Coal roadway support parameter prediction method based on graph network and multi-task learning

Abstract

The invention relates to a coal roadway support parameter prediction method based on a graph network and multi-task learning. The method comprises the following steps: acquiring roadway information of a coal mine roadway; according to the roadway information, obtaining similar roadways to construct a graph sampling aggregation network, sampling and aggregating the graph sampling aggregation network, and extracting key features of the roadway information; according to the key features, multi-task learning is utilized to predict support parameters of the coal mine roadway, the support parameters are obtained, the multi-task learning is obtained based on training of a training set, and the training set comprises actual roadway information and actual support parameters. According to the method, the graph sampling aggregation network and the multi-task learning mechanism are combined, so that the prediction efficiency of the coal roadway support parameters is improved, and a plurality of support parameters can be predicted at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine roadway support, and in particular to a coal mine roadway support parameter prediction method based on graph network and multi-task learning. Background Art

[0002] Coal mine tunnel support is a vital part of coal mining, which is directly related to the safe production and economic benefits of the mine. The appropriate coal mine tunnel support type and parameters should be selected based on the specific geological conditions, service life, use requirements and other factors of the coal mine tunnel. Detailed engineering geological survey and surrounding rock stability analysis should be carried out during the design process to ensure the scientificity and rationality of the support design. Common support forms include bolt support, anchor cable support, shed support, anchor mesh spraying combined support, etc. The selection of support form should fully consider the stability requirements, construction difficulty and cost-effectiveness of the tunnel. After the support parameter design is completed, it is necessary to check and optimize. The purpose of the check is to verify whether the support parameters meet the stability requirements and safety specifications of the tunnel. The purpose of the optimization is to reduce the support cost and improve the construction efficiency as much as possible under the premise of meeting the stability requirements. In the process of checking and optimizing, it may be necessary to verify and adjust with the help of numerical simulation, field tests and other means. Therefore, choosing a suitable support method and determining support parameters are crucial to maintaining tunnel stability and preventing accidents.

[0003] In recent years, with the rapid development of deep learning technology, deep learning has achieved remarkable results in the field of smart mines. Graph sampling aggregation network is a graph network model that can learn deep feature representation of data from unlabeled data, providing a new method for learning coal roadway support data. Through the graph sampling aggregation network, the effective features of coal roadway support data can be automatically extracted, the influence of irrelevant information can be reduced, and the stability of the prediction method of roadway support parameters can be improved. However, there are still some challenges in applying graph sampling aggregation network to the prediction of coal roadway support parameters. First, the geological conditions of coal mine roadways are different, and there is a certain correlation between the roadway information of each roadway. How to build an effective model to extract these related features is a key issue. Secondly, coal mine support data usually has nonlinear relationships. How to preprocess and transform the features of these data to improve the effectiveness of the prediction method is an important issue.

[0004] In order to improve the effectiveness of coal roadway support parameter prediction, the present invention proposes a coal roadway support parameter prediction method based on graph network and multi-task learning. Summary of the invention

[0005] The purpose of the present invention is to provide a coal roadway support parameter prediction method based on graph network and multi-task learning, combining graph sampling aggregation network and multi-task learning mechanism to improve the effectiveness of coal roadway support parameter prediction.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A coal roadway support parameter prediction method based on graph network and multi-task learning, including:

[0008] Collecting tunnel information of coal mine tunnels;

[0009] According to the lane information, similar lanes are obtained to construct a graph sampling aggregation network, and the graph sampling aggregation network is sampled and aggregated to extract key features of the lane information;

[0010] According to the key features, the support parameters of the coal mine tunnel are predicted by multi-task learning to obtain the support parameters; wherein the multi-task learning is obtained based on training set training, and the training set includes actual tunnel information and actual support parameters.

[0011] Optionally, constructing the graph sampling aggregation network includes: taking lanes as nodes, connection relationships between similar lanes as edges, and taking the actual lane information as features of the nodes to construct the graph sampling aggregation network, wherein the similar lanes are different lanes with the same longitude and latitude or the same lane uses.

[0012] Optionally, the key features for extracting the lane information include:

[0013] Sampling neighbor nodes of the node in the graph sampling aggregation network, and aggregating the sampled neighbor nodes using an aggregation function to extract aggregated neighbor nodes;

[0014] Perform feature concatenation on the node and the aggregated neighboring nodes, and update the feature of the node through nonlinear transformation to obtain a feature representation of the node;

[0015] The feature representation of the node is normalized by L2 norm to obtain the key feature.

[0016] Optionally, a method of aggregating the sampled neighbor nodes using an aggregation function is:

[0017]

[0018] in, is the output of the aggregation, N(i) is the set of neighbor nodes of node i, and h j (k-1) is the feature representation of node j at layer k-1.

[0019] Optionally, the method of performing feature concatenation of the node and the aggregated neighboring nodes and updating the feature of the node through nonlinear transformation is:

[0020]

[0021] Among them, W (k) is the weight matrix of the kth layer, σ is the activation function, concat represents the feature concatenation operation, is the feature representation of node i at layer k.

[0022] Optionally, obtaining the training set includes:

[0023] Acquiring support data, wherein the support data includes actual tunnel information and actual support parameters of different coal mine areas;

[0024] The actual tunnel information and the actual support parameters are preprocessed to obtain the training set.

[0025] Optionally, preprocessing the actual support data includes: preprocessing of numerical data and preprocessing of text information;

[0026] The preprocessing of the numerical data includes: cleaning and normalizing the actual tunnel information and the actual support parameters;

[0027] The preprocessing of the text information includes: constructing a text representation model to select keywords from the actual lane information, and digitizing the keywords using one-hot encoding.

[0028] Optionally, according to the key features, predicting the support parameters of the coal mine tunnel using a multi-task learning network includes:

[0029] The prediction task is divided into several subtasks, and a fully connected layer is added to each subtask. The prediction results of each subtask are output respectively according to the key features, wherein the several subtasks are anchor spacing, anchor row spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing and shotcrete thickness.

[0030] Optionally, the loss function of the multi-task learning network is:

[0031]

[0032] Among them, ω j represents the weight of the jth support parameter, L j represents the error of the jth task, h1,…,h n is the key feature of the lane information output by the graph sampling aggregation network, y1,…,y n is the actual support parameter, L[h1,…,h n , y1,…,y n ] is the loss function of n lanes.

[0033] The beneficial effects of the present invention are as follows: (1) The present invention is a coal roadway support parameter prediction method based on graph network and multi-task learning, which uses the sampling and aggregation technology of graph aggregation network to process coal roadway information and realizes the extraction of key features of roadway information. Through the neighbor sampling mechanism, the efficiency of graph sampling aggregation network on large-scale graphs is improved, so that complex coal roadway support data can be effectively processed.

[0034] (2) This method adopts a multi-task learning mechanism, divides the prediction problem into nine subtasks (anchor spacing, anchor row spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable row spacing, and shotcrete thickness), and makes full use of the information sharing and correlation between tasks to improve the generalization ability and efficiency of the prediction method. This prediction method not only improves the prediction effectiveness, but also enhances its ability to adapt to different support scenarios.

[0035] (3) The stochastic gradient descent algorithm is used for training. The training efficiency is improved by partitioning the graph and dividing the data into small batches. In each training iteration, small batches of data are randomly extracted from the dataset and the network parameters are updated using an iterative algorithm to ensure efficient training of the prediction method on large-scale data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A method for predicting coal roadway support parameters based on graph network and multi-task learning according to an embodiment of the present invention;

[0038] Figure 2 It is the loss function curve on the training set and the validation set in the training process of the embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] This embodiment provides a method for predicting coal roadway support parameters based on graph network and multi-task learning, including:

[0042] Collecting tunnel information of coal mine tunnels;

[0043] According to the lane information, similar lanes are obtained to build a graph sampling aggregation network, and the graph sampling aggregation network is sampled and aggregated to extract the key features of the lane information;

[0044] According to the key features, multi-task learning is used to predict the support parameters of coal mine tunnels and obtain the support parameters. The multi-task learning is based on the training set, which includes actual tunnel information and actual support parameters.

[0045] Furthermore, constructing a graph sampling aggregation network includes: taking lanes as nodes, connection relationships between similar lanes as edges, and actual lane information as features of nodes to construct a graph sampling aggregation network, wherein similar lanes are different lanes with the same longitude and latitude or the same lane purpose.

[0046] Furthermore, extracting the key features of the lane information includes: sampling the neighbor nodes of the nodes in the graph sampling aggregation network graph, and aggregating the sampled neighbor nodes using an aggregation function to extract the aggregated neighbor nodes; concatenating the features of the node with the aggregated neighbor nodes, and updating the features of the node through nonlinear transformation to obtain the feature representation of the node; and normalizing the feature representation of the node by the L2 norm to obtain the key features. Among them, the method of aggregating the sampled neighbor nodes using an aggregation function is as shown in formula (2), and the method of concatenating the features of the node with the aggregated neighbor nodes and updating the features of the node through nonlinear transformation is as shown in formula (3).

[0047] Specifically, a graph sampling aggregation network is used to extract the key features of the tunnel information: a graph structure is constructed using the collected tunnel information, in which each node represents a tunnel, and tunnels with the same longitude and latitude or tunnel purpose constitute the adjacent edges of the graph; in order to meet the needs of batch processing, the feature representation of the nodes in the graph is effectively learned by sampling and aggregating neighbor features; the graph sampling aggregation network uses the information of similar tunnels to update the feature representation of the nodes, thereby obtaining a node feature representation with context-awareness; in this way, the graph sampling aggregation network is applied to the processing of support data to obtain the key feature representation of each tunnel information.

[0048] Furthermore, obtaining the training set includes: obtaining support data, wherein the support data includes actual tunnel information and actual support parameters of different coal mining areas; and preprocessing the actual tunnel information and actual support parameters to obtain the training set.

[0049] Specifically, the tunnel information includes tunnel depth, tunnel longitude, tunnel latitude, coal seam thickness, direct roof lithology, direct roof thickness, old roof lithology, old roof thickness, direct bottom lithology, direct bottom thickness, old bottom lithology, old bottom thickness, coal pillar width, cross-sectional shape (rectangular, trapezoidal, circular, semicircular arch, three-center arch), and these tunnel information are used as input information of the graph neural network. The support parameters are anchor spacing, anchor spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing, and shotcrete thickness. The tunnel information and support parameters constitute the support data.

[0050] Furthermore, the preprocessing of the actual support data includes: preprocessing of numerical data and preprocessing of text information; the preprocessing of numerical data includes: cleaning and normalizing the actual support data; the preprocessing of text information includes: constructing a text representation model to select keywords for the actual tunnel information in the actual support data, and using unique hot encoding to digitize the keywords.

[0051] Specifically, after collecting the tunnel information and support parameters of different coal mining areas to form support data, the numerical data preprocessing and text data preprocessing are performed through the support data cleaning and normalization steps to ensure the consistency of the support data.

[0052] Furthermore, based on the key features, multi-task learning is used to predict the support parameters of coal mine tunnels, including: dividing the prediction task into several subtasks, adding a fully connected layer to each subtask, and outputting the prediction results of each subtask according to the key features.

[0053] After obtaining the prediction results, the predicted support parameters are also evaluated. How to evaluate the predicted support parameters is a problem that needs to be considered in engineering. For this problem, expert consultation and engineering experience are common treatment methods. Many scholars use numerical modeling to evaluate the effects of different support parameters. This not only affects the real-time performance of the evaluation, but also fails to include the optimization of various support parameters in the project. The use of hierarchical analysis method to evaluate the predicted support parameters can overcome the problems of insufficient real-time performance and generalization performance of support parameter evaluation to some extent.

[0054] The following is combined with Figure 1, further illustrating the method of this embodiment, first, collect tunnel information and support parameters of multiple coal mining areas to form support data, and preprocess the support data by cleaning and normalizing the support data; input the preprocessed tunnel information into the graph sampling aggregation network, extract the key features of the tunnel information through the graph sampling function and the aggregation function, and use the loss function of multi-task learning to improve the generalization ability and efficiency of the support parameter prediction method; then, combine the stochastic gradient descent algorithm to minimize the loss function, and in order to improve the training efficiency, divide the graph into multiple subgraphs, each subgraph is trained on a different computing node, and during the training process, information is aggregated and updated between the nodes, and the feature representation of the entire graph is learned together until the stopping condition is reached; finally, the predicted support parameters are evaluated using the analytic hierarchy process. The specific execution steps are as follows:

[0055] Step 1: Preprocessing of coal roadway support data: Collect roadway information and support parameters from different coal mining areas to form support data. Through support data cleaning and normalization steps, preprocess numerical data and preprocess text data to ensure the consistency of support data.

[0056] Step 2: Use a graph sampling aggregation network to extract key features of lane information: During training, the collected lane information is used to construct a graph structure, in which each node represents a lane, and the connection relationship between lanes with the same longitude and latitude or lane purpose constitutes the adjacent edges of the graph; in order to meet the needs of batch processing, the feature representation of the node is effectively learned by sampling and aggregating neighbor features of the nodes in the graph; the graph sampling aggregation network uses the information of adjacent lanes to update the feature representation of the node, thereby obtaining a node feature representation with context-awareness; in this way, the graph sampling aggregation network is applied to the processing of lane data to obtain the key feature representation of each lane information.

[0057] During testing, the lane information of the test case and the lane information in the training set are used to obtain similar lanes, thereby establishing a graph sampling aggregation network for the test case, and then the graph sampling aggregation network is sampled and aggregated to obtain the key features of the lane information of the test case; during training, the lane information of the training case and the lane information in the training set are used to obtain similar lanes, thereby establishing a graph sampling aggregation network for the test case, and then the graph sampling aggregation network is sampled and aggregated to obtain the key features of the lane information of the training case, and the key features are input into multi-task learning.

[0058] Step 3, construction of loss function of multi-task learning: the preprocessed tunnel information is input into the graph sampling aggregation network, and the key features closely related to the tunnel information are extracted through the graph sampling function and the aggregation function; the key features of the tunnel are combined with the support parameters to construct the loss function of multi-task learning, and the task is divided into nine subtasks of predicting anchor spacing, anchor spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor spacing, anchor cable spacing and shotcrete thickness, and the information sharing and correlation between tasks are fully utilized to improve the generalization ability and efficiency of the prediction method;

[0059] Step 4: Optimization of the loss function of multi-task learning: The constructed graph is divided into multiple subgraphs, and each subgraph is trained on a different computing node; during the training process, the lane information needs to be aggregated and updated between the computing nodes to jointly learn the representation of the entire graph; the stochastic gradient descent algorithm is combined to minimize the loss function, and the data is divided into small batches for training to improve training efficiency and generalization ability; in each training iteration, small batches of data are randomly extracted from the data set each time, and the network parameters are updated in an iterative manner.

[0060] Step 5: Evaluation of predicted support parameters: Use the analytic hierarchy process to evaluate the quality of the predicted support parameters.

[0061] The preprocessing of the support data in step 1 includes:

[0062] After collecting data through preliminary on-site investigation, field survey, and investigation and interview, this embodiment stores 1,000 support data collected by China University of Mining and Technology into the support database. The tunnel information includes tunnel burial depth, tunnel longitude, tunnel latitude, coal seam thickness, direct roof lithology, direct roof thickness, old roof lithology, old roof thickness, direct bottom lithology, direct bottom thickness, old bottom lithology, old bottom thickness, coal pillar width, cross-sectional shape (rectangular, trapezoidal, circular, semicircular arch, three-center arch), and these tunnel information are used as input information of the graph neural network. The support parameters are anchor spacing, anchor spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing, and shotcrete thickness. The tunnel information and support parameters constitute the support data.

[0063] When processing coal mine roadway support data, data cleaning is required, including replacing missing data items with non-missing data items of roadways with the same longitude and latitude or the same support type, and calculating the sample mean and standard deviation of the support data to determine a reasonable threshold range to remove outliers; then deduplication of data is performed to ensure that each roadway information is unique; finally, data normalization is performed to convert different types of support parameters to the same dimension, and minimum-maximum normalization is used to more accurately compare and analyze the relationship between different features. One-hot encoding is used to process text data.

[0064] The preprocessing module of coal roadway support data includes the preprocessing operations of roadway information and support parameters, which is composed of support data acquisition, support data cleaning, support data normalization and support data division.

[0065] Acquisition of support data: In order to obtain more coal mine support data, support data is obtained by field survey and collection, on-site investigation and reference to geological exploration data, and the collected tunnel information and support parameters are stored in the database. The tunnel information includes tunnel burial depth, tunnel longitude and latitude, coal seam thickness, direct roof lithology, direct roof thickness, old roof lithology, old roof thickness, direct bottom lithology, direct bottom thickness, old bottom lithology, old bottom thickness, coal pillar width, cross-sectional shape (rectangular, trapezoidal, circular, semicircular arch, three-center arch), and these tunnel information are used as input information of the graph sampling aggregation network. The support parameters are anchor spacing, anchor spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing and shotcrete thickness. The tunnel information and support parameters constitute the support data.

[0066] After collecting the original tunnel support data, it needs to be cleaned, including processing missing data items and removing outliers. For missing data items, use tunnels with the same longitude and latitude or the same support type to fill in the missing data items through linear interpolation. If there are many missing data items or direct deletion will result in data reduction, use nonlinear interpolation methods to replace these missing data items. The choice of interpolation method depends on the characteristics of the data. For example, mean or median interpolation is used for continuous support data, while mode interpolation is used for discrete support data.

[0067] Support outliers are observations that are significantly different from other observations in a data set. Outliers may be caused by a variety of reasons, including measurement errors, data entry errors, inherent variability in the data, or may indicate unusual events or behaviors. In order to deal with these outliers, some machine learning methods can be used, such as deleting outliers, replacing outliers, calculating the sample mean and standard deviation of the support data to determine a reasonable threshold range.

[0068] Support data normalization is a method of scaling the data range to a specific interval in order to better process the data in data analysis applications. The purpose of normalization is to ensure that all features have the same numerical scale, thereby avoiding certain features with large numerical ranges from dominating the model training process. In the analysis of tunnel support data, since it contains multiple physical quantities and units, the data is converted to the same dimension in order to more accurately compare and analyze the relationship between different support parameters.

[0069] Since the lane information contains text attributes, the text data needs to be digitized; first, a text representation model is built, that is, keywords are selected to represent the text. The common text representation model is the bag-of-words model, which is a text representation model that divides the entire corpus into words and eventually generates a word set that does not contain repeated objects; it ignores the order in which words appear in the sentence and does not consider the correlation between different words, resulting in a sparse representation.

[0070] After building the text representation model, the next step is to replace each keyword (word, word combination) selected in the model with a numerical value. One-hot encoding is the simplest method to digitize text. Term frequency-inverse document frequency evaluates the importance of a word to a document set or one of the documents in a corpus by introducing the inverse document frequency. Word embedding is different from the frequency-based method. The word embedding method maps each word to a low-dimensional, dense vector space, so that semantically similar words are closer in the vector space.

[0071] In the support data processing stage, the support data set is divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used to optimize the loss function, the validation set is used to adjust the parameters of the prediction method and prevent overfitting, and the test set is used to evaluate the generalization ability of the prediction method. The support data set is divided by stratified sampling to ensure that the support data distribution between the training set, validation set and test set is similar. This can avoid the deviation caused by uneven distribution of support data.

[0072] The key features of the graph sampling aggregation network for extracting lane information in step 2 include:

[0073] The graph sampling aggregation network is a graph network that learns the feature representation of nodes by sampling and aggregating the information of adjacent lanes of nodes. The following is how to design and use the graph sampling aggregation network to process support data and obtain the key feature representation of nodes: In order to construct the structural information of the graph, a lane is used as a node in the graph. If two lanes have the same longitude and latitude or have the same lane purpose, they are considered to be similar, thus forming an edge in the graph structure. Edges represent the association between different lanes. Nodes and edges can contain attributes, such as lane information as node features. Let x1,…,x n Represents the lane information of n lanes, where x i ∈R d , d is the dimension of the input feature. So the feature of the i-th node is x i For each node i, a fixed number of neighbor nodes are sampled. The sampling method is to randomly sample s data points from a set; use an aggregation function (such as mean, maximum, or LSTM) to aggregate the sampled neighbor node features. Assuming that the neighbor node set of node i is N(i), the aggregated neighbor feature is expressed as:

[0074]

[0075] Among them, h j (k-1) is the feature representation of node j at layer k-1, is the output of the aggregation, aggregate() represents the aggregation function, and the initial i-th node feature is X i ;

[0076] The aggregation function used here is:

[0077]

[0078] The node's own features are concatenated with the aggregated neighbor features, and the node's feature representation is updated through nonlinear transformation.

[0079]

[0080] Among them, W (k) is the weight matrix of the kth layer, σ is the activation function (such as ReLU), concat represents the feature concatenation operation, is the feature representation of node i at layer k. After multi-layer sampling and aggregation, the feature representation of node h is obtained. i (i=1,…,n), and perform L2 norm normalization on the features.

[0081] The construction of the loss function for multi-task learning in step 3 includes:

[0082] The key features closely related to the support parameters are extracted through graph sampling function and aggregation function. The transformed features are h1,…,h n . In order to optimize different types of support parameters, the prediction task is divided into nine subtasks: anchor spacing, anchor row spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing and shotcrete thickness, and each subtask corresponds to a regression problem. First, the key features of the nodes are extracted through a three-layer shared graph sampling aggregation network layer, and then a specific fully connected layer is added to each subtask to output the results of each subtask respectively. The single-task learning model predicts a regression result for a specific task. The loss function of single-task learning usually includes two parts: empirical loss and regularization term. The single-task loss function is expressed as:

[0083]

[0084] in, represents the loss of the jth support parameter of the ith lane, usually the predicted value f(h i ) and actual coal roadway support parameters The error between and function f represent a fully connected network. J(f) represents a regularization term, which is used to control the complexity of the prediction model and prevent overfitting. Common regularization terms include L1 regularization and L2 regularization.

[0085] Multi-task learning predicts multiple coal roadway support parameters at the same time. By sharing information and parameters of the fully connected network, multi-task learning can improve the generalization ability and efficiency of the prediction method. Multiple coal roadway support parameters constitute the loss function of multi-task learning, which is expressed as:

[0086]

[0087] Among them, ω j represents the weight of the jth support parameter, indicating the importance of the task in the overall loss. j Measures the error between the predicted value and the actual value of the jth task, h1,…,h n is the key feature of the lane information output by the graph sampling aggregation network, y1,…,y n represents the actual support parameters of n tunnels, y i ∈R m ,m is the number of support parameters. ,L[h1,…,h n , y1,…,y n ] is the loss function of n lanes.

[0088] Through the above steps, a prediction method based on multi-task learning is designed and implemented to predict support parameters. This prediction method uses graph sampling and aggregation techniques to effectively extract key features related to support parameters, and improves the generalization ability and efficiency of the prediction method through a multi-task learning mechanism.

[0089] The optimization of the loss function of multi-task learning in step 4 includes: using the PyTorch toolkit to implement distributed training. In order to adapt to the processing of big data, a distributed training mode is adopted for the overall network composed of the graph sampling aggregation network and multi-task learning. First, the distributed training environment is initialized, and the support data of the training set is distributed to multiple computing nodes. Small batch stochastic gradient descent training is performed on each computing node, and the tunnel information is aggregated and updated between the nodes so that the information of each node can be finally converged. During the training process, the performance of the support parameter prediction method is evaluated through the validation set to ensure the generalization ability of the prediction method. Termination conditions need to be set during the training process, such as reaching the maximum number of iterations or the relative change of the loss function value of adjacent cycles is less than a value.

[0090] After each iteration, the validation set is used to evaluate the performance of the support parameter prediction method. The validation set is a data set different from the training set and is used to view the performance of the prediction method during the training process. The performance of the support parameter prediction method on the unobserved data set is understood by calculating the loss function value on the validation set. If the performance of the support parameter prediction method on the validation set is improved, the parameters of the current model are updated and saved. In this way, after the training is completed, the parameters of the prediction method with the best performance on the validation set will be saved.

[0091] An early stopping strategy is used to prevent overfitting of the prediction method. When the performance of the prediction method on the validation set does not increase or decrease for multiple consecutive iterations, the training will be terminated in advance to avoid over-optimization and damage the characterization ability of the support prediction method. Through this series of iterative training, performance evaluation and early stopping strategy, the optimal prediction model for coal roadway support parameters can be effectively trained and ensured to have good generalization ability in practical applications. Figure 2 The figure shows the change of the loss function value on the training set and the validation set during the training process. The GPU of the computer used for training is an NVIDIA GeForce RTX3090 card, and the number of single training batches during training is 10. The number of epochs of the experiment is 100. During training, the hyperparameters of the prediction method are adjusted according to the validation set fitting. The hyperparameters corresponding to the best performance are selected on the validation set. The number of layers of the graph sampling aggregation network is 3, the output dimension is set to 10, and the sampling method is to randomly sample s = 5 data points from a set; the learning rate of the optimizer is 0.001.

[0092] The evaluation of the support parameters predicted in step 5 is performed by:

[0093] The AHP usually includes three steps: constructing a hierarchical structure, establishing a judgment matrix for each level that satisfies the consistency test, and converting the weights of the solution level. When using the AHP to solve problems, it is first necessary to hierarchize the problem in order to construct a hierarchical structure model. In this way, complex problems are decomposed into multiple groups of elements. These elements form several levels according to their attributes and relationships. The elements of the previous level serve as criteria to dominate the related elements of the next level. The nine subtasks of anchor spacing, anchor spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor spacing, anchor cable spacing, and shotcrete thickness are determined as the final support parameters as the solution level, which is the highest level. The lowest level is the tunnel information.

[0094] In the hierarchical analysis method, two features are taken each time and With a ij express and The ratio of the correlation of the support parameters to the influence of a ijCreate a decision matrix A = (a ij ) d×d , whose elements satisfy the following relations:

[0095]

[0096] If a completely consistent result is required, the elements in the decision matrix A should satisfy the following equation:

[0097] a ij a jk =a ik (7)

[0098] In order to make formula (7) valid, it is necessary to use the consistency test decision matrix. The consistency index CI is obtained through formula (8):

[0099]

[0100] Among them, λ max =4.993.

[0101] After further normalizing each column in the matrix A, the average value of each row is calculated to obtain the corresponding weight matrix A; then the support parameters are normalized by column, and the average value of the support parameters is calculated by row to obtain the characterization vector W of the final results of the predicted and actual support parameters; the weight matrix B of the support parameters is obtained using the support parameters, and the comprehensive evaluation value of the support parameters is calculated using the matrices A and B and the characterization vector W:

[0102] R=AB T W (9)

[0103] The size of the elements in R is used to determine whether the predicted support parameters are better than the actual support parameters. The size of the elements in R is used to determine whether the predicted support parameters are better than the actual support parameters.

[0104] This embodiment uses the sampling and aggregation technology of graph networks to construct a graph sampling aggregation network to learn the key features of tunnel information and uses multi-task learning to construct a loss function, and predicts support parameters with the help of an optimized model. This method uses tunnel information from different coal mining areas to construct a graph sampling aggregation network of coal mine tunnel information, and each node of the network represents a piece of tunnel information. The graph sampling aggregation network improves the efficiency of the graph sampling aggregation network on large-scale graphs through a neighbor sampling strategy. The network allows the use of multiple aggregation functions to better process different graph data. At the same time, this method also adopts a multi-task learning mechanism to divide the support parameter prediction problem into multiple subtasks, and makes full use of information sharing and correlation between tasks to improve the generalization ability and efficiency of the support prediction method. This method can fully consider the topological structure of coal tunnel information, design a safer and more reliable support parameter prediction method, and is scalable, suitable for processing large-scale coal tunnel support data and flexible customization and expansion in practical applications.

[0105] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A coal roadway support parameter prediction method based on graph network and multi-task learning, characterized in that: include: Collecting tunnel information of coal mine tunnels; According to the lane information, similar lanes are obtained to construct a graph sampling aggregation network, and the graph sampling aggregation network is sampled and aggregated to extract key features of the lane information; According to the key features, the support parameters of the coal mine roadway are predicted using multi-task learning to obtain the support parameters; wherein the multi-task learning is obtained based on training set training, and the training set includes actual roadway information and actual support parameters.

2. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 1 is characterized in that: Constructing the graph sampling aggregation network includes: taking lanes as nodes, the connection relationship between similar lanes as edges, and taking the actual lane information as the characteristics of the nodes to construct the graph sampling aggregation network, wherein the similar lanes are different lanes with the same longitude and latitude or the same lane purpose.

3. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 2 is characterized in that: The key features for extracting the lane information include: Sampling neighbor nodes of the node in the graph sampling aggregation network, and aggregating the sampled neighbor nodes using an aggregation function to extract aggregated neighbor nodes; Perform feature concatenation on the node and the aggregated neighboring nodes, and update the feature of the node through nonlinear transformation to obtain a feature representation of the node; The feature representation of the node is normalized by L2 norm to obtain the key feature.

4. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 3 is characterized in that: The method of aggregating the sampled neighbor nodes using the aggregation function is: in, is the output of the aggregation, is the set of neighbor nodes of node i, h j (k-1) is the feature representation of node j at layer k-1.

5. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 4 is characterized in that: The method of concatenating the features of the node and the aggregated neighboring nodes and updating the features of the node through nonlinear transformation is as follows: Among them, W (k) is the weight matrix of the kth layer, σ is the activation function, concat represents the feature concatenation operation, is the feature representation of node i at layer k.

6. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 1 is characterized in that: Acquiring the training set includes: Acquiring support data, wherein the support data includes actual tunnel information and actual support parameters of different coal mine areas; The actual tunnel information and the actual support parameters are preprocessed to obtain the training set.

7. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 6 is characterized in that: Preprocessing the actual support data includes: preprocessing of numerical data and preprocessing of text information; The preprocessing of the numerical data includes: cleaning and normalizing the actual tunnel information and the actual support parameters; The preprocessing of the text information includes: constructing a text representation model to select keywords from the actual lane information, and digitizing the keywords using one-hot encoding.

8. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 1 is characterized in that: According to the key features, the support parameters of the coal mine tunnel are predicted using a multi-task learning network, including: The prediction task is divided into several subtasks, and a fully connected layer is added to each subtask. The prediction results of each subtask are output respectively according to the key features, wherein the several subtasks are anchor spacing, anchor row spacing, anchor diameter, anchor length, anchor cable length, anchor cable diameter, anchor cable spacing, anchor cable spacing and shotcrete thickness.

9. The method for predicting coal roadway support parameters based on graph network and multi-task learning according to claim 1, characterized in that: The loss function of the multi-task learning network is: Among them, ω j represents the weight of the jth support parameter, L j represents the error of the jth task, h1,…,h n is the key feature of the lane information output by the graph sampling aggregation network, y1,…,y n is the actual support parameter, L[h1,…,h n , y1,…,y n ] is the loss function of n lanes.

Images