A mine vectorization representation method considering multi-element information

Abstract

The present application relates to the field of coal mine digitization, and particularly relates to a mine vectorization representation method considering multiple element information, comprising extracting nodes, attributes and relationships in multiple mines to represent as graph structure data, encoding to obtain feature vectors and classifying and normalizing; using a learning parameter matrix to unify dimensions and updating representation through a message passing network; merging and decoding using a multi-layer neural network; and using a multi-layer graph convolution to generate mine vectorization representation. The present application applies a graph neural network model to convert nodes and edges in a mine into vector representation, achieving comprehensive and accurate representation of multiple element information in the mine. The obtained vectorization representation can be applied to subsequent downstream tasks, making the mine digitization transformation and intelligent construction more efficient and accurate. Therefore, the present application has good application prospect and promotional value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of coal mine digitization, and in particular to a mine vectorization characterization method considering multi-element information. BACKGROUND

[0002] The foundation of coal mine digital transformation and intelligent construction is the vectorization characterization of the mine. After completing the vectorization characterization, support for many application scenarios can be realized. As a method of converting complex information into a simple numerical vector, vectorization characterization has become an important technical means for coal mine intelligence. With vectorization characterization, subsequent application scenarios such as rock burst type classification can be supported.

[0003] A mine is composed of many elements, and these elements and combinations of elements can represent the entire mine to some extent. Graph structure data is a very common data structure composed of a set of nodes and edges between them, which can be used to represent various relationship networks such as social networks, road networks, biological networks, etc. The nodes and edges in graph structure data can have different attributes, which can be used to describe the characteristics and relationships of nodes and edges. Graph structure data is widely used in data mining, machine learning, network analysis, etc. Graph neural networks are neural network models based on graph structure data for learning and prediction. Compared with traditional neural network models, graph neural networks not only consider the characteristics of nodes themselves, but also consider the relationships and topological structures between nodes, and can better handle data in non-Euclidean space.

[0004] The present application considers the multi-element information of the mine, combines graph neural network related algorithms to find a vectorization characterization method that can describe the information of the mine, and applies downstream tasks on this basis. SUMMARY

[0005] To provide a basis for mine characterization from data paths, the present application provides a mine vectorization characterization method considering multi-element information, comprising the following steps:

[0006] S11: Extract nodes, node attributes and node relationships from M mines, and represent the nodes and node relationships as graph structure data; M mines have n nodes, Φ node types and P node relationships;

[0007] Wherein, the nodes at least include working face, roadway, fault, fold area and borehole node types, and the node relationships at least include adjacent relationship and containing relationship.

[0008] S12: Feature encoding is performed on the node attributes contained in the nodes to obtain node feature vectors, represented as {χ1, χ2, …, χ i ,…,χ nEach attribute is represented by a feature vector.

[0009] Specifically, for numerical node attributes, the value of its feature vector is its node attribute value; for ordered non-numerical node attributes, a mapped ordered feature encoding is used, and for unordered non-numerical node attributes, a one-hot encoding is used.

[0010] S13: Classify the node feature vectors obtained in step S12 according to node type, and normalize the nodes of each type; the processing method is as follows: assume that the i-th node is in the... In the node types, 1 ≤ i ≤ n, No. The node type contains a set of nodes called V. φ The normalized node feature vector is represented as x i The normalization formula is

[0011]

[0012] S21: By learning the parameter matrix M φ Transform the node feature vectors of each node type into node representations of a uniform dimension {x′1,x′2,…,x′}. i ,…,x′ n The conversion formula is:

[0013] x′ i =M φ ·x i

[0014] Among them, M φ Indicates the first Learning parameter matrices for different node types;

[0015] S22: Update the representation of each node through a message passing network. Specifically: Divide the graph structure data into P graph structure data subgraphs according to node relationships, with one node relationship corresponding to one graph structure data subgraph; update the node representation of each graph structure data subgraph. The update method is as follows: For the p-th graph structure data subgraph, 1≤p≤P, after k updates, the updated value of the node representation of the i-th node is... The updated formula is

[0016]

[0017] MLP stands for Multilayer Perceptron. Let represent the set of neighboring nodes of the i-th node in the p-th subgraph. Let be the node representation update value of the j-th neighbor node among the neighbor nodes of the i-th node in the p-th subgraph after k-1 updates. for The number of nodes in the set, W p To learn the parameter matrix, Let represent the node representation update value of the i-th node after k-1 updates, with the initial value being...

[0018] S23: Update the node representation after step S22 Merging yields the merged node representation Z. i The merging formula is:

[0019]

[0020] Where, β p For learning parameters;

[0021] S3: Decoding the merge node representation Z i The merged node is represented by z. i The input is a multi-layer neural network, where each layer is a fully connected layer, to obtain the decoded node feature vectors. by and x i The cross-entropy is used as the loss function;

[0022] S41: Reconsider all nodes and node relationships as a single graph structure, and then divide the graph into first-level subgraphs based on mines, with one mine corresponding to one first-level subgraph.

[0023] S42: Based on node relationships, further divide each first-level subgraph into P second-level subgraphs, and generate the adjacency matrix A of the second-level subgraphs. mp The node representations in the secondary subgraph are aggregated into a single node representation matrix Z. mp The aggregation formula is

[0024] Z mp =stack i∈nide(m) (z i )

[0025] Where stack represents row-by-row concatenation, node(m) represents all nodes in the m-th mine, 1≤m≤M, and Z mp and A mp These are represented as the node representation matrix and adjacency matrix under the p-th relation in the m-th second-level subgraph, respectively;

[0026] S43: Select the first-level subgraph with the most nodes, take its node count as the maximum node count, and expand the node representation matrix Z. mp The number of rows is the maximum number of nodes, and the expanded row values ​​are filled with 0. The adjacency matrix A mp The number of rows and columns will also be expanded accordingly;

[0027] S44: Transform the node representation matrix Z mp and adjacency matrix A mp Merged into a P-channel matrix Z m and A m And perform multi-layer graph convolution operations, where the convolution formula for the (l+1)th layer is:

[0028]

[0029] Where σ(·) represents a nonlinear activation function; I is the identity matrix; This is the convolution formula for the l-th layer. W (l) It is the learning parameter matrix of the l-th layer. for The degree matrix, It is a diagonal matrix, with elements outside the main diagonal being zero, and elements on the main diagonal being zero. express The ιth row and ιth column, express The ιth row and ξth column, for Sum all columns in the ιth row;

[0030] S45: The matrix H obtained by the convolution formula of the last layer of the multi-layer graph convolution is... m Flattened into vector c m As a vectorized representation of the mine;

[0031] S46: Vectorize the mine representation c m Perform k-means clustering, and use the clustering error as the loss function to fine-tune the learning parameter matrix W. (l) Suppose the k-means clustering algorithm yields a cluster partition S = {S1, S2, ..., S...} Λ Then the clustering error is:

[0032]

[0033] in, It is cluster S λ The mean vector of , 1≤λ≤Λ.

[0034] The application point and beneficial effect of the present application: the present application realizes comprehensive and accurate representation of mine multi-element information by applying a graph neural network model to convert nodes and edges in the mine into vector representation; the present application also considers the relationship and topological structure between nodes, and can better process non-Euclidean space data; and the obtained vector representation can be applied to subsequent downstream tasks, such as rock burst type division, so that the mine digital transformation and intelligent construction are more efficient and accurate. Therefore, the present application has good application prospect and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a mine vectorization representation method framework considering multi-element information of the present application;

[0036] Figure 2 is a brief schematic diagram of graph structure data of the present application;

[0037] Figure 3 is a mine vectorization representation framework of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0039] As shown in Figure 1 , the present application proposes a mine vectorization representation method considering multi-element information, comprising the following steps:

[0040] S11: extracting node, node attribute and node relationship information from M mine geological data and production data, and representing the nodes and node relationships as graph structure data; the nodes at least include working face, roadway, fault, fold area, drill hole and other node types, and the node relationships at least include adjacent relationship, containing relationship and other node relationships; M mines have n nodes, Φ node types and P node relationships.

[0041] Figure 2 is a brief schematic diagram of graph structure data, in which the geometric shapes represent nodes, different geometric shapes represent different node types, and node relationships are described by edges.

[0042] S12: Encode the node attributes contained in the node to obtain the node feature vector. Each attribute is represented by a feature vector. For numerical node attributes, the value of its feature vector is its node attribute value. For non-numerical node attributes, different encoding methods are used depending on whether the node attribute is unordered or ordered. For ordered non-numerical node attributes, mapped ordered feature encoding is used, and for unordered non-numerical node attributes, one-hot encoding is used to obtain the unnormalized node feature vector, represented as {χ1,χ2,…,χ... i ,…,χ n}

[0043] Assuming the i-th node is the 1202 return airway, its node attributes include a width of 4m, a height of 3m, a strike length of 1.5km, an inclination angle of 10°, and a trapezoidal cross-sectional shape. The cross-sectional shape can be arched, trapezoidal, or rectangular, and these are unordered, non-numerical node attributes. Therefore, this node can be coded as follows:

[0044] χ i =[4 3 1.5 10 0 1 0]

[0045] S13: Classify the unnormalized node feature vectors obtained in step S12 according to node type, and normalize the nodes of each type; the processing method is as follows: assume that the i-th node is in the... In the node types, 1 ≤ i ≤ n, This node type contains a set of nodes called V. φ The normalized node feature vector can be represented as x i The normalization formula is as follows:

[0046]

[0047] S21: Since the dimensions of the feature vectors of nodes of different node types are not consistent, the parameter matrix M is learned. φ Transform the node feature vectors of each node type into node representations of a uniform dimension {x′1,x′2,…,x′}. i ,…,x′ n The conversion formula is:

[0048] x′ i =M φ ·x i

[0049] Among them, M φ Indicates the first The learning parameter matrix is ​​for each node type, with different learning parameter matrices corresponding to different node types. The node representation is obtained by learning the node feature vector through some algorithms.

[0050] S22: updating each node representation through the message passing network, specifically: dividing the graph structure data into P sub-graphs according to node relationships, one node relationship corresponds to one sub-graph;

[0051] updating the node representation of each sub-graph, for the pth sub-graph, 1≤p≤P, after k updates, the updated value of the node representation of the ith node is The update formula is

[0052]

[0053] wherein, MLP represents a multi-layer perceptron, represents a set of neighbor nodes of the ith node in the pth sub-graph, represents the updated value of the node representation of the jth neighbor node of the ith node in the pth sub-graph after k-1 updates, is the number of nodes in the set, W p is a learning parameter matrix, represents the updated value of the node representation of the ith node after k-1 updates, the initial value is

[0054] S23: merging the updated node representations of step S22 to obtain a merged node representation z i The merging formula is

[0055]

[0056] wherein, β p is a learning parameter;

[0057] S3: decoding the merged node representation z i inputting the merged node representation z i into a multi-layer neural network, each layer of the neural network is a fully connected layer, to obtain a decoded node feature vector x is equivalent to the reverse operation process of steps S21-S23, taking the cross-entropy of z and x i as a loss function;

[0058] S41: considering all nodes and node relationships as one graph structure data, and then dividing the graph structure data into M first-level sub-graphs according to the mine, one mine corresponds to one first-level sub-graph, the nodes in the first-level sub-graph are all nodes belonging to the mine corresponding to the first-level sub-graph;

[0059] S42: asFigure 3 As shown, based on node relationships, each first-level subgraph is further divided into P second-level subgraphs, and an adjacency matrix A of the second-level subgraphs is generated. mp The adjacency matrix is ​​a two-dimensional matrix where rows and columns represent nodes. The value of each element in the matrix indicates whether there is an edge connecting the corresponding nodes. The node representations in the secondary subgraph are aggregated into a node representation matrix Z. mp The aggregation formula is

[0060] Z mp =stack i∈node(m) (z i )

[0061] Where stack represents row-by-row concatenation, node(m) represents all nodes in the m-th mine, 1≤m≤M, and Z mp and A mp These are represented as the node representation matrix and adjacency matrix under the p-th relation in the m-th second-level subgraph, respectively;

[0062] S43: Select the first-level subgraph with the most nodes, take its node count as the maximum node count, and expand the node representation matrix Z. mp The number of rows is the maximum number of nodes, the padding of the expanded rows is 0, and the adjacency matrix A mp The number of rows and columns will also be expanded accordingly;

[0063] S44: Node representation matrix Z mp and adjacency matrix A mp Merged into a P-channel matrix Z m and A m And perform multi-layer graph convolution (GCN) operations, where the convolution formula (recursive formula) for the (l+1)th layer is as follows:

[0064]

[0065] Where σ(·) represents a nonlinear activation function; I is the identity matrix; This is the convolution formula for the l-th layer. W (l) It is the learning parameter matrix of the l-th layer. for The degree matrix, It is a diagonal matrix, with elements outside the main diagonal being zero, and elements on the main diagonal being zero. express The ιth row and ιth column, express The ιth row and ξth column, for Sum all columns in the ιth row;

[0066] S45: the matrix H obtained by the last layer of convolution formula of multi-layer graph convolution (GCN) m to be a flattened vector c m , as a mine vectorized representation;

[0067] S46: the mine vectorized representation c obtained in step S45 m , k-means clustering is performed, and a clustering error is used as a loss function to fine-tune a learning parameter matrix W (l) , assuming that the k-means clustering algorithm obtains cluster division S={S1, S2,..., S Λ}, the clustering error is:

[0068]

[0069] wherein, is the mean vector of the cluster S λ , and 1≤λ≤Λ.

[0070] Although the present application has been described above with reference to a preferred embodiment, it is not intended to limit the present application. A person with ordinary skill in the art can make various modifications and improvements without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application should be defined by the appended claims.

Claims

1. A mine vectorization representation method considering multi-element information, characterized in that, Comprising the following steps: S11: Extract nodes, node attributes and node relationships from M mines, and represent nodes and node relationships as graph structure data; M mines have n nodes, Φ node types, and P node relationships; the nodes at least include working face, roadway, fault, fold area, and borehole node types, and the node relationships at least include adjacent relationship and containing relationship; S12: feature encode the node attributes contained in the node to obtain a node feature vector, denoted as {x1, x2, …, x i ,…,χ n}, one attribute is represented by one feature vector; For numerical node attributes, The value of the eigenvector is the node attribute value; for ordered non-numerical node attributes, a mapping ordered feature code is used, and for unordered non-numerical node attributes, a one-hot encoding is used; S13: classifying the node feature vectors obtained in step S12 according to node types, and normalizing each type of node; the processing method is as follows: assuming that the ith node is in the φth node type, 1≤i≤n, 1≤φ≤Φ, the node set contained in the φth node type is V φ , and the normalized node feature vector is represented as x i , and the normalization formula is S21: Convert the node feature vector of each node type into a unified dimension node representation {x'1, x'2, …, x'N} φ The conversion formula is as follows: x' = f (x) i n}​ x' i = M φ • x i where M φ denotes the learning parameter matrix of the φ-th node type; S22: update each node representation through the message passing network, specifically: divide the graph structure data into P pieces of graph structure data subgraphs according to node relationships, one node relationship corresponds to one piece of graph structure data subgraph; update the node representation of each piece of graph structure data subgraph, the update method is: for the pth piece of graph structure data subgraph, 1≤p≤P, after k times of update, the updated value of the node representation of the ith node is The update formula is wherein MLP represents a multi-layer perceptron, denotes a set of neighbor nodes of the i-th node in the p-th subgraph, denotes a node representation update value of the j-th neighbor node of the i-th node in the p-th subgraph after k-1 times of updates, is the number of nodes in the set W p is a learning parameter matrix, denotes a node representation update value of the i-th node after k-1 times of updates, and the initial value is S23: updating the node representation of step S22 performing the merging to obtain a merged node representation z i , and the merging formula is where β p is a learning parameter; S3: decoding the merge node representation z i The merge node representation z i Inputting the multi-layer neural network, each layer of the neural network being a fully connected layer, to obtain a decoded node feature vector The cross entropy of z and x i is taken as a loss function; S41: Reconsider all nodes and node relationships as a graph structure data, and then divide a first-level subgraph in units of mines, one mine corresponding to one first-level subgraph; S42: Divide each first-level subgraph into P second-level subgraphs according to the node relationship, and generate an adjacency matrix A of the second-level subgraph mp , aggregate the node representations in the second-level subgraph into a node representation matrix Z mp , and the aggregation formula is Z mp = stack i∈node(m) (z i ) wherein, stack denotes concatenation by row, node(m) is all nodes under the mth mine, 1≤m≤M, Z mp and A mp respectively represent the node representation matrix and the adjacency matrix under the pth type of relationship in the mth secondary subgraph. S43: select a one-level subgraph with the largest number of nodes, and take the number of nodes as the maximum number of nodes to expand the node representation matrix Z mp The number of rows is the maximum number of nodes, and the expansion row value is filled with 0 mp Also correspondingly expand the row and column; S44: merge the node representation matrix Z mp and the adjacency matrix A mp into a P-channel matrix Z m and A m and perform a multi-layer graph convolution operation, wherein the convolution formula of the l+1th layer is Where σ(·) represents a nonlinear activation function; I is the identity matrix; This is the convolution formula for the l-th layer. W (l) It is the learning parameter matrix of the l-th layer. for The degree matrix, It is a diagonal matrix, with elements outside the main diagonal being zero, and elements on the main diagonal being zero. express The ιth row and ιth column, express The l-th row and ξ-th column, for Sum all columns in the ιth row; S45: flattening the matrix H obtained from the last layer of the multi-layer graph convolution formula m flattening the vector c m as a vectorized representation of the mine S46: vectorization of mine characterization c m k-means clustering is performed, and the clustering error is used as a loss function to fine-tune the parameter matrix W (l) , assuming that the k-means clustering algorithm obtains cluster partition S = {S1, S2,..., S Λ}, then the clustering error: wherein is the mean vector of the cluster S λ , 1≤λ≤Λ.

Images