Tunnel surrounding rock structural surface morphology real-time prediction method based on graph attention network
Through the method based on graph attention network, combined with Laplace interpolation, statistical filtering and Gaussian smoothing filtering, the problem of morphology data of the area to be excavated in the existing technology is solved, and the morphology of the surrounding rock structure of the tunnel is realized The high-precision real-time prediction improves the accuracy of rock mass strength judgment and support scheme.
Patent Information
- Application Number
- CN202410925307.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-15
AI Technical Summary
The existing method for predicting the surface morphology data of surrounding tunnel rock structures has shortcomings in data utilization and real-time updates, especially ignoring the morphology data of the area to be excavated, which leads to difficulties in determining the rock mass strength and supporting scheme.
Using a graph attention network method, the point cloud data of the excavated area is obtained through three-dimensional laser scanning, and the point cloud data is aligned, denoised and smoothed, and converted into a graph structure to construct a graph attention network for the excavated area. Forecasting of morphology data, combined with Laplace interpolation, statistical filtering and Gaussian smoothing filtering, the Poisson surface reconstruction method was finally grid-based and visualized.
Improve the accuracy and reliability of the three-dimensional roughness calculation of structural surfaces, enhance the real-time prediction ability of rock structural surface morphology data, reduce the calculation cost, adapt to complex and irregular rock structural surfaces, and meet the needs of different working conditions. Forecast demand in real time.
Smart Images

Figure CN118960603A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tunnel engineering technology, specifically relating to a real-time prediction method for the structural surface morphology of tunnel surrounding rock based on graph attention networks. Background Technology
[0002] In recent years, tunnel construction in my country has gradually expanded to areas with complex surrounding rock geological conditions. To mitigate the impact of potential engineering accidents such as rock mass deformation and collapse on tunnel construction safety, reasonable excavation and support measures are necessary. The three-dimensional roughness of the rock mass's structural surfaces is a key indicator for evaluating rock mass strength and influences the selection of excavation and support schemes. By predicting the three-dimensional morphology data of the structural surfaces, engineers can accurately calculate the three-dimensional roughness of the structural surfaces in the area to be excavated, thereby more accurately assessing rock mass strength and developing more reasonable construction and support plans. Therefore, predicting the morphology of the structural surfaces in the tunnel excavation area and determining their three-dimensional roughness is crucial for determining the support scheme and ensuring construction safety.
[0003] Existing methods for predicting and visualizing structural surface morphology data are mostly based on automatic or semi-automatic characterization of excavated structural surface morphology data. These methods provide different solutions for calculating the three-dimensional roughness of structural surfaces, but they have some shortcomings in terms of data utilization and prediction efficiency, specifically:
[0004] 1. Limitations in data utilization: Existing methods focus on the morphological data of the rock mass structure surface in the excavated part when calculating the surface roughness, while ignoring the morphological data of the part to be excavated, thus affecting the judgment of the rock mass strength in the area to be excavated and the determination of the rock mass support scheme.
[0005] 2. Lack of real-time update capability: Existing methods typically calculate roughness directly for the target structural surface of the tunnel. Because current methods for calculating the roughness of the target structural surface lack in-depth consideration of the relationship between the structural surface morphology data of the excavated and unexcavated areas, the cost of calculating the roughness of a single structural surface is too high. Therefore, dynamically utilizing the structural surface morphology data of the excavated area and applying it to predict the structural surface morphology data of the unexcavated area has significant practical implications.
[0006] Therefore, it is necessary to study a new method that can predict the morphology of the surrounding rock structure in tunnels in real time to guide actual engineering projects. Summary of the Invention
[0007] To address the aforementioned issues, this invention discloses a real-time prediction method for the structural surface morphology of tunnel surrounding rock based on graph attention networks. This method predicts and generates complete morphology data of the structural surface in the area to be excavated based on the point cloud data of the structural surface in the already excavated area, thereby providing an accurate basis for the prediction of the three-dimensional roughness of the structural surface.
[0008] To achieve the above objectives, the technical solution of the present invention is as follows:
[0009] A real-time prediction method for the structural surface morphology of tunnel surrounding rock based on graph attention networks includes the following steps:
[0010] (1) Use a 3D laser scanner to scan the structural surfaces of the excavated area of the tunnel and obtain its point cloud data;
[0011] (2) Align, denoise, and smooth the above point cloud data;
[0012] (3) The processed point cloud data is converted into a graph structure and a graph attention network is constructed to predict the morphology data of the structure to be excavated.
[0013] (4) The above morphological data were optimized using a model ensemble method;
[0014] The optimized morphology data was meshed and visualized using the Poisson surface reconstruction method.
[0015] Furthermore, the point cloud data obtained in step (1) is subjected to three-dimensional coordinate alignment using the Laplacian interpolation method, and its objective equation is:
[0016] Lu = F
[0017] Where L is the graph Laplacian matrix, representing the similarity and connectivity between points in the point cloud; u is the point cloud value vector after interpolation, representing the value to be solved; and F is the vector of the original point cloud data, representing the known value.
[0018] Furthermore, the expression for constructing the Laplace matrix is:
[0019]
[0020] Among them, L ij This represents the element in the i-th row and j-th column of the graph Laplacian matrix, which represents the similarity and connectivity between points in the point cloud.
[0021] Furthermore, statistical filtering is used for point cloud denoising, and its expression is:
[0022] d i >μ+ασ
[0023] Where, d i The coordinate vector p of the midpoint of the point cloud i The mean nearest neighbor distance is μ; μ is the mean nearest neighbor distance of all points; α is a custom threshold parameter used to control the criteria for identifying outliers, with common values ranging from 1 to 3; σ is the standard deviation of the mean nearest neighbor distance of all points.
[0024] Furthermore, the point cloud coordinate vector p i Average nearest neighbor distance d i The calculation method is as follows:
[0025]
[0026] Where, p i p j All are point coordinate vectors in the point cloud, and k is the number of nearest neighbors, which can be 1 to 5 for high-density point clouds.
[0027] Furthermore, Gaussian smoothing filtering is used for point cloud smoothing, and its expression is:
[0028]
[0029] Where, p i ′ represents point p after Gaussian smoothing. i coordinates; w ij Let p be the weight, representing the point p. j For point p i The degree of influence, where n represents the total number of points in the point cloud.
[0030] Furthermore, the point cloud data can be converted into a graph structure by constructing an adjacency matrix.
[0031] Furthermore, the output of the graph attention layer in the graph attention network is represented as follows:
[0032]
[0033] Among them, h i ′ represents the output feature of node i; h j σ represents the input features of node j; σ is the nonlinear activation function. Let be the set of neighboring nodes of node i, W be a trainable weight matrix, and α be the set of neighboring nodes of node i. ij Let be the attention weight between node i and node j.
[0034] Furthermore, the output layer predicted by the graph attention network is represented as follows:
[0035] Z = HW o +b o
[0036] Where Z is the network output; H is the output feature matrix of the graph attention network; W o b is the weight matrix of the output layer; o This is the bias vector for the output layer.
[0037] Furthermore, the graph attention network was trained and tested based on the structural surface point cloud data of the existing tunnel excavation area and the structural surface point cloud data of similar projects, and then optimized using a model ensemble method.
[0038] Furthermore, the expression for meshing using the Possion surface reconstruction method is as follows:
[0039]
[0040] Where f represents the objective function, It is the Laplace operator of f, and V is the normal vector field.
[0041] The beneficial effects of this invention are:
[0042] 1. This invention comprehensively utilizes Laplacian interpolation, statistical filtering, and Gaussian smoothing filtering to process structural surface point clouds. Laplacian interpolation achieves high-precision and continuous 3D coordinate alignment, thus better handling complex point cloud data and reducing computational burden. Statistical filtering effectively removes noise through neighborhood analysis, preserving the effective information of the original point cloud data and improving data quality. Gaussian smoothing filtering smooths the data surface through convolution operations, eliminating high-frequency noise and enhancing data visualization. Combining these methods improves the quality, accuracy, and intuitiveness of point cloud data, enhances the accuracy and reliability of real-time prediction of rock structural surface morphology data, and thus improves the accuracy of structural surface roughness calculation, providing a solid foundation for engineering safety assurance and improved construction efficiency.
[0043] 2. This invention converts processed point cloud data into a graph structure and constructs a graph attention network for topographic data prediction, offering significant advantages over traditional methods. Converting point cloud data into a graph structure preserves the geometric and topological information of the data, thus fully representing the three-dimensional topographic features. The graph attention network can adaptively allocate the importance weights of nodes and capture local and global structural surface topographic data, thereby improving prediction accuracy and robustness. While adapting to the complexity and irregularity of rock structural surfaces, it also possesses good scalability and parallel computing capabilities, meeting the computational power requirements for real-time prediction of the excavated structural surfaces under different working conditions and geological circumstances. Simultaneously, by continuously updating the graph attention network model database with the topographic data of the excavated structural surfaces in real time, dynamic prediction of the topographic data of the structural surfaces in the area to be excavated can be achieved. The prediction effect improves as the number of samples in the database increases. In summary, this method not only enhances the identification and prediction capabilities of rock structural surface features but also improves the efficiency and accuracy of data processing, providing an intelligent solution for engineering applications.
[0044] 3. In this invention, the graph attention network method employs model ensemble for result optimization and uses the Possion surface reconstruction method to mesh the point cloud data. Since the graph attention network model can generate multiple high-probability structural surface shape predictions, the model ensemble method obtains various preliminary prediction results and determines the final prediction result through weighted calculation and voting mechanisms, reducing errors. The Possion surface reconstruction method is used to mesh the point cloud data, generating a continuous and smooth mesh surface, preserving the detailed features of the original point cloud, enhancing the data visualization effect, and making the data more intuitive. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the process of the present invention;
[0046] Figure 2 This is a schematic diagram of the excavation of the real-time prediction method for the structural surface morphology of tunnel surrounding rock based on graph attention network according to the present invention. Detailed Implementation
[0047] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0048] refer to Figure 1 As shown, this embodiment provides a real-time prediction method for the structural surface morphology of tunnel surrounding rock based on graph attention networks, including the following steps:
[0049] Step 1: Use a 3D laser scanner to scan the structural surfaces of the excavated area of the tunnel and obtain its point cloud data.
[0050] A 3D laser scanner is placed directly in front of the tunnel excavation face and its position and orientation are calibrated to ensure accuracy. The scanner captures data reflected from the structural surface, recording the distance and angle of the reflected data point by point during the scan to create a point cloud image. As the scanner moves and rotates, it covers the entire structural surface, ensuring complete point cloud data acquisition. After scanning, the obtained point cloud data is imported into the system.
[0051] Step 2: Align, denoise, and smooth the point cloud data.
[0052] First, data alignment is performed using the Laplacian interpolation method: The system uses a pre-set Laplacian interpolation method program to execute the data alignment command, thereby maintaining a fixed interval in the point cloud data of the structural surface. The specific interval size is determined by the accuracy requirements and the actual dimensions.
[0053] In this embodiment, the expression for point cloud data alignment using the Laplacian interpolation method is:
[0054] Lu=F(1)
[0055]
[0056] Where L is the graph Laplacian matrix, representing the similarity and connectivity between points in the point cloud; u is the point cloud value vector after interpolation, representing the value to be solved; and F is the vector of the original point cloud data.
[0057] Laplacian interpolation offers unique advantages in point cloud data alignment. Its global nature allows it to handle overall point cloud information, maintaining data continuity and consistency, which is beneficial for accurate alignment. Its adaptive performance flexibly handles uneven and sparse data, improving alignment accuracy and stability, thereby ensuring more accurate and stable point cloud alignment results.
[0058] After data alignment, statistical filtering is used to denoise the point cloud data. Statistical filtering is an effective method for denoising point cloud data. For each point in the point cloud, a neighborhood is defined according to specific engineering requirements. This neighborhood is typically all points within a sphere or cube centered on that point.
[0059] In this embodiment, the expression for point cloud denoising using the statistical filtering is:
[0060] d i >μ+ασ(3)
[0061]
[0062] Where, p i p j Both are point coordinate vectors in the point cloud, d i The coordinate vector p of the midpoint of the point cloud i The average nearest neighbor distance is μ; μ is the average nearest neighbor distance of all points; α is a custom threshold parameter used to control the criteria for identifying outliers, with common values ranging from 1 to 3; σ is the standard deviation of the average nearest neighbor distance of all points, and k is the number of nearest neighbors, which can be 1 to 5 for high-density point clouds.
[0063] If data point d i If the value of d is greater than the mean μ plus α times the standard deviation σ, then this data point d i It is considered an outlier.
[0064] Statistical filtering methods preserve the local features of point cloud data, are highly adjustable, and have high computational efficiency, making them suitable for point cloud processing of rock masses of various qualities.
[0065] After denoising, the point cloud data is smoothed using Gaussian smoothing filter. This method has the advantages of adjustable parameters and preservation of local structure.
[0066] In this embodiment, the expression for point cloud smoothing using the Gaussian smoothing filter is:
[0067]
[0068] Where, p i ′ represents point p after Gaussian smoothing. i coordinates; w ij Let p be the weight, representing the point p. j For point p i The degree of influence, where n represents the total number of points in the point cloud.
[0069] After smoothing the point cloud data, it is necessary to evaluate whether the processing effect is excessive or whether details are lost, and adjust the parameters accordingly.
[0070] Step 3: Convert the processed point cloud data into a graph structure and construct a graph attention network to predict the morphological data of the structural surface of the area to be excavated.
[0071] Before constructing a graph attention network model, point cloud data needs to be converted into a graph structure. This process includes the representation of point cloud data and the construction of the adjacency matrix.
[0072] In this embodiment, we assume that the point cloud data P = {p1, p2, ..., p...} n}, where each point p i Includes its three-dimensional coordinates p i =(x i ,y i ,z i ).
[0073] By defining the adjacency relationships between points, an adjacency matrix A is constructed, whose elements are α. ij .
[0074]
[0075] Among them, ||p i -p j || 2 Point p i and point p j The Euclidean distance between them is given by σ, where σ is the standard deviation of the Gaussian kernel function and ε is the distance threshold.
[0076] In this embodiment, the feature representation h of node i j The output of the graph attention layer is represented as:
[0077]
[0078] Among them, h i′ represents the output feature of node i; h j σ represents the input features of node j; σ is the nonlinear activation function. Let be the set of neighboring nodes of node i; W is the trainable weight matrix; α ij The attention weights between node i and node j are calculated as follows:
[0079]
[0080] Where a is a trainable weight vector; || is the vector concatenation operation; LeakyReLU is used as the activation function to introduce nonlinearity.
[0081] In the final layer of the graph attention network, a regression model of a linear layer can be used to predict the topographic data of the structural surfaces. Assuming Z is the output of the network, it is expressed as:
[0082] Z = HW o +b o (9)
[0083] Where H is the output feature matrix of the graph attention network; W o b is the weight matrix of the output layer; o This is the bias vector for the output layer.
[0084] When training the graph attention network, point cloud data of rock tunnel structures with similar geological conditions can be used for model training. The model can be tested using the excavated part of this project to verify the accuracy of the model construction.
[0085] The loss function uses mean squared error loss, specifically expressed as:
[0086]
[0087] Among them, z i Represents true morphological data, This represents the predicted topographic data.
[0088] Graph attention networks have a high efficiency in modeling complex spatial relationships. This method can dynamically adjust the influence between nodes through the attention mechanism, thereby more accurately capturing and expressing the relationship between structural surface morphology data under complex rock mass conditions. At the same time, this method can adaptively learn node features and thereby improve the model's generalization ability and prediction accuracy.
[0089] Step 4: Optimize the above morphological data using a model ensemble method.
[0090] Optimizing prediction results is essential when addressing complex geological and engineering problems. Model ensemble is a method that integrates the prediction results of multiple models to improve overall prediction performance. By integrating multiple models, the bias and variance of individual models can be reduced, thereby enhancing the stability and accuracy of predictions.
[0091] In this embodiment, a soft voting method is used to determine the probability distribution of each model's output category for each sample. Finally, a weighted vote is used to obtain the optimal solution, the expression of which is:
[0092]
[0093] in, For the final prediction result, p m,k Let be the predicted probability of the m-th model for category k.
[0094] Step 5: The optimized morphology information is meshed using the Poisson surface reconstruction method and then visualized.
[0095] The main principle of the Poisson surface reconstruction method is to construct a continuous implicit function around the point cloud by solving the Poisson equation, and then use this function to generate a meshed 3D surface. This method can efficiently process noisy data, generate closed surfaces with consistent topology, and preserve the geometric details, i.e., microscopic morphological features, of the original point cloud.
[0096] In this embodiment, the expression for meshing using the Possion surface reconstruction method is as follows:
[0097]
[0098] Where f represents the objective function, It is the Laplace operator of f, and V is the normal vector field.
[0099] This embodiment selects the Three.js library for real-time display of point cloud data on a webpage. Three.js is a WebGL-based JavaScript library suitable for 3D data visualization in a browser, making data display more convenient and intuitive.
[0100] The specific process of constructing a graph attention network includes:
[0101] 1) Obtain structural surface samples from the excavated area of the tunnel, scan them using a 3D laser scanner, and then align, denoise, and smooth the point cloud data. Alternatively, structural surface morphology data can be obtained from tunnel projects with similar geological conditions, and point cloud processing can be performed in the same manner.
[0102] 2) Divide the above morphological data into training and testing sets, construct a graph attention network model, and debug and optimize its parameters.
[0103] 3) The graph attention network model is used to predict the structural surface morphology data of the area to be excavated, such as... Figure 2 As shown, the accuracy of model predictions is improved by continuously updating and expanding the sample database of structural surface morphology data. Predicting the morphology data of complete structural surfaces helps to accurately evaluate the quality of the rock mass and thus provide appropriate support.
[0104] The real-time prediction method for tunnel surrounding rock structural surface morphology data based on graph attention networks has high predictive visibility. It can reasonably predict the structural surface morphology data of the area to be excavated based on existing structural surface morphology data, providing a basis for the selection of rock mass support schemes and promoting the intelligent acquisition of tunnel surrounding rock structural surface information and the selection of rock mass support schemes.
[0105] The above method employs deep learning techniques to achieve the desired functionality. Graph attention effectively captures the complex spatial relationships and local features of structural surfaces. By representing point cloud data in a graph structure and utilizing an attention mechanism to focus on important neighboring nodes, it improves the accuracy and stability of predictions. Furthermore, because graph attention networks can adaptively learn the weights between nodes and dynamically adjust model parameters during training, they possess strong generalization and adaptability, making them suitable for structural surface data of different types and scales, enabling efficient execution of real-time prediction tasks.
[0106] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.
Claims
1. A real-time prediction method for tunnel surrounding rock structure surface morphology based on graph attention network, characterized in that: The following steps are involved: (1) Use a 3D laser scanner to scan the structural surface of the tunnel excavation area and obtain its point cloud data; (2) Align, denoise, and smooth the above point cloud data; (3) Convert the processed point cloud data into a graph structure and construct a graph attention network to predict the morphological data of the structural surface to be excavated; (4) optimizing the above-mentioned morphological data by using a model integration method; The optimized topographic data are meshed and visualized using the Poisson surface reconstruction method.
2. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1 is characterized in that: The Laplace interpolation method is used to align the three-dimensional coordinates of the acquired point cloud data, and its target equation is: Lu=F Among them, L is the graph Laplacian matrix, which represents the similarity and connection relationship between points in the point cloud; u is the point cloud value vector after interpolation processing, which represents the value to be solved; F is the vector of the original data of the point cloud, which represents the known value.
3. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 2 is characterized in that: The expression for constructing the Laplacian matrix is: Among them, L ij Represents the element in the i-th row and j-th column of the graph Laplacian matrix that shows the similarity and connectivity between points in the point cloud.
4. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1 is characterized in that: Statistical filtering is used to denoise point clouds, and its expression is: d i >m+as Among them, d i is the coordinate vector p of the point in the point cloud i is the average nearest neighbor distance of all points; μ is the average nearest neighbor distance of all points; α is a custom threshold parameter used to control the judgment criteria of outliers, and the common value is 1 to 3; σ is the standard deviation of the average nearest neighbor distance of all points.
5. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 4 is characterized in that: The point cloud midpoint coordinate vector p i The average nearest neighbor distance d i The calculation method is: Among them, p i , p j are all point coordinate vectors in the point cloud, k is the number of nearest neighbors, and is 1 to 5 for high-density point clouds.
6. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1 is characterized in that: Use Gaussian smoothing filter to smooth the point cloud, and its expression is: Among them, p i ′ is the point p after Gaussian smoothing filtering i The coordinates of ij is the weight, indicating point p j Point p i The influence degree of n represents the total number of points in the point cloud.
7. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1 is characterized in that: The point cloud data is converted into a graph structure by constructing an adjacency matrix, and a graph attention network graph attention layer is constructed based on the graph structure. Its output is expressed as: Among them, h i ′ is the output feature of node i; h j is the input feature of node j; σ is a nonlinear activation function; N(i) is the set of neighbor nodes of node i, W is a trainable weight matrix, α ij is the attention weight between node i and node j.
8. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1 is characterized in that: The output layer of the graph attention network prediction is expressed as: Z=HW o +b o Among them, Z is the output of the network; H is the output feature matrix of the graph attention network; W o is the weight matrix of the output layer; b o is the bias vector of the output layer.
9. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 8, characterized in that: Based on the structural surface point cloud data of the excavated area of the existing tunnel and the structural surface point cloud data of similar projects, the graph attention network is trained and tested, and optimized using the model integration method.
10. The method for real-time prediction of tunnel surrounding rock structure surface morphology based on graph attention network according to claim 1, characterized in that: The expression for meshing using the Possion surface reconstruction method is: Where f represents the objective function, is the Laplacian of f, and V is the normal vector field.
