture contrast is lower than 0.2, the confidence of the image's semantic features is reduced, and the system automatically switches to a degradation mode that is geometry-driven and semantically weakly involved. If the theoretical design contour model does not match the mileage number and there is an overall misalignment, it is first recalibrated based on the control points. If the calibration residual is still greater than the set value, for example, more than 20 millimeters, the final risk map is not generated to avoid incorrect guidance for treatment.

[0056] During a night shift in this mountain tunnel, an under-excavation of approximately 5 centimeters was found at the upper right shoulder of the tunnel face. Traditionally, this might be left untreated due to the small deviation. However, in this embodiment, image analysis simultaneously identified intersecting crack textures at this location, and voxel features showed sharp curvature. The risk field results marked it as a high-risk red pulsating area, while the gentle over-excavation of approximately 18 centimeters on the left arch waist was only shown as a high-transparency, low-risk warm-colored area. Based on this, the on-duty technician arranged for localized repairs and temporary reinforcement of the upper right shoulder, without excessive rework on the over-excavated area on the left.

[0057] The purpose of this step is to transform the previously isolated geometric reconstruction results into an engineering decision image with physical interpretation capabilities, thereby achieving integrated processing of three-dimensional representation, risk assessment and visualization output of over-excavation and under-excavation boundaries, and reducing misjudgments caused by relying solely on volume deviation.

[0058] Furthermore, the feature mapping module includes: a semantic extraction unit, used to extract features from a two-dimensional surface texture image based on computer vision algorithms to obtain surface semantic features, wherein the surface semantic features include texture orientation data and crack distribution network; a coordinate registration unit, used to register the two-dimensional pixel coordinates of the two-dimensional surface texture image with the three-dimensional spatial coordinates of the spatial geometric point cloud data to obtain a spatial coordinate registration relationship; and a mesh generation unit, used to assign the surface semantic features to the spatial geometric point cloud data based on the spatial coordinate registration relationship to construct a spatial boundary mesh with semantic features.

[0059] This embodiment provides a mechanism for mapping surface semantic features to three-dimensional boundaries; specifically, in the aforementioned face verification process, if only the point cloud geometry is retained, the system can determine convexity, concavity, depth, and shallowness, but it is difficult to identify which direction the crack extends in and which point is the joint intersection.

[0060] When the surrounding rock integrity parameter is higher than the preset standard, the impact of the defect is lower than the preset error. However, in fracture zones, fault influence zones, or strongly weathered sections, the structural risks of two locations with similar geometry can be completely different. Therefore, this embodiment introduces three consecutive actions: semantic extraction, coordinate registration, and mesh assignment, to stably transfer the geological structure information in the two-dimensional image to the three-dimensional surface.

[0061] Specifically, the semantic extraction unit first performs denoising, illumination equalization, and edge enhancement on the face image; then, it can use a combination of edge detection and texture direction statistics to extract the crack segments and their directions; further, adjacent crack segments can be connected according to distance and angle to form a crack distribution network; for example, in a simplified 6×6 image, if a nearly vertical dark stripe appears between the 2nd and 5th rows, and the gray-level change on both sides of it is greater than the threshold, then the texture is marked as a candidate crack;

[0062] If another oblique dark line intersects with it, and the intersection area is continuous, then the two are recorded as a cross-crack node; the system ultimately obtains not a single pixel, but a set of surface semantic features including the main direction, secondary direction, dense area, and intersection area; the coordinate registration unit maps the two-dimensional pixel coordinates to the three-dimensional point cloud coordinates; here, the transformation between the camera coordinate system and the laser scanning coordinate system can be established first using the device extrinsic parameters, and then the rotation and translation parameters can be obtained through the calibration board or the field control target;

[0063] In a specific calculation example, assume that the pixels in the image are... , , They belong to the same fracture line and are corresponding point cloud points after registration. , , If the system can reconstruct the two-dimensional crack line into a three-dimensional crack trajectory, and if a pixel falls in an area where the point cloud density is lower than the preset density threshold and there is no directly corresponding point, then the nearby triangular facet can be used for centroid interpolation to project the semantics onto the nearest surface grid node.

[0064] After receiving the registration relationship, the mesh generation unit assigns a semantic label to each mesh vertex or facet. For example, if mesh facet M7 covers the fracture intersection area, its attribute value can be recorded as high fracture density, main orientation 60 degrees, and 1 intersection node. If facet M8 only has uniform layered texture without significant fractures, it can be recorded as continuous texture and low fracture density.

[0065] To avoid misleading assignments by a single abnormal pixel, a weighted average of the pixel labels within each grid patch can be taken. If the weights of two types of labels within the same patch are close, for example, the difference between the largest and second largest weights is less than 5%, then the patch is marked as a mixed semantic region for subsequent model calculations of uncertainty measures.

[0066] As an anomaly handling mechanism, if a local area of ​​the image is affected by device reflection, causing cracks to be mistakenly extracted as bright edges, then false cracks with a length less than a preset length threshold and not connected to adjacent pixels can be filtered out through color consistency and shape length threshold.

[0067] If there are insufficient control points during the registration process, such as fewer than 3 sets of stable corresponding points, only coarse registration results are allowed to be output, and semantic mapping is limited to the low-risk auxiliary layer and does not directly participate in high-risk judgment; if the local surface roughness of the mesh surface is greater than the preset roughness threshold, causing a patch to span multiple real geological regions, mesh subdivision is performed first; if the cell area after subdivision is still greater than the preset value, the original geometry of the region is maintained and no high-confidence semantic labels are attached.

[0068] At the upper right shoulder of the same working face, the two-dimensional image shows two sets of intersecting cracks, one set with an orientation of approximately 70 degrees and the other set with an orientation of approximately 145 degrees. After registration, this intersection relationship is projected onto several adjacent faces of the three-dimensional mesh. Thus, the subsequent risk assessment does not face an ordinary under-excavation groove, but rather a sharp under-excavation boundary located on the intersecting crack network. This allows the system to explain why an under-excavation of only 5 centimeters is still worth prioritizing.

[0069] The purpose of this mechanism is to transform geological structure information in two-dimensional texture images into three-dimensional computable attributes, thereby upgrading the boundary mesh from a blank geometric surface to an engineering surface with semantic meaning.

[0070] Furthermore, the voxelization extraction module includes: a spatial discretization unit, used to divide the super-under-dug volume spatial boundary into multiple discrete voxel blocks according to a preset resolution; a curvature calculation unit, used to calculate the local Gaussian curvature for each discrete voxel block; a gradient calculation unit, used to calculate the normal vector space gradient for each discrete voxel block; and a feature aggregation unit, used to aggregate the local Gaussian curvature and the normal vector space gradient to generate voxelized geometric features.

[0071] This embodiment provides a mechanism for transforming irregular over- or under-excavation boundaries into stable geometric features. Specifically, a semantically meaningful 3D boundary mesh has been obtained in the previous stage. However, if a high-density mesh is directly fed into the subsequent model, two problems may easily occur on-site: First, noise with feature sizes smaller than the preset size, such as gravel protrusions and wet spray adhesion, will interfere with the judgment; second, different sampling densities of different devices will lead to inconsistent model input scales. Therefore, this embodiment normalizes complex surfaces to a unified spatial unit through voxelization, curvature calculation, normal gradient solution, and feature aggregation.

[0072] Specifically, the spatial discrete unit first determines a voxel resolution, such as 5 cm cubic or 10 cm cubic. When the voxel size is greater than the first preset size threshold, it will cause the loss of sharp undercut features. When the voxel size is less than the second preset size threshold, irrelevant surface roughness will be introduced into the risk calculation. When the voxel size is greater than or equal to the second preset size threshold and less than or equal to the first preset size threshold, it will be used as the effective voxel size for subsequent geometric feature calculation. Therefore, it can be adaptively selected according to the tunnel cross-section size and design allowable deviation.

[0073] Taking a simplified deduction as an example, if a local boundary is divided into three continuous voxels V1, V2, and V3, with V1 near the gentle over-excavation zone, V2 located in the ordinary under-excavation zone, and V3 located in the sharp-corner under-excavation zone, then each voxel will collect its internal point cloud set and corresponding mesh fragment; the curvature calculation unit calculates the local Gaussian curvature for each voxel.

[0074] Specifically, a local quadratic surface can be fitted within the voxel first, and then the Gaussian curvature can be obtained by multiplying the principal curvatures. If the two principal curvatures of V1 are approximately 0.10 and 0.08, then the local Gaussian curvature is approximately 0.008. If the two principal curvatures of V3 are approximately 0.90 and 0.70, then the local Gaussian curvature is approximately 0.63, indicating that there is a significant geometric abrupt change at this point. The gradient calculation unit then calculates the gradient of the normal vector space based on the normal changes inside the voxel and adjacent voxels.

[0075] Specifically, the system first extracts the average normal vector of the local fitted surface within the current voxel and obtains the corresponding average normal vectors of its directly adjacent surrounding voxels in space; it calculates the angle between the current voxel's normal vector and the normal vectors of each adjacent voxel, divides it by the spatial distance between the center points of adjacent voxels, obtains the rate of change of the normal vector in different spatial directions, and selects the maximum rate of change of the normal vector or the root mean square in each direction as the spatial gradient of the normal vector of the voxel; if the normal vector of V2 changes gently with that of the surrounding voxels, the gradient value can be 0.12; if the angle between V3 and its adjacent voxels changes sharply, the gradient value can reach 0.48; the feature aggregation unit combines the above two geometric quantities into voxelized geometric features;

[0076] The aggregation method can be either splicing or weighted combination; for example, voxelized geometric features can use values ​​that include Gaussian curvature normalization. and normalized value of normal gradient Two-dimensional vectors To represent, the aforementioned voxel V1 can be denoted as V2 can be denoted as V3 can be denoted as If more robust inputs are needed, neighborhood mean features can be added; for example, the average curvature of the current voxel and its 26 neighbors can be used as the third dimension to prevent a single outlier from causing abnormally high values; in this way, the input features received by the subsequent model are not scattered and noisy surface fragments, but a set of comparable, computable and interpretable local geometric descriptions.

[0077] As an anomaly handling mechanism, if the number of points in a voxel is less than the preset point threshold, for example, less than 10 valid points, the local fitting may be unstable. In this case, the point set can be supplemented by expanding the radius of the neighborhood. If it is still insufficient after expansion, the voxel is marked as a low confidence voxel, and only its coarse features consistent with the neighborhood are retained. It is not used alone for high-risk determination.

[0078] If a voxel crosses the boundary between over-drilling and under-drilling, causing feature aliasing, it can be further subdivided into half-size voxels. If the computational cost is too high after subdivision, regions with absolute geometric deviations exceeding a threshold can be subdivided first, while stable regions close to the design contour can be kept at their original scale. If the local Gaussian curvature is greater than the first preset multiple of the average Gaussian curvature of the surrounding neighborhood, for example, more than 3 times the neighborhood average, it can be regarded as pop-point noise and corrected by median filtering.

[0079] In the 5 cm under-excavation area at the upper right shoulder, it appears as only a slight geometric depression in visual observation. However, after voxelization, it was found that the Gaussian curvature and normal gradient of the core voxel in this area are significantly higher than those of the surrounding area. In contrast, although the 18 cm over-excavation area on the left has a larger deviation value, the voxel features are smooth and continuous. In this way, the system has clearly distinguished the sharp gap and the smooth outward expansion at the geometric level, providing an input that is closer to the engineering essence for subsequent risk field calculations.

[0080] The purpose of this mechanism is to transform complex boundaries into stable features suitable for model inference by unifying spatial discrete scales and local geometric description methods, thereby achieving accurate capture of high-risk geometric shapes.

[0081] Furthermore, the risk field generation module includes: a data splicing unit, used to splice surface semantic features and voxelized geometric features into a multidimensional feature vector;

[0082] The model inference unit is used to input multi-dimensional feature vectors into the physical information neural network model for forward inference, and output the local stress concentration coefficient and structural instability probability as physical and mechanical parameters.

[0083] The risk assessment unit is used to calculate the comprehensive risk index by weighting and summing the normalized local stress concentration coefficient and the structural instability probability based on preset weight coefficients. If the comprehensive risk index is greater than or equal to the preset risk threshold, the corresponding area is marked as high-risk data, and if the comprehensive risk index is less than the preset risk threshold, the corresponding area is marked as low-risk data, so as to generate physical response risk field data.

[0084] This embodiment provides a mechanism for generating a physical response risk field from geometric and semantic features. Specifically, if the score is directly based on the curvature and fracture distribution output from the previous stage, the scoring criteria are prone to drift when the surrounding rock grade, burial depth, or support parameters change, resulting in poor applicability across work sites. To solve this problem, this embodiment further introduces a physical information neural network model to map local geometric and semantic features to parameters with physical meaning, and then generates a comprehensive risk index.

[0085] Specifically, the data stitching unit first organizes the multiple features corresponding to a single voxel or voxel cluster into a multidimensional vector. For example, if the semantic features extracted from a voxel region include the main fracture strike angle of 60 degrees, fracture density of 0.7, and whether it intersects (1), and the geometric features include Gaussian curvature of 0.88 and normal gradient of 0.79, then a vector can be formed. ;

[0086] To eliminate dimensional differences, angles, densities, curvatures, etc., can be normalized separately. For example, an angle of 60 degrees can be mapped to 0.33, and a curvature of 0.88 can be kept in the range of 0 to 1. When inferring multiple adjacent voxels, the input can be voxel by voxel, or the features of a 3×3×3 neighborhood can be concatenated into a higher-dimensional vector to preserve the spatial context. After receiving the vector, the model inference unit performs forward calculation and outputs the local stress concentration coefficient and the probability of structural instability.

[0087] In a specific calculation example, the input vector of voxel A is The model outputs a stress concentration factor of 2.4 and an instability probability of 0.74; the input vector for voxel B is... The model outputs a stress concentration factor of 1.1 and an instability probability of 0.06; the risk assessment unit then performs a weighted summation based on preset weight coefficients; the mechanism for determining the preset weight coefficients here can be established by combining the analytic hierarchy process with a matrix of historical engineering experience.

[0088] Specifically, the system pre-analyzes historical data on face collapses and spalling under similar geological conditions to determine the relative contributions of local stress concentration coefficients and structural instability probability to the final disaster event. For example, in surrounding rock sections where the burial depth is greater than a preset burial depth threshold and the in-situ stress is greater than a preset stress threshold, the weight of stress concentration-related terms will be dynamically increased by the system. In shallowly buried or extremely jointed fractured sections, the weight of the gravity-controlled instability probability of the structure will be assigned a higher proportion.

[0089] The specific logic is as follows: A judgment matrix is ​​constructed that includes factors such as burial depth, in-situ stress, and surrounding rock integrity. The eigenvectors of each factor are calculated using the analytic hierarchy process (AHP) and a consistency check is performed. The eigenvectors that pass the check are then normalized and used as the weighting coefficients. and This is to ensure the objectivity and feasibility of the weighted calculation;

[0090] This allows the calculation of the comprehensive risk index to better align with the main disaster-causing mechanisms under different operating conditions. The system's formula for calculating the comprehensive risk index is as follows:

[0091]

[0092] in, This represents the overall risk index. and These represent the preset stress concentration term weights and instability probability weights, respectively; This represents the normalized value of the stress concentration factor. The method for obtaining the stress concentration factor is as follows: the upper and lower limits of the stress concentration factor distribution are preset, and the local stress concentration factor is linearly mapped to the interval between 0 and 1 using the maximum and minimum value normalization algorithm to eliminate the difference in dimensions. Represents the probability of structural instability; where, due to and Since all parameters are dimensionless, the comprehensive risk index obtained by weighted summation is... It is also a dimensionless parameter;

[0093] In the specific simulation, if the comprehensive risk index R is calculated as 0.6 × normalized stress concentration coefficient + 0.4 × instability probability, then voxel A will have R = 0.82 and voxel B will have R = 0.21. If the preset risk threshold is 0.65, then voxel A will be marked as high risk and voxel B will be marked as low risk. After all voxels have been inferred, physical response risk field data covering the tunnel face boundary can be formed. The weighting method here is not limited to a fixed ratio.

[0094] In construction organizations prioritizing safety, the weight of instability probability can be increased; in scenarios prioritizing support optimization, the weight of stress concentration-related items can be increased; in addition, different thresholds can be set for different parts, such as appropriately lowering the threshold for the arch crown and arch shoulder areas, and appropriately increasing the threshold for the sidewall areas, to conform to the actual support sensitivity; if multiple adjacent high-risk voxels are connected into a single risk cluster, the cluster center coordinates, area, and recommended treatment level can be output.

[0095] In some special scenarios, if there are missing features in the input, such as the image semantics cannot be reliably extracted due to dust, the corresponding dimension can be set to a default value and a confidence level label can be attached, so that the model outputs in a degraded mode; if the confidence level is lower than the preset value, such as lower than 0.4, only a status awaiting manual review will be generated without directly giving treatment suggestions.

[0096] If the model output exceeds a reasonable physical range, such as an instability probability greater than 1 or less than 0, boundary truncation should be performed first and the anomaly recorded. If the comprehensive risk index falls right near the threshold, such as in the critical transition range between 0.62 and 0.68, the system can add a medium-risk unconfirmed label instead of forcibly dividing it into two categories to avoid oscillations in on-site treatment.

[0097] At the aforementioned upper right shoulder location, the input vector contains both cross-crack labels and high curvature and high normal gradient, resulting in a high stress concentration coefficient and instability probability output by the model. This area ultimately forms a continuous high-risk cluster. In contrast, although the geometric deviation of the gently over-excavated area on the left is obvious, the semantic features show that there are few cracks and the geometric curvature is low, resulting in a lower overall risk output by the model. The on-site engineers thus know that the treatment logic for the two areas should be different: the former focuses on timely repair and reinforcement, while the latter focuses on adjusting support parameters and conventional backfilling.

[0098] The purpose of this mechanism is to transform the original geometric anomalies into risk field data with engineering significance, thereby enabling rapid quantification of the risk of local instability.

[0099] Furthermore, the visualization rendering module includes: a geometric channel rendering unit, which is used to calculate the spatial depth deviation based on spatial geometric point cloud data and tunnel theory design contour model, and to map the spatial depth deviation into transparency channel data by calculating the ratio of the absolute value of the spatial depth deviation to the preset deviation limit through a preset normalization function.

[0100] The Physical Channel Rendering Unit is used to map physical response risk field data to color texture channel data;

[0101] The image fusion unit is used to fuse transparency channel data and color texture channel data to generate a multi-channel 3D visualization image;

[0102] An adaptive adjustment unit is used to adjust the color saturation of the corresponding high-risk image region in the multi-channel 3D visualization image to a first preset saturation threshold and add pulsating texture data when the physical response risk field data is characterized as high-risk state data; and to adjust the color saturation of the corresponding low-risk image region to a second preset saturation threshold and maintain static texture data when the physical response risk field data is characterized as low-risk state data. The first preset saturation threshold is greater than the second preset saturation threshold.

[0103] This embodiment provides a mechanism for jointly rendering geometric information and physical risk information. Specifically, in field applications, if only a single color is used to represent over-excavation and under-excavation, technicians can easily misinterpret the color intensity as the level of risk. If a separate risk map is generated, it is necessary to compare back and forth between two interfaces, which affects the efficiency of handling. Therefore, this embodiment uses the transparency channel to express geometric deviation, uses the color texture channel to express the risk status, and then uses adaptive adjustment to improve the visual priority of high-risk areas.

[0104] Specifically, the geometry channel rendering unit first calculates the spatial depth deviation of the measured point cloud relative to the theoretical contour; it assumes that the absolute value of the permissible spatial depth deviation on site is controlled within a preset deviation limit. For example, within 0.30 meters, this can be calculated using a preset relationship. The measured spatial depth deviation Convert to transparency channel data ;in, This represents a function that takes the minimum value. This represents the absolute value of the measured spatial depth deviation, and its unit is meters; This indicates the preset deviation limit, and its unit is meters;

[0105] As can be seen from the formula, and With the same length dimension, the dimensions cancel each other out when the two are divided, resulting in the calculated transparency channel data. For dimensionless parameters; in specific examples, when When the value is 0.30, the correspondence is specified as follows: Thus, when the deviation Transparency channel data for this area at a value of 0.03 meters. The corresponding value is 0.1; when the deviation... When the value is 0.18 meters, the corresponding transparency channel data for the area is... The corresponding value is 0.6; the greater the geometric deviation, the more striking the image can be, but it does not directly convey a sense of risk.

[0106] The physical channel rendering unit maps risk field data to color and texture, for example, low risk is yellow-green, medium risk is orange, and high risk is red, and can overlay stripes or pulsating textures in high-risk areas; the image fusion unit overlays the transparency channel and color texture channel onto the same three-dimensional surface.

[0107] In a specific calculation example, area C1 is a gentle over-excavation with a depth deviation of +0.18 meters, corresponding to a transparency of 0.6, a risk index of 0.22, and a yellow-green color. Area C2 is a sharp under-excavation with a depth deviation of -0.05 meters, corresponding to a transparency of 0.17, a risk index of 0.82, and a red color. Area C3 is a normal under-excavation with a depth deviation of -0.04 meters, corresponding to a transparency of 0.13, a risk index of 0.48, and an orange color. After merging, C1 appears as a relatively obvious but soft low-risk area, while C2, although its geometric deviation is less than the preset tolerance threshold, attracts attention primarily due to its red color and enhanced texture. C3 falls between the two. Technicians can simultaneously read the deviation magnitude and engineering hazard in a single view.

[0108] Furthermore, the adaptive adjustment unit can employ different rendering strategies for different risk states. For example, when a region is in a high-risk state, the saturation is adjusted to a first preset threshold of 0.9, and a pulse texture is superimposed twice per second. When a region is in a low-risk state, the saturation is maintained at a second preset threshold of 0.4, and a static texture is preserved. In this way, even if the area of ​​a high-risk region is smaller than the preset area threshold, it will not be visually masked by a low-risk overcut region with an area larger than the corresponding preset area threshold. If there are multiple adjacent high-risk clusters, the outline can be enhanced by edge tracing to quickly locate the spray paint marking position on site.

[0109] In certain special scenarios, areas where the entire face area exceeds a preset proportion are considered high-risk, and color saturation exceeding a preset threshold may cause visual crowding. In such cases, risk levels can be further stratified, with only the top 10% of the area using pulsating textures, while the remaining high-risk areas use high-saturation but static displays. If the contrast in a dark environment after the transparency and color blending is lower than a preset contrast threshold, a high-contrast night mode can be automatically switched. If the point cloud surface is too sparse, causing texture stretching, patch interpolation rendering is prioritized. If the interpolation error exceeds the threshold, a grid outline and risk label are displayed for that area instead of continuous texture mapping to avoid artificial smoothing.

[0110] In the rendering results of the same working face, the 18 cm over-excavation area on the left side has a clear outward outline due to its high transparency, but the color remains a low-saturation warm color; the 5 cm under-excavation area on the upper right shoulder has low transparency, but is highlighted by high-saturation red and pulsating texture; when the on-duty personnel check the large screen, they can deal with the truly dangerous local under-excavation first without switching between multiple report pages, and then decide whether to take conventional support parameter adjustments for the low-risk over-excavation area.

[0111] The purpose of this step is to enable three-dimensional images to maintain geometric realism and have the ability to express risks through dual-channel fusion and dynamic visual enhancement, thereby achieving more direct on-site interpretation.

[0112] Furthermore, the target area of ​​the tunnel to be tested is the tunnel excavation face. The multimodal data acquisition module includes: a three-dimensional scanning unit, used to acquire high-density three-dimensional point cloud data of the tunnel excavation face through a three-dimensional laser scanning device, as spatial geometric point cloud data; an image acquisition unit, used to acquire high-resolution images of the tunnel excavation face through a high-definition camera device, as two-dimensional surface texture images; and a model acquisition unit, used to acquire the theoretical design contour model corresponding to the tunnel excavation face, as the theoretical design contour model of the tunnel.

[0113] This embodiment provides a multimodal acquisition mechanism for tunnel excavation faces. Specifically, the above system can be applied to various target areas, but in the tunnel face scenario, the acquisition time is short, there is a lot of dust, the lighting is complex, and the flow of people and vehicles is frequent. If the acquisition process is not specially designed, any subsequent analysis will be constrained by the quality of the source data. Therefore, this embodiment limits the area to be measured to the tunnel excavation face and sets up a collaborative process of 3D scanning, image acquisition and model acquisition for this scenario.

[0114] Specifically, the 3D scanning unit can be positioned at a safe distance in front of the working face, such as 8 to 15 meters away, and uses a rotating or solid-state 3D laser scanning device to acquire a high-density point cloud. A single scan can be completed in tens of seconds to several minutes, outputting a point set containing 3D spatial coordinates and reflection intensity. The image acquisition unit can use industrial high-definition camera equipment and work in conjunction with supplementary lighting to acquire high-resolution images covering the working face.

[0115] To improve registration stability, several fixed reflective markers or natural feature control points can be set in the field of view; the model acquisition unit then retrieves the theoretical design profile model corresponding to the excavation face based on the current mileage, cross-section type, lead traverse, and design offset parameters.

[0116] In a specific computational example, the tunnel face can be simplified to an area 10 meters wide and 8 meters high; the 3D scanning unit obtains approximately 500,000 points, the image acquisition unit obtains an image with a resolution of 4000×3000, and the model acquisition unit retrieves the stepped cross-sectional outline corresponding to the current mileage marker; specifically, the system uses control points as bridges to complete the unified coordinate connection of point cloud, image, and design model; for example, if the on-site marker points... The pixel position in the image is The three-dimensional coordinates in the point cloud are The corresponding position in the design model can also be determined, thus providing a stable anchor point for subsequent registration;

[0117] Compared with conventional single acquisition methods, this setup avoids a typical drawback: while laser scanning can see surface undulations, it is difficult to identify fine cracks; while camera images can see textures, they cannot accurately calculate spatial deviations; and design models alone cannot reflect the actual shape after excavation. Only when the three types of data are combined under the same coordinate reference can they support subsequent geometric and risk calculations.

[0118] As an anomaly handling mechanism, if a 3D scan has a shadowed area due to temporary obstruction by large machinery, a rescan and multi-station stitching can be performed after the equipment is moved; if there are still holes after the rescan, the area is marked as a blind spot; if strong backlight or water film reflection occurs during image acquisition, the impact can be reduced by multiple exposure shooting or using a polarizing filter; if the effective coverage of the image is lower than the set ratio, semantic extraction will not be initiated; if the current mileage information is inconsistent with the construction log when the model is acquired, the station number after measurement verification will be used as the index first; if a match still cannot be found, Boolean operations will be paused to prevent the use of the wrong cross-section model;

[0119] During the night shift construction of the mountain tunnel, after the blasting was completed and ventilation was finished, the muck loading equipment was temporarily removed from the site. The scanning device first completed the point cloud acquisition, and at the same time, the camera took pictures from multiple angles. The system automatically pulled the design section corresponding to the cycle. After all three types of data were in place, the subsequent semantic mapping and risk rendering entered the calculation process. This can ensure that the closed loop from acquisition to interpretation is completed on site within the same cycle.

[0120] The purpose of this mechanism is to acquire fusionable, traceable, and computable basic data for the high-time-sensitivity and high-interference scenario of the tunnel face, so as to achieve stable input for subsequent 3D image construction.

[0121] Furthermore, the physical information neural network model is a deep learning model that integrates prior knowledge of rock mechanics. The loss function of the physical information neural network model includes a data-driven loss term and a physical equation constraint loss term, wherein the physical equation constraint loss term is constructed based on the rock yield criterion.

[0122] This embodiment provides a model training and inference mechanism that integrates prior knowledge of rock mechanics. Specifically, in the aforementioned risk generation process, if the model only relies on historical samples for parameter optimization, although it can achieve a good fit under known working conditions, once the surrounding rock type, burial depth stress, or support conditions change, the model may output results that are inconsistent with the basic mechanical laws, such as predicting the geometric change point as a low-stress zone. To reduce this bias, this embodiment introduces a physical constraint loss term constructed from the rock mass yield criterion into the deep learning model.

[0123] Specifically, during the model training phase, two types of constraints can be used to update parameters together. One type is the data-driven loss term, which measures the difference between the model output and historical labeled values. For example, it is known that stress concentration coefficients and instability state labels of certain regions have been obtained through numerical analysis or monitoring inversion. The other type is the physical equation constraint loss term, which restricts the prediction results from violating the yield law of rock mass.

[0124] Specifically, the constraint term can be constructed based on the generalized Hawke-Brown criterion or the Mohr-Coulomb yield criterion; the system combines the local stress concentration parameters output by the model with the initial ground stress, maps them to the equivalent principal stress state at the corresponding location, and substitutes this stress state into the yield function; if the substitution result shows that the predicted stress state is completely inside the yield envelope, the physical constraint loss is zero; once the predicted stress state deviates significantly from the yield envelope, the system calculates the degree of deviation and adds it as a residual penalty term to the total loss function;

[0125] To simplify the understanding, if a training sample is in a region with high curvature and cross cracks, but the local stress state predicted by the model is far below the allowable range of the yield criterion, then the physical constraint term will increase, thereby guiding the model prediction space to a more reasonable direction in the parameter update; in this way, the model does not simply fit what kind of image corresponds to high risk, but rather, to a certain extent, characterizes what kind of geometry and structure is more likely to trigger mechanical instability.

[0126] In a specific operational example, assume there are two sample units D1 and D2 in the training set; D1 comes from a gently over-excavated area with a historical stress concentration factor of 1.1; D2 comes from an under-excavated area with intersecting cracks and sharp corners with a historical stress concentration factor of 2.5; when the model initially predicts, D1 outputs 1.3 and D2 outputs 1.7; if only the data difference is considered, the calculation deviation of D2 exceeds the preset error range; after introducing mechanical constraints, the system calculates that the geometric and crack conditions of D2 are close to the yield boundary, but the predicted value of 1.7 is still too low, so an additional penalty is added; after training iterations, the model output is gradually corrected to approximately 1.15 for D1 and approximately 2.4 for D2; the model obtained in this way is less likely to produce abnormal underestimations that violate engineering experience when facing new working faces;

[0127] In actual deployment, the inference phase still uses the aforementioned feature vector input method, but the model parameters used have been calibrated by both data and physics. If it is necessary to adapt to different strata, different prior parameter sets can be set for surrounding rock types such as granite, sandstone, and slate. If a change in the surrounding rock type is detected on site, the corresponding parameters are switched or the subdivision model is reloaded. This way, the fast inference capability of deep learning is preserved, while avoiding complete detachment from the background of rock mechanics.

[0128] As an anomaly handling mechanism, if a work site lacks sufficient labeled samples, resulting in weak data-driven terms, the weight of physical constraint terms can be increased to ensure the model first meets basic mechanical consistency. After accumulating sufficient field samples, the weight of data terms can be gradually increased. If some samples have low-quality labels, such as those from low-quality numerical analysis or manual experience estimation, their sample weights can be reduced to prevent poor labeling from dominating training. If physical constraints and data labels conflict for a long time, it indicates that there may be errors in the selection of surrounding rock parameters or abnormal data sources. In this case, the model verification process should be triggered instead of unconditionally continuing the training iteration.

[0129] When the mountain tunnel transitioned from a intact rock section to a jointed section, the pure data-driven model prediction results tended to classify some small-scale under-excavation as low risk, because similar geometric deviations in historical samples often corresponded to low treatment levels. After introducing mechanical constraints, the model calculation will take additional consideration of fracture intersection and yield sensitivity, so that the small-scale sharp under-excavation of the upper right arch shoulder is no longer underestimated, thus better meeting the actual on-site treatment needs.

[0130] The purpose of this mechanism is to apply reasonable boundaries to the model output using rock mechanics priors, thereby improving the stability and reliability of risk prediction results under different working conditions.

[0131] Furthermore, the physical response risk field data includes a heatmap of the stress concentration risk index.

[0132] This embodiment provides a mechanism for generating and expressing a stress concentration risk index heatmap. Specifically, in the aforementioned process, the risk field can include various data types, such as stress concentration coefficient, instability probability, and comprehensive risk index. In order to enable on-site personnel to make a judgment in the shortest possible time, this embodiment defines the most intuitive type of output as a stress concentration risk index heatmap and uses it as one of the core physical layers in the three-dimensional image.

[0133] Specifically, the stress concentration risk index heatmap can be mapped from stress concentration-related indicators of each voxel, mesh patch, or sampling point; the mapping method can use continuous color bands, such as a five-level gradient of blue-green-yellow-orange-red; the preset normalized calculation formula for the stress concentration risk index is:

[0134]

[0135] in, This indicates the stress concentration risk index. This represents the local stress concentration factor output by the model. and These represent the preset lower limit and preset upper limit of the local stress concentration factor, respectively; because , , All are dimensionless mechanical ratio coefficients, hence the stress concentration risk index It is also a dimensionless parameter;

[0136] In a specific calculation example, let the local stress concentration factor be... Lower limit of distribution 1.0, upper limit of distribution If the value is 3.0, then the above normalization logic can be specified as follows: Perform calculations;

[0137] When the local stress concentration factor When the risk index is 1.1, =0.05, corresponding to blue-green; when the local stress concentration factor is... When the risk index is 1.8, the risk index is... =0.40, corresponding to yellow; when the local stress concentration factor is... When the risk index is 2.6, the risk index is... =0.80, corresponding to red; if it is also necessary to highlight the critical area, the risk index can also be adjusted. A grid-like flickering border is overlaid with pixels larger than 0.75;

[0138] Heatmaps can be displayed independently or superimposed on the three-dimensional boundary surface as part of the aforementioned color texture channels. To ensure spatial continuity, discrete voxel values ​​can be interpolated so that the same risk clusters visually form continuous hot spots instead of scattered block noise. If a patch-level display is used, adjacent patches with shared boundaries can be smoothly transitioned to avoid color band breaks affecting interpretation.

[0139] During on-site inspections, technicians can click on hotspot areas to see the deviation value, crack density, risk level, and treatment recommendations for that location. This heat map is an important supplement to the simple geometric deviation map. Because on the actual working face, the location with the largest deviation is not necessarily the most dangerous location, and stress concentration hotspots are often coupled with factors such as geometric sharp corners, crack intersections, and abrupt changes in local boundaries. Through the heat map, the system can intuitively display the spatial distribution of this level of danger.

[0140] As an anomaly handling mechanism, if a region can only obtain a low-confidence prediction due to missing input features, the heatmap can use diagonal fill or gray mask to indicate that the estimated value is insufficient, preventing users from mistakenly believing it to be low-risk; if the overall index distribution range is smaller than the preset range threshold, resulting in indistinct color differences, quantile stretching can be used instead of fixed range mapping to improve visual resolution; if there are individual extreme high-value voxels that raise the global color band upper limit, the upper limit can be truncated, for example, mapping values ​​above the 95th percentile to the highest red level, preventing most medium- and high-risk areas from being compressed into narrow color segments that are difficult to identify;

[0141] In the rendering results of the working face, a compact red hot spot is formed at the upper right shoulder, the middle of the arch is an orange-yellow transition zone, and the gentle over-excavation area on the left is mainly green and yellow. The on-duty technician does not need to read the detailed numerical table first. He can quickly determine the spatial location and expansion range of the risk hot spot by simply looking at the heat map. Then, combined with the geometric transparency, he decides whether to make local repairs, add temporary support, or carry out subsequent monitoring and densification.

[0142] The purpose of this mechanism is to visually present the spatial distribution of stress concentration risks in a heat map format, thereby enabling rapid identification and location of risk information.

[0143] Furthermore, the system also includes a data compression module, which is used to compress multi-channel 3D visualization images based on a preset encoding algorithm, and transmit the compressed multi-channel 3D visualization images to a remote monitoring terminal based on a block transmission strategy, so as to display an interactive 3D digital twin on the remote monitoring terminal.

[0144] This embodiment provides a compression and segmented transmission mechanism for remote monitoring terminals. Specifically, after generating multi-channel 3D images at the tunnel face, if the complete point cloud, mesh, texture, and risk field data are directly transmitted, data transmission delays or even interruptions are likely to occur due to fluctuations in network bandwidth within the tunnel, making it impossible for the remote monitoring center to see the latest tunnel face status in a timely manner.

[0145] Especially under continuous nighttime tunneling conditions, a stable data link is required between the site and the rear technical support. Therefore, this embodiment adds a data compression module to encode and compress multi-channel 3D images, and adopts a block transmission strategy to achieve remote interactive display. Specifically, the compression process can process geometric data and texture data separately. For the geometric part, voxel downsampling, predictive coding, or mesh simplification can be performed on the point cloud to retain the cross-sectional outline and hotspot area details. Texture and risk field layers can be compressed using image encoding methods.

[0146] In a specific computational example, it is assumed that the original 3D data contains 1 million points, 1 high-resolution texture map, and 1 risk heat map, with a total data volume of approximately 120 megabytes. After processing, the point cloud density in stable areas far from hotspots is reduced to 30% of the original, while the density in high-risk areas remains above 80%. The texture map is compressed to 25% of its original volume, and the final total volume is reduced to approximately 28 megabytes.

[0147] Furthermore, the block transmission strategy can divide the entire tunnel face into several blocks according to spatial location, such as the upper dome block, left shoulder block, right shoulder block, left wall block, right wall block, and central core block; it can also be divided according to risk priority, sending high-risk hotspot blocks first, and then sending the remaining background blocks.

[0148] For example, if the available bandwidth of the network is lower than the preset bandwidth threshold, the system will prioritize transmitting the high-risk block in the upper right arch to the remote terminal within 3 seconds, while only sending a low-resolution preview of the low-risk area; after the link is restored, the remaining detailed data will be gradually supplemented; when the remote terminal opens the digital twin interface, it will first see the overall outline and hot spot area, and the rest of the background details will be loaded gradually, without affecting the remote expert's preliminary judgment.

[0149] To support interactive viewing, hierarchical index information can be included during transmission; when the user rotates or zooms in on a certain area, the remote terminal requests high-resolution data blocks for that area from the field; for example, when an expert zooms in on the upper right shoulder, the system only incrementally transmits the fine mesh and risk texture of that block, instead of repeatedly transmitting the entire face data; this can significantly reduce the bandwidth required for continuous interaction.

[0150] As an anomaly handling mechanism, if a sudden network downtime causes some data block verification failures, the terminal retains the most recent complete version and displays a timestamp and a "to be synchronized" mark for the unupdated areas; if excessive compression causes reconstruction deviations in high-risk area details, such as hotspot boundary offsets exceeding a set threshold, the system automatically reverts to a higher fidelity level to re-encode the block; if visual seams appear at block boundaries, an overlap buffer can be set between adjacent blocks, and a fusion transition can be performed on the terminal side; if the remote terminal has limited computing power, simplified geometry and critical risk layers are transmitted first, while non-core texture animations are paused.

[0151] During a nighttime cycle in the mountain tunnel, the on-site system had just completed risk rendering of the hotspot area on the upper right arch shoulder when the wireless link inside the tunnel experienced a short-term fluctuation. The data compression module first sent high-risk data blocks from the central and upper right areas to the project monitoring center. Remote experts could then see the red pulsating hotspots and their corresponding cross-section locations and immediately direct on-site temporary reinforcement via telephone. After the link was restored a few minutes later, the detailed blocks for the remaining areas were gradually transmitted, and the remote end finally formed a complete and interactive 3D digital twin display.

[0152] The purpose of this mechanism is to ensure that critical risk information reaches remote terminals in a priority and stable manner under the complex network conditions of tunnels, thereby achieving highly efficient collaborative interpretation between the field and the rear.

[0153] 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.

Claims

1. A three-dimensional image reconstruction system for the spatial boundary of tunnel over- and under-excavation, characterized in that, The system includes: The multimodal data acquisition module is used to acquire spatial geometric point cloud data and two-dimensional surface texture images of the target area of ​​the tunnel to be tested, and to acquire the theoretical design contour model of the tunnel in the target area. The feature mapping module is connected to the multimodal data acquisition module and is used to extract surface semantic features from the two-dimensional surface texture image, obtain the spatial coordinate registration relationship between the two-dimensional surface texture image and the spatial geometric point cloud data, and map the surface semantic features to the spatial geometric point cloud data based on the spatial coordinate registration relationship to generate a spatial boundary mesh with semantic features. A voxel extraction module, which is connected to the feature mapping module, is used to perform spatial Boolean operations on the spatial boundary mesh with semantic features and the tunnel theoretical design contour model to extract the over-excavation and under-excavation spatial boundary, perform voxelization processing on the over-excavation and under-excavation spatial boundary, extract multiple discrete voxel blocks, and calculate voxelized geometric features based on the multiple discrete voxel blocks. A risk field generation module, which is connected to the voxelization extraction module, is used to take the surface semantic features and the voxelized geometric features as inputs to a preset physical information neural network model, and to calculate and generate physical response risk field data based on the physical mechanical parameters output by the physical information neural network model. A visualization rendering module, which is connected to the risk field generation module, is used to generate a multi-channel three-dimensional visualization image based on the physical response risk field data and the spatial geometric point cloud data, and to adjust the rendering parameters according to the physical response risk field data in order to output the multi-channel three-dimensional visualization image. The physical information neural network model is a deep learning model that integrates prior knowledge of rock mechanics. The loss function of the physical information neural network model includes a data-driven loss term and a physical equation constraint loss term, wherein the physical equation constraint loss term is constructed based on the rock yield criterion.

2. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The feature mapping module includes: A semantic extraction unit is used to extract features from the two-dimensional surface texture image based on computer vision algorithms to obtain the surface semantic features, wherein the surface semantic features include texture orientation data and crack distribution network; A coordinate registration unit is used to register the two-dimensional pixel coordinates of the two-dimensional surface texture image with the three-dimensional spatial coordinates of the spatial geometric point cloud data to obtain the spatial coordinate registration relationship. A mesh generation unit is used to assign the surface semantic features to the spatial geometric point cloud data based on the spatial coordinate registration relationship, so as to construct the spatial boundary mesh with semantic features.

3. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation as described in claim 1, characterized in that, The voxelization extraction module includes: Spatial discrete unit, used to divide the spatial boundary of the under-excavated body into multiple discrete voxel blocks according to a preset resolution; A curvature calculation unit is used to calculate the local Gaussian curvature for each of the discrete voxels; The gradient calculation unit is used to calculate the gradient of the normal vector space for each of the discrete voxels; The feature aggregation unit is used to aggregate the local Gaussian curvature and the gradient of the normal vector space to generate the voxelized geometric features.

4. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The risk field generation module includes: The data splicing unit is used to splice the surface semantic features and the voxelized geometric features into a multidimensional feature vector. The model inference unit is used to input the multidimensional feature vector into the physical information neural network model for forward inference, and output the local stress concentration coefficient and the structural instability probability as the physical and mechanical parameters. The risk assessment unit is used to calculate a comprehensive risk index by weighting and summing the normalized local stress concentration coefficient and the structural instability probability based on preset weighting coefficients. If the comprehensive risk index is greater than or equal to a preset risk threshold, the corresponding area is marked as high-risk state data, and if the comprehensive risk index is less than the preset risk threshold, the corresponding area is marked as low-risk state data, so as to generate the physical response risk field data.

5. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The visualization rendering module includes: The geometric channel rendering unit is used to calculate the spatial depth deviation based on the spatial geometric point cloud data and the tunnel theoretical design contour model, and to calculate the ratio of the absolute value of the spatial depth deviation to the preset deviation limit through a preset normalization function, and to map the spatial depth deviation into transparency channel data. The physical channel rendering unit is used to map the physical response risk field data into color texture channel data; An image fusion unit is used to fuse the transparency channel data and the color texture channel data to generate the multi-channel three-dimensional visualization image.

6. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 5, characterized in that, The visualization rendering module also includes: An adaptive adjustment unit is configured to, when the physical response risk field data is characterized as high-risk state data, adjust the color saturation of the corresponding high-risk image region in the multi-channel three-dimensional visualization image to a first preset saturation threshold and add pulsating texture data; and when the physical response risk field data is characterized as low-risk state data, adjust the color saturation of the corresponding low-risk image region to a second preset saturation threshold and maintain static texture data, wherein the first preset saturation threshold is greater than the second preset saturation threshold.

7. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The target area of ​​the tunnel to be tested is the tunnel excavation face. The multimodal data acquisition module includes: A three-dimensional scanning unit is used to acquire high-density three-dimensional point cloud data of the tunnel excavation surface through a three-dimensional laser scanning device, so as to serve as the spatial geometric point cloud data; An image acquisition unit is used to acquire a high-resolution image of the tunnel excavation face using a high-definition camera device, as the two-dimensional surface texture image; The model acquisition unit is used to acquire the theoretical design profile model corresponding to the tunnel excavation face, so as to serve as the theoretical design profile model of the tunnel.

8. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The physical response risk field data includes a stress concentration risk index heatmap.

9. The three-dimensional image construction system for the spatial boundary of tunnel over-excavation and under-excavation bodies according to claim 1, characterized in that, The system also includes: The data compression module is used to compress the multi-channel 3D visualization image based on a preset encoding algorithm, and transmit the compressed multi-channel 3D visualization image to a remote monitoring terminal based on a block transmission strategy, so as to display an interactive 3D digital twin on the remote monitoring terminal.

Images