Rock structural surface shear failure area and type identification method

Abstract

The invention discloses a rock structural surface shear failure area and type identification method, which comprises the following steps: acquiring rock joint surface information after numerical simulation direct shear test under different test conditions, performing normalization processing on node coordinates of joint surface grids and numerical variables in test condition parameters, and obtaining a joint surface shear failure area and type identification result; performing one-hot coding on classification variables in the test condition parameters, and splicing the numerical variables and one-hot coding vectors to form parameter feature vectors of preset dimensions; converting node coordinates of the joint surface grids into three-dimensional voxel coordinates and then generating a voxel label matrix; expanding the parameter feature vectors, and injecting the expanded parameter feature vectors into a constructed 3D convolutional network model through channel splicing; training the model by adopting a cross entropy loss function, and verifying the model to obtain an optimal training model; and preprocessing the joint surface to be predicted, inputting the preprocessed joint surface to the optimal training model, and outputting a voxel-level damage area and type prediction result, thereby improving the recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of layered rock slope stability analysis, and in particular to a method for identifying shear failure areas and types of rock structural surfaces. Background Art

[0002] In geological hazard assessments of layered rock slopes, rock structural surface morphology is a key factor influencing the shear strength of jointed rock masses. Accurately identifying potential failure zones and types is crucial for improving the accuracy of shear strength predictions. Traditional methods rely on indoor direct shear tests and numerical simulations to identify failure zones, but these methods suffer from low efficiency, high cost, and poor timeliness, and also make it difficult to identify failure types.

[0003] Currently, existing patents provide methods for automatically identifying shear failure zones on structural surfaces. For example, Chinese patent publication number CN113284092A discloses a method and device for automatically identifying shear failure zones on structural surfaces based on feature matching. The method's main steps include: acquiring images P1 and P2 of the rock mass structural surface before and after a direct shear test; aligning P1 and P2 using the ORB algorithm; converting the registered images into grayscale images P'1 and P'2; and calculating a difference image P from the grayscale images. The OSTU algorithm is used to convert the difference image into a binary image, and a failure zone label map PL is generated based on an 8-connection analysis method. The perimeter and area of each failure zone are then calculated from the PL.

[0004] This method relies on image comparison after direct shear testing, which is a post-analysis approach. It cannot proactively predict potential damage areas and types based on rock mass conditions and stress states at the construction site, making it difficult to meet the need for advanced geological disaster warning. Furthermore, this method relies on two-dimensional image analysis, making it difficult to characterize three-dimensional damage characteristics and unable to predict damage types.

[0005] Therefore, how to quickly and accurately identify the shear failure area and failure type of rock structural surfaces is a technical problem that needs to be solved urgently in this field. Summary of the Invention

[0006] In order to solve the above technical problems, this application proposes the following technical solutions:

[0007] The present application provides a method for identifying shear failure areas and types of rock structural surfaces, including:

[0008] Obtaining rock joint surface information after numerical simulation direct shear tests under different test conditions, wherein the rock joint surface information includes node coordinates, unit states, and test condition parameters of the joint surface mesh;

[0009] Normalizing the node coordinates of the joint surface grid and multiple numerical variables in the test condition parameters, and performing one-hot encoding on multiple categorical variables in the test condition parameters to form multiple categorical variable vectors;

[0010] Multiple numerical variables in the normalized experimental condition parameters are concatenated with multiple categorical variable vectors after one-hot encoding to form an experimental condition parameter feature vector of a preset dimension;

[0011] Set the voxel resolution parameters, convert the node coordinates of the normalized joint surface mesh into three-dimensional voxel coordinates, and then generate the voxel label matrix;

[0012] After expanding the experimental condition parameter feature vector to the same dimension as the voxel label matrix, it is injected into the first encoder layer of the pre-built 3D convolutional network model through channel splicing;

[0013] The model is trained using a cross entropy loss function, the loss is calculated only for valid voxels, and the model is verified based on macro-F1 to obtain the best training model;

[0014] The joint surface to be predicted is pre-processed and then input into the optimal training model, and the prediction results of the voxel-level damage area and type are output.

[0015] In a possible implementation, obtaining rock joint surface information after numerical simulation direct shear tests under different test conditions includes:

[0016] A three-dimensional numerical model of a rock joint surface specimen is established based on numerical software;

[0017] Tetrahedron or hexahedron grids are used to divide the unit body and multiple groups of direct shear tests under different test conditions are carried out;

[0018] After the test is completed, the rock joint surface information under each set of test conditions is extracted to obtain the rock joint surface information after the numerical simulation direct shear test under different test conditions.

[0019] In a possible implementation, the calculation formula for normalizing the node coordinates of the joint surface grid is:

[0020]

[0021] Among them, x′ is the normalized coordinate value; x is the original coordinate value; x max and x min are the maximum and minimum values of the node coordinates, respectively.

[0022] In one possible implementation, multiple numerical variables in the normalized test condition parameters are concatenated with multiple categorical variable vectors after one-hot encoding to form a test condition parameter feature vector of a preset dimension. The calculation formula is:

[0023]

[0024] One-hot encoding is:

[0025]

[0026] in, are k normalized numerical variables; c j For each categorical variable, the number of categories is generated as vectors after one-hot encoding. m is the number of categories of a categorical variable; R is a set of real numbers.

[0027] In one possible implementation, the calculation formula for converting the node coordinates of the normalized joint surface mesh into three-dimensional voxel coordinates is:

[0028] v=[x′×N],N∈Z +

[0029] Where x′ is the normalized coordinate value; N is the resolution parameter of the voxel grid; v is the three-dimensional voxel coordinate; Z + is a set of positive integers.

[0030] In one possible implementation, after the experimental condition parameter feature vector is expanded to the same dimension as the voxel label matrix, it is injected into the first layer of the encoder of the pre-built 3D convolutional network model through channel splicing. The calculation formula is:

[0031]

[0032] Where V∈Z D×H×W is the voxel label matrix; For tensor broadcast; 1 D×H×W It is a tensor of all 1s; D, H, and W are the number of voxels in the z, y, and x-axis directions respectively.

[0033] In one possible implementation, the cross entropy loss function is used to train the model, and the loss is calculated only for valid voxels. The calculation formula is:

[0034]

[0035] Among them, N v is the total number of valid voxels; i is the index of the valid voxel; y i,c is the one-hot encoding of the true label; pi,c is the Softmax probability of the model output; c is the unit state category.

[0036] In one possible implementation, the calculation formula for macro-F1 is:

[0037]

[0038] Among them, Pc is the precision of unit state category c; Rc is the recall rate of unit state category c.

[0039] In a possible implementation, the pre-processing of the joint surface to be predicted is input into the optimal training model, and the output of the voxel-level damage area and type prediction results includes:

[0040] After normalizing the node coordinates of the mesh of the joint surface to be predicted, the normalized node coordinates of the mesh of the joint surface to be predicted are converted into three-dimensional voxel coordinates by setting the voxel resolution parameter to generate a voxel label matrix in the test phase;

[0041] Inputting the voxel label matrix of the test phase and the normalized numerical variables in the experimental condition parameters into the optimal training model, and outputting a voxel-level prediction matrix;

[0042] Extracting voxel position labels in the voxel-level prediction matrix;

[0043] The mode of vertex labels of each element is counted as the failure type.

[0044] In a possible implementation, counting the mode of vertex labels of each unit as the damage type includes:

[0045] If multiple labels appear the same number of times, the category with the smaller value is given priority;

[0046] If all tags are invalid, the destruction type is marked as not destroyed.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] This application achieves efficient fusion of three-dimensional geometric data and physical test parameters through voxelization and conditional parameter injection mechanism, improves the adaptability of the model to complex working conditions, and at the same time ensures accurate alignment of the predicted results with the original geometric structure through mode statistics and inverse coordinate mapping, and has a high accuracy rate in identifying the type of damage. This application can be applied to the stability analysis of jointed rock masses with different lithologies and stress conditions, thereby improving engineering practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1A schematic flow chart of a method for identifying shear failure areas and types of rock structural surfaces provided in an embodiment of the present application;

[0050] Figure 2 A three-dimensional numerical model of a rock joint surface specimen provided in an embodiment of the present application;

[0051] Figure 3 Joint surface information under certain experimental conditions provided in the embodiments of this application;

[0052] Figure 4 The joint surface to be predicted provided in the embodiment of the present application;

[0053] Figure 5 The predicted potential damage areas and types provided in the embodiments of this application;

[0054] Figure 6 The FLAC3D numerical simulation results of the joint surface to be predicted are provided in the embodiments of this application. DETAILED DESCRIPTION

[0055] The present invention will be described below with reference to the accompanying drawings and specific implementation methods.

[0056] Figure 1 A flow chart of a method for identifying shear failure areas and types of rock structural surfaces provided in an embodiment of the present application is provided in FIG. Figure 1 In this embodiment, a method for identifying shear failure areas and types of rock structural surfaces includes:

[0057] S101, obtaining rock joint surface information after numerical simulation direct shear tests under different test conditions, wherein the rock joint surface information includes node coordinates, unit states, and test condition parameters of the joint surface mesh.

[0058] In this embodiment, a three-dimensional numerical model of a rock joint surface sample is established based on numerical software, and a tetrahedron or hexahedron grid is used to divide the unit body, and n groups of direct shear tests under different test conditions are carried out. The joint surface grid is generally composed of triangular units or quadrilateral units. For a grid composed of triangular units, the node coordinate storage format of the joint surface grid is the three-directional coordinate values of the three vertices of point A, point B and point C, and the specific format is: xA, yA, zA, xB, yB, zB, xC, yC, zC. For a grid composed of quadrilateral units, the node coordinate storage format of the joint surface grid is the three-directional coordinate values of the four vertices of point A, point B, point C and point D, and the specific format is: xA, yA, zA, xB, yB, zB, xC, yC, zC, xD, yD, zD. There are four types of unit states: undamaged, shear failure, tensile failure and combined failure, which are represented by the values 0, 1, 2 and 3 respectively.

[0059] See also Figure 2 The three-dimensional numerical model of the rock joint surface sample provided in the embodiment of the present application is divided into units using a hexahedral grid, and 20 groups of direct shear tests of rock joint surface samples with different normal stress, uniaxial compressive strength, tensile strength, and shear direction are carried out. After the test is completed, the joint surface information of each group of tests is extracted using the Fish language, including the node coordinates and unit status of the joint surface grid, and the joint surface information of each group of tests is stored in the file F i (i=1,2,…,20), see Figure 3 , the joint surface information under a certain test condition provided by the embodiment of the present application, in the figure, the joint surface grid is composed of quadrilateral units, and Table 1 is the storage format of each group of joint surface information.

[0060] Table 1 Storage format of each group of joint surface information

[0061]

[0062] S102 , normalizing the node coordinates of the joint surface mesh and the numerical variables in the test condition parameters, and performing one-hot encoding on the categorical variables in the test condition parameters to form a one-hot encoding vector.

[0063] In this embodiment, the node coordinates of the joint surface mesh, the three numerical variables in the test condition parameters (normal stress σ n , uniaxial compressive strength UCS and tensile strength σ T ) all need to be normalized. The normalization formula is:

[0064]

[0065] Among them, x′ is the normalized coordinate value; x is the original coordinate value; x max and x min are the maximum and minimum values of the node coordinates, respectively.

[0066] The categorical variable (shear direction S) in the experimental condition parameter needs to be one-hot encoded. There are 4 categories of shear direction S, so the one-hot vector s onehot ∈{0,1} 4 for:

[0067]

[0068] S103: Concatenate the numerical variables in the normalized test condition parameters with the one-hot encoding vector to form a test condition parameter feature vector of a preset dimension.

[0069] In this embodiment, the numerical variables in the splicing test condition parameters and the shear direction one-hot encoding vector are concatenated to form a 7-dimensional test condition parameter feature vector p c , which has the form:

[0070] p c =[σ′ n ,UCS′,σ′ T ,s0,s1,s2,s3]∈R 7

[0071] Among them, σ′ n ,UCS′,σ′ T are the normalized values corresponding to the normal stress, uniaxial compressive strength and tensile strength of the rock wall, respectively, and s is a one-hot encoding vector.

[0072] Save the node coordinate normalization parameters and the normalization parameters of the numerical variables in the test condition parameters to the file scaler.pkl for reuse during the testing phase. See Table 2 for the storage format of the test condition parameters.

[0073] Table 2 Test condition parameter storage format

[0074] Joint surface information file name Normal stress / MPa Uniaxial compressive strength / MPa Tensile strength / MPa Shear direction <![CDATA[F1]]> 1.0 10 2.0 0 ... 1.5 20 2.5 1 <![CDATA[F 20 ]]> 3.0 30 3.0 3

[0075] S104 , setting voxel resolution parameters, converting the node coordinates of the normalized joint surface mesh into three-dimensional voxel coordinates and generating a voxel label matrix.

[0076] In this embodiment, the voxel resolution parameter N is set to 64, and the node coordinates in the joint surface information are converted into three-dimensional voxel coordinates through normalization and discretization formulas to generate a voxel label matrix. Each element in the matrix stores the cell state label of the corresponding voxel (0: undamaged; 1: shear damage; 2: tensile damage; 3: combined damage). Voxels not occupied by joint surfaces are marked as -1. The discretization formula is:

[0077] v=[x′×N],N∈Z +

[0078] Where x′ is the normalized coordinate value; N is the resolution parameter of the voxel grid; v is the three-dimensional voxel coordinate; Z + is a set of positive integers.

[0079] S105: After expanding the experimental condition parameter feature vector to the same dimension as the voxel label matrix, it is injected into the first layer of the encoder of the pre-built 3D convolutional network model through channel splicing.

[0080] In this embodiment, a 3D convolutional network model with a parameter conditional injection mechanism is constructed, which includes an encoder and a decoder. The encoder is composed of multiple levels of downsampling convolutional layers, each of which includes three-dimensional convolution, batch normalization, and activation functions. The downsampling formula is:

[0081] F l =Conv3D(F l-1 , W l , stride = 2)

[0082] Among them, F l is the feature map of the lth layer; W l is the convolution kernel parameter of the lth layer; stride is the stride, and stride = 2 can halve the size of the feature map.

[0083] The decoder implements upsampling by the transposed convolution layer, and the formula is:

[0084] F l =TransposedConv3D(F l-1 ,W l , stride=2)

[0085] The experimental condition parameter feature vector is expanded to the same dimension as the voxel label matrix and injected into the first layer of the encoder through channel splicing. The formula is:

[0086]

[0087] Where V∈Z D×H×W is the voxel label matrix; Indicates tensor broadcast; 1 D×H×W It is a tensor of all 1s; D, H, and W are the number of voxels in the z, y, and x-axis directions respectively.

[0088] S106 , using a cross entropy loss function to train the model, calculating the loss only for valid voxels, and verifying the model based on macro-F1 to obtain the best training model.

[0089] In this embodiment, the cross entropy loss function is used to train the model, and the loss is calculated only for valid voxels. The calculation formula is:

[0090]

[0091] Among them, N v is the total number of valid voxels; i is the index of the valid voxel; y i,c is the one-hot encoding of the true label; p i,c is the Softmax probability of the model output; c is the unit state category.

[0092] Macro-F1 is used for model verification to obtain the best training model. The calculation formula of macro-F1 is:

[0093]

[0094] Among them, Pc is the precision of unit state category c; Rc is the recall rate of unit state category c

[0095] S107, pre-processing the joint surface to be predicted and inputting it into the optimal training model, and outputting the voxel-level damage area and type prediction results.

[0096] See also Figure 4 In this embodiment, Figure 4 The joint surface to be predicted in the training phase is used as the test set. The geometric coordinates are converted into the voxel label matrix of the test phase according to the same normalization parameters and discretization rules as the training phase. At the same time, the normalization parameters scaler.pkl saved during training are used to preprocess the test conditions and generate a 7-dimensional feature vector consistent with the training data format.

[0097] The voxel label matrix and the experimental condition parameter vector in the test phase are input into the optimal training model, and the voxel-level prediction matrix is output.

[0098] Extract the voxel position label from the voxel-level prediction matrix as the prediction result of the point;

[0099] label point =pred_grid[v x ,v y ,v z ]

[0100] For each unit, the majority of the unit vertex labels is counted as the unit's final damage type. The majority calculation rule is: if multiple labels appear the same number of times, the category with the smaller value is preferred. If all labels are invalid, it is marked as undamaged. Figure 5 The predicted potential damage areas and types.

[0101] See also Figure 6 The FLAC3D numerical simulation results of the joint surface to be predicted show that the predicted failure area and type are very consistent with the numerical results, and the prediction accuracy is over 90%.

[0102] In the embodiments of this application, "multiple" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean that A exists alone, A and B exist simultaneously, or B exists alone. A and B can be singular or plural. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0103] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0104] The above description is merely a specific embodiment of the present application. Any person skilled in the art may easily conceive of variations or substitutions within the technical scope disclosed in this application, and such variations or substitutions shall be within the scope of protection of this application. The scope of protection of this application shall be subject to the scope of protection of the claims.

Claims

1. A method for identifying shear failure areas and types of rock structural surfaces, characterized in that: include: Obtaining rock joint surface information after numerical simulation direct shear tests under different test conditions, wherein the rock joint surface information includes node coordinates, unit states, and test condition parameters of the joint surface mesh; Normalizing the node coordinates of the joint surface grid and multiple numerical variables in the test condition parameters, and performing one-hot encoding on multiple categorical variables in the test condition parameters to form multiple categorical variable vectors; Multiple numerical variables in the normalized experimental condition parameters are concatenated with multiple categorical variable vectors after one-hot encoding to form an experimental condition parameter feature vector of a preset dimension; Set the voxel resolution parameters, convert the node coordinates of the normalized joint surface mesh into three-dimensional voxel coordinates, and then generate the voxel label matrix; After expanding the experimental condition parameter feature vector to the same dimension as the voxel label matrix, it is injected into the first encoder layer of the pre-built 3D convolutional network model through channel splicing; The model is trained using a cross entropy loss function, the loss is calculated only for valid voxels, and the model is verified based on macro-F1 to obtain the best training model; The joint surface to be predicted is pre-processed and then input into the optimal training model, and the prediction results of the voxel-level damage area and type are output.

2. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The obtaining of rock joint surface information after numerical simulation direct shear tests under different test conditions includes: A three-dimensional numerical model of a rock joint surface specimen is established based on numerical software; Tetrahedron or hexahedron grids are used to divide the unit body and multiple groups of direct shear tests under different test conditions are carried out; After the test is completed, the rock joint surface information under each set of test conditions is extracted to obtain the rock joint surface information after the numerical simulation direct shear test under different test conditions.

3. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The calculation formula for normalizing the node coordinates of the joint surface grid is: Among them, x′ is the normalized coordinate value; x is the original coordinate value; x max and x min are the maximum and minimum values of the node coordinates, respectively.

4. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The calculation formula for concatenating multiple numerical variables in the normalized experimental condition parameters with multiple categorical variable vectors after one-hot encoding to form the experimental condition parameter feature vector of the preset dimension is: One-hot encoding is: in, are k normalized numerical variables; c j For each categorical variable, the number of categories is generated as vectors after one-hot encoding. m is the number of categories of a categorical variable; R is a set of real numbers.

5. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The calculation formula for converting the node coordinates of the normalized joint surface mesh into three-dimensional voxel coordinates is: v=[x′×N],N∈Z + Where x′ is the normalized coordinate value; N is the resolution parameter of the voxel grid; v is the three-dimensional voxel coordinate; Z + is a set of positive integers.

6. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: After the experimental condition parameter feature vector is expanded to the same dimension as the voxel label matrix, the calculation formula for injecting it into the first layer of the encoder of the pre-built 3D convolutional network model through channel splicing is: Where V∈Z D×H×W is the voxel label matrix; For tensor broadcast; 1 D×H×W It is a tensor of all 1s; D, H, and W are the number of voxels in the z, y, and x-axis directions respectively.

7. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The cross entropy loss function is used to train the model, and the loss is calculated only for valid voxels as follows: Among them, N v is the total number of valid voxels; i is the index of the valid voxel; y i,c is the one-hot encoding of the true label; p i,c is the Softmax probability of the model output; c is the unit state category.

8. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The calculation formula of macro-F1 is: Among them, Pc is the precision of unit state category c; Rc is the recall rate of unit state category c.

9. The method for identifying shear failure areas and types of rock structural surfaces according to claim 1, characterized in that: The joint surface to be predicted is pre-processed and then input into the optimal training model, and the voxel-level damage area and type prediction results are output, including: After normalizing the node coordinates of the mesh of the joint surface to be predicted, the normalized node coordinates of the mesh of the joint surface to be predicted are converted into three-dimensional voxel coordinates by setting the voxel resolution parameter to generate a voxel label matrix in the test phase; Inputting the voxel label matrix of the test phase and the normalized numerical variables in the experimental condition parameters into the optimal training model, and outputting a voxel-level prediction matrix; Extracting voxel position labels in the voxel-level prediction matrix; The mode of vertex labels of each element is counted as the failure type.

10. The method for identifying shear failure areas and types of rock structural surfaces according to claim 9, characterized in that: The method of counting the mode of vertex labels of each unit as the damage type includes: If multiple labels appear the same number of times, the category with the smaller value is given priority; If all tags are invalid, the destruction type is marked as not destroyed.

Images