Accurate and intelligent identification method for lithology of surrounding rock of underground engineering

Abstract

The invention relates to an accurate and intelligent identification method for lithology of surrounding rock of underground engineering, and belongs to the technical field of lithology identification. The method comprises the following steps: S1, preparing a rock test piece: synthesizing a rock sample through field sampling or artificial simulation, and preparing the test piece meeting experimental requirements; s2, an indoor drilling experiment is conducted, and drilling parameters are collected; s3, performing a borehole peeping experiment and storing a peeping image as a data set; s4, carrying out data preprocessing and fusion; s5, carrying out machine learning model training: selecting a graph neural network to carry out training and lithology classification identification; s6, evaluating and verifying the machine learning model; and S7, performing field optimization and application. According to the method, the advantages of the drilling peeping technology and the machine learning technology are combined, the drilling parameters are obtained based on the indoor drilling experiment, the drilling image is obtained through drilling peeping, and accurate and intelligent lithology identification is achieved by coupling data of the drilling image and the drilling parameters.

Description

Technical Field

[0001] The invention relates to a method for accurately and intelligently identifying the lithology of surrounding rocks in underground engineering, and belongs to the technical field of lithology identification. Background Art

[0002] Lithology identification is a key link in mine support design. It provides a scientific basis for support design and is of great significance for ensuring safe and efficient mining. Different lithologies have different mechanical properties. By accurately identifying lithologies, appropriate support methods can be selected to improve support effects and reduce costs. At the same time, lithology identification helps to assess the stability of surrounding rocks. By analyzing the brittleness of rock formations and the development of cracks, potential dangers such as rock bursts and water inrush can be identified in advance, and corresponding preventive measures can be taken.

[0003] Borehole peeping technology provides important technical support for mining, and provides accurate data for surrounding rock stability assessment and support design. It can help engineers accurately identify vulnerable areas and fracture zones in surrounding rocks, optimize support schemes, and ensure the effectiveness of support measures. A large amount of borehole image data can be obtained through borehole peeping. In addition, the corresponding drilling parameters can be obtained during the drilling process. The drilling parameters are also important reference data for lithology identification. If the two-dimensional data of drilling parameters and borehole images are fused, and the model is trained by fusion data, better identification effects can be obtained without increasing the workload of on-site construction. To this end, the present invention is proposed, combining the advantages of borehole peeping technology and machine learning technology, obtaining drilling parameters based on indoor drilling experiments, obtaining borehole images through borehole peeping, coupling the data of borehole images and drilling parameters, and realizing accurate and intelligent identification of lithology. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for accurate and intelligent identification of the lithology of surrounding rocks of underground engineering. Model data is collected through indoor drilling experiments, and the mechanical parameters of rocks are obtained by controlling the experimental conditions. The indoor drilling experiments have the advantages of controllable experimental environment, accurate data acquisition, high repeatability and reliability, and the ability to take a large number of samples for testing. Accurate and intelligent identification of lithology is achieved by coupling the data of drilling images and drilling parameters.

[0005] The technical solution of the present invention is as follows:

[0006] A method for accurate intelligent identification of the lithology of surrounding rocks in underground engineering, the steps are as follows:

[0007] Step S1: rock specimen preparation: prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples;

[0008] Step S2: Conducting indoor drilling experiments and collecting drilling parameters;

[0009] Step S3: Conduct a drilling peek experiment and save the peek image as a data set;

[0010] Step S4: perform data preprocessing and fusion;

[0011] Step S5: Perform machine learning model training: select graph neural network for training and lithology classification and identification;

[0012] Step S6: evaluating and verifying the machine learning model;

[0013] Step S7: On-site optimization and application.

[0014] Preferably, according to the present invention, in step S1, the specific steps of preparing the rock specimen are as follows:

[0015] S11: Prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples. The size of each single specimen is 70mm×70mm×50mm, that is, a rectangular specimen with a length of 70mm, a width of 70mm, and a height of 50mm, to ensure that multiple specimens are easy to splice and have enough space for drilling operations and to generate valid experimental data;

[0016] S12: According to the experimental objectives, select representative lithology combinations to prepare combined specimens. Each combined specimen is usually divided into several layers. The rock mechanical parameters such as strength and internal friction angle of each layer are different to simulate the different lithologies that may be encountered in the actual mining process. According to the actual geological conditions, select several representative lithology combinations, such as the alternating combination of coal seams and sandstone, clay, limestone, etc.; the alternating combination of soft and hard rock layers. For each combination, prepare multiple specimens to ensure the diversity and reliability of the experimental results.

[0017] Preferably, according to the present invention, in step S2, the specific steps are as follows:

[0018] S21: Select a small drilling rig as the drilling equipment, simulate the operation of actual mine drilling operation, start the drilling equipment, and perform drilling operation on the test piece;

[0019] S22: Drilling process and data collection: When drilling on the test piece, real-time monitoring and adjustment of drilling parameters, including drilling speed, drill bit pressure, etc., to ensure the stability and representativeness of the drilling process, and set the drilling parameters to be saved in time series.

[0020] Preferably, according to the present invention, in step S3, the specific steps are as follows:

[0021] S31: Install the viewing device: According to the needs of the experimental site, use the viewing device such as endoscope or small camera to view the borehole. The viewing device can observe the rock formation characteristics in the borehole through the camera and lens. Push the viewing device into the borehole and require the device to be able to enter different depths of the borehole to ensure that all relevant rock formations can be observed.

[0022] S32: Image recording: Use the peep device to take images or videos, record the characteristics of rock formations at different depths, and mark the observed depth. Use high-resolution equipment to ensure that high-quality images can be taken and changes in rock layers can be recorded every time a predetermined depth is passed or a significant rock change occurs.

[0023] Preferably, according to the present invention, in step S4, the specific steps are as follows:

[0024] S41: preprocessing the acquired drilling image for subsequent analysis and feature extraction, the preprocessing steps include: denoising, cropping and scaling, and feature extraction, and feature extraction uses a convolutional neural network to extract high-dimensional features from the image;

[0025] S42: preprocessing the drilling parameters. The drilling parameters are structured data. Through normalization processing, the data is mapped into the range of [0, 1] to ensure that the data of different parameters are compared on the same scale;

[0026] S43: The image and the drilling parameters are fused through a graph structure representation method, and the image and the drilling parameters are represented as node features of the graph respectively, the drilling image features are used as features of the image nodes, and the drilling parameters are used as features of the non-image nodes, and the similarity relationship between the graph nodes is defined.

[0027] The goal of the graph structure representation method is to transform the two heterogeneous data, image data and structured drilling parameters, into node features of the graph, and to uniformly model the features of the image and drilling parameters in the form of a graph, thereby realizing the fusion of multimodal data and ultimately achieving the purpose of mapping the data to the node features of the graph.

[0028] According to the present invention, further preferably, in step S43, the lithology identification method proposed in the present invention involves data of two dimensions, namely, borehole image and drilling parameters. Therefore, a multimodal feature fusion mechanism is introduced into the convolutional neural network to merge different types of node features and enhance the expression ability of the model. The formula is:

[0029]

[0030] in: is the feature of node i at layer l+1, calculated from the data at layer l; is the image feature represented by node j; is the drilling parameter characteristic represented by node j, W img and W param is the weight matrix for linear transformation of image and drilling parameter features respectively, N(i) is the set of neighbor nodes of node i; α ij is the weight of the edge between node i and node j; b (l) is the bias term; σ is the activation function, which is used to introduce nonlinearity. In this way, the graph convolution operation can simultaneously process information from different data sources and fuse different node features through the weighted edge, thereby improving the accuracy of lithology identification.

[0031] Preferably, according to the present invention, in step S5, the specific steps are as follows:

[0032] S51: Graph structure design: In order to integrate the drilling image and drilling parameters into the graph neural network, we first need to build a graph structure. Each node of the graph represents a sample, that is, a drilling hole, and each edge of the graph represents the relationship or similarity between different samples.

[0033] The task of graph structure design is to define the topology of the graph, clarify the connection relationship between nodes, determine which nodes are connected by edges, and the weight or similarity of the edges. The core of graph neural network is the information transmission between nodes, which depends on the topology of the graph. The topology of the graph, that is, the connection relationship between nodes, determines how information is transmitted and aggregated in the graph.

[0034] S52: Model training of graph neural network: During the training process of graph neural network, node features are propagated through the structure of the graph. The training process is divided into the following steps:

[0035] (1) Message passing: Each node transmits information from its neighbor nodes through the message passing mechanism. This process will be iterated multiple times in the graph neural network, so that the features of each node are continuously updated;

[0036] (2) Node feature update: Node features are continuously updated during the message transmission process. Specifically, after the message is transmitted, the new features of node i will be updated based on the messages sent by its neighboring nodes. This update process is achieved by aggregating the messages of neighboring nodes. The update formula is as follows:

[0037]

[0038] Among them, W (l) is the weight matrix of the lth layer, representing the linear transformation of the neighbor features, and the meanings of the remaining characters are the same as those of the formula in step S43;

[0039] (3) Output layer and loss function: After multiple layers of message passing, each node in the graph neural network will have a final feature vector. These feature vectors can be used for classification tasks. The difference between the predicted result and the true label is calculated and optimized through the cross entropy loss function. The cross entropy loss function measures the difference between the category probability distribution predicted by the node and the true label. By calculating the cross entropy loss of all nodes, the model can obtain an overall loss to guide parameter optimization. The loss function is:

[0040]

[0041] Where N is the number of samples, y i is the true label of sample i, p i is the predicted probability.

[0042] Preferably, according to the present invention, in step S6, the specific steps are as follows:

[0043] S61: Validation model: At the end of each training cycle, apply the current model to the validation data set, calculate the validation loss and accuracy, and determine whether the model is overfitting on the training set by comparing the training set loss with the validation set loss. If the validation loss does not decrease or increase significantly over multiple consecutive cycles, it means that the model may begin to overfit. Consider using an early stopping strategy to stop the training process, thereby retaining the best model parameters and preventing the model from continuing to optimize on the training set while performing poorly on the validation set.

[0044] S62: Model optimization and adjustment: Use the learning rate scheduler to dynamically adjust the learning rate to help the model converge better. In terms of regularization, use the weight decay method to reduce overfitting. Adjust the batch size, learning rate, number of layers and other hyperparameters through cross-validation to improve model performance.

[0045] The above learning rate scheduler is a cosine annealing scheduler. As the number of training rounds increases, the learning rate is gradually reduced, which helps the training process to converge smoothly and avoid overfitting.

[0046] The above cross-validation selects the K-fold cross-validation method: the data set is randomly divided into K subsets, and K-1 subsets are used to train the model each time, and the remaining subset is used for verification. This process is repeated K times, each time a different verification set is selected, and finally the average performance index of K times is calculated. In the present invention, K=5 is selected, and a five-fold cross-validation is performed. By verifying the generalization ability of the model, the best hyperparameters are selected. After that, the average accuracy of each cross-validation is calculated, and the best model is selected based on its stability and performance.

[0047] According to the preferred embodiment of the present invention, in step S7, the borehole images and drilling parameters collected from the field are preprocessed and data fused and then input into the machine learning model. The model performs a new round of learning training on the data collected on the field based on the original data training, and automatically adjusts the model parameters according to the training results to adapt to the new data input. This adaptive adjustment mechanism can ensure that the model is continuously optimized in the continuous accumulation of engineering practice, thereby improving the prediction accuracy and reliability and meeting the application requirements in different engineering environments.

[0048] The beneficial effects of the present invention are:

[0049] (1) Multi-dimensional data fusion: The present invention combines borehole images with drilling parameter data, comprehensively utilizing image information and physical parameters during the drilling process to provide more comprehensive feature input for lithology identification. This data fusion can improve the accuracy of model recognition and overcome the limitations of a single data source.

[0050] (2) Improved accuracy of rock type identification: By combining borehole images (which directly reflect rock characteristics) and drilling parameters (dynamic information such as drilling pressure and rotation speed), the machine learning model can learn rock characteristics more accurately and improve the accuracy of intelligent rock type identification, especially in complex geological environments.

[0051] (3) Indoor experiments can accurately simulate the drilling process of different rock combinations and generate high-quality experimental data, thereby providing a reliable source of training data for machine learning models and further improving the robustness and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the present invention.

[0053] Figure 2 It is a machine learning flow chart of the present invention.

[0054] Figure 3 is an example diagram of the experimental specimen combination in the embodiment of the present invention, wherein: Figure 3 (a) is an example of a coal-sandstone-limestone specimen, simulating actual geological conditions; Figure 3 (b) is an example of a limestone-coal-limestone specimen, simulating the alternation of soft and hard rock layers. DETAILED DESCRIPTION

[0055] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0056] Embodiment 1:

[0057] like Figure 1-Figure 2 As shown, this embodiment provides a method for accurate intelligent identification of the lithology of surrounding rocks in underground engineering, and the steps are as follows:

[0058] Step S1: rock specimen preparation: prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples;

[0059] The specific steps are as follows:

[0060] S11: Prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples. The size of each single specimen is 70mm×70mm×50mm, that is, a rectangular specimen with a length of 70mm, a width of 70mm, and a height of 50mm, to ensure that multiple specimens are easy to splice and have enough space for drilling operations and to generate valid experimental data;

[0061] S12: According to the experimental objectives, select representative lithology combinations to prepare combined specimens. Each combined specimen is usually divided into several layers. The rock mechanical parameters such as strength and internal friction angle of each layer are different to simulate the different lithologies that may be encountered in the actual mining process. According to the actual geological conditions, select several representative lithology combinations, such as the alternating combination of coal seams and sandstone, clay, limestone, etc.; the alternating combination of soft and hard rock layers. For each combination, prepare multiple specimens to ensure the diversity and reliability of the experimental results.

[0062] Step S2: Conducting indoor drilling experiments and collecting drilling parameters;

[0063] The specific steps are as follows:

[0064] S21: Select a small drilling rig as the drilling equipment, simulate the operation of actual mine drilling operation, start the drilling equipment, and perform drilling operation on the test piece;

[0065] S22: Drilling process and data collection: When drilling on the test piece, real-time monitoring and adjustment of drilling parameters, including drilling speed, drill bit pressure, etc., to ensure the stability and representativeness of the drilling process, and set the drilling parameters to be saved in time series.

[0066] Step S3: Conduct a drilling peek experiment and save the peek image as a data set;

[0067] The specific steps are as follows:

[0068] S31: Install the viewing device: According to the needs of the experimental site, use the viewing device such as endoscope or small camera to view the borehole. The viewing device can observe the rock formation characteristics in the borehole through the camera and lens. Push the viewing device into the borehole and require the device to be able to enter different depths of the borehole to ensure that all relevant rock formations can be observed.

[0069] S32: Image recording: Use the peep device to take images or videos, record the characteristics of rock formations at different depths, and mark the observed depth. Use high-resolution equipment to ensure that high-quality images can be taken and changes in rock layers can be recorded every time a predetermined depth is passed or a significant rock change occurs.

[0070] Step S4: perform data preprocessing and fusion;

[0071] The specific steps are as follows:

[0072] S41: preprocessing the acquired drilling image for subsequent analysis and feature extraction, the preprocessing steps include: denoising, cropping and scaling, and feature extraction, and feature extraction uses a convolutional neural network to extract high-dimensional features from the image;

[0073] S42: preprocessing the drilling parameters. The drilling parameters are structured data. Through normalization processing, the data is mapped into the range of [0, 1] to ensure that the data of different parameters are compared on the same scale;

[0074] S43: The image and drilling parameters are fused by graph structure representation, and the image and drilling parameters are represented as node features of the graph respectively. The drilling image features are used as features of image nodes, and the drilling parameters are used as features of non-image nodes. The similarity relationship between graph nodes is defined. Specifically:

[0075] ①Define nodes: Each node in the graph represents a specific object or time point, which can be defined as: image node and parameter node. The image node can regard each area or each local feature of the static image data as a node in the graph. The drilling parameter data is usually time series data, and the drilling parameter of each time step can be regarded as a drilling parameter node in the graph.

[0076] ② Define edges: The edges in the graph represent the relationship or dependency between nodes. In order to fuse the image and drilling parameter data, the following types of edges can be defined:

[0077] Time series dependency edges: Drilling parameters are usually time series data, so edges can be established in the graph based on the continuity of time. For a time series data, the nodes adjacent in time are connected to form a chain structure. These edges describe the dependency of drilling parameters over time.

[0078] Spatially dependent edges: For image data, there may be spatial dependencies between local regions or adjacent regions in the image. Adjacent pixels or image blocks of the image can be connected by edges to describe the spatial relationship between them.

[0079] Edges between images and drilling parameters: In order to combine image data with drilling parameters, the association between them can be represented by connecting image nodes and drilling parameter nodes in the graph. According to the time step, the image features at a certain moment are connected to the drilling parameter features at the corresponding moment through edges.

[0080] ③Build an adjacency matrix: The adjacency matrix A of the graph structure is used to represent the connection relationship between nodes. Each element A in the matrix ij Represents the connection strength between node i and node j, that is, the weight of the edge. Through the adjacency matrix, the graph neural network can propagate information between nodes.

[0081] The goal of the graph structure representation method is to transform the two heterogeneous data, image data and structured drilling parameters, into node features of the graph, and to uniformly model the features of the image and drilling parameters in the form of a graph, thereby realizing the fusion of multimodal data and ultimately achieving the purpose of mapping the data to the node features of the graph.

[0082] The lithology identification method proposed in the present invention involves data in two dimensions: borehole images and drilling parameters. Therefore, a multimodal feature fusion mechanism is introduced into the convolutional neural network to merge different types of node features and enhance the expression ability of the model. The formula is:

[0083]

[0084] in: is the feature of node i at layer l+1, calculated from the data at layer l; is the image feature represented by node j; is the drilling parameter characteristic represented by node j, W img and W param is the weight matrix for linear transformation of image and drilling parameter features respectively, N(i) is the set of neighbor nodes of node i; α ij is the weight of the edge between node i and node j; b (l) is the bias term; σ is the activation function, which is used to introduce nonlinearity. In this way, the graph convolution operation can simultaneously process information from different data sources and fuse different node features through the weighted edge, thereby improving the accuracy of lithology identification.

[0085] Step S5: Perform machine learning model training: select graph neural network for training and lithology classification and identification;

[0086] The specific steps are as follows:

[0087] S51: Graph structure design: In order to integrate the drilling image and drilling parameters into the graph neural network, we first need to build a graph structure. Each node of the graph represents a sample, that is, a drilling hole, and each edge of the graph represents the relationship or similarity between different samples.

[0088] The basic building block of a graph neural network is the graph convolutional layer. In each layer, the features of a node are weighted averaged by the features of its neighboring nodes, and its own features are updated:

[0089] H (k+1) =σ(AH k W (k) )

[0090] Among them, H (k+1) is the node feature matrix of the k+1th layer, W (k) is the weight matrix of the kth layer, A is the adjacency matrix, and σ is the activation function. Through multi-layer graph convolution operations, the features of the nodes will gradually aggregate the information of the surrounding neighbors and capture the complex relationship between the nodes in the graph.

[0091] Since more fine-grained attention to neighbor nodes is required, the attention mechanism is introduced through the graph attention network to give each neighbor node a different weight:

[0092]

[0093] Among them, α ij is the attention weight of node i to neighbor node j, W is the weight matrix, h j is the feature of neighbor node j, and the attention weight α ij It is obtained by calculating the node features, using a shared attention weight and non-linear activation function LeakyReLU to calculate and update:

[0094]

[0095] In order to further extract global information, a graph pooling layer is added to the graph neural network to aggregate the information of all nodes in the graph. Finally, the lithology classification is performed through the fully connected layer.

[0096] The task of graph structure design is to define the topology of the graph, clarify the connection relationship between nodes, determine which nodes are connected by edges, and the weight or similarity of the edges. The core of graph neural network is the information transmission between nodes, which depends on the topology of the graph. The topology of the graph, that is, the connection relationship between nodes, determines how information is transmitted and aggregated in the graph.

[0097] S52: Model training of graph neural network: During the training process of graph neural network, node features are propagated through the structure of the graph. The training process is divided into the following steps:

[0098] (1) Message passing: Each node transmits information from its neighbor nodes through the message passing mechanism. This process will be iterated multiple times in the graph neural network, so that the features of each node are continuously updated;

[0099] (2) Node feature update: Node features are continuously updated during the message transmission process. Specifically, after the message is transmitted, the new features of node i will be updated based on the messages sent by its neighboring nodes. This update process is achieved by aggregating the messages of neighboring nodes. The update formula is as follows:

[0100]

[0101] Among them, W (l) is the weight matrix of the lth layer, representing the linear transformation of the neighbor features, and the meanings of the remaining characters are the same as those of the formula in step S43;

[0102] (3) Output layer and loss function: After multiple layers of message passing, each node in the graph neural network will have a final feature vector. These feature vectors can be used for classification tasks. The difference between the predicted result and the true label is calculated and optimized through the cross entropy loss function. The cross entropy loss function measures the difference between the category probability distribution predicted by the node and the true label. By calculating the cross entropy loss of all nodes, the model can obtain an overall loss to guide parameter optimization. The loss function is:

[0103]

[0104] Where N is the number of samples, y i is the true label of sample i, p i is the predicted probability.

[0105] Step S6: evaluating and verifying the machine learning model;

[0106] The specific steps are as follows:

[0107] S61: Validation model: At the end of each training cycle, apply the current model to the validation data set, calculate the validation loss and accuracy, and determine whether the model is overfitting on the training set by comparing the training set loss with the validation set loss. If the validation loss does not decrease or increase significantly over multiple consecutive cycles, it means that the model may begin to overfit. Consider using an early stopping strategy to stop the training process, thereby retaining the best model parameters and preventing the model from continuing to optimize on the training set while performing poorly on the validation set.

[0108] S62: Model optimization and adjustment: Use the learning rate scheduler to dynamically adjust the learning rate to help the model converge better. In terms of regularization, use the weight decay method to reduce overfitting. Adjust the batch size, learning rate, number of layers and other hyperparameters through cross-validation to improve model performance.

[0109] The above learning rate scheduler is a cosine annealing scheduler. As the number of training rounds increases, the learning rate is gradually reduced, which helps the training process to converge smoothly and avoid overfitting.

[0110] The above cross-validation selects the K-fold cross-validation method: the data set is randomly divided into K subsets, and K-1 subsets are used to train the model each time, and the remaining subset is used for verification. This process is repeated K times, each time a different verification set is selected, and finally the average performance index of K times is calculated. In the present invention, K=5 is selected, and a five-fold cross-validation is performed. By verifying the generalization ability of the model, the best hyperparameters are selected. After that, the average accuracy of each cross-validation is calculated, and the best model is selected based on its stability and performance.

[0111] Step S7: On-site optimization and application: The borehole images and drilling parameters collected on-site are pre-processed and data fused before being input into the machine learning model. The model conducts a new round of learning and training on the data collected on-site based on the original data training, and automatically adjusts the model parameters according to the training results to adapt to the new data input. This adaptive adjustment mechanism can ensure that the model is continuously optimized in the continuous accumulation of engineering practice, thereby improving the prediction accuracy and reliability and meeting the application needs in different engineering environments.

Claims

1. A method for accurate intelligent identification of rock properties of underground engineering surrounding rocks, characterized in that: Here are the steps: Step S1: rock specimen preparation: prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples; Step S2: Conducting indoor drilling experiments and collecting drilling parameters; Step S3: Conduct a drilling peek experiment and save the peek image as a data set; Step S4: perform data preprocessing and fusion; Step S5: Perform machine learning model training: select graph neural network for training and lithology classification and identification; Step S6: evaluating and verifying the machine learning model; Step S7: On-site optimization and application.

2. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 1, characterized in that: In step S1, the specific steps of preparing the rock specimen are as follows: S11: Prepare specimens that meet the experimental requirements through on-site sampling or artificial simulation of synthetic rock samples. The size of each single specimen is 70mm×70mm×50mm; S12: According to the experimental objectives, select the lithology combination to prepare the combined specimen.

3. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 2, characterized in that: In step S2, the specific steps are as follows: S21: Select a drilling machine as the drilling equipment, simulate the operation of actual mine drilling operation, start the drilling equipment, and perform drilling operation on the test piece; S22: Drilling process and data collection: When drilling on the test piece, the drilling parameters are monitored and adjusted in real time, including drilling speed and drill bit pressure, and the drilling parameters are saved in time series.

4. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 3, characterized in that: In step S3, the specific steps are as follows: S31: Install the peeping device: according to the requirements of the experimental site, use the peeping device to peek into the borehole and push the peeping device into the borehole; S32: Image recording: Take images or videos through the peep device to record the characteristics of rock formations at different depths and mark the observed depth.

5. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 4, characterized in that: In step S4, the specific steps are as follows: S41: preprocessing the acquired drilling image, the preprocessing steps include: denoising, cropping and scaling, and feature extraction, the feature extraction uses a convolutional neural network to extract high-dimensional features from the image; S42: pre-processing the drilling parameters, mapping the data into the range of [0,1] through normalization processing, and ensuring that the data of different parameters are compared on the same scale; S43: The image and drilling parameters are fused through a graph structure representation method, and the image and drilling parameters are represented as node features of the graph respectively, the drilling image features are used as features of the image nodes, and the drilling parameters are used as features of the non-image nodes, and the similarity relationship between the graph nodes is defined.

6. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 5, characterized in that: In step S43, a multimodal feature fusion mechanism is introduced to merge different types of node features to enhance the expression ability of the model. The formula is: in: is the feature of node i at layer l+1, calculated from the data at layer l; is the image feature represented by node j; is the drilling parameter characteristic represented by node j, W img and W param is the weight matrix for linear transformation of image and drilling parameter features respectively, N(i) is the set of neighbor nodes of node i; α uj is the weight of the edge between node i and node j; b (l) is the bias term; σ is the activation function, which is used to introduce nonlinearity.

7. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 6, characterized in that: In step S5, the specific steps are as follows: S51: Graph structure design: In order to integrate the drilling image and drilling parameters into the graph neural network, we first need to build a graph structure. Each node of the graph represents a sample, that is, a drilling hole, and each edge of the graph represents the relationship or similarity between different samples. S52: Model training of graph neural network: During the training process of graph neural network, node features propagate information through the structure of the graph.

8. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 7, characterized in that: In step S52, the training process is divided into the following steps: (1) Message passing: Each node transmits information from its neighbor nodes through the message passing mechanism; (2) Node feature update: Node features are continuously updated during the message transmission process. Specifically, after the message is transmitted, the new features of node i will be updated based on the messages sent by its neighboring nodes. This update process is achieved by aggregating the messages of neighboring nodes. The update formula is as follows: Among them, W (l) is the weight matrix of the lth layer, representing the linear transformation of neighbor features; (3) Output layer and loss function: After multiple layers of message passing, each node in the graph neural network will have a final feature vector. These feature vectors are used for classification tasks. The difference between the predicted result and the true label is calculated and optimized through the cross entropy loss function. The cross entropy loss function measures the difference between the category probability distribution predicted by the node and the true label. By calculating the cross entropy loss of all nodes, the model obtains an overall loss to guide parameter optimization. The loss function is: Where N is the number of samples, y i is the true label of sample i, p i is the predicted probability.

9. The method for accurate intelligent identification of rock properties of surrounding rocks in underground engineering according to claim 8, characterized in that: In step S6, the specific steps are as follows: S61: Validation model: At the end of each training cycle, apply the current model to the validation data set, calculate the validation loss and accuracy, and determine whether the model is overfitting on the training set by comparing the training set loss with the validation set loss. If the validation loss does not drop or rise significantly over multiple consecutive cycles, it means that the model is beginning to overfit. Consider using an early stopping strategy to stop the training process and retain the best model parameters. S62: Model optimization and adjustment: Use the learning rate scheduler to dynamically adjust the learning rate to help the model converge better. In terms of regularization, use the weight decay method to reduce overfitting, and adjust the batch size, learning rate, and number of layers through cross-validation.

10. The method for accurate intelligent identification of surrounding rock properties of underground engineering according to claim 9, characterized in that: In step S7, the drilling images and drilling parameters collected from the field are preprocessed and data fused before being input into the machine learning model. The model performs a new round of learning and training on the data collected on site based on the original data training, and adjusts the model parameters according to the training results to adapt to the new data input.

Images