A feature map generation method for rock mass trace direction continuous structure expression

Abstract

The application discloses a feature map generation method for rock mass trace direction continuous structure expression. The application promotes the rock mass trace direction continuous structure from implicit geometric attribute to explicit feature map expression through direction consistency field modeling, direction continuous relationship graph construction and graph iteration propagation update, so that the voxels continuously distributed along the trace direction are enhanced in direction consistency features, the propagation of the direction discontinuous and noise areas is inhibited, the discontinuous traces are effectively connected, the weak trace response is enhanced, the noise interference is inhibited, and the recognition accuracy and robustness of the weak fissure and low saliency structure surface in the complex rock mass scene are improved.

Description

Technical Field

[0001] This invention relates to the technical field of three-dimensional point cloud data processing and geological engineering information, specifically to a method for generating feature maps for continuous structural representation of rock mass traces. Background Technology

[0002] The spatial distribution characteristics of rock mass structural surfaces (such as joints, faults, and bedding) are the core foundation for evaluating rock mass stability, conducting geological hazard risk assessments, and designing engineering support systems. With the rapid maturation of acquisition technologies such as 3D laser scanning, photogrammetry, and ground-based lidar, the volume and accuracy of rock mass point cloud data have achieved a leap forward, providing technical support for high-precision identification of rock mass traces. In rock mass trace identification tasks, the essential geometric characteristic of the trace is its significant directional continuity along the direction of the structural surfaces—along the trace's extension direction, the local principal geometric direction remains stable and consistent, while in areas where the trace terminates, intersects, or is affected by noise, the directional distribution tends to be disordered. Therefore, the ability to effectively represent this directional continuity directly determines the accuracy and robustness of trace identification. Existing point cloud feature representation methods mainly fall into three categories: The first category is based on local geometric features to select salient points. This method performs neighborhood search and principal component analysis on the point cloud, calculates multiple criteria such as the angle between normal vectors and the flatness coefficient, and then weights and selects feature points. This type of method stops at feature point extraction; after the selection is completed, the directional information is discarded, and it cannot further express the directional continuity structure of the trace. The second category is based on local neighborhood encoding and deep networks, through... The first type of method constructs local neighborhoods and concatenates and encodes features such as coordinates and density before inputting them into a deep network. This type of method is geared towards blocky target detection, where each neighborhood is encoded independently, without considering directional relationships between neighborhoods or voxels. The second type is based on voxelization and feature encoding, which divides the point cloud into a regular voxel grid. Each voxel independently encodes its coordinates and positional information such as relative center offset to generate a feature map. In this type of method, voxels are isolated from each other, and there are no directional constraints. The common challenge faced by the above methods in the rock mass trace recognition scenario is that they all belong to the isolated encoding paradigm. Whether it is isolated keypoints, isolated neighborhoods, or isolated voxels, none of them explicitly model and represent the directional continuity structure of the rock mass trace. Especially in complex surface environments, traces often exhibit discontinuous and blurred characteristics due to weathering, noise interference, uneven density, and severe occlusion. In existing feature representation methods, the directional signals of weak traces are easily submerged by noise, and discontinuous traces cannot form a continuous representation due to the lack of directional correlation mechanisms. Therefore, it is necessary to propose a feature map generation method for continuous structural representation of rock mass traces to meet the practical needs of intelligent identification of rock mass structural surfaces. Summary of the Invention

[0003] The technical problem this invention aims to solve is that existing technologies employ an isolated coding paradigm for point cloud data, failing to explicitly model and represent the inherent directional continuity structure of rock mass traces. This results in the directional signals of weak fractures being submerged by noise, and discontinuous traces failing to form a continuous representation due to the lack of directional correlation mechanisms. The purpose is to provide a feature map generation method oriented towards representing the directional continuity structure of rock mass traces, thereby solving the aforementioned technical problem.

[0004] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for generating feature maps for expressing continuous structure along rock mass traces, comprising: Acquire three-dimensional point cloud data of the rock mass, determine multiple target points from the three-dimensional point cloud data of the rock mass, obtain the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point; Based on the local principal directions of each target point, a direction consistency field is generated; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location; The directional consistency field is sampled onto a three-dimensional voxel mesh. With each voxel as a node, for each pair of adjacent voxels, the directional propagation weight between the two pairs of adjacent voxels is determined based on the angle between the local principal directions corresponding to the two voxels, the difference in the directional consistency field values, and the spatial distance. The directional propagation weight is used as the edge weight to construct a directional continuity graph. On the directional continuity graph, voxels with directional consistency field values ​​higher than a preset threshold are identified as propagation source nodes. Along the edges of the directional continuity graph, the propagation intensity is controlled by directional propagation weights, and the directional consistency characteristics of each voxel are iteratively updated in multiple rounds. Based on the iteratively updated directional consistency features, channels are combined with various geometric attributes corresponding to each voxel to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

[0005] Secondly, the present invention provides a feature map generation device for expressing continuous structure in the direction of rock mass traces, comprising: The acquisition module is used to acquire three-dimensional point cloud data of rock mass, determine multiple target points from the three-dimensional point cloud data of rock mass, acquire the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point; The first generation module is used to generate a direction consistency field based on the local principal directions of each target point; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location; The module is used to sample the orientation consistency field onto a three-dimensional voxel mesh. With each voxel as a node, for each pair of adjacent voxels, the orientation propagation weight between the pair of adjacent voxels is determined based on the angle between the local principal directions of the two voxels, the difference in orientation consistency field values, and the spatial distance. The orientation propagation weight is used as the edge weight to construct an orientation continuity graph. The iterative module is used to identify voxels with directional consistency field values ​​higher than a preset threshold as propagation source nodes on the directional continuity graph, and to perform multiple rounds of iterative updates on the directional consistency characteristics of each voxel by controlling the propagation intensity along the edges of the directional continuity graph with directional propagation weights. The second generation module is used to combine the directional consistency features with various geometric attributes corresponding to each voxel based on the iteratively updated directional consistency features to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

[0006] Thirdly, the present invention provides an electronic device, comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, the instructions being executed by the one or more processors to cause the one or more processors to implement the method described above.

[0007] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0008] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention elevates the directional continuity structure of rock mass traces from implicit geometric attributes to explicit feature map representation through directional consistency field modeling, directional continuity relationship graph construction, and graph-based iterative propagation updates. This enhances the directional consistency features of voxels continuously distributed along the trace direction, suppresses the propagation of directional discontinuities and noise regions, effectively connects discontinuous traces, enhances the response of weak traces, and suppresses noise interference, thereby improving the recognition accuracy and robustness of weak fractures and low-significance structural surfaces in complex rock mass scenarios. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is one of the flowcharts for a method of generating feature maps for expressing continuous structure in the direction of rock mass traces, as provided in the embodiments of this specification. Figure 2 This is the second flowchart of a method for generating feature maps for expressing continuous structures in the direction of rock mass traces, as provided in the embodiments of this specification. Figure 3 This is a block diagram of a feature map generation device for expressing continuous structure in the direction of rock mass traces, as provided in the embodiments of this specification. Figure 4 This is a block diagram of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0011] like Figure 1 and Figure 2 As shown, this embodiment provides a method for generating feature maps for continuous structural representation of rock mass traces. The execution subject of the method can be an electronic device with point cloud data processing capabilities and graph calculation capabilities. Specifically, it can include: a general-purpose computer or workstation with a 3D point cloud processing software platform, an industrial control computer with a ground laser scanning data post-processing module, a portable reinforcement computer with a rock mass structure surface analysis program, and an embedded processing terminal that is used in conjunction with a 3D laser scanning device and has built-in program instructions corresponding to this method.

[0012] The method may include: Step S11: Obtain three-dimensional point cloud data of the rock mass, determine multiple target points from the three-dimensional point cloud data of the rock mass, obtain the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point.

[0013] In this embodiment, the executing entity can acquire three-dimensional point cloud data of the rock mass. Specifically, the three-dimensional point cloud data of the rock mass may include the three-dimensional coordinate information of each sampling point. More specifically, in a particular implementation, the three-dimensional point cloud data of the rock mass may also include the normal vector information of each sampling point. The executing entity can determine multiple target points from the three-dimensional point cloud data of the rock mass. The target points serve as the basic data objects for subsequent local principal direction calculation and direction consistency field construction.

[0014] In this embodiment, the executing entity can first calculate the geometric saliency score of each point in the three-dimensional point cloud data of the rock mass. The geometric saliency score can be obtained by weighting curvature, normal gradient, and the standard deviation of neighborhood point density. Specifically, curvature is used to characterize the degree of curvature of the local surface. The normal gradient is used to characterize the rate of change of the normal direction of the local surface. The standard deviation of neighborhood point density is used to characterize the uniformity of the local point cloud distribution. The executing entity can determine points with geometric saliency scores higher than a preset saliency threshold as saliency target points. The executing entity then performs farthest point sampling on the three-dimensional point cloud data of the rock mass to obtain spatially uniform sampling target points. The executing entity combines the saliency target points with the sampling target points to obtain the plurality of target points.

[0015] In this embodiment, after obtaining the target point, the executing entity can acquire a set of points within its local neighborhood for each target point. Specifically, the executing entity can first search for a preset number of nearest neighbors for each target point to obtain an initial neighborhood point set. The executing entity performs principal component analysis on the initial neighborhood point set to preliminarily determine the local principal direction and the local normal direction. Then, the executing entity can reconstruct the local neighborhood point set corresponding to the target point using a larger neighborhood scale along the principal direction of the structural surface and a smaller neighborhood scale along the normal direction, thereby matching the local neighborhood point set with the spatial morphology of the rock mass structural surface. The executing entity also performs principal component analysis on the reconstructed local neighborhood point set to obtain the local principal direction corresponding to the target point. The local principal direction serves as the input direction information for the subsequent generation of the direction consistency field.

[0016] Step S12: Generate a direction consistency field based on the local principal directions of each target point; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location.

[0017] In this embodiment, the executing entity can determine a local reference region centered on any spatial location within the rock mass surface. The executing entity can determine the local principal direction corresponding to a target point falling within the local reference region. Then, a regional reference direction is determined based on each local principal direction within the local reference region. Next, the absolute value of the inner product between each local principal direction within the local reference region and the regional reference direction is calculated, and the mean of the absolute values ​​of the inner products is used as the directional consistency field value for that spatial location.

[0018] Step S13: Sample the directional consistency field to the three-dimensional voxel mesh. Using each voxel as a node, for each pair of adjacent voxels, determine the directional propagation weight between the pair of adjacent voxels based on the angle between the local principal directions of the two voxels, the difference in directional consistency field values, and the spatial distance. Use the directional propagation weight as the edge weight to construct a directional continuity graph.

[0019] In this embodiment, after generating the directional consistency field, the executing entity can sample the directional consistency field onto a three-dimensional voxel mesh. Each voxel in the three-dimensional voxel mesh corresponds to a spatial node. The executing entity can establish voxel adjacency relationships with each voxel as a node. For each pair of adjacent voxels, the directional propagation weight between the pair of adjacent voxels is determined based on the angle between the local principal directions corresponding to the two voxels, the difference in directional consistency field values, and the spatial distance. Specifically, attenuation terms can be constructed for the angle between the local principal directions, the absolute value of the difference in directional consistency field values, and the spatial distance, and the attenuation terms can be combined to obtain the directional propagation weight between the pair of adjacent voxels. Using the directional propagation weight as edge weights, a directional continuity graph is constructed. Thus, the edge weights in the directional continuity graph simultaneously characterize the degree of directional continuity, the degree of directional consistency proximity, and the degree of spatial proximity between voxels.

[0020] Step S14: On the direction continuity graph, voxels with direction consistency field values ​​higher than a preset threshold are identified as propagation source nodes. Along the edges of the direction continuity graph, the propagation intensity is controlled by the direction propagation weight, and the direction consistency characteristics of each voxel are iteratively updated in multiple rounds.

[0021] In this embodiment, after the directional continuity graph is constructed, the execution entity performs multiple rounds of iterative updates on the directional consistency features of each voxel on the directional continuity graph. The execution entity first identifies voxels with directional consistency field values ​​higher than a preset threshold as propagation source nodes. Propagation source nodes maintain their directional consistency features unchanged during iteration. For non-propagation source nodes, the execution entity obtains the current round directional consistency feature values ​​and corresponding directional propagation weights of each voxel adjacent to the non-propagation source node in each iteration. The execution entity normalizes the directional propagation weights between the non-propagation source node and each adjacent voxel to obtain the normalized propagation weights corresponding to each adjacent voxel. Using the normalized propagation weights as weighting coefficients, the execution entity performs a weighted summation of the current round directional consistency feature values ​​of each adjacent voxel to obtain a neighborhood aggregation feature value. The execution entity then fuses the neighborhood aggregation feature value with the current round directional consistency feature value of the non-propagation source node according to a preset update step size coefficient to obtain the next round directional consistency feature value of the non-propagation source node. The executing entity stops iterating after reaching the preset number of iterations, and obtains the directional consistency characteristics of each voxel after iterative updates.

[0022] Step S15: Based on the iteratively updated directional consistency features, channel combinations are performed with various geometric attributes corresponding to each voxel to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

[0023] In this embodiment, the executing entity combines the directional consistency feature with various geometric attributes corresponding to each voxel to generate a rock mass trace feature map.

[0024] Specifically, the multiple geometric properties are at least two of point density, average curvature, normal gradient, and elevation variation. The executing entity uses the directional consistency feature after iterative updates of each voxel as one feature channel, and uses the at least two geometric properties as other feature channels, forming multi-channel feature data corresponding to each voxel. The executing entity then arranges the multi-channel feature data corresponding to each voxel according to its spatial position in the three-dimensional voxel grid to generate a rock mass trace feature map.

[0025] In this embodiment, the directional consistency feature channel in the rock mass trace feature map is used to characterize the directional continuity structure of the rock mass trace. Voxels continuously distributed along the trace direction exhibit a continuous high value distribution in the directional consistency feature channel. Voxels at the trace termination or intersection exhibit a decreasing distribution in the directional consistency feature channel. Thus, the generated rock mass trace feature map can express the directional continuity structure of the rock mass trace.

[0026] In some embodiments, the step of determining multiple target points from the three-dimensional point cloud data of the rock mass includes: Step S111: Calculate the geometric significance score of each point in the three-dimensional point cloud data of the rock mass; wherein, the geometric significance score is the weighted sum of the curvature, normal gradient and neighborhood point density standard deviation of the point; wherein, curvature represents the degree of surface bending, normal gradient represents the rate of change of the surface normal direction, and neighborhood point density standard deviation represents the uniformity of the local point cloud distribution.

[0027] In this embodiment, the execution entity traverses all points in the three-dimensional point cloud data of the rock mass and searches for a preset number of spatial nearest neighbors centered on each point, forming a neighborhood point set for that point. Based on the three-dimensional coordinates and normal vector information of each point in the neighborhood point set, the execution entity calculates the curvature, normal gradient, and standard deviation of the neighborhood point density for that point.

[0028] Specifically, curvature is used to characterize the degree of bending of the local surface at that point. More specifically, during calculation, the execution entity fits a local quadratic surface based on the set of neighborhood points of that point, and extracts the principal curvature of that point as the curvature value from the fitted surface. The larger the curvature value, the more severe the surface bending at that point, which may correspond to the edge of a structural surface or the tip of a crack.

[0029] Specifically, the normal gradient is used to characterize the rate of change of the normal direction of the local surface at that point. More specifically, the executing entity obtains the normal vectors of the point and its nearest neighbors, calculates the magnitude of the change in the direction of the normal vector between adjacent points in space, and uses the statistical value of the magnitude of the change as the normal gradient. The larger the normal gradient, the more significant the change in the normal direction near that point.

[0030] Specifically, the standard deviation of neighborhood point density is used to characterize the uniformity of point cloud distribution within a local neighborhood of a point. The executing entity divides the neighborhood of the point into multiple sub-regions, counts the number of points in each sub-region, calculates the deviation between the number of points in each sub-region and the average number of points in the sub-region, and uses the statistical value of the deviation as the standard deviation of neighborhood point density. A larger standard deviation indicates a more uneven distribution of point cloud within the neighborhood of that point. This indicator can effectively capture the trace signals implied by uneven point cloud distribution in complex surface environments with weathering, fracturing, and severe shading.

[0031] In this embodiment, after obtaining the three statistical quantities mentioned above, the executing entity performs a weighted summation of curvature, normal gradient, and neighborhood point density standard deviation according to preset weights to obtain the geometric significance score corresponding to each point. All three preset weight coefficients for the weighted summation are positive and their sum is 1. The specific values ​​of the three weight coefficients are determined based on the rock mass type, point cloud quality, and target structural surface characteristics. In a specific implementation, the weight coefficient for curvature is 0.4, the weight coefficient for normal gradient is 0.35, and the weight coefficient for neighborhood point density standard deviation is 0.25. Therefore, each point in the three-dimensional point cloud data of the rock mass corresponds to a geometric significance score; a higher score indicates a greater probability that the point is located in a potential trace area such as the edge of a structural surface, the tip of a fracture, or a lithological boundary.

[0032] Step S112: Determine the points with geometric significance scores higher than the preset significance threshold as significance target points.

[0033] In this embodiment, the executing entity can compare the geometric saliency scores of each point obtained in step S111 with a preset saliency threshold point by point. When the geometric saliency score of a point is higher than the preset saliency threshold, the executing entity determines that point as a saliency target point. When the geometric saliency score of a point is not higher than the preset saliency threshold, the executing entity does not include that point in the set of saliency target points.

[0034] In this embodiment, a preset saliency threshold is used to distinguish between geometrically salient regions and geometrically flat regions. Its value can be set according to the rock mass type, point cloud sampling density, and noise level. In a specific implementation, the executing entity sorts all points by their geometric saliency scores from highest to lowest, and takes the score corresponding to the 90th percentile after sorting as the preset saliency threshold. This ensures that salient target points with geometric saliency scores higher than the threshold account for approximately 10% of the total number of points. Through this step, the executing entity filters out points that are more prominent in local geometry from the rock mass 3D point cloud data, forming a set of salient target points.

[0035] Step S113: Perform farthest point sampling on the three-dimensional point cloud data of the rock mass to obtain sampling target points that are uniformly distributed in space.

[0036] In this embodiment, the executing entity performs farthest point sampling in the 3D point cloud data of the rock mass. The executing entity first randomly selects a point from the 3D point cloud data of the rock mass as the first sampling target point and adds it to the sampling target point set. This selection is then iterative. Specifically, in each iteration, the executing entity calculates the Euclidean distance from each remaining point to the nearest point in the sampling target point set, and selects the point with the largest distance value as the new sampling target point and adds it to the set. The executing entity repeats the above sampling process until the number of points in the sampling target point set reaches the preset sampling quantity, resulting in a spatially uniformly distributed set of sampling target points. The sampling target points generated by farthest point sampling do not depend on geometric saliency and are covered in both the flat background area and the geometrically salient area of ​​the rock mass surface, thus compensating for the problem of insufficient background coverage caused by salient target points being concentrated only in the potential area of ​​the trace.

[0037] Step S114: Combine the salient target points and the sampling target points according to a preset ratio to obtain multiple target points.

[0038] In this embodiment, the executing entity reads the saliency target point set obtained in step S112 and the sampling target point set obtained in step S113, and merges the points in the two sets into the final target point set. The ratio of the number of points in the two sets has been pre-controlled in steps S112 and S113 by setting a saliency threshold and a sampling number.

[0039] Understandably, the preset ratio determines the ratio of salient target points to sampled target points in the final target point set. When the proportion of salient target points is too low, the trace region coverage is insufficient, affecting the construction quality of the subsequent directional consistency field. When the proportion of salient target points is too high, the background region coverage is insufficient, resulting in large blank areas in the background region of the directional consistency field, affecting the topological integrity of the subsequent directional continuity graph.

[0040] In some implementations, obtaining the set of points in the local neighborhood of each target point includes: Step S115: Search for a preset number of nearest neighbors for each target point to obtain an initial neighborhood point set, and perform principal component analysis on the initial neighborhood point set to preliminarily determine the local principal direction and local normal direction.

[0041] In this embodiment, for each target point, the executing entity searches for a preset number of points in the three-dimensional point cloud data of the rock mass that are spatially closest to the target point, with the target point as the center, and then uses these points together with the target point to form an initial neighborhood point set.

[0042] In this embodiment, after obtaining the initial neighborhood point set, the executing entity performs principal component analysis on the initial neighborhood point set. Specifically, the executing entity first constructs a covariance matrix based on the three-dimensional coordinates of each point in the initial neighborhood point set. The covariance matrix can be a 3x3 real symmetric matrix, where each matrix element is obtained by summing the products of the deviations of the coordinate values ​​of each point in the corresponding coordinate dimension from the mean. The executing entity performs eigenvalue decomposition on the covariance matrix to obtain three eigenvalues ​​and their corresponding three eigenvectors. These three eigenvalues ​​can be arranged in descending order. The executing entity determines the eigenvector corresponding to the largest eigenvalue as the preliminary local principal direction and the eigenvector corresponding to the smallest eigenvalue as the preliminary local normal direction. The preliminary local principal direction indicates the preliminary extension direction of the rock mass structure surface trace at the target point, and the preliminary local normal direction is approximately perpendicular to the rock mass surface at the target point.

[0043] Step S116: Based on the initially determined local principal direction and local normal direction, a first neighborhood scale is adopted along the principal direction of the structural surface, and a second neighborhood scale is adopted along the normal direction, wherein the neighborhood range corresponding to the first neighborhood scale is larger than the neighborhood range corresponding to the second neighborhood scale.

[0044] In this embodiment, the executing entity can use the preliminary local principal direction and preliminary local normal direction obtained in step S115 as directional references to determine two differentiated neighborhood scales for the target point.

[0045] Specifically, along the initial local principal direction, the implementing entity adopts the first neighborhood scale. The neighborhood range corresponding to the first neighborhood scale extends along the strike direction of the trace, and its value is relatively large to fully track the continuity of the trace along the strike direction. The trace of the rock mass structural surface can extend several meters or even tens of meters along the strike direction. If the neighborhood range along this direction is too small, the number of points in the neighborhood will be insufficient to stably estimate the principal direction.

[0046] Specifically, along the initial local normal direction, the implementing entity adopts a second neighborhood scale. The neighborhood range corresponding to the second neighborhood scale narrows along the normal direction perpendicular to the rock mass surface, and its value is smaller than that of the first neighborhood scale. The rock mass structural plane is only a few centimeters to tens of centimeters wide along the normal direction.

[0047] In this embodiment, the differentiated setting of the first neighborhood scale and the second neighborhood scale changes the neighborhood shape of the target point from a conventional spherical isotropic shape to an anisotropic shape that is elongated along the trace and narrowed along the normal, which matches the geometric characteristics of the rock mass structure surface that is long along the strike and narrow in the normal.

[0048] Step S117: Based on the first neighborhood scale and the second neighborhood scale, reconstruct the local neighborhood point set corresponding to each target point to obtain the anisotropic neighborhood point set.

[0049] In this embodiment, the executing entity takes the target point as the center, determines the cutoff range of the neighborhood in the main direction according to the first neighborhood scale along the initial local main direction, determines the cutoff range of the neighborhood in the normal direction according to the second neighborhood scale along the initial local normal direction, and uses the corresponding scale to determine the cutoff range in the direction that is orthogonal to both the initial local main direction and the initial local normal direction.

[0050] The executing entity searches for point cloud data points within their respective truncation ranges in three directions, and then uses the searched points to form a reconstructed local neighborhood point set. Due to the different neighborhood scales in the main direction and the normal direction, the reconstructed neighborhood point set exhibits anisotropy in spatial morphology, meaning that the range of points taken along the trace direction is large, while the range of points taken along the normal direction is small.

[0051] In this embodiment, the anisotropic neighborhood point set obtained in this step, compared with the isotropic initial neighborhood point set in step S115, can accurately capture the crack boundary while tracing the continuity of the trajectory, and avoid mixing of directional information across structural surfaces.

[0052] In some implementations, the step of generating a direction-consistent field based on the local principal directions of each target point includes: Step S121: For any spatial location in the orientation consistency field, determine a local reference region centered on that location, and extract the local principal directions of each target point within the local reference region from the obtained local principal directions of each target point.

[0053] In this embodiment, the executing entity selects a spatial location to be calculated within the space of the rock mass surface covered by the directional consistency location. The executing entity can use this spatial location as the center and define a three-dimensional spatial range with a preset radius. This three-dimensional spatial range is the local reference area of ​​the spatial location.

[0054] In this embodiment, after determining the local reference region, the execution entity traverses each obtained target point and determines whether the spatial coordinates of each target point fall within the local reference region. The execution entity forms a reference point set from the target points whose spatial coordinates fall within the local reference region, and extracts the local principal direction corresponding to each target point in the reference point set from the stored local principal direction data of each target point.

[0055] Step S122: Determine the regional reference direction of the local reference area based on the local principal direction of each target point within the local reference area.

[0056] In this embodiment, the executing entity can obtain the local principal directions of each target point in the reference point set extracted in step S121. Each local principal direction is a three-dimensional unit vector. The executing entity performs statistical summation on all local principal direction vectors in the reference point set to obtain a reference direction representing the overall direction of the local reference region, which is denoted as the region reference direction.

[0057] In one specific implementation, the executing entity calculates the vector sum of all local principal direction vectors in the reference point set, divides the vector sum by its magnitude to obtain a unit vector, which is the regional reference direction. In this method, the regional reference direction represents the average trend of the local principal directions of each target point within the local region.

[0058] Step S123: Calculate the absolute value of the inner product of the local principal direction and the regional reference direction of each target point within the local reference region, and use the average of the absolute values ​​of the inner products as the orientation consistency value of the spatial location.

[0059] In this embodiment, the executing entity can extract the local principal direction vector of each target point in the reference point set one by one, calculate the inner product of each local principal direction vector and the regional reference direction vector obtained in step S122, and take the absolute value of the inner product result. The value of the absolute value of the inner product is between 0 and 1. Specifically, an absolute value of 1 indicates that the local principal direction of the target point is completely aligned with the regional reference direction, and an absolute value of 0 indicates that the two are orthogonal.

[0060] The executing entity can sum the absolute values ​​of the inner products corresponding to all target points in the reference point set, divide the sum by the number of target points in the reference point set, and the average value is the directional consistency value of the current spatial position.

[0061] Specifically, the geometric alignment of the local principal directions of each target point in the reference point set with the regional reference direction is averaged. The closer the average value is to 1, the more consistent the local principal directions of the target points in that local area are, corresponding to areas with continuous rock mass structure surface trace directions. The closer the average value is to 0, the more disordered the distribution of local principal directions is, corresponding to areas where rock mass traces terminate, traces intersect, or there is noise interference.

[0062] Step S124: Traverse the spatial range covered by the directional consistency field, calculate the directional consistency field value at each position, and generate the directional consistency field.

[0063] In this embodiment, the executing entity traverses each spatial location to be calculated row by row, column by column, and layer by layer within the spatial area of ​​the rock mass surface covered by the directional consistency field, according to a preset spatial sampling step size. For each spatial location, the executing entity repeats steps S121 to S123 to obtain the directional consistency value of that spatial location. After the directional consistency values ​​of all spatial locations have been calculated, the executing entity outputs each spatial location and its corresponding directional consistency value as a whole data volume, completing the generation of the directional consistency field.

[0064] In this embodiment, the directional consistency values ​​at various spatial locations in the directional consistency field implicitly encode the directional continuity structure of the rock mass traces through their overall spatial distribution and gradient variations. Specifically, along the trace's strike, the directional consistency values ​​exhibit a continuous distribution of high values. At the trace's termination or intersection, the directional consistency values ​​show a precipitous drop. In noisy or unstructured areas, the directional consistency values ​​exhibit a disordered distribution of low values.

[0065] In some implementations, the step of determining the directional propagation weights between adjacent voxels based on the angle between the local principal directions corresponding to the two voxels, the difference in directional consistency field values, and the spatial distance includes: Step S131: Obtain the local principal directions corresponding to two adjacent voxels, calculate the angle between the two local principal directions, and use the angle as the attenuation variable of the first exponential attenuation factor to construct the first attenuation factor.

[0066] In this embodiment, the executing entity, within the constructed three-dimensional voxel mesh, obtains the local principal direction vector corresponding to the first voxel and the local principal direction vector corresponding to the second voxel for any pair of adjacent voxels in the direction continuity diagram. Both local principal direction vectors are three-dimensional unit vectors, each indicating the extension direction of the rock mass structure surface trace at the location of the voxel.

[0067] In this embodiment, the executing entity can calculate the angle between two local principal direction vectors. Specifically, the angle can be calculated as follows: the executing entity calculates the inner product of the two unit vectors, takes the inverse cosine of the inner product result, and obtains the angle value expressed in radians or degrees. The smaller the angle, the more consistent the orientation of the structural surfaces at the two voxels, and the higher the probability that the traces are continuous along this direction. The larger the angle, the more the trace orientation at the two voxels deviates, and the traces may terminate or turn at this point.

[0068] In this embodiment, the executing entity can use the included angle as the decay variable of the first exponential decay factor to construct the first decay factor. Specifically, the included angle can be multiplied by a first positive adjustment parameter, the product can be negatively taken, and the exponent of the negative value can be calculated with the natural constant e as the base. The result is the first decay factor.

[0069] The first attenuation factor takes a maximum value of 1 when the included angle is 0, and decays exponentially as the included angle increases, approaching 0. The first positive adjustment parameter controls the sensitivity of this attenuation process. The larger the value of the first positive adjustment parameter, the more sensitive the first attenuation factor is to the increase of the included angle, and the more rapid the attenuation is at the same included angle.

[0070] Step S132: Obtain the directional uniformity field value of each of the two adjacent voxels, calculate the absolute value of the difference between the two directional uniformity field values, and use the absolute value as the decay variable of the second exponential decay factor to construct the second decay factor.

[0071] In this embodiment, the executing entity can read the orientation consistency field values ​​of the first voxel and the second voxel from the generated orientation consistency field. The orientation consistency field values ​​range from 0 to 1, representing the degree of consistency of the local principal orientation within the neighborhood of each voxel's location.

[0072] In this embodiment, the executing entity can calculate the absolute value of the difference between the two directional consistency field values. This absolute value reflects the difference in the degree of directional consistency between the two voxels. Specifically, an absolute value of 0 indicates that the two voxels have exactly the same degree of directional consistency, are both in trace regions or both in background regions, and the reliability of directional information propagation from one voxel to the other is high. An absolute value close to 1 indicates that the two voxels have a significant difference in directional consistency, possibly one voxel is located in a trace region while the other voxel is located in a noise region or at the end of a trace, and the reliability of directional propagation is low.

[0073] In this embodiment, the executing entity can use the absolute value as the decay variable of the second exponential decay factor to construct the second decay factor. Specifically, the absolute value can be multiplied by the second positive value adjustment parameter, the product can be negatively taken, and the exponent of the negative value can be calculated with the natural constant e as the base. The result is the second decay factor. The second decay factor takes a maximum value of 1 when the absolute value is zero, and decays exponentially as the absolute value increases, approaching 0. The larger the value of the second positive value adjustment parameter, the more sensitive the second decay factor is to consistency differences.

[0074] Step S133: Calculate the spatial distance between two adjacent voxels, and use the spatial distance as the decay variable of the third exponential decay factor to construct the third decay factor.

[0075] In this embodiment, the executing entity can obtain the spatial center coordinates of the first voxel and the spatial center coordinates of the second voxel. The spatial center coordinates of the voxel are the position coordinates of the voxel's geometric center in three-dimensional space.

[0076] In this embodiment, the executing entity can calculate the Euclidean distance between the coordinates of two spatial centers, which is the spatial distance between two adjacent voxels. The spatial distance reflects the proximity of two voxels in physical space. Specifically, the smaller the spatial distance, the higher the reliability of inferring directional continuity from a geometric proximity perspective. The larger the spatial distance, the lower the reliability of directional continuity inference even if the directional angle is small and the consistency is close.

[0077] The executing entity can use this spatial distance as the decay variable of the third exponential decay factor to construct the third decay factor.

[0078] In one specific implementation, the spatial distance can be multiplied by a third positive adjustment parameter, the product is negative, and the exponent of this negative value is calculated with the natural constant e as the base. The result is the third attenuation factor. The third attenuation factor has a maximum value of 1 when the spatial distance is zero, and decays exponentially with increasing spatial distance, approaching 0. The third positive adjustment parameter controls the sensitivity of this attenuation process; the larger the value of the third positive adjustment parameter, the more sensitive the third attenuation factor is to the increase in spatial distance.

[0079] Step S134: Multiply the first attenuation factor, the second attenuation factor and the third attenuation factor to obtain the directional propagation weight between the pair of adjacent voxels; wherein the attenuation sensitivity of the first attenuation factor, the second attenuation factor and the third attenuation factor are controlled by preset positive adjustment parameters.

[0080] In this embodiment, the executing entity can perform a multiplication operation on the first attenuation factor obtained in step S131, the second attenuation factor obtained in step S132, and the third attenuation factor obtained in step S133. The product is the directional propagation weight between the pair of adjacent voxels.

[0081] Understandably, since the values ​​of the three attenuation factors all range from 0 to 1, the value of their product also ranges from 0 to 1. The three attenuation factors are coupled in a product form, making the directional propagation weight highly sensitive to the three conditions of directional angle, consistency difference, and spatial distance.

[0082] The directional propagation weight approaches 1 only when all three conditions are met simultaneously: the angle between the local principal directions of the two voxels is small, their directional consistency values ​​are close, and their spatial distance is short; and all three attenuation factors are close to 1. This indicates a high degree of confidence in the directional continuity between voxels. If any one of the conditions is not met, the corresponding attenuation factor approaches 0, and the directional propagation weight is pulled towards 0 through multiplication, almost blocking propagation.

[0083] In this embodiment, the attenuation sensitivity of the first attenuation factor is controlled by a first positive value adjustment parameter, the attenuation sensitivity of the second attenuation factor is controlled by a second positive value adjustment parameter, and the attenuation sensitivity of the third attenuation factor is controlled by a third positive value adjustment parameter.

[0084] Specifically, the three positive adjustment parameters are independent of each other and can be set separately according to the point cloud density, noise level and structural surface morphology characteristics in the actual rock mass scene.

[0085] In one specific implementation, when it is necessary to emphasize the constraint of the directional angle on propagation, the first positive adjustment parameter is increased. When it is necessary to emphasize the constraint of consistency differences on propagation, the second positive adjustment parameter is increased. When it is necessary to emphasize the constraint of spatial distance on propagation, the third positive adjustment parameter is increased.

[0086] The executing entity calculates the directional propagation weight for each pair of adjacent voxels in the directional continuity graph according to steps S131 to S134 above, and uses the calculation results as the edge weights of the corresponding edges in the directional continuity graph to complete the assignment of edge weights in the directional continuity graph.

[0087] In some implementations, the step of iteratively updating the orientation consistency features of each voxel includes: Step S141: In the direction continuity graph, the direction consistency characteristics of each propagation source voxel remain unchanged during the iteration process.

[0088] In this embodiment, the executing entity can mark voxels with directional consistency field values ​​higher than a preset propagation source threshold as propagation source voxels, and mark the remaining voxels as non-propagation source voxels. The propagation source voxels correspond to regions on the rock mass surface with consistent directional height, and are anchor points in the directional continuity relationship diagram of the trace direction continuity structure.

[0089] The executing entity can set iterative behaviors for each propagation source voxel and non-propagation source voxel. For propagation source voxels, the executing entity locks their directional consistency eigenvalue as the initial value, which is the directional consistency field value corresponding to the propagation source voxel in the directional consistency field. In all subsequent iterations, the directional consistency eigenvalue of the propagation source voxel does not participate in the update calculation and remains unchanged. For non-propagation source voxels, the executing entity uses their directional consistency field value in the directional consistency field as the initial value for iteration and updates it according to steps S142 to S144 in each subsequent iteration.

[0090] It is understandable that the mechanism of keeping the source voxels unchanged ensures that there is a stable source of directional information injection in the directional continuity graph, and avoids the directional consistency features of all voxels from becoming homogenized due to mutual influence during iterative propagation, thus losing the ability to distinguish the directional continuity structure.

[0091] Step S142: In each iteration, for each non-propagating source voxel, calculate the ratio of the directional propagation weight between the non-propagating source voxel and each of its neighboring voxels to the sum of the directional propagation weights between the non-propagating source voxel and all its neighboring voxels, and use this ratio as the normalized propagation weight of each neighboring voxel.

[0092] In this implementation, at the beginning of each iteration, the execution entity traverses all non-propagating source voxels in the directional continuity graph. For the currently traversed non-propagating source voxel, the execution entity obtains all its adjacent voxels in the directional continuity graph. These adjacent voxels may include propagating source voxels or other non-propagating source voxels.

[0093] In this embodiment, the executing entity acquires the calculated directional propagation weights between the non-propagation source voxel and each adjacent voxel one by one, and sums the directional propagation weights corresponding to all adjacent voxels to obtain a total weight. Then, for each adjacent voxel, the executing entity calculates the ratio of the directional propagation weight corresponding to that adjacent voxel to the total weight; this ratio is the normalized propagation weight of that adjacent voxel at the current non-propagation source voxel. The sum of the normalized propagation weights of all adjacent voxels is 1.

[0094] The normalization of propagation weights addresses the issue of uneven aggregated signal intensity caused by varying numbers of neighboring voxels among different non-propagating source voxels. Without normalization, non-propagating source voxels with more neighboring voxels receive significantly stronger total signal intensity during aggregation than those with fewer neighboring voxels, resulting in uneven propagation. After normalization, the aggregated neighboring signal received by each non-propagating source voxel is on the same order of magnitude, and the propagation of directional information is controlled solely by the relative magnitude of the directional propagation weights, unaffected by differences in neighborhood size.

[0095] Step S143: Using the normalized propagation weights of each neighboring voxel as weighting coefficients, perform a weighted summation of the directional consistency feature values ​​of each neighboring voxel in the current round to obtain the neighborhood aggregation feature value.

[0096] In this embodiment, after completing the normalized propagation weight calculation in step S142, the executing entity performs neighborhood aggregation on the current non-propagation source voxel.

[0097] Specifically, the executing entity acquires the directional consistency feature value of each neighboring voxel in the current iteration round. For propagation source voxels, the feature value in the current round is the initial value. For other non-propagation source voxels, the feature value in the current round is the value updated in the previous iteration. The executing entity multiplies the directional consistency feature value of each neighboring voxel in the current round by the normalized propagation weight corresponding to each neighboring voxel, and then sums all the product results. The summation result is the neighborhood aggregation feature value of the non-propagation source voxel.

[0098] Understandably, the neighborhood aggregation feature value integrates the directional consistency information of all neighboring voxels of the non-propagating source voxel. Among them, neighboring voxels with high directional propagation weights (corresponding to voxel pairs with small directional angles, small consistency differences, and close spatial distances) make a dominant contribution to the aggregation result, while neighboring voxels with low directional propagation weights contribute little. Therefore, the neighborhood aggregation feature value is mainly determined by neighboring voxels that are directionally continuous and have similar consistency with the non-propagating source voxel in the trace direction.

[0099] Step S144: Weight the neighborhood aggregation feature value with a preset update step size coefficient, and weight the direction consistency feature value of the non-propagating source voxel in the current round by subtracting the preset update step size coefficient from 1. The sum of the two weighted results is used as the direction consistency feature value of the non-propagating source voxel in the next round.

[0100] In this embodiment, the executing entity can fuse the neighborhood aggregation feature value obtained in step S143 and the directional consistency feature value of the non-propagation source voxel in the current round to generate the directional consistency feature value for the next round.

[0101] Specifically, the fusion process is controlled by a preset update step size coefficient. The preset update step size coefficient is a positive number greater than 0 and less than 1. The executing agent multiplies the neighborhood aggregation feature value by the preset update step size coefficient to obtain the update term. The directional consistency feature value of the non-propagating source voxel in the current round is multiplied by 1 and subtracted from the preset update step size coefficient to obtain the retention term. The executing agent adds the update term and the retention term; the sum is the directional consistency feature value of the non-propagating source voxel in the next iteration.

[0102] It should be noted that in this fusion mechanism, the retention term ensures that non-propagating source voxels retain a certain proportion of their original directional information in each iteration, preventing features from being completely assimilated by the neighborhood after multiple rounds of propagation and losing their positional specificity. The update term introduces the propagation information from neighborhood aggregation in proportion to a preset update step size coefficient, allowing directionally consistent features to gradually diffuse along a continuous graph path. The preset update step size coefficient controls the step size of each round of propagation; that is, a larger value results in a wider propagation range and faster convergence speed in a single round, but may excessively smooth out directional details. A smaller value results in finer propagation and a smoother cumulative effect over multiple rounds, but requires more iterations to fully propagate directionally consistent features across the entire field.

[0103] Step S145: Stop the iteration after a preset number of iterations to obtain the directional consistency features of each voxel after iterative updates.

[0104] In this embodiment, in each iteration, after completing steps S142 to S144 of the update process for all non-propagating source voxels in the direction continuity graph, the execution subject determines whether the current iteration number has reached the preset iteration number. If not, the execution subject uses the updated direction consistency feature value as the current feature value for the next iteration and continues to execute the next iteration. If the target has been reached, the execution subject stops iterating.

[0105] In this embodiment, the specific value of the preset number of iteration rounds can be determined comprehensively based on the propagation distance requirement of the directional consistency feature and the preset update step size coefficient. Specifically, when the preset update step size coefficient is small, the propagation range of a single round is limited, requiring a larger number of iteration rounds to cover a sufficient propagation distance. When the preset update step size coefficient is large, the propagation speed is fast, and the number of iteration rounds can be reduced accordingly.

[0106] In this embodiment, after the iteration stops, the orientation consistency eigenvalues ​​of each voxel are the final results after multiple rounds of iterative propagation and updates. For propagation source voxels, the final eigenvalue is the initial orientation consistency field value. For non-propagation source voxels, the final eigenvalue is the result after multiple rounds of retain-update fusion, which integrates the orientation information of the voxel itself and the propagation source voxels reachable along the orientation continuity graph path.

[0107] In this embodiment, after multiple iterations, the directional consistency feature is propagated and accumulated along the graph paths with large directional propagation weights and continuous directions in the directional continuity graph. Between adjacent voxels that are continuously distributed along the trace, due to the small angle between the local principal directions, the small difference in directional consistency values, and the short spatial distance, the directional propagation weight is close to 1, and the directional consistency feature is propagated with sufficient weight.

[0108] For discontinuous trace regions where point clouds are locally sparse due to weathering, occlusion, or scanning angle, although the initial directional consistency value of the voxels in these regions is low, the directional consistency from the source voxels is enhanced after multiple iterations through weighted propagation along the trace direction to adjacent voxels, thus forming a continuous directional signal in the feature map. For noisy regions, the noise signal is suppressed during the iteration process due to the large difference in directional consistency with surrounding voxels and the low directional propagation weight.

[0109] In some implementations, the step of generating a rock mass trace feature map by combining the channel of the iteratively updated directional consistency features with multiple geometric attributes corresponding to each voxel includes: Step S151: Obtain at least two geometric properties corresponding to each voxel; wherein the at least two geometric properties are selected from point density, mean curvature, normal gradient and elevation change.

[0110] In this embodiment, the executing entity extracts geometric attributes for each voxel one by one. The candidate types of geometric attributes are point density, mean curvature, normal gradient, and elevation variation. The executing entity can select at least two of the above four candidate geometric attributes.

[0111] Specifically, point density reflects the density of point cloud data within the voxel, with significantly lower point density in weathered, fractured, or obscured areas. Mean curvature reflects the overall curvature of the rock mass surface within the voxel, with higher curvature values ​​at structural plane edges and fracture tips. Normal gradient reflects the degree of variation in the normal direction of the rock mass surface within the voxel, with larger normal gradients at trace intersections or lithological boundaries. Elevation variation reflects the vertical undulation of the rock mass surface within the voxel, with significant elevation variations at steep scarps and fault scarps.

[0112] In one specific implementation, the executing entity can select all four geometric attributes—point density, mean curvature, normal gradient, and elevation variation—to provide the richest local geometric information. In another specific implementation, the executing entity selects two or three of these attributes based on the geological characteristics and trace type of the target rock mass. The executing entity then assigns values ​​to each selected geometric attribute one by one according to voxels, forming a geometric attribute data set corresponding to each voxel.

[0113] Step S152: Take the directional consistency feature updated by each voxel as one channel, and take at least two geometric attributes as the remaining channels, and combine them according to the channel dimension to form the multi-channel feature vector corresponding to each voxel.

[0114] In this implementation, the executing entity can construct a multi-channel feature vector for each voxel. The multi-channel feature vector can be formed by concatenating multiple independent channels along the channel dimension.

[0115] Specifically, the executing entity can use the directional consistency features of each voxel after iterative updates in steps S141 to S145 as one channel in the multi-channel feature vector, denoted as the directional consistency feature channel. The value of the directional consistency feature channel is the final directional consistency feature value corresponding to that voxel, ranging from 0 to 1. Its value reflects the positional attribute of the voxel in the directional continuity structure, that is, high-value voxels are located in the continuous region of the trace direction, while low-value voxels are located in the trace termination, intersection, or noise region.

[0116] In this embodiment, the executing entity can use at least two geometric attributes selected in step S151 as the remaining channels in the multi-channel feature vector. Each geometric attribute corresponds to an independent channel.

[0117] The executing entity can sequentially concatenate the values ​​of each channel along the channel dimension according to a preset channel order to form a multi-channel feature vector corresponding to that voxel. The length of the multi-channel feature vector is the total number of channels. Among them, the directional consistency feature channel provides directional continuous structural information optimized by global propagation, and the geometric attribute channel provides local geometric detail information. The two types of channels complement each other and coexist in the feature vector.

[0118] Step S153: For voxels without point distribution in the voxel grid, fill each channel of the voxel with the preset default value.

[0119] In this embodiment, after generating the multi-channel feature vectors of each voxel, the executing entity can check whether there are voxels with no point distribution in the three-dimensional voxel mesh. Specifically, a voxel with no point distribution is one whose spatial range does not include any three-dimensional point cloud data points of the rock mass.

[0120] For each voxel with no point distribution, the execution entity fills each channel of the voxel's multi-channel feature vector with a preset default value. One specific value of the preset default value is 0. After filling, the voxel with no point distribution has a default value of 0 in both the orientation consistency feature channel and each geometric attribute channel, forming a numerical comparison with the voxels in the trace region with point distribution in the feature map.

[0121] Step S154: Arrange the multi-channel feature vectors of all voxels according to their spatial positions in the three-dimensional voxel grid to generate a rock mass trace feature map.

[0122] In this embodiment, the executing entity can arrange the multi-channel feature vectors of each voxel according to the three-dimensional spatial index of each voxel in the three-dimensional voxel grid. Each voxel in the three-dimensional voxel grid has a unique row index, column index, and layer index, which correspond to the position of the voxel in three orthogonal directions in three-dimensional space, respectively.

[0123] The execution entity arranges the multi-channel feature vectors of each voxel in ascending order of index value, using row index, column index, and layer index as the basis for arrangement, thereby generating a three-dimensional multi-channel tensor structure. This three-dimensional multi-channel tensor is the rock mass trace feature map.

[0124] In the rock mass trace feature map, the spatial distribution pattern of the directional consistency feature channel reflects the directional continuity structure of the rock mass trace. Voxels that are continuously distributed along the trace direction show a high continuous distribution on this channel, while the voxel values ​​at the end or intersection of the trace drop sharply on this channel.

[0125] The geometric attribute channel provides supplementary local geometric discrimination criteria for trace identification, the point density channel can distinguish between sparsely occluded areas of point clouds and densely exposed rock surfaces, the average curvature channel can distinguish between flat rock surfaces and structural surface edges, the normal gradient channel can distinguish between homogeneous rock surfaces and trace intersection areas, and the elevation change channel can distinguish between flat rock surfaces and steep faults.

[0126] In one possible and specific implementation scheme, such as Figure 3 As shown, this implementation scheme provides a feature map generation method for continuous structural representation of rock mass traces, targeting point cloud data of the Jinbaoshi rock mass in Mianyang. This method may include, but is not limited to, the following steps: Step S110: Obtain the three-dimensional point cloud data of the rock mass and calculate the geometric saliency score of each point in the three-dimensional point cloud data of the rock mass.

[0127] Step S110 includes, but is not limited to, steps S111 to S113.

[0128] Step S111: Construct a spatial neighborhood for each point in the three-dimensional point cloud data of the rock mass, and calculate the neighborhood geometric statistics of each point based on the spatial positional relationship between each point and its neighboring points.

[0129] Step S112: Calculate the curvature value of each point based on the neighborhood geometric statistics. Normal gradient and neighborhood point density standard deviation .

[0130] Step S113: The curvature value, normal gradient, and standard deviation of neighborhood point density are weighted and fused to obtain the geometric significance score. : in, The geometric significance score of a point is represented. Represents the curvature of a point. Represents the normal gradient. This represents the standard deviation of the neighborhood point density. , , The weighting coefficients are preset. In some embodiments of this application, the weighting coefficients are... Possible values ​​are [0.2, 0.5]. Possible values ​​are [0.3, 0.6]. The possible values ​​are [0.1, 0.4], and they satisfy the condition. + + =1; In the example of the Mianyang Jinbaoshi rock mass point cloud data, it can be taken as 1. =0.4、 =0.35、 =0.25.

[0131] Through steps S111 to S113, the geometric significance score is calculated for each point in the three-dimensional point cloud data of the rock mass. First, the three-dimensional point cloud data of the rock mass is obtained. The three-dimensional point cloud data of the rock mass is point cloud format data obtained by ground laser scanning of the rock mass of Mianyang Jinbaoshi. The three-dimensional point cloud data of the rock mass contains the three-dimensional coordinates and normal vector information of each point on the surface of the rock mass. Second, a spatial neighborhood is constructed for each point in the three-dimensional point cloud data of the rock mass. Based on the spatial positional relationship between each point and its neighboring points, the neighborhood geometric statistics of the point are calculated. Then, based on the neighborhood geometric statistics, the curvature value, normal gradient and neighborhood point density standard deviation of each point are calculated. Finally, the curvature value, normal gradient and neighborhood point density standard deviation are weighted and fused to obtain the geometric significance score. Among them, the curvature value characterizes the degree of surface bending, the normal gradient characterizes the rate of change of the surface normal direction, and the standard deviation of the neighborhood point density characterizes the uniformity of the local point cloud distribution. This index can effectively capture the trace signals implied by the uneven distribution of point clouds in complex surface environments with weathering, fracturing, and severe shading. Thus, by calculating the geometric saliency score, the degree of geometric abrupt change of each point in the rock mass point cloud data can be comprehensively characterized, providing a quantitative basis for the subsequent selection of key points.

[0132] Step S120: Select a set of key points based on geometric saliency scores.

[0133] Step S120 includes, but is not limited to, steps S121 to S123.

[0134] Step S121: Compare the geometric saliency score with a preset saliency threshold, and determine the points with geometric saliency scores higher than the preset saliency threshold as salient key points. In some embodiments of this application, the preset saliency threshold is adaptively set according to the rock mass type, point cloud sampling density, and noise level, preferably the 90th percentile of the geometric saliency score, so that salient key points account for approximately 10% of the total number of points; in this embodiment, for the Mianyang Jinbaoshi rock mass point cloud data, the preset saliency threshold is the 90th percentile.

[0135] Step S122: Perform farthest point sampling on the three-dimensional point cloud data of the rock mass to obtain sampling key points that are uniformly distributed in space.

[0136] Step S123: Combine the salient key points and the sampled key points according to a preset ratio to obtain the plurality of key points. In some embodiments of this application, the value of the preset ratio is based on the following: when the proportion of salient key points is too low, the trace area coverage is insufficient, and the lack of key points in areas rich in directional information will affect the construction quality of the subsequent directional consistency field; when the proportion of salient key points is too high, the spatial distribution of key points is uneven, the background area coverage is insufficient, and the directional consistency field has a large area of ​​blank space in the background area, affecting the topological integrity of the subsequent directional continuity relationship graph. Therefore, by setting this ratio to balance the priority coverage of salient areas and the global uniform coverage, the preset ratio in this embodiment is 1:1.

[0137] Through steps S121 to S123, a set of key points is selected based on geometric saliency scores. First, the geometric saliency scores are compared with a preset saliency threshold, and points with geometric saliency scores higher than the preset threshold are identified as salient key points. Second, farthest-point sampling is performed on the 3D point cloud data of the rock mass to obtain sampled key points that are evenly distributed in space. Then, the salient key points and sampled key points are combined according to a preset ratio to obtain a set of key points. In this way, the key point set is selected through a hybrid sampling strategy, which ensures both priority coverage of the trace area and uniform distribution of the background area, providing a complete and representative directional data foundation for the subsequent construction of the directional consistency field.

[0138] Step S130: Construct an anisotropic multi-scale local neighborhood point set for each key point.

[0139] Step S130 includes, but is not limited to, steps S131 to S135.

[0140] Step S131: Search for K nearest neighbors for each key point to obtain an initial neighborhood point set. Perform principal component analysis on the initial neighborhood to preliminarily estimate the local principal direction and local normal direction.

[0141] Step S132: Based on the preliminary estimated local principal direction and local normal direction, a first neighborhood scale is used along the principal direction of the structural surface to trace the continuity of the trace direction, and a second neighborhood scale is used along the normal direction to capture the fracture boundary. The neighborhood range corresponding to the first neighborhood scale is larger than the neighborhood range corresponding to the second neighborhood scale to accommodate the anisotropic geometric characteristics of the rock mass structural surface, which extends long along the strike direction and is narrow in the normal direction. The trace can extend for several meters or even tens of meters along the strike direction, but is only a few centimeters to tens of centimeters wide along the normal direction. This differentiated design, with a large scale along the principal direction and a small scale along the normal direction, can accurately capture the fracture boundary while tracing the continuity of the trace direction, avoiding the mixing of directional information across structural surfaces.

[0142] Step S133: Reconstruct the anisotropic local neighborhood point set based on the first neighborhood scale and the second neighborhood scale.

[0143] Step S134: A multi-scale neighborhood strategy is adopted for the anisotropic local neighborhood point set to construct at least two neighborhoods of different scales. Among them, the fine-scale neighborhood has 32 to 128 nearest neighbors and is used to capture fine traces and narrow fractures; the medium-scale neighborhood has 64 to 256 nearest neighbors and is used to capture regular joints; the coarse-scale neighborhood has 128 to 512 nearest neighbors and is used to capture large faults and wide fracture zones.

[0144] Step S135: Standardize the number of neighboring points to unify the number of neighboring points to a preset target number.

[0145] In some embodiments of this application, the range of the number of nearest neighbor points in the fine-scale neighborhood, medium-scale neighborhood, and coarse-scale neighborhood in step S134 is determined after conducting multiple sets of comparative experiments on the point cloud data of the Mianyang Jinbaoshi rock mass. The basis for the value is that different scale structural surfaces have different requirements for the number of neighborhood support points. Too few neighborhood points will cause the local principal direction estimation to be sensitive to noise and the estimation to be unstable. Too many neighborhood points will cause cross-structural surface mixing, weaken the boundary of abrupt change in direction, and reduce the ability to identify the direction of fine traces.

[0146] Through steps S131 to S135, an anisotropic multi-scale local neighborhood point set is constructed for each key point. First, an initial neighborhood is constructed and the principal direction is initially estimated. Second, an anisotropic neighborhood is constructed based on the initial principal direction and normal direction. Then, a multi-scale neighborhood strategy is used to construct anisotropic neighborhoods at three scales: fine, medium, and coarse, and the number of points is standardized. In this way, the anisotropic and multi-scale neighborhood construction strategy considers both the geometric anisotropy of the rock mass structural plane and adapts to the wide-span structural planes ranging from millimeter-level micro-fractures to meter-level large faults, providing an accurate basis for the estimation of the principal direction for the subsequent directional consistency field.

[0147] Step S140: Estimate the local principal direction through principal component analysis, establish a local coordinate system, and perform coordinate transformation and normalization.

[0148] Step S140 includes, but is not limited to, steps S141 to S145.

[0149] Step S141: Construct a covariance matrix for the local neighborhood point set and perform eigenvalue decomposition to obtain three eigenvalues ​​and their corresponding three eigenvectors.

[0150] Step S142: Based on the magnitude of the eigenvalues, determine the eigenvector corresponding to the largest eigenvalue as the local principal direction, the eigenvector corresponding to the second largest eigenvalue as the local secondary direction, and the eigenvector corresponding to the smallest eigenvalue as the local normal direction.

[0151] Step S143: Establish a local coordinate system based on the local principal direction, local secondary direction, and local normal direction. By normalizing the local coordinate system, the attitude differences in the neighborhood of different key points are eliminated, and all local principal directions are expressed under a unified reference frame, providing a unified direction reference benchmark for the subsequent construction of the direction consistency field, and ensuring the rotation invariance and comparability of the direction consistency calculation.

[0152] Step S144: Transform the neighborhood points from the global coordinate system to the local coordinate system based on the rotation matrix R.

[0153] Step S145: Perform scale normalization processing on the local neighborhood point set, scaling all point coordinates to the range of [0,1], so that the local neighborhoods corresponding to different key points have a consistent spatial scale.

[0154] Through steps S141 to S145, local principal directions are estimated and spatially normalized using principal component analysis. First, the covariance matrix is ​​constructed and eigenvalue decomposition is performed to determine the local principal direction, secondary direction, and normal direction, and a local coordinate system is established. Second, neighborhood points are transformed to the local coordinate system based on a rotation matrix. Finally, scale normalization is performed. This eliminates the differences in position, orientation, and scale between the neighborhoods of different key points, and all local principal directions are expressed within a unified reference frame, laying the foundation for comparable calculation of the orientation consistency field across the entire space.

[0155] Step S150: Construct a direction-consistent field based on the local principal directions of each key point.

[0156] Step S150 includes, but is not limited to, steps S151 to S154.

[0157] Step S151: For any spatial location in the orientation-consistent field Determine a region centered on that location. Local reference areas for key points.

[0158] Step S152: Obtain the local principal direction of each key point within the local reference region. Calculate the statistical principal directions of each local principal direction within the local reference region, and determine them as the regional reference directions. .

[0159] Step S153: Calculate the directional consistency value of the spatial location using the following formula: in, This represents the absolute value of the inner product between the local principal direction and the regional reference direction at each key point. The value range is [0, 1]. The closer it is to 1, the more consistent the directional distribution is within the neighborhood of that location, corresponding to a region where the traces of the rock mass structure surface are continuous in direction; The closer to 0, the more disordered the directional distribution is in the neighborhood of that location, corresponding to areas where rock mass traces terminate, intersect, or are subject to noise interference.

[0160] Step S154: Calculate all locations within the spatial range of the rock mass surface. This completes the construction of the directional consistency field. The directional consistency field is used to describe the spatial distribution and directional correlation of the local principal directions on the rock mass surface.

[0161] Orientational uniformity field This is the core structural model of this application. Unlike existing technologies that only statistically analyze isolated statistics such as directional variance or directional entropy within voxels, the directional consistency field is a continuous scalar field constructed across the entire rock mass surface space. The directional consistency value at each spatial location in the field not only reflects the degree of directional order within that location's neighborhood but also implicitly encodes the continuous directional structure of the rock mass trace—along the trace's strike—through the field's spatial distribution and gradient variations. It exhibits a continuous distribution of high values; at the termination or intersection of the traces... A sharp drop occurs; in noisy areas, It exhibits a low-value, disordered distribution. This application uses the mean absolute value of the inner product of the local principal direction and the regional reference direction of each key point as a measure of directional consistency. Compared with traditional statistics such as directional variance or directional entropy, the mean inner product directly measures the degree of geometric alignment between directional vectors, has a clear geometric meaning and a normalized value range of [0,1], which is more conducive to the unified calculation of propagation weights in the subsequent continuous directional relationship graph.

[0162] Through steps S151 to S154, by constructing a directional consistency field, firstly, a local reference region is determined for each location in space, and the local principal orientation of each key point is obtained; secondly, the regional reference orientation is calculated; and then, the directional consistency value is calculated based on the inner product mean, completing the construction of the entire field. In this way, the discrete principal orientation information of key points on the rock mass surface is transformed into a spatially continuous directional consistency scalar field, elevating the directional continuity structure from an implicit geometric attribute to an explicit spatial structural expression.

[0163] Step S160: Discretize the directional uniformity field to a voxel grid, construct a directional continuity graph, and perform iterative propagation updates.

[0164] Step S160 includes, but is not limited to, steps S161 to S166.

[0165] Step S161: Discretize the directional uniformity field to a regular three-dimensional voxel grid, that is, for each voxel, take the voxel center position as the sampling point, and extract the directional uniformity field... The directional consistency value of the voxel was obtained by sampling. .

[0166] Step S162: Using each voxel as a node, establish edges between adjacent voxels to construct a directional continuity graph. Adjacent voxels are defined as voxels that are adjacent along a local principal direction, a local secondary direction, or a local normal direction in the 3D voxel mesh.

[0167] Step S163: Calculate the direction propagation weight for each edge in the direction continuity graph. The directional propagation weights are determined by adjacent voxels. and The angle between the local principal directions Differences in directional consistency values and spatial distance Jointly determined: in, , , Preset positive adjustment parameters are used to control the attenuation sensitivity of the propagation weights to the directional angle, consistency difference, and spatial distance, respectively. The design of the directional propagation weights described above has a clear physical meaning: directional angle... The smaller the value, the more consistent the orientation of the structural planes at the two voxels, and the higher the propagation weight; the difference in directional consistency values... The smaller the value, the higher the reliability of propagation, indicating that both voxels are either trace regions or background regions; spatial distance The smaller the value, the higher the reliability of the directional continuity inference. The three factors are coupled in an exponentially decaying form as a product, making... It exhibits high sensitivity to directional discontinuities, abrupt changes in consistency, or spatial jumps simultaneously; however, it only becomes truly sensitive when all three conditions are met. It is close to 1; if any condition is not met, All decay rapidly, approaching 0.

[0168] Step S164: Select the propagation source node. Set the direction consistency value. Higher than the preset propagation source threshold The voxels were identified as the propagation source nodes. A value of 0.7 to 0.9 can be used in the example of point cloud data from the Mianyang Jinbaoshi rock mass. =0.8. The region on the rock mass surface corresponding to the propagation source node has a consistent height in the direction and is the "anchor point" of the continuous structure of the trace direction.

[0169] Step S165: In each iteration In the middle, for each non-propagation source voxel node Get all its adjacent voxel nodes Current direction consistency feature and the corresponding directional propagation weights Perform weighted aggregation update as follows: in, ∈(0, 1) represents the preset propagation update step size coefficient. Voxel representation In the The directional consistency eigenvalue in each iteration. This update formula consists of retained terms. It consists of two parts: a retention term to ensure that the original information is not lost during propagation, and a neighborhood aggregation term to perform a weighted average of the directional consistency features of adjacent voxels using normalized directional propagation weights as weighting coefficients. Control the step size of each update round. Larger sizes result in faster propagation speeds but may lead to over-smoothing of details. When the size is small, the propagation is fine but requires more iterations. In this embodiment, Take values ​​between 0.3 and 0.5, for the total number of iterations. Take 10 to 50 rounds.

[0170] Step S166: via After one round of iterations, the process stops, yielding the directional consistency feature updated by propagation. After multiple iterations, the directional consistency feature is continuously correlated along the directional relationship graph. Propagation and accumulation occur along large, directionally continuous paths. Between voxels continuously distributed along the path, due to small directional angles, minimal consistency differences, and short spatial distances, When the value is close to 1, the directional consistency feature is fully weighted and enhanced; at edges where the direction is discontinuous (such as trace terminations, intersections, or boundaries of noisy regions), due to a sudden increase in the directional angle or large differences in consistency, Approaching zero, transmission is almost completely blocked.

[0171] The core technical effect of the aforementioned iterative propagation update mechanism lies in enhancing directional continuity along the trajectory. For discontinuous trajectory regions where the point cloud is locally sparse due to weathering, occlusion, or scanning angle, although the initial directional consistency value of the voxels in this region is low, through weighted propagation along the trajectory of adjacent voxels, enhanced directional consistency from the propagation source node can be obtained after multiple iterations, thus forming a continuous directional signal in the feature map. Simultaneously, for noisy regions, due to the large difference in directional consistency with surrounding voxels and low propagation weight, the noise signal is naturally suppressed during iteration.

[0172] Through steps S161 to S166, the directional consistency field is discretized, a continuous directional relationship graph is constructed, and iterative propagation updates are performed. First, the continuous directional consistency field is sampled onto a voxel grid. Second, a graph structure is constructed with voxels as nodes and adjacent voxels as edges, and the directional propagation weight of each edge is calculated based on three factors: directional angle, consistency difference, and spatial distance. Then, voxels with high consistency are selected as propagation source nodes. Finally, multiple rounds of weighted aggregation iterative updates are performed on the continuous directional relationship graph. In this way, the directional consistency feature is propagated and accumulated along the continuous directional graph path, achieving effective connection of discontinuous traces, significant enhancement of weak trace response, and suppression of noise.

[0173] Step S170: Encode multi-channel geometric attributes to generate a rock mass trace feature map with directional continuous structure expression.

[0174] Step S170 includes, but is not limited to, steps S171 to S174.

[0175] Step S171: Encode multi-channel geometric attributes for each voxel. The multi-channel geometric attributes include orientation consistency features updated via propagation. And at least two of the following geometric properties: point density, mean curvature, normal gradient, and elevation variation.

[0176] Step S172: Fill the voxels with default values ​​without point distribution to generate a complete feature map.

[0177] Step S173: Arrange the multi-channel geometric attributes of all voxels according to their spatial location to generate a rock mass trace feature map.

[0178] Step S174: The rock mass trace feature map not only includes local geometric attributes, but also explicitly characterizes the directional continuity structure, spatial topological continuity relationship along the trace direction, and directional propagation relationship through the edge weight distribution and propagation-updated directional consistency feature distribution of the directional continuity relationship map.

[0179] Compared to existing feature maps based on isolated voxel encoding, the rock mass trace feature map generated in this application has the following essential differences: First, the directional consistency feature of each voxel. It is the global optimization result after iterative propagation and update on the directional continuity graph, rather than a local isolated statistic. It integrates the directional information of the voxel itself and all propagation sources within the directional connectivity region along the trace. Second, the feature map inherently contains the directional continuity relationship and propagation path information between voxels, which can be obtained through the edge weights of the directional continuity graph. Third, the directional continuity structure becomes an explicit and quantifiable feature channel, exhibiting a high-value continuous distribution along the trajectory and a cliff-like drop at the end or intersection of the trajectory, making the spatial topology of the trajectory directly readable from the feature map.

[0180] Through steps S171 to S174, a rock mass trace feature map is generated by encoding multi-channel geometric attributes. First, each voxel is encoded with a multi-channel attribute containing the directional consistency feature after propagation and update. Second, empty voxels are filled. Then, the feature map is generated by arranging them according to their spatial positions. In this way, the generated feature map explicitly expresses the directional continuity structure of the rock mass trace, providing a complete expression with both geometric attributes and directional structure information for subsequent trace recognition.

[0181] Through steps S110 to S170, by utilizing the feature map generation method for continuous structural expression of rock mass trace direction provided in this application, firstly, three-dimensional point cloud data of the rock mass is acquired, geometric saliency scores are calculated, and key points are selected; secondly, an anisotropic multi-scale local neighborhood is constructed, local principal directions are estimated through principal component analysis, and coordinate system normalization is performed; then, a direction consistency field is constructed based on the local principal directions of each key point, transforming the discrete local principal direction information into a spatially continuous direction structure expression; subsequently, the direction consistency field is discretized to a voxel grid, and a direction continuity relationship graph is constructed with voxels as nodes and three-factor coupled direction propagation weights as edge weights. On this graph structure, the direction consistency features are iteratively propagated and updated using highly consistent voxels as propagation sources to achieve enhanced direction continuity along the trace direction; finally, multi-channel geometric attributes are encoded to generate a rock mass trace feature map with a continuous structural expression of direction. Thus, by following the core technical route of "directional consistency field modeling → directional continuous relation graph construction → iterative propagation update → directional continuous structure feature graph generation", isolated points, neighborhoods and voxels in existing technologies are uniformly incorporated into a graph structure with directional propagation weights as edge weights for continuous propagation and enhancement, realizing a paradigm shift from isolated encoding to directional continuous structure expression.

[0182] In summary, to address the technical challenges of explicitly modeling and representing the continuous directional structure of rock mass traces in existing technologies, the susceptibility of weak fracture directional signals to noise, and the lack of directional correlation mechanisms for discontinuous traces, this implementation proposes a feature map generation method for representing the continuous directional structure of rock mass traces. This method can automatically process 3D point cloud data of rock masses in complex surface environments. Through steps such as hybrid keypoint sampling, anisotropic multi-scale neighborhood construction, directional consistency field modeling, directional continuity graph construction, iterative propagation and updating, and multi-channel geometric attribute encoding, it achieves an explicit representation of the continuous directional structure of rock mass traces. Compared to existing feature representation methods based on isolated encoding, this embodiment can effectively connect discontinuous traces, enhance the response of weak traces, suppress noise interference, and improve the recognition accuracy and robustness of weak fractures and low-salience structural surfaces in complex rock mass scenarios, demonstrating good engineering practical value and promising prospects for widespread application.

[0183] According to an embodiment of the present invention, a feature map generation device for expressing continuous structure in the direction of rock mass traces is provided. Please refer to [link to relevant documentation]. Figure 3 ,include: The acquisition module is used to acquire three-dimensional point cloud data of rock mass, determine multiple target points from the three-dimensional point cloud data of rock mass, acquire the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point; The first generation module is used to generate a direction consistency field based on the local principal directions of each target point; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location; The module is used to sample the orientation consistency field onto a three-dimensional voxel mesh. With each voxel as a node, for each pair of adjacent voxels, the orientation propagation weight between the pair of adjacent voxels is determined based on the angle between the local principal directions of the two voxels, the difference in orientation consistency field values, and the spatial distance. The orientation propagation weight is used as the edge weight to construct an orientation continuity graph. The iterative module is used to identify voxels with directional consistency field values ​​higher than a preset threshold as propagation source nodes on the directional continuity graph, and to perform multiple rounds of iterative updates on the directional consistency characteristics of each voxel by controlling the propagation intensity along the edges of the directional continuity graph with directional propagation weights. The second generation module is used to combine the directional consistency features with various geometric attributes corresponding to each voxel based on the iteratively updated directional consistency features to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

[0184] According to an embodiment of the present invention, an electronic device is provided; please refer to... Figure 4 The electronic device in this embodiment may include one or more of the following components: a processor, a network interface, memory, non-volatile memory, and one or more application programs, wherein the one or more application programs may be stored in non-volatile memory and configured to be executed by one or more processors, and the one or more programs are configured to perform the methods as described in the foregoing method embodiments.

[0185] According to embodiments of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a computer, causes the computer to perform the method described in any of the above embodiments.

[0186] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating feature maps for expressing continuous structure along rock mass traces, characterized in that, include: Acquire three-dimensional point cloud data of the rock mass, determine multiple target points from the three-dimensional point cloud data of the rock mass, obtain the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point; Based on the local principal directions of each target point, a direction consistency field is generated; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location; The directional consistency field is sampled onto a three-dimensional voxel mesh. With each voxel as a node, for each pair of adjacent voxels, the directional propagation weight between the two pairs of adjacent voxels is determined based on the angle between the local principal directions corresponding to the two voxels, the difference in the directional consistency field values, and the spatial distance. The directional propagation weight is used as the edge weight to construct a directional continuity graph. On the directional continuity graph, voxels with directional consistency field values ​​higher than a preset threshold are identified as propagation source nodes. Along the edges of the directional continuity graph, the propagation intensity is controlled by directional propagation weights, and the directional consistency characteristics of each voxel are iteratively updated in multiple rounds. Based on the iteratively updated directional consistency features, channels are combined with various geometric attributes corresponding to each voxel to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

2. The method according to claim 1, characterized in that, The step of determining multiple target points from the three-dimensional point cloud data of the rock mass includes: Calculate the geometric significance score of each point in the three-dimensional point cloud data of the rock mass; wherein the geometric significance score is the weighted sum of the curvature, normal gradient and neighborhood point density standard deviation of the point; wherein curvature represents the degree of surface bending, normal gradient represents the rate of change of the surface normal direction, and neighborhood point density standard deviation represents the uniformity of the local point cloud distribution. Points with geometric significance scores higher than a preset significance threshold are identified as significance target points; Perform farthest point sampling on the three-dimensional point cloud data of the rock mass to obtain sampling target points that are uniformly distributed in space; By combining salient target points and sampling target points according to a preset ratio, multiple target points are obtained.

3. The method according to claim 2, characterized in that, The step of obtaining the set of points in the local neighborhood of each target point includes: For each target point, a preset number of nearest neighbors are searched to obtain an initial neighborhood point set. Principal component analysis is then performed on the initial neighborhood point set to preliminarily determine the local principal direction and local normal direction. Based on the initially determined local principal direction and local normal direction, a first neighborhood scale is adopted along the principal direction of the structural surface, and a second neighborhood scale is adopted along the normal direction. The neighborhood range corresponding to the first neighborhood scale is larger than the neighborhood range corresponding to the second neighborhood scale. Based on the first and second neighborhood scales, the local neighborhood point sets corresponding to each target point are reconstructed to obtain anisotropic neighborhood point sets.

4. The method according to claim 1, characterized in that, The step of generating a direction-consistent field based on the local principal directions of each target point includes: For any spatial location in the orientation consistency field, a local reference region is determined with that location as the center. From the obtained local principal directions of each target point, the local principal directions of each target point within the local reference region are extracted. Based on the local principal direction of each target point within the local reference area, determine the regional reference direction of the local reference area; Calculate the absolute value of the inner product between the local principal direction and the regional reference direction of each target point within the local reference region, and use the mean of the absolute values ​​of the inner products as the orientation consistency value of that spatial location. Traverse the spatial range covered by the directional consistency field, calculate the directional consistency field value at each position, and generate the directional consistency field.

5. The method according to claim 4, characterized in that, The step of determining the directional propagation weight between adjacent voxels based on the angle between the local principal directions of the two voxels, the difference in directional consistency field values, and the spatial distance includes: Obtain the local principal directions corresponding to two adjacent voxels, calculate the angle between the two local principal directions, and use the angle as the decay variable of the first exponential decay factor to construct the first decay factor. Obtain the directional uniformity field values ​​of two adjacent voxels, calculate the absolute value of the difference between the two directional uniformity field values, and use the absolute value as the decay variable of the second exponential decay factor to construct the second decay factor; Calculate the spatial distance between two adjacent voxels, and use the spatial distance as the decay variable of the third exponential decay factor to construct the third decay factor; Multiplying the first attenuation factor, the second attenuation factor, and the third attenuation factor together yields the directional propagation weight between the pair of adjacent voxels; wherein the attenuation sensitivity of the first attenuation factor, the second attenuation factor, and the third attenuation factor is controlled by preset positive adjustment parameters.

6. The method according to claim 5, characterized in that, The process of iteratively updating the directional consistency features of each voxel includes: In the direction continuity graph, the direction consistency characteristics of each propagation source voxel remain unchanged during the iteration process; In each iteration, for each non-propagating source voxel, the ratio of the directional propagation weight between the non-propagating source voxel and each of its neighboring voxels to the sum of the directional propagation weights between the non-propagating source voxel and all its neighboring voxels is calculated and used as the normalized propagation weight of each neighboring voxel. Using the normalized propagation weights of each neighboring voxel as weighting coefficients, the directional consistency feature values ​​of each neighboring voxel in the current round are weighted and summed to obtain the neighborhood aggregation feature value. The neighborhood aggregation feature value is weighted by a preset update step size coefficient, and the directional consistency feature value of the non-propagating source voxel in the current round is weighted by 1 minus the preset update step size coefficient. The sum of the two weighted results is taken as the directional consistency feature value of the non-propagating source voxel in the next round. After a preset number of iterations, the iteration stops, and the directional consistency characteristics of each voxel after iteration are obtained.

7. The method according to claim 6, characterized in that, The step of generating a rock mass trace feature map by combining the iteratively updated directional consistency features with multiple geometric attributes corresponding to each voxel through channels includes: Obtain at least two geometric properties corresponding to each voxel; wherein, the at least two geometric properties are selected from point density, mean curvature, normal gradient and elevation change; The directional consistency feature after iterative update of each voxel is used as one channel, and at least two geometric attributes are used as the remaining channels. The channels are combined according to the channel dimension to form the multi-channel feature vector corresponding to each voxel. For voxels with no point distribution in the voxel grid, fill each channel of the voxel with the preset default value; Arrange the multi-channel feature vectors of all voxels according to their spatial positions in the three-dimensional voxel grid to generate a rock mass trace feature map.

8. A feature map generation device for expressing continuous structure along rock mass traces, characterized in that, include: The acquisition module is used to acquire three-dimensional point cloud data of rock mass, determine multiple target points from the three-dimensional point cloud data of rock mass, acquire the point set in the local neighborhood of each target point, perform principal component analysis on the point set, and obtain the local principal direction corresponding to each target point; The first generation module is used to generate a direction consistency field based on the local principal directions of each target point; wherein, the direction consistency field is a continuous scalar field defined on the surface space of the rock mass; wherein, the field value at any spatial location in the continuous scalar field represents the degree of consistency of each local principal direction within the neighborhood centered on that location; The module is used to sample the orientation consistency field onto a three-dimensional voxel mesh. With each voxel as a node, for each pair of adjacent voxels, the orientation propagation weight between the pair of adjacent voxels is determined based on the angle between the local principal directions of the two voxels, the difference in orientation consistency field values, and the spatial distance. The orientation propagation weight is used as the edge weight to construct an orientation continuity graph. The iterative module is used to identify voxels with directional consistency field values ​​higher than a preset threshold as propagation source nodes on the directional continuity graph, and to perform multiple rounds of iterative updates on the directional consistency characteristics of each voxel by controlling the propagation intensity along the edges of the directional continuity graph with directional propagation weights. The second generation module is used to combine the directional consistency features with various geometric attributes corresponding to each voxel based on the iteratively updated directional consistency features to generate a rock mass trace feature map; wherein, the directional consistency feature channel in the rock mass trace feature map represents the directional continuity structure of the rock mass trace.

9. An electronic device, characterized in that, include: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images