A method for identifying abnormal surrounding rock of a tunnel during excavation of a horizontal layered rock mass

Abstract

The application discloses a kind of horizontal layered rock mass excavation tunnel surrounding rock anomaly identification method, belong to three-dimensional point cloud processing technical field.The application obtains three kinds of scale original point cloud graph based on tunnel surrounding rock three-dimensional point cloud graph, then obtains abnormal inclination by calculating the normal vector of each scale point cloud point, and then filters abnormal and normal connected region and generates shape offset degree, reflection offset degree point cloud graph;Subsequently, the same scale inclination, shape offset degree and reflection offset degree point cloud graph are fused to construct corresponding scale weight point cloud graph;Finally, the layered rock mass anomaly identification three-channel fusion network is used to process three kinds of scale original point cloud graph, and the weight point cloud graph is combined to apply weight, to realize the accurate identification of surrounding rock abnormal area.The application improves the accuracy of horizontal layered rock mass tunnel surrounding rock anomaly identification by multi-scale feature extraction and weight fusion, and provides technical support for tunnel construction safety.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional point cloud processing technology, specifically relating to a method for identifying anomalies in the surrounding rock of a horizontally layered rock mass excavation tunnel. Background Technology

[0002] During tunnel excavation, the stability of the surrounding rock directly affects the safety and progress of the project. For horizontally layered rock masses, due to their distinct stratification characteristics and significant differences in mechanical properties between different layers, they are highly susceptible to interlayer displacement, spalling, and localized abnormal damage under excavation disturbance. Therefore, accurately identifying abnormal areas in the surrounding rock during the excavation of horizontally layered rock masses is a crucial foundation for dynamic monitoring and early warning of construction risks.

[0003] Currently, point cloud data of tunnel surrounding rock obtained based on 3D scanning technology has been widely used in rock mass structure analysis and deformation monitoring. The 3D point cloud image is filtered and then input into a target recognition neural network for identification. Due to the large amount of 3D point cloud data and the indistinct features of abnormal areas, the target recognition neural network needs to adaptively search for abnormal areas.

[0004] However, in practical applications, the spatial structure of surrounding rock point clouds is complex and unevenly distributed. Anomaly regions often account for only a very small proportion of the overall point cloud, and their spatial features are easily overwhelmed by the overall morphological information. Existing neural networks, when extracting point cloud features, typically rely on global feature aggregation or feature scanning methods with fixed convolutional kernels. This makes it difficult to extract key features from anomaly regions, resulting in insufficient model response to changes in anomaly structures and low anomaly recognition accuracy. Summary of the Invention

[0005] To address the aforementioned shortcomings in the existing technology, this invention discloses a method for identifying anomalies in the surrounding rock of tunnels excavated in horizontally layered rock masses. The purpose of this method is to solve the problem of low accuracy in anomaly identification in the existing technology.

[0006] The technical solution adopted in this invention is as follows:

[0007] A method for identifying anomalies in the surrounding rock of a tunnel excavated in a horizontally layered rock mass includes:

[0008] Based on the three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass, three scales of original point cloud maps were obtained.

[0009] Based on the normal vector of each point in the original point cloud map at each scale, calculate the abnormal tilt of each point to obtain the tilt point cloud map at the corresponding scale.

[0010] Based on the abnormal tilt of the point cloud points, abnormal and normal connected regions are filtered out. Based on the difference in shape and reflection value between the abnormal and normal connected regions, the corresponding scale shape offset point cloud map and reflection offset point cloud map are obtained.

[0011] Based on the tilt point cloud map, shape offset point cloud map, and reflection offset point cloud map of the same scale, generate a weighted point cloud map of the corresponding scale.

[0012] A three-channel fusion network for anomaly identification of layered rock masses is used to process the original point cloud maps at three scales. Weights are applied to the point cloud maps at the corresponding scales to obtain the anomaly regions.

[0013] Furthermore, the process of obtaining the original point cloud maps at three scales includes:

[0014] The three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass is used as the high-scale original point cloud map.

[0015] The high-scale original point cloud map is divided into multiple three-dimensional spaces. The center of the three-dimensional space is selected as the representative point. The representative points are arranged according to their positions in the three-dimensional spaces to obtain the medium-scale original point cloud map.

[0016] The original point cloud map at the medium scale is divided into multiple three-dimensional spaces. The center of each three-dimensional space is selected as a representative point. The representative points are arranged according to their positions in each three-dimensional space to obtain the original point cloud map at the low scale.

[0017] Furthermore, the formula for calculating the abnormal tilt of each point cloud point is:

[0018] ,

[0019] Among them, Υ i Let n be the anomalous tilt of the i-th point in the original point cloud map at each scale. i Let n be the normal vector of the i-th point cloud point. i,j Let be the normal vector of the j-th point in the neighborhood of the i-th point in the point cloud, arccos be the angle between the two normal vectors, N be the number of point clouds in the neighborhood of the i-th point in the point cloud, and i and j be positive integers.

[0020] Furthermore, the process of filtering out abnormal and normal connected regions includes:

[0021] Calculate the mean of the abnormal tilt of each point in the original point cloud map at each scale to obtain the overall tilt.

[0022] Point cloud points with an abnormal tilt greater than the overall tilt are marked as abnormal points, and other points are marked as normal points;

[0023] The connected regions formed by each outlier point are considered as outlier connected regions.

[0024] The connected regions formed by each normal point are considered as normal connected regions.

[0025] Furthermore, the process of obtaining the point cloud map with the corresponding scale shape offset includes:

[0026] Extract each normal connected region that contacts the abnormal connected region, and select the normal connected region with the largest area from the contacting normal connected regions as the reference connected region.

[0027] Calculate the distance from each anomalous point in the anomalous connected region to the plane equation of the corresponding reference connected region to obtain the shape distance;

[0028] The shape distance of each outlier point is normalized to obtain the shape offset of the outlier point, and the shape offset of the normal point is set to 0.

[0029] The shape offset of the original point cloud map belonging to the corresponding scale is used to construct a point cloud map with shape offset at the corresponding scale.

[0030] Furthermore, the process of obtaining the corresponding scale reflection offset point cloud map includes:

[0031] Extract each normal connected region that contacts the abnormal connected region, and select the normal connected region with the largest area from the contacting normal connected regions as the reference connected region.

[0032] The overall reflection value is obtained by averaging the reflection values ​​of each region in the reference connected region.

[0033] Calculate the distance from the reflection value of each abnormal point in the abnormal connected region to the overall reflection value to obtain the reflection distance;

[0034] The reflection distance is normalized to obtain the reflection offset of abnormal points, and the reflection offset of normal points is set to 0.

[0035] The reflection offset of the original point cloud map belonging to the corresponding scale is used to construct the reflection offset point cloud map of the corresponding scale.

[0036] Furthermore, the process of generating a weighted point cloud map at the corresponding scale includes:

[0037] The first pointwise convolutional layer is used to map the tilt point cloud map pointwise to obtain the tilt feature point cloud map.

[0038] The shape offset point cloud is mapped point by point using a second pointwise convolutional layer to obtain the shape offset feature point cloud.

[0039] A third pointwise convolutional layer is used to map the reflection offset point cloud map pointwise to obtain the reflection offset feature point cloud map.

[0040] The tilt feature point cloud map, shape offset feature point cloud map, and reflection offset feature point cloud map of the same scale are added element by element to obtain the weighted point cloud map of the corresponding scale.

[0041] Furthermore, the three-channel fusion network for layered rock mass anomaly identification includes: a first enhanced feature extraction channel, a second enhanced feature extraction channel, a third enhanced feature extraction channel, a first downsampling layer, a second downsampling layer, adder A1, adder A2, and an output unit;

[0042] The first input of the first enhanced feature extraction channel is used to input the high-scale original point cloud map, its second input is used to input the high-scale weighted point cloud map, and its output is connected to the input of the first downsampling layer.

[0043] The first input of the second enhanced feature extraction channel is used to input the original mesoscale point cloud map, and its second input is used to input the mesoscale weighted point cloud map.

[0044] The first input of the third enhanced feature extraction channel is used to input the low-scale original point cloud map, and its second input is used to input the low-scale weighted point cloud map.

[0045] The first input of adder A1 is connected to the output of the first downsampling layer, its second input is connected to the output of the second enhanced feature extraction channel, and its output is connected to the input of the second downsampling layer.

[0046] The first input of adder A2 is connected to the output of the second downsampling layer, its second input is connected to the output of the third enhanced feature extraction channel, and its output is connected to the input of the output unit.

[0047] The output end of the output unit serves as the output end of the three-channel fusion network for layered rock mass anomaly identification.

[0048] Furthermore, the first enhanced feature extraction channel, the second enhanced feature extraction channel, and the third enhanced feature extraction channel have the same structure, each including: a point cloud feature enhancement module, a first stacked convolution module, a second stacked convolution module, and a third stacked convolution module;

[0049] The first input of the point cloud feature enhancement module serves as the first input of the three enhanced feature extraction channels, the second input serves as the second input of the three enhanced feature extraction channels, and the output is connected to the input of the first stacked convolution module.

[0050] The input of the second stacked convolutional module is connected to the output of the first stacked convolutional module, and its output is connected to the input of the third stacked convolutional module.

[0051] The output of the third stacked convolutional module serves as the output of the three enhanced feature extraction channels.

[0052] Furthermore, the point cloud feature enhancement module includes: a fourth pointwise convolutional layer and a multiplication layer;

[0053] The input of the fourth pointwise convolutional layer serves as the first input of the point cloud feature enhancement module;

[0054] The first input of the multiplication layer is connected to the output of the fourth pointwise convolutional layer, and its second input serves as the second input of the point cloud feature enhancement module.

[0055] The output of the multiplication layer serves as the output of the point cloud feature enhancement module.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. The identification method of the present invention divides the three-dimensional point cloud data of the tunnel surrounding rock into multiple scales and calculates the inclination of the normal vector of the point cloud points at each scale. It can adaptively capture the abnormal morphological change features at different spatial scales and effectively improve the model's sensitivity to the identification of small abnormal areas.

[0058] 2. The identification method of this invention uses three-dimensional indicators—abnormal tilt, shape offset, and reflection offset—to characterize the features of abnormal regions from different dimensions, avoiding the abnormal features being submerged by the overall shape. It also generates a weighted cloud map based on multi-dimensional indicators, giving higher weight to abnormal regions, allowing the network to focus on key areas, and solving the problems of small abnormal proportions and difficulty in feature extraction.

[0059] 3. Compared with traditional methods that rely on global feature aggregation, the identification method of this invention calculates and fuses local offset features at multiple scale levels, making it less likely for abnormal regions to be covered by overall structural information, thus significantly improving the network's response capability to local abnormal regions.

[0060] 4. The recognition method of the present invention introduces a weighted point cloud map generated based on tilt, shape offset and reflection offset in the network input stage. Weights are applied to the corresponding features of the original point cloud map at the same scale, so that the network can capture key anomaly features at different scales and combine key anomaly features at multiple scales to obtain anomaly regions, thereby improving the accuracy of anomaly recognition. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of a three-dimensional model of a stone particle according to an embodiment of the present invention.

[0062] Figure 2 This is a model diagram of the boulders stacking platform used in the simulation of an embodiment of the present invention.

[0063] Figure 3 This is an intermediate state diagram of particle accumulation during the discrete element simulation process in an embodiment of the present invention.

[0064] Figure 4 This is a three-dimensional model of the pile of stones after the stacking is completed in an embodiment of the present invention.

[0065] Figure 5 This is a diagram showing the result of segmentation and annotation of the stacking model after topological analysis in an embodiment of the present invention.

[0066] Figure 6 This is a complete point cloud image of the particle model after uniform sampling in an embodiment of the present invention.

[0067] Figure 7 This is a point cloud image after occlusion removal processing in an embodiment of the present invention. Detailed Implementation

[0068] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0069] like Figure 1 As shown, a method for identifying anomalies in the surrounding rock of a tunnel excavated in a horizontally layered rock mass includes:

[0070] Based on the three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass, three scales of original point cloud maps were obtained.

[0071] Based on the normal vector of each point in the original point cloud map at each scale, calculate the abnormal tilt of each point to obtain the tilt point cloud map at the corresponding scale.

[0072] Based on the abnormal tilt of the point cloud points, abnormal and normal connected regions are filtered out. Based on the difference in shape and reflection value between the abnormal and normal connected regions, the corresponding scale shape offset point cloud map and reflection offset point cloud map are obtained.

[0073] Based on the tilt point cloud map, shape offset point cloud map, and reflection offset point cloud map of the same scale, generate a weighted point cloud map of the corresponding scale.

[0074] A three-channel fusion network for anomaly identification of layered rock masses is used to process the original point cloud maps at three scales. Weights are applied to the point cloud maps at the corresponding scales to obtain the anomaly regions.

[0075] In this embodiment, after acquiring a 3D laser point cloud image of the tunnel surrounding rock during the excavation of horizontally layered rock mass, the normal vector of each point in the 3D point cloud image is first obtained. The process of obtaining the normal vector includes: acquiring a neighborhood point cloud set for each target point in the 3D point cloud image; calculating the geometric center (i.e., the mean coordinate) of the neighborhood point cloud set; and subtracting the coordinates of the geometric center from the coordinates of the target point and all its neighbors to obtain a decentralized point cloud set, thus eliminating the influence of translation on local geometric features. Based on the decentralized point set, a 3×3 covariance matrix is ​​calculated. M is the covariance matrix, P i,k Let C be the coordinates of the k-th point in the neighborhood of the i-th target point, T be the transpose, and K be the number of point points in the neighborhood of the target point. Perform eigenvalue decomposition on the covariance matrix M to obtain three eigenvalues ​​λ1≥λ2≥λ3 and their corresponding eigenvectors v1, v2, v3. The eigenvector v3 corresponding to the smallest eigenvalue λ3 is the normal vector of the target point.

[0076] In this embodiment, the process of obtaining the original point cloud maps at three scales includes:

[0077] The three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass is used as the high-scale original point cloud map.

[0078] The high-scale original point cloud map is divided into multiple three-dimensional spaces. The center of the three-dimensional space is selected as the representative point. The representative points are arranged according to their positions in the three-dimensional spaces to obtain the medium-scale original point cloud map.

[0079] The original point cloud map at the medium scale is divided into multiple three-dimensional spaces. The center of each three-dimensional space is selected as a representative point. The representative points are arranged according to their positions in each three-dimensional space to obtain the original point cloud map at the low scale.

[0080] In this invention, since a normal vector is obtained for each point in the three-dimensional point cloud map, if a point exists in the center of the three-dimensional space of the high-scale original point cloud map, the normal vector of that point is used as the normal vector of the representative point in the mesoscale original point cloud map. If no point exists in the center of the three-dimensional space, the normal vector of the point closest to the center of the three-dimensional space in the high-scale original point cloud map is used as the normal vector of the representative point.

[0081] In the low-scale original point cloud map, if there is a point cloud point at the center of the three-dimensional space in the meso-scale original point cloud map, then the normal vector of that point cloud point is used as the normal vector of the representative point in the low-scale original point cloud map. If there is no point cloud point at the center of the three-dimensional space, then the normal vector of the point cloud point closest to the center of the three-dimensional space in the meso-scale original point cloud map is used as the normal vector of the representative point.

[0082] Therefore, for each scale of the original point cloud map, there is a corresponding normal vector. The original point cloud map at one scale will generate a tilt point cloud map, a shape offset point cloud map, and a reflection offset point cloud map.

[0083] At the original point cloud scale, the calculation of the abnormal tilt between adjacent points often only reflects local minor changes. If the abnormal area is large, its overall tilt trend is difficult to be effectively reflected.

[0084] This invention divides the original point cloud into multiple three-dimensional spaces at different scales, and uses the center point of each three-dimensional space as a representative point, which is equivalent to performing a downsampling in space. This allows anomalous regions to be "compressed" into smaller local structures after downsampling. Therefore, large-scale anomalous regions that are difficult to reflect the overall trend in high-density point clouds can be clearly revealed at medium or low scales.

[0085] Medium- and low-scale point clouds expand their neighborhood range through stereo sampling (increasing the spacing between representative points), allowing the neighborhood to cover a larger space: medium-scale points are adapted to medium-sized anomalies, and low-scale points are adapted to large-scale anomalies. At this point, the anomaly tilt calculation of representative points can integrate normal vector information over a larger range, avoiding the segmentation of large-scale anomalies into "locally normal segments" due to an excessively small neighborhood (e.g., local points of large-scale anomalies may be misjudged as normal because the normal vectors in the neighborhood are consistent, but after expanding the scale, the overall tilt difference can be captured).

[0086] In this embodiment, the side length of the three-dimensional space is set to 5 times the point cloud pitch.

[0087] In this embodiment, the formula for calculating the abnormal tilt of each point cloud point is:

[0088] ;

[0089] Among them, Υ i Let n be the anomalous tilt of the i-th point in the original point cloud map at each scale. i Let n be the normal vector of the i-th point cloud point. i,j Let be the normal vector of the j-th point in the neighborhood of the i-th point in the point cloud, arccos be the angle between the two normal vectors, N be the number of point clouds in the neighborhood of the i-th point in the point cloud, and i and j be positive integers.

[0090] In this invention, the abnormal tilt degree is calculated for each point in the original point cloud map at each scale, and the abnormal tilt degrees are arranged according to the position of the point cloud points to obtain the tilt point cloud map of the corresponding scale.

[0091] This invention calculates the average angle between the normal vectors of a target point and its neighboring points, and normalizes it to an interval. This allows for a direct and quantitative reflection of the dispersion of the normal vectors in the local area surrounding the target point. For horizontally layered rock masses, the normal vectors in normal bedding regions are highly consistent, with abnormal dip values ​​close to 0. However, in abnormal regions such as fractures and breccia zones, the normal vectors exhibit significant abrupt changes, and the abnormal dip values ​​increase significantly, thus enabling accurate identification of abnormal regions.

[0092] In this embodiment, the process of filtering abnormal and normal connected regions includes:

[0093] Calculate the mean of the abnormal tilt of each point in the original point cloud map at each scale to obtain the overall tilt.

[0094] In the original point cloud map at the same scale, point cloud points with abnormal tilt greater than the corresponding overall tilt are marked as abnormal points, and other points are marked as normal points;

[0095] The connected regions formed by each outlier point are considered as outlier connected regions.

[0096] The connected regions formed by each normal point are considered as normal connected regions.

[0097] Simple single-point labeling may misclassify noise points as anomalous points (such as abrupt changes in the normal vector of an isolated point). However, by using "connected region aggregation" (defining the region formed by adjacent anomalous points as an anomalous connected region), isolated noise points can be filtered out. Real rock mass anomalies (such as fracture zones and breccia zones) often have spatial continuity, and their anomalous points will form continuous regions, while noise points are mostly isolated. Isolated anomalous points are classified as normal points.

[0098] In this embodiment, the process of obtaining the point cloud map of the corresponding scale shape offset includes:

[0099] Extract each normal connected region that contacts the abnormal connected region, and select the normal connected region with the largest area from the contacting normal connected regions as the reference connected region.

[0100] Calculate the distance from each anomalous point in the anomalous connected region to the plane equation of the corresponding reference connected region to obtain the shape distance;

[0101] The shape distance of each outlier point is normalized to obtain the shape offset of the outlier point, and the shape offset of the normal point is set to 0.

[0102] Arrange the shape offsets of the original point cloud maps belonging to the corresponding scale according to the position of the point cloud points to form a point cloud map with shape offsets of the corresponding scale.

[0103] This invention selects the "normal connected region that is in contact with the abnormal connected region and has the largest area" as the reference benchmark. This ensures the spatial correlation (contact) between the reference region and the abnormal region, and also ensures that the "largest area" can represent the mainstream normal bedding characteristics at this scale (in horizontally layered rock masses, the bedding surfaces of large normal regions are flatter, more continuous, and have higher plane fitting accuracy), thus avoiding the problem of spatial disconnect or insufficient representativeness between the reference region and the abnormal region.

[0104] The plane equation of the reference connected region is: Ax + By + Cz + D = 0, where A, B, and C are the components of the plane normal vector, D is the plane constant term, and x, y, and z represent the spatial coordinates of any point in three-dimensional space.

[0105] The formula for calculating shape distance is:

[0106] ;

[0107] Where, d S,i x is the shape distance of the i-th outlier. i y i z i Let be the coordinates of the i-th outlier point, and || be the absolute value operation.

[0108] In this embodiment, the process of obtaining the point cloud map of the corresponding scale reflection offset includes:

[0109] Extract each normal connected region that contacts the abnormal connected region, and select the normal connected region with the largest area from the contacting normal connected regions as the reference connected region.

[0110] The overall reflection value is obtained by averaging the reflection values ​​of each region in the reference connected region.

[0111] Calculate the distance from the reflection value of each abnormal point in the abnormal connected region to the overall reflection value to obtain the reflection distance;

[0112] The reflection distance is normalized to obtain the reflection offset of abnormal points, and the reflection offset of normal points is set to 0.

[0113] Arrange the reflection offsets of the original point cloud maps belonging to the corresponding scale according to the point cloud point positions to form a reflection offset point cloud map of the corresponding scale.

[0114] In horizontally layered rock masses, the reflectance values ​​of normal areas (intact rock strata) usually have strong consistency (e.g., the reflectance of homogeneous rock masses is stable), while the reflectance values ​​of abnormal areas (e.g., fracture zones, water-bearing bodies) will show significant differences due to factors such as porous material and changes in water content (e.g., the reflectance of fractured rock masses decreases, and the reflectance of water-bearing bodies changes abruptly); the magnitude of the reflection distance directly reflects the degree of this difference.

[0115] The formula for calculating the reflection distance is: d S,i =|R i -R ref,avg |, where d S,i R is the reflection distance of the i-th anomaly point. i Let R be the reflection value of the i-th anomaly point. ref,avg is the overall reflection value, and | is the absolute value operation.

[0116] like Figure 2 As shown, the process of generating a weighted point cloud map at the corresponding scale includes:

[0117] The first pointwise convolutional layer is used to map the tilt point cloud map pointwise to obtain the tilt feature point cloud map.

[0118] The shape offset point cloud is mapped point by point using a second pointwise convolutional layer to obtain the shape offset feature point cloud.

[0119] A third pointwise convolutional layer is used to map the reflection offset point cloud map pointwise to obtain the reflection offset feature point cloud map.

[0120] The tilt feature point cloud map, shape offset feature point cloud map, and reflection offset feature point cloud map of the same scale are added element by element to obtain the weighted point cloud map of the corresponding scale.

[0121] This invention employs point-by-point convolutional layers to perform feature mapping on point clouds of tilt, shape offset, and reflection offset, respectively. This allows for the extraction of deep representational information for each feature without disrupting the spatial structure of the point cloud. Furthermore, by fusing the three types of feature point clouds at the same scale element-wise, each feature complements and reinforces the others within the same spatial domain. Tilt features reflect geometric abrupt changes, shape offsets reflect structural distortions, and reflection offsets reflect differences in surface reflection characteristics. Weighted fusion of these multiple features can simultaneously capture structural and optical anomalies, making anomalous regions more prominent. Through the nonlinear mapping of point-by-point convolution, each feature exhibits higher response values ​​at anomalous points and near-zero response values ​​at normal points, thereby enhancing the weight of anomalous regions.

[0122] like Figure 3 As shown, the three-channel fusion network for layered rock mass anomaly identification includes: a first enhanced feature extraction channel, a second enhanced feature extraction channel, a third enhanced feature extraction channel, a first downsampling layer, a second downsampling layer, adder A1, adder A2, and an output unit;

[0123] The first input of the first enhanced feature extraction channel is used to input the high-scale original point cloud map, its second input is used to input the high-scale weighted point cloud map, and its output is connected to the input of the first downsampling layer.

[0124] The first input of the second enhanced feature extraction channel is used to input the original mesoscale point cloud map, and its second input is used to input the mesoscale weighted point cloud map.

[0125] The first input of the third enhanced feature extraction channel is used to input the low-scale original point cloud map, and its second input is used to input the low-scale weighted point cloud map.

[0126] The first input of adder A1 is connected to the output of the first downsampling layer, its second input is connected to the output of the second enhanced feature extraction channel, and its output is connected to the input of the second downsampling layer.

[0127] The first input of adder A2 is connected to the output of the second downsampling layer, its second input is connected to the output of the third enhanced feature extraction channel, and its output is connected to the input of the output unit.

[0128] The output end of the output unit serves as the output end of the three-channel fusion network for layered rock mass anomaly identification.

[0129] The high-scale enhanced feature extraction channel preserves local geometric details and can sensitively detect minute anomalies; the medium-scale and low-scale enhanced feature extraction channels capture structural deformation features over a larger range, reflecting overall layering and deformation trends. The hierarchical fusion of the three scales enables the network to possess both local and global perception capabilities.

[0130] This invention achieves spatial dimension matching of features at different scales through a first downsampling layer and a second downsampling layer, enabling high-scale features to be fused with medium- and low-scale features at the same semantic level.

[0131] like Figure 4 As shown, the first enhanced feature extraction channel, the second enhanced feature extraction channel, and the third enhanced feature extraction channel have the same structure, all including: a point cloud feature enhancement module, a first stacked convolution module, a second stacked convolution module, and a third stacked convolution module;

[0132] The first input of the point cloud feature enhancement module serves as the first input of the three enhanced feature extraction channels, the second input serves as the second input of the three enhanced feature extraction channels, and the output is connected to the input of the first stacked convolution module.

[0133] The input of the second stacked convolutional module is connected to the output of the first stacked convolutional module, and its output is connected to the input of the third stacked convolutional module.

[0134] The output of the third stacked convolutional module serves as the output of the three enhanced feature extraction channels.

[0135] The first and second input terminals can respectively connect to the original point cloud map and the weighted point cloud map (or anomaly features of different dimensions). The point cloud feature enhancement module fuses and enhances the dual input information, preserving the spatial structure information of the original point cloud while incorporating the guiding information of the weights, so that the subsequent convolutional modules can prioritize high-value feature regions (such as anomaly core regions). Multiple stacked convolutional modules gradually extract deep features.

[0136] like Figure 5 As shown, the point cloud feature enhancement module includes: a fourth pointwise convolutional layer and a multiplication layer;

[0137] The input of the fourth pointwise convolutional layer serves as the first input of the point cloud feature enhancement module;

[0138] The first input of the multiplication layer is connected to the output of the fourth pointwise convolutional layer, and its second input serves as the second input of the point cloud feature enhancement module.

[0139] The output of the multiplication layer serves as the output of the point cloud feature enhancement module.

[0140] The multiplication layer is used for element-wise multiplication.

[0141] The fourth pointwise convolutional layer maps the point cloud features from the first input point pointwise. The weighted point cloud map input from the second input point is multiplied pointwise with the features after pointwise convolution in the multiplication layer. This significantly enhances the features of high-weight regions (core anomaly regions) and effectively suppresses the features of low-weight regions (normal or noisy regions). This allows the network to automatically focus on the real anomaly features, avoids ineffective learning of irrelevant information, and significantly improves the accuracy and specificity of anomaly recognition.

[0142] like Figure 6 As shown, the first stacked convolutional module, the second stacked convolutional module, and the third stacked convolutional module have the same structure, each including a first convolutional layer, a second convolutional layer, and a third convolutional layer connected in sequence, and the convolutional kernel size of the first convolutional layer, the second convolutional layer, and the third convolutional layer is 3×3×3.

[0143] This invention divides the three-dimensional point cloud data of tunnel surrounding rock into multiple scales and calculates the tilt of the normal vector of the point cloud points at each scale. This allows it to adaptively capture the abnormal morphological change features at different spatial scales, thereby effectively improving the model's sensitivity to the identification of small abnormal areas.

[0144] like Figure 7As shown, the output unit consists of a fourth, fifth, and sixth convolutional layer connected in sequence, and a classification layer. The kernel size of the fourth and fifth convolutional layers is 3×3×3, and the kernel size of the sixth convolutional layer is 1×1×1. The sixth convolutional layer is used to reduce the number of channels to 1. The classification layer uses the Sigmoid activation function and outputs the anomaly probability of each feature point. A threshold is set (e.g., 0.5); when the anomaly probability exceeds the threshold, it is considered an anomalous region.

[0145] This invention uses three-dimensional indicators—abnormal tilt, shape offset, and reflection offset—to characterize the features of abnormal regions from different dimensions, preventing abnormal features from being submerged by the overall shape. It also generates a weighted cloud map based on multi-dimensional indicators, assigning higher weights to abnormal regions and allowing the network to focus on key areas, thus solving the problems of small abnormal proportions and difficulty in feature extraction.

[0146] Compared to traditional methods that rely on global feature aggregation, this invention calculates and fuses local offset features at multiple scale levels, making it less likely for abnormal regions to be masked by overall structural information, thus significantly improving the network's response capability to local abnormal regions.

[0147] This invention introduces a weighted point cloud map generated based on tilt, shape offset, and reflection offset during the network input stage. Weights are applied to the corresponding features of the original point cloud map at the same scale, enabling the network to capture key anomaly features at different scales. By combining key anomaly features at multiple scales, the anomaly region is obtained, thereby improving the accuracy of anomaly recognition.

[0148] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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 identifying anomalies in the surrounding rock of a tunnel excavated in a horizontally layered rock mass, characterized in that... include: Based on the three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass, three scales of original point cloud maps were obtained. Based on the normal vector of each point in the original point cloud map at each scale, calculate the abnormal tilt of each point to obtain the tilt point cloud map at the corresponding scale. Based on the abnormal tilt of the point cloud points, abnormal and normal connected regions are filtered out. Based on the differences in shape and reflectance values ​​between the abnormal and normal connected regions, corresponding scale shape offset point cloud maps and reflectance offset point cloud maps are obtained; where: The process of obtaining the shape offset point cloud map at the corresponding scale includes: extracting each normal connected region that contacts the abnormal connected region; selecting the normal connected region with the largest area from the contacting normal connected regions as the reference connected region; calculating the distance from each abnormal point in the abnormal connected region to the plane equation of the corresponding reference connected region to obtain the shape distance; normalizing the shape distance of each abnormal point to obtain the shape offset of the abnormal point; setting the shape offset of the normal points to 0; and constructing the shape offset point cloud map at the corresponding scale from the shape offsets of the original point cloud map at the corresponding scale. The process of obtaining the reflection offset point cloud map at the corresponding scale includes: extracting each normal connected region that contacts the abnormal connected region; selecting the normal connected region with the largest area from the contacting normal connected regions as the reference connected region; taking the average of each reflection value in the reference connected region to obtain the overall reflection value; calculating the distance from the reflection value of each abnormal point in the abnormal connected region to the overall reflection value to obtain the reflection distance; normalizing the reflection distance to obtain the reflection offset of the abnormal point, and setting the reflection offset of the normal point to 0; and constructing the reflection offset point cloud map at the corresponding scale from the reflection offsets of the original point cloud map at the corresponding scale. Based on the tilt point cloud map, shape offset point cloud map, and reflection offset point cloud map of the same scale, generate a weighted point cloud map of the corresponding scale. A three-channel fusion network for anomaly identification of layered rock masses is used to process the original point cloud maps at three scales. Weights are applied to the point cloud maps at the corresponding scales to obtain the anomaly regions.

2. The method for identifying anomalies in the surrounding rock of a horizontally layered rock mass tunnel as described in claim 1, characterized in that, The process of obtaining the original point cloud images at three scales includes: The three-dimensional point cloud map of the surrounding rock of the tunnel during the excavation of the horizontal layered rock mass is used as the high-scale original point cloud map; The high-scale original point cloud map is divided into multiple three-dimensional spaces. The center of the three-dimensional space is selected as the representative point. The representative points are arranged according to their positions in the three-dimensional spaces to obtain the medium-scale original point cloud map. The original point cloud map at the medium scale is divided into multiple three-dimensional spaces. The center of each three-dimensional space is selected as a representative point. The representative points are arranged according to their positions in each three-dimensional space to obtain the original point cloud map at the low scale.

3. The method for identifying anomalies in the surrounding rock of tunnels excavated in horizontally layered rock masses according to claim 1, characterized in that, The formula for calculating the abnormal tilt of each point in the point cloud is: ， Among them, Υ i Let n be the anomalous tilt of the i-th point in the original point cloud map at each scale. i Let n be the normal vector of the i-th point cloud point. i,j Let be the normal vector of the j-th point in the neighborhood of the i-th point in the point cloud, arccos be the angle between the two normal vectors, N be the number of point clouds in the neighborhood of the i-th point in the point cloud, and i and j be positive integers.

4. The method for identifying anomalies in the surrounding rock of a horizontally layered rock mass tunnel as described in claim 1, characterized in that, The process of filtering out abnormal and normal connected regions includes: Calculate the mean of the abnormal tilt of each point in the original point cloud map at each scale to obtain the overall tilt. Point cloud points with an abnormal tilt greater than the overall tilt are marked as abnormal points, and other points are marked as normal points; The connected regions formed by each outlier point are considered as outlier connected regions. The connected regions formed by each normal point are considered as normal connected regions.

5. The method for identifying anomalies in the surrounding rock of a horizontally layered rock mass tunnel as described in claim 1, characterized in that, The process of generating a weighted point cloud map at the corresponding scale includes: The first pointwise convolutional layer is used to map the tilt point cloud map pointwise to obtain the tilt feature point cloud map. The shape offset point cloud is mapped point by point using a second pointwise convolutional layer to obtain the shape offset feature point cloud. A third pointwise convolutional layer is used to map the reflection offset point cloud map pointwise to obtain the reflection offset feature point cloud map. The tilt feature point cloud map, shape offset feature point cloud map, and reflection offset feature point cloud map of the same scale are added element by element to obtain the weighted point cloud map of the corresponding scale.

6. The method for identifying anomalies in the surrounding rock of a tunnel excavated in a horizontally layered rock mass according to claim 1, characterized in that, The three-channel fusion network for layered rock mass anomaly identification includes: a first enhanced feature extraction channel, a second enhanced feature extraction channel, a third enhanced feature extraction channel, a first downsampling layer, a second downsampling layer, adder A1, adder A2, and an output unit; The first input of the first enhanced feature extraction channel is used to input the high-scale original point cloud map, its second input is used to input the high-scale weighted point cloud map, and its output is connected to the input of the first downsampling layer. The first input of the second enhanced feature extraction channel is used to input the original mesoscale point cloud map, and its second input is used to input the mesoscale weighted point cloud map. The first input of the third enhanced feature extraction channel is used to input the low-scale original point cloud map, and its second input is used to input the low-scale weighted point cloud map. The first input of adder A1 is connected to the output of the first downsampling layer, its second input is connected to the output of the second enhanced feature extraction channel, and its output is connected to the input of the second downsampling layer. The first input of adder A2 is connected to the output of the second downsampling layer, its second input is connected to the output of the third enhanced feature extraction channel, and its output is connected to the input of the output unit. The output end of the output unit serves as the output end of the three-channel fusion network for layered rock mass anomaly identification.

7. The method for identifying anomalies in the surrounding rock of a horizontally layered rock mass tunnel as described in claim 6, characterized in that, The first enhanced feature extraction channel, the second enhanced feature extraction channel, and the third enhanced feature extraction channel have the same structure, all including: a point cloud feature enhancement module, a first stacked convolution module, a second stacked convolution module, and a third stacked convolution module; The first input of the point cloud feature enhancement module serves as the first input of the three enhanced feature extraction channels, the second input serves as the second input of the three enhanced feature extraction channels, and the output is connected to the input of the first stacked convolution module. The input of the second stacked convolutional module is connected to the output of the first stacked convolutional module, and its output is connected to the input of the third stacked convolutional module. The output of the third stacked convolutional module serves as the output of the three enhanced feature extraction channels.

8. The method for identifying anomalies in the surrounding rock of a horizontally layered rock mass tunnel as described in claim 7, characterized in that, The point cloud feature enhancement module includes: a fourth pointwise convolutional layer and a multiplication layer; The input of the fourth pointwise convolutional layer serves as the first input of the point cloud feature enhancement module; The first input of the multiplication layer is connected to the output of the fourth pointwise convolutional layer, and its second input serves as the second input of the point cloud feature enhancement module. The output of the multiplication layer serves as the output of the point cloud feature enhancement module.

Images