A method for predicting rock mass structural plane contact characteristics based on artificial neural network

Abstract

The application discloses a method for predicting rock mass structural plane contact characteristics based on an artificial neural network, and relates to the field of rock mass structural plane contact state measurement, and comprises the following steps: step 1, determining the uniaxial compression test of the structural plane to be predicted, and obtaining the contact condition of the structural plane by using pressure-sensitive paper; step 2, binarizing the obtained pressure-sensitive paper, packing the orthogonal one square millimeter of pixels, and respectively calculating the contact area ratio of each square millimeter; step 3, setting a reference threshold value, and defining as contact if the reference threshold value exceeds the set value, and vice versa; and step 4, performing preliminary processing on the point cloud of the upper and lower plates of the rock mass structural plane. The application estimates the contact condition of the structural plane based on the point cloud of the rock mass structural plane, the geometry and geometric parameters of the microconvex body on the structural plane can be more accurately described by using the point cloud to describe the appearance of the structural plane, and the precision can be selected according to actual requirements, and the calculation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of rock mass structural surface contact state measurement, and specifically to a method for predicting rock mass structural surface contact characteristics based on artificial neural networks. Background Technology

[0002] Rock masses in nature often contain numerous discontinuous surfaces, known as rock mass structural surfaces. These structural surfaces exert a controlling influence on the mechanical properties and overall stability of the rock mass. The presence of these structural surfaces disrupts the integrity of the rock mass, leading to a significant reduction in its strength and stability. The separation and slippage of these structural surfaces can cause rock and soil disasters such as collapses and landslides, seriously threatening people's lives and property. Furthermore, rock mass structural surfaces also have a significant impact on tunnels, water conservancy and hydropower projects, and oil and gas resource extraction. The contact state of rock mass structural surfaces often determines the seepage characteristics of water, oil, and gas, which is of great significance for the extraction of fluid or gaseous resources and energy. Typically, even with well-matched structural surfaces, the actual contact portion of the rock mass structural surfaces is only a small part. Therefore, past studies have led to an overestimation of rock mass strength, which has had a significant negative impact on engineering construction or underground resource and energy development.

[0003] Many scholars have proposed methods for calculating the contact area of ​​structural surfaces, mainly including methods for directly measuring the contact area. However, the process of measuring the contact area between structural surfaces of rock masses is relatively complex. Therefore, this paper starts with tension-type structural surfaces, uses artificial neural networks with point cloud geometric features as input parameters, and uses contact and non-contact areas as training results. By adjusting the network parameters, it forms a prediction of the contact situation of structural surfaces of rock masses of the same lithology under the same normal stress. Furthermore, by combining image processing technology, the prediction results are visualized, providing a basis for subsequent calculations and analysis.

[0004] Given the practical requirements of engineering applications, there is an urgent need for a method to predict the indirect contact between rock mass structural surfaces. Therefore, this invention proposes a method based on artificial neural networks to predict the contact between rock mass structural surfaces under normal stress.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the contact characteristics of rock mass structural surfaces based on artificial neural networks. This method innovatively combines image processing and point cloud processing to predict the contact conditions of different rock mass structural surfaces based on trained artificial neural networks. The aim is to gain a better understanding of the contact conditions of rock mass structural surfaces and to accurately estimate the shear strength of the rock mass structural surfaces, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the surface contact characteristics of rock mass structure based on artificial neural networks, comprising the following steps:

[0008] Step 1: Determine the uniaxial compression test of the structural surface to be predicted, and use pressure-sensitive paper to obtain the contact condition of the structural surface;

[0009] Step 2: Binarize the obtained pressure-sensitive paper, pack the pixels of one square millimeter orthogonal, and calculate the contact area ratio of each square millimeter.

[0010] Step 3: Set a reference threshold. If the reference threshold exceeds the set value, it is defined as contact; otherwise, it is defined as non-contact.

[0011] Step 4: Perform preliminary processing on the top and bottom point clouds of the rock mass structure surface;

[0012] Step 5: Divide the point cloud into units at intervals of one millimeter along the horizontal and vertical coordinates, and associate each point cloud unit with a pixel unit from Steps 1-4.

[0013] Step 6: Calculate the geometric parameters within each point cloud unit;

[0014] Step 7: Based on the initial contact ratio, filter between contact and non-contact, and set up training and test sets;

[0015] Step 8: Build an artificial neural network, train it on the data, and adjust the training parameters until the accuracy of the test results reaches 80% and then stop the adjustment.

[0016] Step 9: Use the new point cloud to make predictions through this artificial neural network and draw a binarized image.

[0017] Preferably, the specific implementation steps of step 1 are as follows: select rock mass structural surfaces with the same lithology to conduct a uniaxial compression test, place pressure-sensitive paper between the structural surfaces, the normal loading speed is 0.07kN / s, when the normal stress reaches the target value, maintain for half a minute and remove the pressure-sensitive paper, and then upload it to the computer.

[0018] Preferably, the specific implementation steps of step 2 are as follows:

[0019] Step 2.1: Convert the color image into a black and white binary image using image processing techniques;

[0020] Step 2.2: Calculate the contact area ratio of the black and white binary image per square millimeter using the following formula:

[0021] Preferably, the specific implementation steps of step 3 are as follows:

[0022] Based on the contact area ratio of each unit calculated in step 2, select the threshold a. Whether there is contact or not is determined according to the following formula: A(i,j)%>a→contact, A(i,j)%≤a→non-contact.

[0023] Preferably, the specific implementation steps of step 4 are as follows:

[0024] The point cloud obtained by the 3D scanner is processed by the point cloud processing software to align the point clouds of the upper and lower plates on the structural surface to match the positions of the upper and lower surfaces, and the point cloud file is saved.

[0025] Preferably, the specific implementation steps of step 5 are as follows:

[0026] The point cloud is partitioned along two dimensions, horizontal and vertical, according to the following formula:

[0027] cell(i,j)={x,y,z|i≤x≤i+1,j≤y≤(j+1)}

[0028] In the formula: x, y, z are the three-dimensional coordinates of the point;

[0029] i,j∈[1,100] and are integers;

[0030] cell(i,j) represents the name and location of the point cloud grid region, and each cell contains the coordinates of several points.

[0031] Preferably, the specific implementation steps of step 6 are as follows:

[0032] Step 6.1:

[0033] Based on the aforementioned point cloud units, relevant input parameters are calculated. One function of the chamfer distance is as an indicator for evaluating the similarity of 3D point clouds, and it is calculated using the following formula:

[0034]

[0035] In the formula: cell1 and cell2 represent the point cloud units at corresponding positions on the upper and lower disks of the structure, respectively;

[0036] x is the three-dimensional coordinate of a point in the point cloud cell of cell1.

[0037] y is the three-dimensional coordinate of a point in the point cloud cell of cell2;

[0038] Step 6.2:

[0039] The average curvature of the points within each cell is calculated, and the maximum value is taken as the curvature of this region. First, a quadratic surface is fitted using the least squares method, and the partial derivatives of the surface are calculated: rx, ry, rxx, ryy, rxy. The average curvature H is then calculated using the following formula:

[0040]

[0041] In the formula: E = r x r x F = r x r y G = r y r y ;

[0042] L = r xx n;N=r yy n;M=r xy n;

[0043] n is the number of points in the point cloud;

[0044] Step 6.3:

[0045] After registering the point cloud, the initial aperture is obtained by taking the average of the z coordinates in each cell and then subtracting them.

[0046] Step 6.4:

[0047] For a point cloud surface that is smooth everywhere, any local point cloud can be fitted with a planar point cloud. The same applies to the point cloud of a local rock mass structure surface. For each point in the local point cloud, search for its k neighborhood points and fit a plane P using the least squares method, expressed as follows:

[0048] In the formula: N is the normal vector of plane P; d is the distance from the origin to P; P0 is the coordinate of the centroid of the local point cloud;

[0049] Taking N as the normal vector of the point, the normal vector of plane P is obtained by principal component analysis, i.e., PCA. P passes through its neighborhood centroid P0 and satisfies that the normal vector magnitude is 1. Perform singular value decomposition on M to obtain its eigenvalues. The normal vector corresponding to the smallest eigenvalue is the normal vector of P.

[0050] Preferably, the number of contacts and non-contacts is calculated and counted, and then associated with each cell, so that each cell corresponds to a result: contact or non-contact.

[0051] Preferably, the specific implementation steps of step 8 are as follows:

[0052] Step 8.1:

[0053] Build an artificial neural network and train it based on the data selected in step 7. Select the amount of contact and non-contact training based on the measured contact area ratio, with the amount being 80% of the number of cells.

[0054] Step 8.2:

[0055] Adjust the basic parameters of the network: number of training iterations, learning rate, training accuracy, and maximum number of failures, so that the classification accuracy of the artificial neural network reaches more than 80%.

[0056] Preferably, contact is recorded as "0" and non-contact as "1", the contact area ratio is calculated, and a binary image is drawn based on the prediction results.

[0057] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0058] This invention estimates the contact condition of structural surfaces based on point clouds of rock mass structural surfaces. Using point clouds to describe the morphology of structural surfaces can more accurately describe the geometric shape and geometric parameters of micro-protrusions on them, and the accuracy can be selected according to actual needs, thereby improving computational efficiency.

[0059] This invention uses pressure-sensitive paper obtained from experiments as training samples to estimate the current contact status of different structural surfaces. By dividing the pressure-sensitive paper into pixel partitions and corresponding them with point cloud coordinates, the training volume of the neural network is guaranteed, and more results can be predicted with a small number of experiments, thus reducing labor costs.

[0060] The calculation process of this invention can be transformed into a specific flow using a programming language, and is easy to develop into software, facilitating subsequent operations. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0062] Figure 1 This is a binarized image of the pressure-sensitive paper according to the present invention.

[0063] Figure 2 This is a flowchart illustrating the generation of training samples for this invention.

[0064] Figure 3 This is a point cloud map of the rock mass structure surface after registration according to the present invention.

[0065] Figure 4 This is a structural diagram of the artificial neural network of the present invention.

[0066] Figure 5This is a flowchart of the method of the present invention. Detailed Implementation

[0067] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0068] This invention provides, for example Figures 1 to 5 The method shown here for predicting the surface contact characteristics of rock mass structure based on artificial neural networks includes the following steps:

[0069] Step 1: Determine the uniaxial compression test of the structural surface to be predicted, and use pressure-sensitive paper to obtain the contact condition of the structural surface;

[0070] The specific implementation steps of step 1 are as follows: Select a rock mass structural surface with the same lithology to conduct a uniaxial compression test, and place pressure-sensitive paper between the structural surfaces. The normal loading speed is 0.07 kN / s. When the normal stress reaches the target value, maintain it for half a minute and remove the pressure-sensitive paper. Then upload the data to the computer.

[0071] Step 2: Binarize the obtained pressure-sensitive paper, pack the pixels of one square millimeter orthogonal, and calculate the contact area ratio of each square millimeter.

[0072] The specific implementation steps of step 2 are as follows:

[0073] Step 2.1: Use image processing techniques to convert the color image into a black and white binary image. The processed image is as follows: Figure 1 As shown, this operation can be performed using software such as MATLAB and ImageJ;

[0074] Step 2.2: Calculate the contact area ratio of the black and white binary image per square millimeter using the following formula:

[0075] Step 3: Set a reference threshold. If the reference threshold exceeds the set value, it is defined as contact; otherwise, it is defined as non-contact.

[0076] The specific implementation steps of step 3 are as follows:

[0077] Based on the contact area ratio of each unit calculated in step 2, a threshold 'a' is selected. Whether or not there is contact is determined according to the following formula: A(i,j)% > a → contact; A(i,j)% ≤ a → non-contact. The processing procedure is as follows: Figure 2 As shown;

[0078] Step 4: Perform preliminary processing on the point cloud of the upper and lower disks of the rock mass structure, including noise reduction and coarse registration.

[0079] The specific implementation steps of step 4 are as follows:

[0080] Point clouds acquired using a 3D scanner, such as Figure 1 As shown, point cloud processing software, such as CloudCompare or Geomagic Studio, is used to perform denoising, encapsulation, and removal of isolated points. The point clouds on the upper and lower disks of the structural surface are then registered to match their positions. The point cloud file is saved, and the processed point cloud is shown below. Figure 3 As shown;

[0081] Step 5: Divide the point cloud into units at intervals of one millimeter along the horizontal and vertical coordinates, and associate each point cloud unit with a pixel unit from Steps 1-4.

[0082] The specific implementation steps of step 5 are as follows:

[0083] The point cloud is partitioned along two dimensions, horizontal and vertical, as shown in the figure. The partitioning is based on the following formula:

[0084] cell(i,j)={x,y,z|i≤x≤i+1,j≤y≤(j+1)}

[0085] In the formula: x, y, z are the three-dimensional coordinates of the point;

[0086] i,j∈[1,100] and are integers;

[0087] cell(i,j) is the name and location of this point cloud grid region, and each cell contains the coordinates of several points;

[0088] Step 6: Calculate the geometric parameters within each point cloud unit, including chamfer distance, average curvature, components of the normal vector, and initial aperture;

[0089] The specific implementation steps of step 6 are as follows:

[0090] Step 6.1:

[0091] Based on the aforementioned point cloud units, relevant input parameters are calculated. One function of the chamfer distance is as an indicator for evaluating the similarity of 3D point clouds. It is calculated using the following formula:

[0092]

[0093] In the formula: cell1 and cell2 represent the point cloud units at corresponding positions on the upper and lower disks of the structure, respectively;

[0094] x is the three-dimensional coordinate of a point in the point cloud cell of cell1.

[0095] y is the three-dimensional coordinate of a point in the point cloud cell of cell2;

[0096] Step 6.2:

[0097] The average curvature of the points within each cell is calculated, and the maximum value is taken as the curvature of this region. A quadratic surface is first fitted using the least squares method, and the partial derivative of the surface is calculated: r x r y r xx r yy r xy And calculate the mean curvature H according to the following formula:

[0098]

[0099] In the formula: E = r x r x F = r x r y G = r y r y ;

[0100] L = r xx n;N=r yy n;M=r xy n;

[0101] n is the number of points in the point cloud;

[0102] Step 6.3:

[0103] After registering the point cloud, the initial aperture is obtained by taking the average of the z coordinates in each cell and then subtracting them.

[0104] Step 6.4:

[0105] For a point cloud surface that is smooth everywhere, any local point cloud can be fitted with a planar point cloud. The same applies to the point cloud of a local rock mass structure surface. For each point in the local point cloud, search for its k neighborhood points and fit a plane P using the least squares method, expressed as follows:

[0106] In the formula: N is the normal vector of plane P; d is the distance from the origin to P; P0 is the coordinate of the centroid of the local point cloud;

[0107] Taking N as the normal vector of the point, the normal vector of plane P is obtained by principal component analysis, i.e., PCA. P passes through its neighborhood centroid P0 and satisfies that the normal vector magnitude is 1. Perform singular value decomposition on M to obtain its eigenvalues. The normal vector corresponding to the smallest eigenvalue is the normal vector of P.

[0108] Step 7: Based on the initial contact ratio, filter between contact and non-contact, and set up training and test sets;

[0109] The number of contacts and non-contacts is calculated and counted, and then associated with each cell, so that each cell corresponds to a result: contact or non-contact.

[0110] Step 8: Build an artificial neural network, train it on the data, and adjust the training parameters until the accuracy of the test results reaches 80% and then stop the adjustment.

[0111] The specific implementation steps of step 8 are as follows:

[0112] Step 8.1:

[0113] Artificial neural networks can be built using software such as Python or MATLAB, and their structure is as follows: Figure 4 As shown, training is performed based on the data selected in step 7. The amount of contact and non-contact training is selected according to the measured contact area ratio. The ratio of the two should be similar to that of the contact area ratio, and the number should be 80% of the number of cells.

[0114] Step 8.2:

[0115] Adjust the basic parameters of the network: number of training iterations, learning rate, training accuracy, and maximum number of failures, so that the classification accuracy of the artificial neural network can reach more than 80%.

[0116] Step 9: Use the new point cloud to make predictions through this artificial neural network and draw a binarized image;

[0117] Contact is denoted as "0" and non-contact as "1". The contact area ratio is calculated, and a binary image is plotted based on the prediction results.

[0118] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the surface contact characteristics of rock mass structures based on artificial neural networks, characterized in that, Includes the following steps: Step 1: Determine the uniaxial compression test of the structural surface to be predicted, and use pressure-sensitive paper to obtain the contact condition of the structural surface; Step 2: Binarize the obtained pressure-sensitive paper, pack the pixels of one square millimeter orthogonal, and calculate the contact area ratio of each square millimeter. Step 3: Set a reference threshold. If the reference threshold exceeds the set value, it is defined as contact; otherwise, it is defined as non-contact. Step 4: Perform preliminary processing on the top and bottom point clouds of the rock mass structure surface; Step 5: Divide the point cloud into units at intervals of one millimeter along the horizontal and vertical coordinates, and associate each point cloud unit with a pixel unit from Steps 1-4. Step 6: Calculate the geometric parameters within each point cloud unit; Step 7: Based on the initial contact ratio, filter between contact and non-contact, and set up training and test sets; Step 8: Build an artificial neural network, train it on the data, and adjust the training parameters until the accuracy of the test results reaches 80% and then stop the adjustment. Step 9: Use the new point cloud to make predictions through this artificial neural network and draw a binarized image.

2. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The specific implementation steps of step 1 are as follows: Select a rock mass structural surface with the same lithology to conduct a uniaxial compression test, and place pressure-sensitive paper between the structural surfaces. The normal loading speed is 0.07 kN / s. When the normal stress reaches the target value, maintain it for half a minute and remove the pressure-sensitive paper. Then upload the data to the computer.

3. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The specific implementation steps of step 2 are as follows: Step 2.1: Convert the color image into a black and white binary image using image processing techniques; Step 2.2: Calculate the contact area ratio of the black and white binary image per square millimeter using the following formula:

4. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 3, characterized in that, The specific implementation steps of step 3 are as follows: Based on the contact area ratio of each unit calculated in step 2, select the threshold a. Whether there is contact or not is determined according to the following formula: A(i,j)%>a→contact, A(i,j)%≤a→non-contact.

5. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The specific implementation steps of step 4 are as follows: The point cloud obtained by the 3D scanner is processed by the point cloud processing software to align the point clouds of the upper and lower plates on the structural surface to match the positions of the upper and lower surfaces, and the point cloud file is saved.

6. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The specific implementation steps of step 5 are as follows: The point cloud is partitioned along two dimensions, horizontal and vertical, according to the following formula: cell(i,j)={x,y,z|i≤x≤i+1,j≤y≤(j+1)} In the formula: x, y, z are the three-dimensional coordinates of the point; i,j∈[1,100] and are integers; cell(i,j) represents the name and location of the point cloud grid region, and each cell contains the coordinates of several points.

7. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The specific implementation steps of step 6 are as follows: Step 6.1: Based on the aforementioned point cloud units, relevant input parameters are calculated. One function of the chamfer distance is as an indicator for evaluating the similarity of 3D point clouds, and it is calculated using the following formula: In the formula: cell1 and cell2 represent the point cloud units at corresponding positions on the upper and lower disks of the structure, respectively; x is the three-dimensional coordinate of a point in the point cloud cell of cell1. y is the three-dimensional coordinate of a point in the point cloud cell of cell2; Step 6.2: The average curvature of the points within each cell is calculated, and the maximum value is taken as the curvature of this region. First, a quadratic surface is fitted using the least squares method, and the partial derivatives of the surface are calculated: rx, ry, rxx, ryy, rxy. The average curvature H is then calculated using the following formula: In the formula: E = r x r x F = r x r y G = r y r y ; L=r xx n;N=r yy n;M=r xy n; n is the number of points in the point cloud; Step 6.3: After registering the point cloud, the initial aperture is obtained by taking the average of the z coordinates in each cell and then subtracting them. Step 6.4: For a point cloud surface that is smooth everywhere, any local point cloud can be fitted with a planar point cloud. The same applies to the point cloud of a local rock mass structure surface. For each point in the local point cloud, search for its k neighborhood points and fit a plane P using the least squares method, expressed as follows: In the formula: N is the normal vector of plane P; d is the distance from the origin to P; P0 is the coordinate of the centroid of the local point cloud; Taking N as the normal vector of the point, the normal vector of plane P is obtained by principal component analysis, i.e., PCA. P passes through its neighborhood centroid P0 and satisfies that the normal vector magnitude is 1. Perform singular value decomposition on M to obtain its eigenvalues. The normal vector corresponding to the smallest eigenvalue is the normal vector of P.

8. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, The number of contacts and non-contacts is calculated and counted, and then associated with each cell, so that each cell corresponds to a result: contact or non-contact.

9. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 8, characterized in that, The specific implementation steps of step 8 are as follows: Step 8.1: Build an artificial neural network and train it based on the data selected in step 7. Select the amount of contact and non-contact training based on the measured contact area ratio, with the amount being 80% of the number of cells. Step 8.2: Adjust the basic parameters of the network: number of training iterations, learning rate, training accuracy, and maximum number of failures, so that the classification accuracy of the artificial neural network reaches more than 80%.

10. The method for predicting surface contact characteristics of rock mass structure based on artificial neural networks according to claim 1, characterized in that, Contact is denoted as "0" and non-contact as "1". The contact area ratio is calculated, and a binary image is plotted based on the prediction results.

Images