Rock mass structure evaluation method based on siamese network

Abstract

The application discloses a rock mass structure evaluation method based on a Siamese network, fully utilizes an image recognition technology, constructs a rock mass structure standard image dataset, classifies images through a Siamese network model, then adopts a quantitative division method of the rock mass structure, accurately and quantitatively calculates the rock mass structure value through an output similarity value of the Siamese network, and through a normalization method and a high matching degree weight value calculation method. The quantitative evaluation method can more accurately classify the rock mass structure and provide more objective parameters for quantitative calculation of GSI. Based on the above technical scheme, the application can efficiently and accurately evaluate the rock mass structure.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and more specifically to a quantitative calculation method for the geological strength index (GSI) based on the Siamese network in the study of soil and rock stability in geotechnical engineering. Background Technology

[0002] In geotechnical engineering stability evaluation studies, determining the parameters of the engineering rock mass is crucial. The Geological Strength Index (GSI) is an important parameter reflecting rock mass quality and is directly related to the Hoek-Brown strength criterion parameters. GSI is a method for quantitatively analyzing the engineering geological information of rock masses. As an important parameter of the Hoek-Brown strength criterion, it enables the quantitative calculation of rock mass strength parameters. Compared with other surrounding rock classification methods, it is the only method that can directly determine the mechanical parameters of the rock mass. The GSI surrounding rock classification system was developed based on years of practical experience, aiming to modify the Hoek-Brown rock mass failure criterion and estimate the Hoek-Brown criterion parameters s, a, and m of the rock mass. b To correct the strength of rock masses under different geological conditions.

[0003] The GSI value of a rock mass is estimated based on its structural characteristics and surface conditions of structural surfaces. Hoek et al. classified rock mass structures into six categories: intact / massive, massive, fragmented, layered fractured / disturbed, fractured / loose, and sheet-like / shear-deformed. Surface conditions of structural surfaces were classified into five categories: very good, good, fair, poor, and very poor. The classification chart is shown below. Figure 1 As shown, accurate judgment of rock mass structural characteristics and surface conditions is crucial for obtaining GSI values ​​that closely approximate the actual rock mass when using GSI tables. This often relies on engineers' extensive experience and sound judgment, thus introducing a degree of subjectivity. This limitation restricts the widespread application of GSI. Therefore, reducing the subjectivity of GSI values ​​is of great significance for improving the accuracy of rock mass parameters.

[0004] In order to reduce the subjectivity of GSI value, experts and scholars have introduced quantitative evaluation indicators to convert GSI values. Common methods include: (1) Sonmez and Ulusay introduced a method based on the rock mass volume joint number J. v The “rock mass structural grade SR” and the “rock mass surface grade SCR” calculated based on roughness, weathering degree and filling condition are used to quantitatively describe GSI by matching SR value and SCR value and looking up a table (International Journal of Rock Mechanics and Mining Sciences 36 (1999) 743±760). (2) Cai and Kaiser introduced the block volume V band structural surface condition factors J c Quantitative description of GSI: the block volume Vb is calculated through joint spacing, and structural surface condition factors J... c (0~12) Calculated by roughness, weathering degree, and filling condition, and finally by V b Value and J c Value matching and table lookup are used to quantitatively describe GSI. (3) Hu Shengming and Hu Xiuwen introduced an improved method to calculate the volumetric joint number J of rock mass. v Calculate the structural surface condition factors J based on roughness, weathering degree, and infill condition. c This further quantifies GSI. A key point is the proposal to estimate the volume joint number J using the surface number of the debris mass and the average size of the mass (usually considered equal to the average joint spacing). v Statistical methods. (4) Wang Keyi and Xu Dongqiang introduced the rock mass integrity coefficient K from the highway tunnel design code. v The weathering coefficient K in geotechnical engineering investigation specifications f To quantitatively describe GSI, while K v The calculation also requires using the rock mass volume joint number J. v Let's convert it.

[0005] From the above common quantitative evaluation methods for reducing the subjectivity of GSI values, it can be seen that the quantitative conversion methods for structural surface conditions are largely consistent, namely, calculations based on roughness, weathering degree, and filling condition. While the quantitative conversion methods for rock mass structures differ in process, they all ultimately follow the same principle, using direct and indirect parameters to calculate the rock mass volume joint number J. v The conversion. In actual engineering, the volumetric joint number J of rock mass. v Obtaining the joint count requires dividing the survey area and manually measuring the spacing between each joint group and counting the number of joint groups. This method also has shortcomings. Firstly, manual measurement and statistics are inefficient and have poor safety in underground engineering. Secondly, for highly jointed, fractured, or folded rock masses, there are difficulties in joint identification, joint spacing statistics, and incomplete statistics, thus limiting the ability to obtain the number of joints J from rock mass volume. v Quantitative evaluation of the application of GSI.

[0006] Chinese invention patent application, publication number CN114676777A, discloses a self-supervised learning fine-grained image classification method based on Siamese networks. First, it utilizes an attention encoder to extract deep convolutional features from the image, obtaining an attention map containing semantic information, and then encodes image features using bilinear pooling. Second, it locates salient regions with high response values ​​on the attention map from the original image, performs cropping and erasing operations on these salient regions, thereby forming views from different perspectives, and learns view-invariant features in a self-supervised manner. Finally, it combines a center loss function and a consistency loss function to explicitly constrain the features from different perspectives, maintaining their intra-class consistency. This invention enables the network to achieve significant performance gains, significantly improving classification accuracy on the baseline of fine-grained image classification. Summary of the Invention

[0007] The purpose of this invention is to provide a rock mass structure evaluation method based on Siamese networks. Specifically, based on image recognition to obtain a network image of the rock mass structure surface, this method innovatively uses a Siamese network to identify the target rock mass structure category. Furthermore, in some preferred schemes, it can output a quantitative value (0-100) of the rock mass structure, which corresponds to the "Rock Mass Structure Grade (SR)" value proposed by Sonmez and Ulusay. The calculation method of this patent can efficiently and accurately quantitatively characterize the rock mass structure features in the GSI (Geometry Indicators).

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A rock mass structure evaluation method based on Siamese network includes the following steps:

[0010] A. Network image representation of rock mass structural surfaces;

[0011] B. Input the network image representation results into the trained twin neural network model for rock mass structure category recognition to quantitatively classify the rock mass structure, and output the similarity value between the predicted image and the standard images of various rock mass structures. Define the category to which the predicted image belongs as the rock mass structure with the largest similarity value.

[0012] The training steps for the trained twin neural network model for rock mass structure category recognition are as follows:

[0013] 1) Establish a standard image dataset of rock mass structures. The standard rock mass structures are categorized according to the six types of rock mass structures given in the geological intensity chart: intact / massive structure, massive structure, fragmented structure, layered fractured / disturbed structure, fractured / loose structure, and flaky /

[0014] Shear deformation structure;

[0015] 2) Input the standard image dataset of the rock mass structure into the Siamese network for training to obtain the trained twin neural network model for rock mass structure category recognition.

[0016] Some preferred embodiments of the present invention are as follows:

[0017] Preferably, in step A, the U-net structural surface recognition model with VGG16 as the backbone network is used to extract the structural surface network from the rock image through image segmentation. The extracted structural surface network is further processed by a skeletonization algorithm to obtain a skeleton image of the rock mass structural surface network.

[0018] Preferably, for the aforementioned steps, for images with poor extraction results, a threshold segmentation method is used to remove noise points in the rock mass structure network skeleton, thereby obtaining a structure network image that better represents the actual rock mass structure.

[0019] Preferably, in step B, the quantitative classification of the six types of rock mass structures includes the following steps:

[0020] ① Six types of rock mass structures are quantitatively assigned to six equal parts within the range of 0 to 100, and the structural surface category weight parameter Q = 100 / 12 = 8.33 is defined. Each category is represented by i (i = 1, 2, 3, 4, 5, 6), and the median is represented by M. i express:

[0021] Class 1, complete / massive structure (83.31–100, median M1 is 91.66);

[0022] Class 2, blocky structure (66.67–83.30, median M2 is 75);

[0023] Type 3, fragmented structure (50.01–66.66, median M3 is 58.33);

[0024] Type 4, layered fragmentation / disturbance structure (33.34–50, median M4 is 41.66);

[0025] Type 5, fragmented / dispersed structure (16.67–33.33, median M5 is 25);

[0026] Type 6, sheet-like / shear-deformed structures (0–16.66, median M6 number 8.33);

[0027] ② When the twin neural network identifies rock mass structural surfaces, it outputs six similarity values: S1, S2, S3, S4, S5, and S6.

[0028] ③ Find the maximum similarity S j=max(S1, S2, S3, S4, S5, S6), and its corresponding structural plane category j;

[0029] ④ Exclude S j from (S1, S2, S3, S4, S5, S6), and denote the remaining 5 similarity values as (P1, P2, P3, P4, P5);

[0030] ⑤ Normalize the above 5 similarity values:

[0031]

[0032] ⑥ Divide the normalized similarity values (N1, N2, N3, N4, N5) into two groups: classify N where i<j k into one group, assign a positive sign to the corresponding N k values, classify N where i>j k into another group, and assign a negative sign to the corresponding N k values;

[0033] ⑦ Sum all the normalized values assigned with positive and negative signs to obtain the weight coefficient W;

[0034] ⑧ Calculate the rock mass structure quantitative evaluation value V:

[0035] V=M j +Q×W j

[0036] wherein V is the rock mass structure quantitative evaluation value,

[0037] M j is the median of the corresponding structural plane category j,

[0038] Q is the structural plane category weight parameter, with a value of 8.33,

[0039] W j is the weight coefficient.

[0040] Preferably, in step 1 of the training process of the siamese neural network model, establishing the rock mass structure standard image dataset includes: intercepting corresponding rock mass structure images from the GSI chart, manually drawing rock mass structures consistent with the standard images, constructing discrete fracture network models consistent with various types of rock mass structures, and intercepting images consistent with the standard rock mass structures from the models.

[0041] Preferably, the convolutional layer of the siamese neural network model adopts a VGG16 convolutional neural network.

[0042] Preferably, the loss function of the siamese neural network model is a contrastive loss function based on Euclidean distance.

[0043] Preferably, the training step 2 of the Siamese neural network model includes the following steps:

[0044] ① Calculate the feature vectors of the two input samples using the VGG16 backbone network;

[0045] ② Calculate the Euclidean distance between the two samples in the feature space;

[0046] ③ Calculate the loss function and adjust the loss value based on the similarity of the samples;

[0047] ④ Optimize the neural network parameters through backpropagation and gradient descent to minimize the loss function and obtain the optimal model.

[0048] In summary, this invention provides a new scheme for the quantitative calculation of the Geological Strength Index (GSI), significantly reducing the subjectivity of GSI value determination and improving the accuracy of rock mass parameters. Traditional methods rely on engineers' subjective experience and judgment, while this invention employs twin neural network technology, enabling a more objective and accurate evaluation of rock mass structure. Furthermore, this invention, through image recognition technology, greatly improves efficiency and reduces manual measurement and statistical work, making it particularly suitable for situations involving fractured or complex rock masses. The method of this invention can also be applied to the quantitative evaluation of various types of rock mass structures, exhibiting broad applicability and improving the accuracy and safety of obtaining engineering rock mass parameters. Attached Figure Description

[0049] Figure 1 A classification diagram of surface conditions for soil and rock structures;

[0050] Figure 2 This is a flowchart of an embodiment of the present invention;

[0051] Figure 3 This is a dataset of rock mass structure surface images from an embodiment of the present invention;

[0052] Figure 4 This is a twin neural network for GSI rock mass structure classification based on VGG16 in this embodiment of the invention;

[0053] Figure 5 This is a schematic diagram of rock mass structure surface network identification in an embodiment of the present invention;

[0054] Figure 6 This is a data list showing the similarity between the rock mass structure surface network diagram and various standard rock mass structures in the embodiments of the present invention;

[0055] Figure 7 This is a schematic diagram of the twin neural network framework in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings.

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] Example 1

[0059] This embodiment obtains a network image of rock mass structural surfaces using image recognition technology. This technology can intuitively represent the characteristics of rock mass structural surfaces through images, and the method is unaffected by the degree of development of structural surfaces, the degree of rock mass fragmentation, or the degree of rock mass deformation. Furthermore, based on the network image of structural surfaces obtained through image recognition, a standard image dataset of various rock mass structures is established, and a Siamese network model is trained. The model achieves accurate identification of the network image of rock mass structural surfaces, effectively reducing the subjectivity of conventional manual comparison methods and improving the accuracy of rock mass structure classification.

[0060] The method in this embodiment is as follows: Figure 2 The steps shown are as follows:

[0061] 1. Construction of a standard image dataset for rock mass structures

[0062] Standard rock mass structure images refer to the six types of rock mass structure images given in the standard Geological Strength Index (GSI) chart: intact / massive structure, massive structure, fragmented structure, layered fractured / disturbed structure, fractured / loose structure, and sheet-like / shear-deformed structure. There are three main methods for collecting these six types of images:

[0063] (1) The most direct method is to directly extract the corresponding rock mass structure image from the GSI chart. The rock mass structure images obtained by this method are standard images. However, the number of standard images obtained is small. However, the number of samples can be increased by rotating, scaling, cropping and flipping.

[0064] (2) The second method is to manually draw rock mass structures that are consistent with standard images. The drawing requirements are to keep the line density constant and the number and shape errors small. This method can construct a certain number of rock mass structure images, but the efficiency is low.

[0065] (3) The third method is to construct a discrete fracture network model consistent with various rock mass structures, and extract a large number of images consistent with standard rock mass structures from the model. Among them, the discrete fracture network is a three-dimensional network model used to describe joints, fissures and faults in rocks. By inputting and adjusting the dip, dip angle, length and distribution parameters of the structural surface, a discrete fracture network model close to the standard rock mass structure can be established.

[0066] Training a Siamese network model relies on a large amount of image data. The three methods described above can be used to quickly and efficiently acquire a large number of standard images of six types of rock mass structures. The constructed image dataset is as follows: Figure 3 As shown.

[0067] 2. Construction of the Siamese network model

[0068] A Siamese network-based model framework is suitable for GSI rock mass structure classification, such as... Figure 7 As shown.

[0069] Establish a feature extraction network VGG16, such as Figure 4 As shown, VGG16 is a classic deep convolutional neural network widely used for image recognition and feature extraction. It consists of 16 convolutional and fully connected layers, effectively extracting high-level features from images. This invention uses a VGG16 backbone feature extraction network (with the global average pooling and classification layers of the VGG16 model removed) as two identical sub-networks in a Siamese neural network, sharing the same weights. Each sub-network receives an input image of a rock mass structure and extracts features through the backbone feature extraction network (VGG16), resulting in two sets of one-dimensional feature vectors.

[0070] We choose the contrastive loss function. The goal of contrastive loss is to make samples from the same class (positive samples) closer together in the feature space, while separating samples from different classes (negative samples). This helps ensure that similar samples in the feature space converge, while dissimilar samples are separated. We select two sets of one-dimensional feature vectors (X... a X b The Euclidean distance between these two samples is expressed as... Contrastive Loss can then be expressed as:

[0071]

[0072] Where Y represents (X) a X b ) Whether it matches, 1 for a match and 0 for no match, m is the set safety distance, when (X a X b When the distance of X is less than m, the Contrastive Loss will become 0, which makes X a and X b Similarity, rather than identicality, ensures the generalization ability of the algorithm.

[0073] Model training steps

[0074] ① Calculate the feature vectors of the two input samples using the VGG16 backbone network;

[0075] ② Calculate the Euclidean distance between the two samples in the feature space;

[0076] ③ Calculate the contrastive loss and adjust the loss value based on the similarity of the samples;

[0077] ④ Optimize the neural network parameters through backpropagation and gradient descent to minimize the contrastive loss and obtain the optimal model.

[0078] This lays the model foundation for the next step of rock mass structure identification and quantitative evaluation.

[0079] 3. Network image representation of rock mass structural surfaces

[0080] To address the issues of low efficiency and poor security in manually measuring and statistically analyzing structural surface parameters, this invention constructs a U-net structural surface recognition model with VGG16 as the backbone network. The structural surface network is extracted from the rock image using image segmentation. A skeletonization algorithm is then applied to process the extracted network, obtaining a skeleton image of the rock mass structural surface network. For images with poor extraction results, threshold segmentation can be used to remove noise from the rock mass structural surface network skeleton, thus obtaining a structural surface network image that better represents the actual rock mass structure. This lays a data foundation for the subsequent identification and quantitative evaluation of the rock mass structure. An example of representing a rock mass image according to the method of this embodiment is shown below. Figure 5 As shown.

[0081] 4. Rock mass structure identification and quantitative evaluation

[0082] (1) Rock mass structure identification

[0083] After obtaining a high-performance twin neural network model for rock mass structure category recognition through model training in point 2, the network image of the structural surface (predicted image) obtained through image recognition in point 3 is input into the twin neural network model. The output result is the similarity between the predicted image and the standard images of 6 types of rock mass structures. The similarity value is 0 to 1. The larger the value, the more similar the predicted image is to that type of rock mass structure. Then, the category to which the predicted image belongs is defined as the rock mass structure with the highest similarity value.

[0084] (2) Quantitative evaluation of rock mass structure

[0085] In GSI quantitative evaluation, the quantitative classification of six types of rock mass structures is a key point. Based on the above rock mass structure identification, in order to achieve the best results in the quantitative classification of rock mass structures, this invention provides a new method for the quantitative classification of rock mass structures:

[0086] Six types of rock mass structures were quantitatively assigned values ​​in a range of 0 to 100, divided into six equal parts. These values ​​correspond to the "Rock Mass Structure Grade (SR)" proposed by Sonmez and Ulusay, and can be applied to the quantitative characterization of rock mass structure characteristics. Furthermore, they can be accurately and reliably used for calculating the Geological Strength Index (GSI) value. In addition, a structural surface category weight parameter Q = 100 / 12 = 8.33 was defined, with each category represented by i (i = 1, 2, 3, 4, 5, 6), and the median represented by M. i express,:

[0087] Class 1 - Complete / massive structures (83.31–100, median M1 is 91.66)

[0088] Type 2 - Massive structure (66.67–83.30, median M2 is 75)

[0089] Type 3 - fragmented structures (50.01–66.66, median M3 58.33)

[0090] Type 4 - Layered fragmentation / disturbance structure (33.34–50, median M4 is 41.66)

[0091] Type 5 - fragmented / dispersed structures (16.67–33.33, median M5 is 25)

[0092] Type 6 - sheet-like / shear-deformed structures (0–16.66, median M6 number 8.33)

[0093] ② When the twin neural network identifies rock mass structural surfaces, it outputs six similarity values: S1, S2, S3, S4, S5, and S6.

[0094] ① Find the maximum similarity S j =max(S1, S2, S3, S4, S5, S6), and its corresponding structural surface category j;

[0095] ②S j After removing from (S1, S2, S3, S4, S5, S6), the remaining 5 similarity values ​​are denoted as (P1, P2, P3, P4, P5).

[0096] ⑤ Normalize the above 5 similarity values:

[0097]

[0098] ⑥ Divide the normalized similarity values (N1, N2, N3, N4, N5) into two groups: the N where i<j k is classified as one group, and the corresponding N k value is assigned a positive sign, and the N where i>j k is classified as the other group, and the corresponding N k value is assigned a negative sign;

[0099] ⑦ Sum all the normalized values assigned with positive and negative signs to obtain the weight coefficient W j ;

[0100] ⑧ Calculate the rock mass structure quantitative evaluation value V:

[0101] V=M j +Q×W j

[0102] V is the rock mass structure quantitative evaluation value

[0103] M j is the median of the corresponding discontinuity category j

[0104] Q is the discontinuity category weight parameter, with a value of 8.33

[0105] W j is the weight coefficient.

[0106] According to the above method, input the constructed rock mass structure standard image dataset into the VGG16-based siamese neural network to carry out model training tasks, and obtain the VGG16-based GSI rock mass structure category discrimination siamese neural network;

[0107] Input one discontinuity network image obtained by the U-net discontinuity recognition model (as shown in the discontinuity grid on the left side for example) Figure 6 into the above-mentioned siamese neural network to perform prediction, and output the similarity between the image and each type of standard rock mass structure respectively, as shown by the data on the right side in the table Figure 6 .

[0108] According to Figure 6 the data in the table, find the maximum similarity value S j =max(0.1,0.784,0.043,0.431,0.009,0.291)=0.784(S2), judge that the discontinuity category is the second type of block structure, and the median value M2 of the rock mass structure of its category is 75;

[0109] Eliminate S2 to obtain the remaining 5 category probabilities P=(0.1,0.043,0.431,0.009,0.291), which are normalized to obtain N=(0.114, 0.049, 0.493, 0.010, 0.333);

[0110] (5) Assign positive signs to the normalized values ​​corresponding to rock mass structure types below the second category (i.e., blocky structure) and negative signs to those above, i.e. (0.114, -0.049, -0.493, -0.010, -0.333), and sum them to obtain the weight coefficient W2 = 0.114-0.049-0.493-0.010-0.333 = -0.771;

[0111] (6) Calculate the quantitative evaluation value of rock mass structure V=M2+Q×W2=75+8.33×(-0.771)=68.58.

[0112] In Sonmez and Ulusay’s “Modified Geological Strength Index (GSI) and its applicability to slope stability”, the quantitative evaluation value of rock mass structure obtained by step (6) above, 68.58, can be used to obtain the geological strength index GSI by combining the structural surface characteristic value.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rock mass structure evaluation method based on Siamese network, comprising the following steps: A. Network image representation of rock mass structural surfaces; B. Input the network image representation results into the trained twin neural network model for rock mass structure category recognition to quantitatively classify the rock mass structure, and output the similarity value between the predicted image and the standard images of various rock mass structures. Define the category to which the predicted image belongs as the rock mass structure with the largest similarity value. in, The training steps for the trained twin neural network model for rock mass structure category recognition are as follows: 1) Establish a standard image dataset of rock mass structure. The standard images of rock mass structure are divided into six types of rock mass structure given in the geological intensity chart, namely, intact / massive structure, blocky structure, fragmented structure, layered fractured / disturbed structure, fractured / loose structure, and sheet-like / shear-deformed structure. The quantitative classification of the six types of rock mass structures includes the following steps: ① The six types of rock mass structures are quantitatively assigned to six equal parts within the range of 0 to 100, and the structural surface category weight parameter Q = 100 / 12 = 8.33 is defined. Each category is represented by i, i∈{1,2,3,4,5,6}, and the median is represented by M. i express, ② When the twin neural network identifies rock mass structural surfaces, it outputs six similarity values: S1, S2, S3, S4, S5, and S6. ③ Find the maximum similarity value Sj=max(S1, S2, S3, S4, S5, S6) and its corresponding structural surface category j; ④ After removing Sj from (S1, S2, S3, S4, S5, S6), the remaining 5 similarity values ​​are denoted as (P1, P2, P3, P4, P5). ⑤ Normalize the above 5 similarity values: ，k∈{1,2,3,4,5}, in, This represents the normalized similarity value of the k-th remaining similarity, where k represents the index number of the similarity value; ⑥ Divide the normalized similarity values (N1, N2, N3, N4, N5) into two groups: the N where i<j k are classified as one group, and the corresponding N k values are assigned a positive sign, the N where i>j k are classified as another group, and the corresponding N k values are assigned a negative sign; ⑦ The weight coefficients are obtained by summing all normalized values ​​assigned positive and negative signs. ; ⑧ Calculate the quantitative evaluation value V of the rock mass structure: Where V is the quantitative evaluation value of rock mass structure, and M j Let W be the median of the corresponding structural surface category j, Q be the structural surface category weight parameter with a value of 8.33, and W be the median of the corresponding structural surface category j. j These are the weighting coefficients; 2) Input the standard image dataset of the rock mass structure into the Siamese network for training to obtain the trained twin neural network model for rock mass structure category recognition.

2. The rock mass structure evaluation method based on Siamese network according to claim 1, characterized in that, It also includes the U-net structural surface recognition model with VGG16 as the backbone network in step A, which extracts the structural surface network from the rock image through image segmentation, and further processes the extracted structural surface network using a skeletonization algorithm to obtain the rock mass structural surface network skeleton image.

3. The rock mass structure evaluation method based on Siamese network according to claim 2, characterized in that, Noise points in the rock mass structure surface network skeleton are removed by threshold segmentation.

4. The rock mass structure evaluation method based on Siamese network according to claim 1, characterized in that, In step 1 of the training steps of the Siamese neural network model, the establishment of a standard image dataset of rock mass structure includes: extracting the corresponding rock mass structure image from the GSI chart, manually drawing the rock mass structure consistent with the standard image, constructing a discrete fracture network model consistent with various rock mass structures, and extracting the image consistent with the standard rock mass structure from the model.

5. The rock mass structure evaluation method based on Siamese network according to claim 1, characterized in that, The convolutional layers of the twin neural network model use a VGG16 convolutional neural network.

6. The rock mass structure evaluation method based on Siamese network according to claim 1, characterized in that, The loss function of the twin neural network model is a contrastive loss function based on Euclidean distance.

7. The rock mass structure evaluation method based on Siamese network according to claim 1, characterized in that, The training step 2 of the twin neural network model includes the following steps: ① Calculate the feature vectors of the two input samples using the VGG16 backbone network; ② Calculate the Euclidean distance between the two samples in the feature space; ③ Calculate the loss function and adjust the loss value based on the similarity of the samples; ④ Optimize the neural network parameters through backpropagation and gradient descent to minimize the loss function and obtain the optimal model.

Images