Rock mass structural surface recognition and 3D network modeling method based on point cloud deep learning

Abstract

The present invention discloses a rock structure surface recognition and three-dimensional network modeling method based on point cloud deep learning, which belongs to the technical field of rock structure surface recognition. The method comprises the following steps: S1, obtaining fine three-dimensional point cloud data of the rock structure of the rock slope; S2, extracting point cloud feature data from the fine three-dimensional point cloud data using an improved OCM-PointNet2 deep learning model, segmenting the rock structure surface based on the point cloud feature data, obtaining segmentation data, and enhancing the segmentation data; extracting geometric parameters of the rock structure surface based on the enhanced segmentation data; S3, calculating the probability function of the geometric parameters of the structure surface, predicting the spatial distribution of the structure surface based on the probability function, and constructing a three-dimensional network model; The present invention is helpful to promote the research on the stability of engineering slopes, and help solve the problems of safety prediction and control of rock slopes during construction and operation and maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of rock mass structural surface recognition, and specifically relates to a rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning. Background Art

[0002] Engineering practice shows that a large number of weak structural surfaces in unstable or nearly unstable rock slopes provide geometric boundaries and mechanical parameter conditions for slope failure. The geometric characteristics of the structural surfaces (thickness, occurrence, contact area, etc.), the mechanical characteristics of the fillings (cementation, strength, etc.), and the spatial combination characteristics play a vital role in the deformation and failure mechanism of the slope rock mass. In the structural analysis of the slope rock mass, it is very important to conduct a detailed analysis and characterization of these structural surfaces. Therefore, a structural surface identification method suitable for high and steep slopes under complex geological conditions is proposed, which helps to clarify the spatial effect of the structural surface and predict the distribution characteristics of the structural surface inside the slope. It is the basis for conducting engineering slope stability research and is a key link in solving problems such as safety prediction and control during the construction and operation and maintenance period of rock slopes. Summary of the invention

[0003] The purpose of the present invention is to propose a method which is simple to use, scientific and effective and adaptable to engineering needs for rock mass structural surface identification and three-dimensional network model construction.

[0004] To achieve the above object, the present invention provides the following scheme: a rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning, comprising the following steps:

[0005] S1. Obtain fine 3D point cloud data of rock slope rock mass structure;

[0006] S2. Using an improved OCM-PointNet2 deep learning model to extract point cloud feature data from the refined three-dimensional point cloud data, segmenting the rock mass structural surface based on the point cloud feature data to obtain segmentation data, and enhancing the segmentation data; extracting geometric parameters of the rock mass structural surface based on the enhanced segmentation data;

[0007] The improved OCM-PointNet2 deep learning model includes: a point cloud feature fusion module, a point cloud segmentation module and a point cloud enhancement module;

[0008] The point cloud feature fusion module is used to extract the point cloud feature data;

[0009] The point cloud segmentation module is used to segment the rock mass structure surface to obtain the segmentation data; the point cloud segmentation module includes a sampling layer, a grouping layer, a feature extraction layer, a feature propagation layer and a fully connected layer;

[0010] The point cloud enhancement module is used to enhance the segmentation data based on the optimal color mapping;

[0011] S3, calculating the probability function of the geometric parameters of the structural surface, predicting the spatial distribution of the structural surface based on the probability function, and constructing a three-dimensional network model;

[0012] The probability function includes: the distribution characteristics of the occurrence of the rock mass structural surface, the distribution characteristics of the trace length of the rock mass structural surface, the density distribution characteristics of the rock mass structural surface and the multi-parameter joint probability density of the rock mass structural surface.

[0013] Further preferably, S1 comprises the following steps:

[0014] S11. Interpret the structural surface distribution characteristics in the remote sensing images of the study area and obtain the geometric images of the structural surfaces of levels I-III;

[0015] S12, obtaining multi-level excavation data of the structural surface, identifying the type of the structural surface, and obtaining measured structural surface development data;

[0016] S13, performing oblique photogrammetry on the study area to obtain point cloud data and color attributes corresponding to the point cloud data and surface texture, and then obtaining a three-dimensional point cloud model of the study area;

[0017] S14. Based on the structural surface geometric image, the structural surface development data and the three-dimensional model, the photography blind area and the image blur and distortion area are determined, and a multi-angle close measurement method is adopted to obtain the panoramic high-resolution fine three-dimensional point cloud data.

[0018] Further preferably, the point cloud feature data in S2 includes: directional features, geometric features and optical features;

[0019] The directional features include: normal vectors in the three directions of X, Y, and Z N x , N y , N z ;

[0020] The geometric characteristics include: roughness, curvature, anisotropy, planarity, linearity, omnidirectional variability, and perpendicularity;

[0021] The optical characteristics include: R, G, B color characteristics.

[0022] Further preferably, the method of obtaining the point cloud feature data by using the point cloud feature fusion module includes:

[0023] Perform feature extraction on the input point cloud data to obtain multi-dimensional features;

[0024] splicing the multi-dimensional features with the coordinates of the original point cloud data to obtain high-dimensional point cloud feature data;

[0025] Slope structural surface identification is performed based on the high-dimensional point cloud feature data.

[0026] Further preferably, S3 comprises the following steps:

[0027] S31, extracting the occurrence, trace length and spacing information of the structural surface;

[0028] S32, calculating the probability function of rock mass structural surface parameters;

[0029] S33, performing Monte Carlo simulation on the probability function to obtain simulated structural surface parameters; comparing the simulated structural surface parameters with the actually obtained structural surface parameters, and taking the model with the most similar structural surface parameters as the three-dimensional network model.

[0030] Further preferably, the occurrence distribution characteristics of the rock mass structural surface include:

[0031] ,

[0032] ,

[0033] In the formula, W i Indicates i The weight of the occurrence of the strip structure surface; w Indicates the sampling window width; h Indicates the sampling window length; l i Indicates i The structural surface is long; α i Indicates i The tendency of a structural surface; β i Indicates i Structural surface inclination angle; α r Represents the sampling window w The tendency of the side; n Indicates the number of structural surfaces; Af i Represents the frequency of the structural surface.

[0034] Further preferably, the rock mass structural surface trace length distribution characteristics include:

[0035] ,

[0036] In the formula, m Indicates the number of survey lines; C m Representation andm The ratio of the number of structural planes intersected by the survey lines to the expected number; w Indicates the sampling window width; Indicates the sampling window length; d i Indicates the measuring line d Length; E Indicates expected value; α i Indicates i The tendency of a structural surface; i Representation of structural surface and sampling window w The apparent inclination of the edge.

[0037] Further preferably, the rock mass structural surface density distribution characteristics include:

[0038] ,

[0039] In the formula, E ( D 2 ) represents the second-order moment of the trace length distribution of the structural surface; Represents the normal line density of the structural surface.

[0040] Further preferably, the multi-parameter joint probability density of the rock mass structural surface includes:

[0041] ,

[0042] In the formula, It represents the probability density of structural surface tendency; represents the probability density of the structural surface inclination; Represents the probability density of the structural surface trace length; Represents the probability density of the structural surface spacing.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] Existing technologies often rely on limited artificial features for structural surface identification, which cannot capture local features and are difficult to apply to complex and large-scale scenarios. The slope rock structural surface identification and three-dimensional network model construction method proposed in the present invention, on the one hand, no longer relies on the calculation of artificial features, can learn rich high-dimensional features, and can more finely characterize local features. And based on high-dimensional features, the trained deep learning network is used to segment the point cloud, which improves the accuracy and efficiency of rock structural surface identification; on the other hand, it can help clarify the spatial effect of the structural surface, predict the distribution characteristics of the structural surface inside the slope, and realize the construction of a three-dimensional network model of the rock mass. The present invention helps to promote the research on the stability of engineering slopes and help solve problems such as safety prediction and control of rock slopes during construction and operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0046] Figure 1 It is a schematic diagram of the flow of a rock structure surface recognition and three-dimensional network modeling method based on point cloud deep learning according to an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the structure of a deep learning model for rock mass structural plane identification according to an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of the process of constructing a three-dimensional model of a structural surface according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Embodiment 1:

[0052] like Figure 1 As shown, this embodiment provides a rock structure surface recognition and three-dimensional network modeling method based on point cloud deep learning, including the following steps:

[0053] S1. Obtain fine 3D point cloud data of rock slope rock mass structure.

[0054] This embodiment uses digital image technology to combine remote sensing interpretation, UAV oblique photogrammetry and multi-angle close aerial photogrammetry, and conducts a detailed survey of slopes at all levels in the study area to obtain three-dimensional dense point clouds and high-precision image data of the study area, thereby compensating for the photographic blind spots and image blur and distortion problems existing in existing measurement technologies.

[0055] Specifically, S1 includes the following steps:

[0056] S11. Remote sensing interpretation: Interpret the structural surface distribution characteristics in the remote sensing images of the study area to obtain the geometric images of level I-III structural surfaces.

[0057] Using optical imaging, InSAR, geophysical detection and other means, we can identify and interpret the distribution characteristics of I-III structural surfaces in low-resolution or greater than meter-level data image areas, and clarify the structural system of the study area. We can refine centimeter-level image data through remote sensing image enhancement and other means to draw geometric images of I-III structural surfaces.

[0058] S12. Detailed investigation of the study area: obtain multi-level excavation data of the structural surface, and obtain the structural surface development data through field observation and actual measurement.

[0059] Carry out intensive investigation of engineering rock slope structural surfaces and observation after multi-level excavation, and record detailed descriptions of their occurrence, trace length, spacing and other information.

[0060] S13. Oblique photogrammetry: Oblique photogrammetry is performed on the study area to obtain point cloud data and color attributes corresponding to the point cloud data and surface texture, and then a three-dimensional point cloud model of the study area is obtained.

[0061] The specific process mainly includes: flight route planning, aerial photography, control point arrangement and process modeling. First of all, the flight route planning mainly takes into account the scope of the survey area, topography, lens performance and scale requirements. In order to meet the data resolution requirements, the heading overlap ratio of the aerial photography drone is greater than 80%, and the lateral overlap ratio is greater than 60%; after completing the aerial photography, select the characteristic positions on the film as the film control points, and measure them on the spot to obtain the coordinates of the selected points; finally, the motion recovery structure algorithm technology is used in the model production process. The specific steps mainly include: image data reading, matching image pair search, incremental multi-view reconstruction and three-dimensional point cloud reconstruction. This technology can process data more quickly, and the acquired point cloud data has color properties close to the surface texture of the object.

[0062] S14. Multi-angle close-up photogrammetry: Based on the geometric image of the structural surface, the structural surface development data and the three-dimensional model, the photography blind area and the image blur and distortion area are determined, and the multi-angle close-up measurement method is used to obtain panoramic high-resolution fine three-dimensional point cloud data.

[0063] According to the terrain features and rock structure surface development obtained in steps S11 and S12, the multi-angle shooting route of the UAV is planned, and the three-dimensional model of the study area established in step S13 is checked to find the existing blind spots in oblique photography and image blur and distortion. For the key parts of the complex areas of high and steep slopes photographed, the shooting angle is designed according to the actual situation, and multi-angle close-up photogrammetry is carried out to finally obtain panoramic high-resolution and fine three-dimensional point cloud data.

[0064] S2. Use the improved OCM-Pointnet2 deep learning model to extract point cloud feature data from the 3D point cloud data, perform rock mass structural plane segmentation based on the point cloud feature data to obtain segmentation data, and perform enhancement processing on the segmentation data based on point cloud enhancement, and extract the geometric parameters of the rock mass structural plane based on the enhanced segmentation data.

[0065] As Figure 2 shown, in this embodiment, the improved OCM-Pointnet2 deep learning model includes: a point cloud feature fusion module, a point cloud segmentation module, and a point cloud enhancement module. Among them, the high-dimensional point cloud features are extracted through the point cloud feature fusion module to increase the feature dimension in the input point cloud data; the point cloud segmentation module is used for rock mass structural plane segmentation, and the point cloud segmentation module includes a sampling layer, a grouping layer, a feature extraction layer, a feature propagation layer, and a fully connected layer; point cloud enhancement is realized based on the optimal color mapping (OCM) to optimize the boundaries or details of the displayed segmentation results and improve the visualization effect of the segmentation results. Use 3D point cloud data and high-precision images to label the rock mass structural planes in the study area slope to create training samples, and divide the training samples into a training set and a test set. Perform cyclic training on the improved OCM-Pointnet2 deep learning model, and its process includes forward propagation, loss calculation, backpropagation, and parameter update. The maximum number of allowable iterations set during model training is determined according to the amount of training data. Set the network generalization ability detection standard. When the error of the continuous multiple training validation sets no longer decreases, terminate the training task to prevent overfitting. The model verification selects accuracy, precision, recall rate, and F1 value as evaluation indicators.

[0066] Specifically, the point cloud feature fusion module is used to extract high-dimensional point cloud features, introduce point cloud direction features, geometric features, and optical features, and form a higher-dimensional input feature vector by splicing these features with the original XYZ coordinate information, and input them into the model together for multi-channel feature learning to obtain a richer local feature representation, thereby enhancing the local feature expression and improving the recognition accuracy.

[0067] Specifically, the point cloud segmentation module is used for rock mass structural plane segmentation. Its sampling layer is used to select a part of the points from the original point cloud to represent the entire point cloud. This helps to reduce the computational amount and more effectively capture global features during downstream processing. The input is the original point cloud data, usually an N×3 matrix, where N is the number of points and 3 represents the three-dimensional coordinates (x, y, z) of each point. The output is a sampled point set with a size of M×3, where M < N. The grouping layer is used to group the sampled points according to certain rules (such as radius or fixed number). Each point group contains several points in the original point cloud, and these points are used for local feature extraction. The input is the sampled point set M×3 and the original point cloud N×3. The output is a group of points around each sampled point, usually represented as M×K×3, whereK is the number of points around each sample point. The feature extraction layer is used to extract local geometric features from the grouped point cloud. This layer usually includes a series of multi-layer perceptrons (MLP) and maximum pooling operations. The input is the grouped point set M×K×3. The output is the local feature of each sample point, with a size of M×D, where D is the feature dimension. The feature propagation layer is used to propagate local features to the global one. It uses upsampling methods (such as interpolation) to propagate features from fewer sampled points back to the original point cloud or a higher resolution point cloud. The input is the local feature M×D and the coordinates of the point set to which the feature needs to be propagated. The output is the global feature of the original point cloud, which is N×D in size. The fully connected layer is used to further process the global features, usually for classification or regression tasks. This layer maps the features to the target output space. The input is the global feature N×D. The output is the final prediction result, and the specific size depends on the task. For example, for a classification task, the output may be N×C, where C is the number of categories. The design and functionality of each layer enables the deep learning model to effectively process complex point cloud data and capture local and global features, thus performing well in the task of rock structure surface recognition.

[0068] Specifically, the point cloud enhancement module optimizes the display of the boundaries or details of the segmentation results through optimal color mapping (OCM), making the visualization of the segmentation results more intuitive and informative. P ={ p i}, its RGB value R i , G i and B i ; For each point p i , use the standard RGB to HSV conversion formula to convert the RGB value to the HSV color space to get the HSV color value H i , S i and V i ; Adjust H (hue), S (saturation) and V (brightness) in the HSV value to optimize the color display effect of the point cloud, and make feedback adjustments based on the observation results to ensure that the boundaries and details are clearly displayed, and obtain the optimized HSV color value.

[0069] In this embodiment, the point cloud feature data includes: directional features, geometric features and optical features; directional features include: normal vectors in the three directions of X, Y and Z N x , N y, N z ; Geometric features include: roughness, curvature, anisotropy, planarity, linearity, edge points and corner points; optical features include: R, G, B color features.

[0070] Among them, the method of obtaining point cloud feature data by the OCM-Pointnet2 deep learning model includes:

[0071] Feature extraction is performed on the input point cloud data to obtain multi-dimensional features.

[0072] Specifically, for the input point cloud data P ={ p i},in, p i Represents the coordinates of the point; for each point p i , find its neighborhood point set N { p i}; Use PCA to calculate N { p i}, and obtain its eigenvector; the eigenvector with the smallest eigenvalue is the normal vector N i .

[0073] For input point cloud data P ={ p i}, its neighborhood point set N { p i}, calculate the sum of square errors of the domain point to the plane fitting of each point; normalize the sum of square errors to obtain the roughness feature R i .

[0074] For input point cloud data P ={ p i}, its neighborhood point set N { p i}, using PCA calculation N { p i}, and obtain the eigenvalues l 1. l 2 and l 3. Calculate curvature Get the curvature .

[0075] For input point cloud data P ={p i}, its neighborhood point set N { p i}, using PCA calculation N { p i}, and obtain the eigenvalues l 1. l 2 and l 3. Calculate anisotropic characteristics: A i = l 1- l 3 l 1. Get anisotropy A i .

[0076] For input point cloud data P ={ p i}, its neighborhood point set N { p i}, using PCA calculation N { p i}, and obtain the eigenvalues l 1. l 2 and l 3. Calculate linear features: P i =( l 2- l 3) / l 1. Obtain planarity features P i .

[0077] For input point cloud data P ={ p i}, its normal vector N i , for each point p i , calculate the angle between its normal vector and the normal vector of the neighboring point; if the angle is greater than a certain set threshold, then p i is the edge point, and the output edge point set O={o i};

[0078] For input point cloud data P ={ p i}, its normal vector N i , for each point p i, calculate its curvature; if the curvature is greater than a certain set threshold, then p i is the corner point, output the corner point set C ={ c i};

[0079] The coordinates of the multi-dimensional features and the original point cloud data are spliced ​​to obtain the point cloud feature data F i : F i =[ x i , y i , z i , N ix , N iy , N iz , k i , R i , A i , P i , O i , C i , R i , G i , B i ].

[0080] The point cloud feature data is dimensionally upgraded to obtain a high-dimensional point cloud feature representation.

[0081] This embodiment uses a multi-layer perceptron (MLP) to perform feature dimensionality increase. Each layer performs linear transformation and nonlinear activation function to map the input features to a higher dimensional space. The specific process is as follows:

[0082] First layer MLP: The input feature dimension is D, and the goal is to increase it to 64. Input: 16-dimensional features. Weight matrix: W 1∈ R D×64 . Bias vector: b 1∈ R 64 . Linear transformation: h1= W 1 x + b 1. Activation function: a1=ReLU(h1). Output: 64-dimensional features.

[0083] The second layer of MLP: Upgrade the 64-dimensional features to 128. Input: 64-dimensional features. Weight matrix: W 1∈ R 64×128 . Bias vector: b 2∈ R 128 . Linear transformation: h2= W 2 a 1+ b 2. Activation function:

[0084] a 2=ReLU(h2). Output: 128-dimensional features.

[0085] Continue in this way and gradually increase the feature dimension to 256, 512, 1024, etc. Through the gradual dimension increase of multi-layer MLP, a high-dimensional feature representation can be finally obtained. Slope structural surface recognition is performed based on high-dimensional geometric features.

[0086] In this embodiment, the cross entropy loss function is used to process the rock mass structural surface classification task. The cross entropy loss function is defined as:

[0087] ,

[0088] In the formula, N represents the total number of samples; C Indicates the total number of categories; y i,c Representation sample i The true label of (a one-hot encoded vector where the position corresponding to the true category is 1 and the rest of the positions are 0); Samples predicted by the model i Belongs to category c probability.

[0089] The model uses the AdamW optimizer to separate weight decay and gradient update, making weight decay independent of gradient update and avoiding weight decay errors. At the same time, by correcting weight decay, the generalization ability of the model is improved to avoid overfitting.

[0090] S3. Calculate the probability function of the geometric parameters of the structural surface; predict the spatial distribution of the structural surface based on the probability function, and construct a three-dimensional network model.

[0091] Specifically, Figure 3 As shown, S3 includes the following steps:

[0092] S31. After the point cloud data is classified and segmented, the individual structural surfaces are clustered, the plane normal vectors are fitted, and the geometric features of the structural surfaces are extracted. The geometric features include: surface occurrence, trace length, and spacing information.

[0093] S32. Analyze the geometric features of the acquired structural surface using probability statistics methods, determine the probability distribution characteristics of each element of the structural surface, and calculate the probability function of the structural surface parameters.

[0094] In this embodiment, the probability function includes: the distribution characteristics of the occurrence of the rock mass structural surface, the distribution characteristics of the trace length of the rock mass structural surface, the distribution characteristics of the density of the rock mass structural surface, and the multi-parameter joint probability density of the rock mass structural surface.

[0095] Among them, the distribution characteristics of rock mass structural surface include:

[0096] ,

[0097] ,

[0098] In the formula, W i Indicates i The weight of the occurrence of the strip structure surface; w Indicates the sampling window width; h Indicates the sampling window length; l i Indicates i The structural surface is long; α i Indicates i The tendency of a structural surface; β i Indicates i Structural surface inclination angle; α r Represents the sampling window w The tendency of the side; n Indicates the number of structural surfaces; Af i Represents the frequency of the structural surface.

[0099] The trace length distribution characteristics of rock mass structural surface include:

[0100] ,

[0101] In the formula, m Indicates the number of survey lines; C m Representation and m The ratio of the number of structural planes intersected by the survey lines to the expected number; w Indicates the sampling window width; Indicates the sampling window length; d iIndicates the measuring line d Length; E Indicates expected value; α i Indicates i The tendency of a structural surface; i Representation of structural surface and sampling window w The apparent inclination of the edge.

[0102] The characteristics of rock mass structural surface density distribution include:

[0103] ,

[0104] In the formula, E ( D 2 ) represents the second-order moment of the trace length distribution of the structural surface; Represents the normal line density of the structural surface.

[0105] The multi-parameter joint probability density of rock mass structural surface includes:

[0106] ,

[0107] In the formula, It represents the probability density of structural surface tendency; represents the probability density of the structural surface inclination; Represents the probability density of the structural surface trace length; Represents the probability density of the structural surface spacing.

[0108] S33, performing Monte Carlo simulation on the probability function to obtain simulated structural surface parameters; comparing the simulated structural surface parameters with the actually obtained structural surface parameters, and taking the model with the most similar structural surface parameters as the three-dimensional network model.

[0109] According to the above method, the vector method is used to analyze the occurrence distribution of the structural surface, the trace length distribution of the structural surface is determined by using the corrected trace length distribution, and the probability distribution characteristics of each element of the structural surface are determined by the conditional probability density distribution. Based on the probability distribution model of the inclination, dip, trace length and spacing of the structural surface, the Monte Carlo method is used to simulate, and the simulated structural surface parameters are compared with the structural surface parameters obtained from the interpretation of the real-life three-dimensional model, and the model with the most similar parameters between the two is selected as the final three-dimensional network model.

[0110] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning, characterized in that: The following steps are involved: S1. Obtain fine 3D point cloud data of rock slope rock mass structure; S2. Using an improved OCM-PointNet2 deep learning model to extract point cloud feature data from the refined three-dimensional point cloud data, performing rock mass structural surface segmentation based on the point cloud feature data to obtain segmentation data, and performing enhancement processing on the segmentation data; Extracting geometric parameters of rock mass structural surfaces based on the enhanced segmented data; The improved OCM-PointNet2 deep learning model includes: a point cloud feature fusion module, a point cloud segmentation module and a point cloud enhancement module; The point cloud feature fusion module is used to extract the point cloud feature data, including: extracting features from the input point cloud data to obtain multi-dimensional features; splicing the multi-dimensional features with the coordinates of the original point cloud data to obtain high-dimensional point cloud feature data; and identifying the slope structure surface based on the high-dimensional point cloud feature data; The point cloud segmentation module is used to segment the rock mass structure surface to obtain the segmentation data; the point cloud segmentation module includes a sampling layer, a grouping layer, a feature extraction layer, a feature propagation layer and a fully connected layer; The point cloud enhancement module is used to enhance the segmentation data based on the optimal color mapping; S3, calculating the probability function of the geometric parameters of the structural surface, predicting the spatial distribution of the structural surface based on the probability function, and constructing a three-dimensional network model; The probability function includes: the distribution characteristics of the occurrence of the rock mass structural surface, the distribution characteristics of the trace length of the rock mass structural surface, the density distribution characteristics of the rock mass structural surface and the multi-parameter joint probability density of the rock mass structural surface.

2. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: S1 includes the following steps: S11. Interpret the structural surface distribution characteristics in the remote sensing images of the study area and obtain the geometric images of the structural surfaces of levels I-III; S12, obtaining multi-level excavation data of the structural surface, identifying the type of the structural surface, and obtaining measured structural surface development data; S13, performing oblique photogrammetry on the study area to obtain point cloud data and color attributes corresponding to the point cloud data and surface texture, and then obtaining a three-dimensional point cloud model of the study area; S14. Based on the structural surface geometric image, the structural surface development data and the three-dimensional point cloud model, the photography blind area and the image blur and distortion area are determined, and a multi-angle close measurement method is adopted to obtain the panoramic high-resolution fine three-dimensional point cloud data.

3. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: The point cloud feature data in S2 includes: directional features, geometric features and optical features; The directional features include: normal vectors N in the three directions of X, Y, and Z x 、N y 、N z ; The geometric characteristics include: roughness, curvature, anisotropy, planarity, linearity, omnidirectional variability, and perpendicularity; The optical characteristics include: R, G, B color characteristics.

4. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: S3 includes the following steps: S31, extracting the occurrence, trace length and spacing information of the structural surface; S32, calculating the probability function of rock mass structural surface parameters; S33, performing Monte Carlo simulation on the probability function to obtain simulated structural surface parameters; comparing the simulated structural surface parameters with the actually obtained structural surface parameters, and taking the model with the most similar structural surface parameters as the three-dimensional network model.

5. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: The occurrence distribution characteristics of the rock mass structural surface include: ， ， Where W i represents the weight of the occurrence of the i-th structural surface; w represents the sampling window width; h represents the sampling window length; l i represents the length of the i-th structural surface; α i represents the inclination of the i-th structural surface; β i represents the inclination angle of the i-th structural surface; α r represents the inclination of the sampling window w edge; n represents the number of structural surfaces; Af i Represents the frequency of the structural surface.

6. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: The rock mass structural surface trace length distribution characteristics include: ， In the formula, m represents the number of survey lines; C m It represents the ratio of the number of structural surfaces intersecting m survey lines to the expected number; w represents the sampling window width; Indicates the sampling window length; d i represents the length of the measuring line d; E represents the expected value; α i represents the inclination of the i-th structural surface; θ represents the apparent inclination angle between the structural surface and the w edge of the sampling window.

7. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: The rock mass structural surface density distribution characteristics include: ， In the formula, E(D 2 ) represents the second-order moment of the trace length distribution of the structural surface; Represents the normal line density of the structural surface.

8. The rock mass structural surface recognition and three-dimensional network modeling method based on point cloud deep learning according to claim 1 is characterized in that: The multi-parameter joint probability density of the rock mass structural surface includes: ， In the formula, It represents the probability density of structural surface tendency; represents the probability density of the structural surface inclination; Represents the probability density of the structural surface trace length; Represents the probability density of the structural surface spacing.

Images