Method for identifying and forecasting unstable block of tunnel face

Abstract

The invention relates to the field of tunnel engineering, and provides a method for identifying and forecasting an unstable block of a tunnel face in order to improve the accuracy and efficiency of identifying the unstable block of the tunnel face and realize advanced forecasting of the unstable block of the tunnel face. The main structural surface of the tunnel face is effectively extracted, and the efficiency and accuracy of structural surface recognition are greatly improved; according to the method, the spatial relationship between the structural plane and the tunnel direction is analyzed, potential unstable blocks can be accurately identified, the unstable block distribution of the future tunnel face is predicted in advance, and the safety risk in the tunneling process is reduced.

Description

Technical Field

[0001] The invention relates to the field of tunnel engineering, and in particular to a method for identifying and predicting unstable blocks on a tunnel face. Background Art

[0002] Rapid identification and advance prediction of unstable blocks on the tunnel face in tunnel engineering is the key to ensuring construction safety. Traditional methods mainly rely on manual experience and on-site measurements, usually using laser scanning, total station measurement or structural surface extraction technology based on image processing. These methods identify potential structural surfaces by performing two-dimensional or three-dimensional geological cataloging of the tunnel face, and perform stability analysis in combination with geomechanical models. However, traditional methods have the problems of limited accuracy, low processing efficiency, and poor adaptability to complex geological conditions.

[0003] Although laser scanning and total station measurements can provide high-precision three-dimensional data, the equipment is expensive, the operation is complicated, and the data processing time is long, which cannot meet the needs of real-time forecasting. Structural surface extraction technology based on image processing relies on high-quality images and complex algorithms, and it is often difficult to obtain stable results under field conditions. In addition, traditional methods lack an in-depth understanding of the spatial distribution and interrelationships of structural surfaces during data analysis, which leads to greater uncertainty in identifying unstable blocks. In general, existing technologies are costly, difficult to operate, and have poor real-time performance. They are not adaptable enough under complex geological conditions, and their intelligent analysis capabilities are limited, making it difficult to fully meet the efficiency and safety requirements of modern tunnel construction. Summary of the invention

[0004] In order to improve the accuracy and efficiency of the identification of unstable blocks on the tunnel face and realize the advance prediction of unstable blocks on the tunnel face, the present application provides a method for identifying and predicting unstable blocks on the tunnel face.

[0005] The technical solution adopted by the present invention to solve the above problems is: The identification method of unstable blocks on the tunnel face includes: Step 1: Obtaining point cloud data of the tunnel face; Step 2: Build and train a structural surface extraction model based on the machine learning model; Step 3: pre-process the point cloud data, and use the trained structural surface extraction model to process and analyze the pre-processed point cloud data to extract the structural surface of the current tunnel face; Step 4: Calculate the attitude of the structural planes, the angle between the attitudes of the structural planes and the angle between the attitude of the structural planes and the direction of the hole, and use block theory to identify potential unstable blocks.

[0006] Furthermore, step 1 specifically includes: Step 11: Using a mobile device to collect tunnel face images from multiple angles; Step 12: register and stitch the multi-view images to generate a two-dimensional cataloging surface; Step 13: Perform three-dimensional reconstruction based on the two-dimensional catalog surface to generate point cloud data of the tunnel face.

[0007] Furthermore, in step 11, the mobile device is a camera or a laser scanning device.

[0008] Furthermore, the specific steps of registering and stitching the multi-view images in step 12 are: using the scale-invariant feature transform algorithm SIFT to extract and match feature points, using the random sampling consensus algorithm RANSAC to perform registration, and using the stitching algorithm in the OpenCV library to stitch the multi-view images into a two-dimensional cataloging surface.

[0009] Furthermore, step 13 uses a 3D reconstruction algorithm SfM to perform 3D reconstruction, which specifically includes: Feature extraction and matching based on two-dimensional cataloging surfaces; Determine the camera parameters, including the intrinsic matrix and the extrinsic matrix; Sparse reconstruction: determine the initial camera pair, gradually add new images, and update the camera parameters and the position of the 3D points; Dense reconstruction: Use a multi-view stereo matching algorithm to generate a depth map for each view, and fuse the depth maps from multiple views to generate a dense point cloud; Optimization and post-processing: In the sparse reconstruction and dense reconstruction processes, global BA is used to optimize the camera parameters and the positions of 3D points to minimize the reprojection error; the generated dense point cloud is filtered to remove noise points and outliers.

[0010] Furthermore, in step 4, the calculation method of the structural surface occurrence is: Select three points at random from the point cloud data , , Perform plane fitting and calculate two vectors: , , calculate the normal vector of the plane : , normalize the normal vector to get , according to the direction of the normal vector, calculate the occurrence of the structural surface: , .

[0011] Furthermore, in step 4, the angle calculation formula is: , In the formula, It is the angle between the structural surface attitudes or between the structural surface attitude and the hole direction. is the normal vector of the structural surface 、 is the normal vector of the structural surface or hole.

[0012] Furthermore, the method further includes step 5, monitoring the structural surface data of the tunnel face in real time through sensors, and optimizing the structural surface extraction model based on the monitoring data of the sensors.

[0013] Furthermore, the sensor includes an inertial sensor and a lidar sensor.

[0014] The prediction method of unstable blocks on the tunnel face includes: Step 1: using a method for identifying unstable blocks on the tunnel face to identify the structural surface of the tunnel face within the continuous excavation distance; Step 2: Based on the structural surface and distribution of unstable blocks on the tunnel face within the continuous excavation distance, the time series analysis method is used to predict the possibility and distribution of unstable blocks on the tunnel face in the future.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. Fast and accurate identification: By building a structural surface extraction model and combining it with high-precision point cloud data, the main structural surfaces of the tunnel face can be effectively extracted, greatly improving the efficiency and accuracy of structural surface identification.

[0016] 2. Advanced prediction: By analyzing the spatial relationship between the structural surface and the tunnel direction, it is possible to accurately identify potential unstable blocks and predict the distribution of unstable blocks on the future tunnel face in advance, thereby reducing safety risks during excavation.

[0017] 3. Real-time dynamic adjustment: Using real-time monitoring data, the model and prediction results are dynamically updated to ensure the continued effectiveness of the prediction and adapt to changes in geological conditions during tunneling. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the identification method of unstable blocks on the tunnel face; Figure 2 Schematic diagram of the ARIMA model prediction structure surface coordinate distribution. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0020] like Figure 1 As shown in FIG. 1 , the method for identifying unstable blocks on the tunnel face includes: Step 1: Obtain the point cloud data of the tunnel face.

[0021] Use mobile devices to collect tunnel face images from multiple angles, register and stitch the collected multi-view images to ensure the alignment and fusion of images from different perspectives to generate a high-precision two-dimensional catalog surface. In terms of equipment selection, use a high-resolution camera. When registering images, use the scale-invariant feature transform (SIFT) algorithm to extract feature points, and match based on the extracted feature points, and use the random sampling consensus (RANSAC) algorithm for registration to ensure the accuracy of image stitching; when stitching images, use the stitching algorithm in the OpenCV library to stitch multi-view images into a high-precision two-dimensional catalog surface, and then use the three-dimensional reconstruction algorithm to perform three-dimensional reconstruction based on the high-precision two-dimensional catalog surface to generate point cloud data of the tunnel face.

[0022] The key parameters that need to be set in the SIFT algorithm during feature point extraction and matching are shown in Table 1: Table 1 Key parameters of SIFT algorithm parameter Parameter name default value meaning nfeatures Number of feature points 0 Specify the number of feature points to be returned. nOctaveLayers The number of layers in each group of the pyramid 3 Gaussian pyramid structure that affects images contrastThreshold Contrast threshold 0.04 Used to filter out poor feature points edgeThreshold Edge Threshold 10 Used to filter out edge effects Sigma Gaussian filter coefficients 1.6 Gaussian filtering for the image at level 0 of the pyramid The parameter settings of the RANSAC algorithm for image registration are shown in Table 2: Table 2 Key parameters of RANSAC algorithm parameter Parameter name Common settings meaning maxIters Maximum number of iterations 1000~2000 Specify the maximum number of iterations for the RANSAC algorithm. confidence Confidence 0.99 Indicates the confidence of the model threshold Inlier distance threshold 1.0~3.0 Used to determine whether a point is an interior point The 3D reconstruction algorithm uses SfM (Structure from Motion) technology. The following are some key parameter setting details: ①Feature extraction and matching Feature detector: Use SIFT feature detector to improve the accuracy of feature matching.

[0023] Feature matching algorithm: Use BF Matcher (brute force matching) for feature matching. Set a distance threshold to filter out unreliable matching points during matching.

[0024] ②Camera parameter estimation Intrinsic parameter matrix: The intrinsic parameter matrix of the camera (including focal length, principal point coordinates, etc.) is obtained through camera calibration.

[0025] Extrinsic parameter matrix: The extrinsic parameter matrix (including rotation matrix and translation vector) is estimated through the geometric relationship of feature matching points. The RANSAC algorithm is used to remove outliers to improve the robustness of the estimation.

[0026] ③Sparse reconstruction Initial Camera Pair Selection: Image pairs with larger baselines are selected as initial camera pairs to improve the quality of the initial sparse point cloud.

[0027] Incremental reconstruction: gradually add new images and update the camera parameters and the positions of 3D points.

[0028] ④Dense reconstruction Depth map estimation: The multi-view stereo matching algorithm Semi-Global Matching is used to generate the depth map of each view. The key parameters of the Semi-Global Matching algorithm are shown in Table 3.

[0029] Table 3 Key parameters of Semi-Global Matching algorithm parameter Parameter name Common settings meaning Max Disparity Maximum parallax range 64、128、256 Set the maximum value used to calculate the disparity, affecting the calculation complexity and disparity range Min Disparity Minimum parallax 0 Sets the minimum value of the disparity search range, usually 0 or a negative value, to accommodate smaller disparities in certain scenes Num Disparities The total step length of the parallax range Multiples of 16 Part of the parallax range Block Size The size of the cost calculation window 3~9 The window size used for cost calculation Point cloud fusion: Fusion of depth maps from multiple perspectives to generate a dense point cloud. Use TSDF (Truncated Signed Distance Function) for point cloud fusion.

[0030] The key parameters of the TSDF algorithm are shown in Table 4.

[0031] Table 4 TSDF algorithm key parameters parameter Parameter name Common settings meaning voxel_length Voxel edge length 0.01~0.02 or 4.0 / 512.0 The side length of a single voxel sdf_trunc TSDF cutoff distance 0.01~0.05 Cutoff distance of TSDF value color_type Color Data Type TSDFVolumeColorType.RGB8 or TSDFVolumeColorType.None Set whether to merge color data in voxels ⑤Optimization and post-processing Global BA optimization: In the sparse and dense reconstruction process, global BA is used to optimize the camera parameters and the positions of 3D points to minimize the reprojection error.

[0032] Point cloud filtering: Filter the generated dense point cloud to remove noise points and outliers and improve the quality of the point cloud.

[0033] Post-process the point cloud data, including denoising, filtering and precision improvement, to improve data quality. Statistical Outlier Removal algorithm is used for denoising, Voxel Grid Filter algorithm is used for filtering, and ICP algorithm is used for precision improvement. The following are some key parameter setting details: ① Statistical Outlier Removal Mean K: Sets the number of neighborhood points used to calculate the average distance of each point, set to 50.

[0034] Std dev Mul Thresh: Sets the standard deviation multiplier from the mean, set to 1.0.

[0035] ② Filter (Voxel Grid Filter) Leaf Size: Sets the leaf size of the voxel grid, set to 0.01 meters depending on the density of the point cloud and the required accuracy.

[0036] ③ Improved accuracy (ICP algorithm) Max Correspondence Distance: Set the maximum distance between corresponding points, set to 0.05 meters.

[0037] Maximum Iterations: Set the maximum number of iterations of the ICP algorithm to 50.

[0038] Transformation Epsilon: Set the convergence threshold of the transformation matrix, set to 1e-6.

[0039] Euclidean Fitness Epsilon: Set the convergence threshold of Euclidean fitness, set to 1e-5.

[0040] In addition to the above-mentioned method for acquiring point cloud data, other methods may also be used to acquire point cloud data of the tunnel face, which are not limited here.

[0041] Step 2: Build and train a structural surface extraction model based on the machine learning model.

[0042] In this embodiment, the GNN algorithm (based on Graph SAGE) is used to establish a machine learning model, with the input being the adjacency matrix and node features of the point cloud data, and the output being the node embedding vector and the node classification result. The number of layers of the Graph SAGE model is set to 2. The hidden layer dimension, i.e., the number of hidden units in each layer, is set to 64. The ReLU function is used as the activation function.

[0043] In the hyperparameter settings, the learning rate is set to 0.001, the batch size is 32, the number of training rounds is 100, and the Adam optimizer is used.

[0044] A large data set of tunnels under different geological conditions and construction scenarios is used for training. The data set is divided into a training set, a validation set, and a test set in a ratio of 70%:15%:15%. K-fold cross validation (usually 5-fold) is used to evaluate model performance to ensure the generalization ability of the model.

[0045] Step 3: pre-process the point cloud data, and use the trained structural surface extraction model to process and analyze the pre-processed point cloud data to extract the structural surface of the current tunnel face.

[0046] The preprocessing of point cloud data includes downsampling and normal vector estimation. The point cloud data is downsampled using voxel grid filtering to reduce the amount of calculation. The voxel size is set to 0.01 meters based on the density of the point cloud and the required accuracy. The number of neighborhood points used to estimate the normal vector is set to 30 neighborhood points, and the KD tree is used to search the neighborhood points to improve the calculation efficiency.

[0047] Step 4: Calculate the attitude of the structural planes, the angle between the attitudes of the structural planes and the angle between the attitude of the structural planes and the direction of the hole, and use block theory to identify potential unstable blocks.

[0048] The calculation method of the structural surface is as follows: select three points at random from the point cloud data , , Perform plane fitting and calculate two vectors: , , calculate the normal vector of the plane by vector cross multiplication : , normalize the normal vector for subsequent calculations: . According to the direction of the normal vector, calculate the occurrence of the structural surface: , , In the formula, is the normal vector of the structural surface.

[0049] The angle calculation formula is: , In the formula, It is the angle between the structural surface attitudes or between the structural surface attitude and the hole direction. is the normal vector of the structural surface 、 is the normal vector of the structural surface or hole.

[0050] Furthermore, the structural surface data of the tunnel face can be monitored in real time through sensors (such as IMU and LiDAR), and the structural surface extraction model can be optimized and adjusted based on the monitoring data of the sensors. Real-time monitoring and optimization can be achieved based on embedded AI chips and edge computing technology.

[0051] The present application also provides a method for predicting unstable blocks on a tunnel face, comprising: using a method for identifying unstable blocks on a tunnel face to identify the structural surface of the tunnel face within a continuous excavation distance; based on the structural surface of the tunnel face within the continuous excavation distance and the distribution of unstable blocks, using a time series analysis method (ARIMA) to predict the possibility and distribution of the formation of unstable blocks on the tunnel face in the future.

[0052] Based on point cloud data and GNN, the rapid identification and advance prediction of unstable blocks on the face of a tunnel under construction is taken as an example. The implementation steps and details are as follows: Step 1: Construct a 2D visualization surface for the tunnel face.

[0053] A mobile device equipped with a high-resolution camera was used to shoot the tunnel face from multiple perspectives. The shooting distance was controlled within the range of 1-3 meters, and the angles covered all directions (horizontal, vertical, and oblique). The SIFT algorithm was used for feature point extraction and matching, and the specific parameters were set as follows: nfeatures=500, nOctaveLayers=3, contrastThreshold=0.04, edgeThreshold=10, and sigma=1.6. The RANSAC algorithm was used for registration, and the specific parameters were set as follows: threshold =3.0, maxIters = 2000, and confidence = 0.99. The multi-perspective images were stitched into a high-precision two-dimensional cataloging surface using the stitching algorithm in the OpenCV library.

[0054] Use the SIFT algorithm to extract features from the two-dimensional catalog surface, and set the specific parameters in the same way as step 1. Use BFMatcher for feature matching, set the distance threshold to 0.75, and the reconstruction accuracy to 0.1 mm. Obtain the intrinsic parameter matrix through camera calibration, set the focal length to 1000, the principal point coordinates to (640, 480), and the intrinsic parameter matrix is ​​as follows: , Use the RANSAC algorithm to estimate the extrinsic parameter matrix, and set the specific parameters in the same way as step 1. Select the image pair with a larger baseline as the initial camera pair, gradually add new images, and update the camera parameters and the position of the 3D points. Each time a new image is added, use BA (Bundle Adjustment) to optimize the system parameters. Use Semi-Global Matching to generate a depth map, and set the specific parameters to: maxDisparity=64, minDisparity=0, numDisparities=16, blockSize=3.

[0055] Use the TSDF method to generate a dense point cloud. The specific parameters are set to: voxel_length=4.0 / 512.0, sdf_trunc=0.04, and color_type is TSDFVolumeColorType.RGB8.

[0056] Use global BA to optimize camera parameters and 3D point positions to minimize reprojection errors. Use Statistical Outlier Removal for denoising and Voxel Grid Filter for filtering. The specific parameters are set as: Mean K is 50, Std devMul Thresh is 1.0, and Leaf Size is 0.01.

[0057] Use the ICP algorithm for fine registration, and the specific parameters are set as follows: max_correspondence_distance = 0.05, maximum_iterations = 50, transformation_epsilon = 1e-6, euclidean_fitness_epsilon = 1e-5.

[0058] Step 2: Build and train a structural surface extraction model based on the machine learning model.

[0059] The input of the structural surface extraction model is the adjacency matrix and node features of the point cloud data, and the output is the node embedding vector and node classification results.

[0060] The model input is the adjacency matrix and node features of the point cloud data, and the output is the node embedding vector and node classification result. The model is set to 2 layers, the hidden layer dimension is 128, and the ReLU function is used as the activation function. The learning rate is set to 0.001, the batch size is 32, the number of training rounds is 100, and the Adam optimizer is used. A large data set of tunnels under different geological conditions and construction scenarios is used for training. The data set is divided into training set, validation set and test set in a ratio of 70%:15%:15%. A 5-fold cross validation is used to evaluate the model performance to ensure the generalization ability of the model.

[0061] Step 3: Preprocess the point cloud data and use the trained structural surface extraction model to process and analyze the preprocessed point cloud data to extract the structural surface of the current tunnel face.

[0062] The preprocessing includes downsampling the point cloud data using voxel grid filtering to reduce the amount of calculation. The voxel size is set to 0.02 meters.

[0063] The number of neighborhood points used to estimate the normal vector is set to 40, and the KD tree is used to search the neighborhood points to improve computational efficiency.

[0064] Step 4: Calculate the attitude of the structural planes, the angle between the attitudes of the structural planes and the angle between the attitude of the structural planes and the direction of the hole, and use block theory to identify potential unstable blocks.

[0065] The calculation method of the structural surface occurrence is as follows: select three points at random from the point cloud data such as , , Perform plane fitting and calculate two vectors: , , calculate the normal vector of the plane by vector cross multiplication : , normalize the normal vector for subsequent calculations: . According to the direction of the normal vector, calculate the occurrence of the structural surface, including the inclination and tendency : , .

[0066] The direction of the tunnel axis is (along the x-axis), the angle between the structural surface and the hole direction is: .

[0067] This angle is relatively large, so according to the block theory, the probability of generating unstable blocks is relatively small. The calculation method of the angle between the structural planes is the same as the angle between the structural plane and the hole direction, which will not be repeated here.

[0068] Step 5: Use the ARIMA model for advance prediction. Collect the structural surface and unstable block distribution data of the tunnel face within the continuous excavation distance, use the time series analysis algorithm ARIMA to model the data, and predict the possibility and distribution of unstable blocks on the tunnel face in the future based on the model. The prediction results are as follows: Figure 2 shown.

[0069] In order to improve the accuracy of recognition and prediction, IMU and LiDAR sensors can also be used to monitor the structural surface data of the tunnel face in real time, dynamically integrate new data into the GNN model, and continuously update the prediction results.

[0070] In addition to being used in tunnel construction, the method of the present application can also be used for rock stability analysis and early warning in mining, foundation stability analysis and structural health monitoring in construction projects, and for monitoring and early warning of geological disasters such as landslides and collapses.

Claims

1. A method for identifying unstable blocks on a tunnel face, characterized in that: include: Step 1: Obtaining point cloud data of the tunnel face; Step 2: Build and train a structural surface extraction model based on the machine learning model; Step 3: pre-process the point cloud data, and use the trained structural surface extraction model to process and analyze the pre-processed point cloud data to extract the structural surface of the current tunnel face; Step 4: Calculate the attitude of the structural planes, the angle between the attitudes of the structural planes and the angle between the attitude of the structural planes and the direction of the hole, and use block theory to identify potential unstable blocks.

2. The method for identifying unstable blocks at a tunnel face according to claim 1, characterized in that: Step 1 specifically includes: Step 11: Using a mobile device to collect tunnel face images from multiple angles; Step 12: register and stitch the multi-view images to generate a two-dimensional cataloging surface; Step 13: Perform three-dimensional reconstruction based on the two-dimensional catalog surface to generate point cloud data of the tunnel face.

3. The method for identifying unstable blocks at a tunnel face according to claim 2, characterized in that: In step 11, the mobile device is a camera or a laser scanning device.

4. The method for identifying unstable blocks on a tunnel face according to claim 2, characterized in that: The specific steps of registering and stitching the multi-view images in step 12 are: using the scale-invariant feature transform algorithm SIFT to extract and match feature points, using the random sampling consensus algorithm RANSAC for registration, and using the stitching algorithm in the OpenCV library to stitch the multi-view images into a two-dimensional cataloging surface.

5. The method for identifying unstable blocks at a tunnel face according to claim 2, characterized in that: Step 13 uses the 3D reconstruction algorithm SfM to perform 3D reconstruction, which specifically includes: Feature extraction and matching based on two-dimensional cataloging surfaces; Determine the camera parameters, including the intrinsic matrix and the extrinsic matrix; Sparse reconstruction: determine the initial camera pair, gradually add new images, and update the camera parameters and the position of the 3D points; Dense reconstruction: Use a multi-view stereo matching algorithm to generate a depth map for each view, and fuse the depth maps from multiple views to generate a dense point cloud; Optimization and post-processing: In the sparse reconstruction and dense reconstruction processes, global BA is used to optimize the camera parameters and the positions of 3D points to minimize the reprojection error; the generated dense point cloud is filtered to remove noise points and outliers.

6. The method for identifying unstable blocks at a tunnel face according to claim 1, characterized in that: In step 4, the calculation method of the structural surface occurrence is: Select three points at random from the point cloud data , , Perform plane fitting and calculate two vectors: , , calculate the normal vector of the plane : , normalize the normal vector to get , according to the direction of the normal vector, calculate the occurrence of the structural surface: ， 。 7. The method for identifying unstable blocks at a tunnel face according to claim 1, characterized in that: In step 4, the angle calculation formula is: ， In the formula, It is the angle between the structural surface attitudes or between the structural surface attitude and the hole direction. is the normal vector of the structural surface 、 is the normal vector of the structural surface or hole.

8. The method for identifying unstable blocks at a tunnel face according to claim 1, characterized in that: The method further includes step 5, monitoring the structural surface data of the tunnel face in real time through sensors, and optimizing the structural surface extraction model based on the monitoring data of the sensors.

9. The method for identifying unstable blocks at a tunnel face according to claim 8, characterized in that: Sensors include inertial sensors and lidar sensors.

10. A method for predicting unstable blocks on a tunnel face, characterized in that: include: Step 1, using the method for identifying unstable blocks on the tunnel face according to any one of claims 1 to 9 to identify the structural surface of the tunnel face within the continuous excavation distance; Step 2: Based on the structural surface and distribution of unstable blocks on the tunnel face within the continuous excavation distance, the time series analysis method is used to predict the possibility and distribution of unstable blocks on the tunnel face in the future.

Images