Stope contour deduction method and system based on graph attention network

Abstract

The present application relates to the technical field of profile deduction, in particular to a stope profile deduction method and system based on a graph attention network, in which, on the basis of a hierarchical boundary topology chain, coordinate difference, projection offset, mid-section interlayer edge weight, roof state value, side slope state value and filling body bearing value jointly participate in graph attention network weight distribution, abnormal roof nodes, abnormal side slope nodes and low bearing filling body adjacent connection items are locked or removed before displacement constraint, on the basis of a sequence of attention displacement constraints, A-star algorithm arranges candidate connection points from both ends of a gap, and deletes reverse jump boundary sections by using endpoint distance, path length, adjacent corner and bidirectional path direction difference, so that a closed path is subjected to bidirectional consistency constraint, and the gain is embodied in that the profile connection relationship is more adaptive to irregular goaf, the gap closed error path is less, the roof and side slope risk nodes are more controllable in interfering with the deduction result, and the profile evolution relationship between recovery and filling stages is more continuous.

Description

Technical Field

[0001] This invention relates to the field of contour extrapolation technology, and in particular to a method and system for contour extrapolation of mining sites based on graph attention networks. Background Technology

[0002] The field of contour projection technology aims to calculate and analyze the boundary morphology, contour node positions, contour line connection relationships, and contour change trends of a target object, obtaining the contour extension, offset, closure, and deformation results of the target object within a specified spatial range or time period. Based on existing boundary nodes, contour topological relationships, and boundary change parameters of the mining area, a contour node graph structure is constructed. Then, based on an artificial intelligence computing framework, the spatial dependencies between contour nodes are learned to project the subsequent contour morphology of the mining area, providing a computational basis for determining the mining area boundary, controlling the mining scope, and assessing contour changes.

[0003] The method for predicting mining contours based on graph attention networks combines the graph neural network learning mechanism from artificial intelligence with graph attention networks. Mining contour nodes are used as graph nodes, and the connections between contour nodes are used as graph edges. Node position parameters, boundary change parameters, and node connection states are input into the graph attention network. Attention coefficients are calculated to determine the influence weights of adjacent contour nodes on the target contour node. A multi-head attention aggregation method is used to calculate the node state representation of the target contour node, and nonlinear mapping is used to obtain the boundary node displacement, contour line extension direction, and contour closure state. Iterative predictions are performed on multiple contour node states based on the trained network parameters. Contour node positions and contour line connection states are updated according to node connection relationships to generate mining contour prediction results, thereby improving the continuity of mining contour change prediction, the stability of boundary connections, and the ability to represent local contour changes.

[0004] Existing contour extrapolation techniques typically calculate based on existing boundary nodes, contour topology, and boundary change parameters. In practice, the stope boundary is often treated as a continuous line segment on the same level. There is a lack of unified hierarchical constraints between roof boundaries, floor boundaries, sidewall boundaries, mid-level elevations, segment elevations, ore junctions, and backfill junctions. This leads to boundary misconnections in the extrapolation of goaf penetration sections, caving expansion sections, and segmented misalignment sections. When gap areas are connected only based on the distance between adjacent points or the direction of contour extension, the two ends of the gap may connect to different closure trends, forming reverse jump paths, which in turn cause local closure anomalies in the stope contour. Furthermore, the intervention of existing processing on roof condition, sidewall condition, and backfill bearing condition usually lags behind the contour extrapolation results. Abnormal nodes may participate in contour extension first and then be screened out by subsequent checks, resulting in repeated corrections of the extrapolation path and affecting the consistency of stope boundary determination, mining range control, and contour change assessment. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a method and system for inferring field contours based on graph attention networks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a sampling field contour inference method based on graph attention networks, comprising the following steps: S1: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, mid-section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. S2: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use a graph attention network to lock the abnormal nodes of the top plate and side wall, remove the adjacent connection items of low-bearing infill bodies, and generate an attention displacement constraint sequence. S3: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster. S4: Based on the rule-closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing filling bodies, and establish a constrained retreat contour chain; S5: Based on the constraint backtracking contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the plane boundary points and longitudinal boundary points, filter the boundary state paths, and generate the mining area contour projection map.

[0007] As a further embodiment of the present invention, the hierarchical boundary topology chain includes contour node number, node layer number, boundary type label, and node connection order; the attention displacement constraint sequence includes node attention weight, directional displacement amount and directional displacement limit, dip displacement amount and dip displacement limit, elevation displacement amount and elevation limit; the regular closed path cluster includes candidate path number, endpoint pairing number, path status label, and closure deviation value; the constrained retreat contour chain includes retreat node number, retreat direction label, displacement reduction amplitude, and constraint boundary number; and the mining area contour projection map includes a recovery contour layer, a filling contour layer, and a constrained contour layer.

[0008] As a further aspect of the present invention, the specific steps for generating the hierarchical boundary topology chain are as follows: Based on the planar boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, mid-level elevation, sub-level elevation, mining advance sequence, mining outline gap endpoint, ore boundary point of the goaf, and filling body boundary point, the stratum to which the boundary point belongs is determined according to the mid-level elevation and sub-level elevation, the boundary points of the same stratum are arranged according to the mining advance sequence, the roof and floor points are paired according to the elevation correspondence between the roof and floor boundaries, and the disconnection position is marked according to the mining outline gap endpoint to generate a stratum boundary point sequence; Based on the stratigraphic boundary point sequence, connect the planar boundary points and sidewall endpoints in the adjacent order of the same stratigraphic level, connect the top plate points and bottom plate points according to the cross-sectional boundary positions, set boundary status markers according to the ore boundary points and the filling body boundary points, and supplement the endpoints of the broken segments according to the connection numbers of the gap endpoints to establish a hierarchical boundary topology chain.

[0009] As a further aspect of the present invention, the specific steps for generating the attention displacement constraint sequence are as follows: Based on the hierarchical boundary topology chain, the starting point coordinates and ending point coordinates are extracted according to the adjacent node numbers of the same layer, the direction coordinate difference, dip coordinate difference, and elevation coordinate difference are calculated, the plane projection offset is extracted according to the segmented projection node number, and the contour edges, segmented related edges, and middle-segment related edges of the same stage are collected according to the middle segment layer number to generate a node differential weight table. Based on the node differential weight table, a graph attention network is used to locate abnormal adjacent nodes by marking the top plate state value, side wall state value, filling body bearing capacity value, and node boundary type. Low bearing capacity filling body adjacent connection items are deleted, and top plate stable nodes, side wall stable nodes, and middle section associated nodes are retained to generate an abnormal adjacent node table. Based on the abnormal adjacent node table, node displacement components are selected according to node coordinate difference, segmented projection offset, inter-layer edge weight of middle section, and abnormal node mark. The strike displacement, dip displacement, and elevation displacement are respectively limited. The displacement components and constraint marks are arranged according to the node number to generate a attention displacement constraint sequence.

[0010] As a further embodiment of the present invention, the graph attention network first writes the top plate state value, side wall state value, infill bearing capacity value, and node boundary type marker into the node state matrix; writes the same-stage contour edges, segmented associated edges, and mid-segment associated edges into the adjacency relation matrix; and writes the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight into the edge weight matrix. The node state matrix, adjacency relation matrix, and edge weight matrix are then input into the graph attention network. The adjacency relation matrix is ​​used to read the adjacent node numbers of the current contour node; the node state matrix is ​​used to read the top plate state value, side wall state value, infill bearing capacity value, and node boundary type marker of the adjacent nodes; and the edge weight matrix is ​​used to read the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight between the current contour node and its adjacent nodes. The current contour node state value is then... The state values ​​of adjacent nodes are concatenated to form a node state pair. The node state pair is multiplied by the corresponding edge weight in the edge weight matrix to obtain the attention score of the adjacent node. The attention scores of multiple adjacent nodes corresponding to the same current contour node are normalized to the node attention weight. Nodes whose node attention weight is less than the abnormal weight threshold and whose filling body bearing value is less than the bearing threshold are marked as low bearing filling body adjacent nodes. Nodes whose node attention weight is greater than the abnormal state threshold and whose top plate state value is greater than the top plate state threshold are marked as top plate abnormal nodes. Nodes whose node attention weight is greater than the abnormal state threshold and whose side wall state value is greater than the side wall state threshold are marked as side wall abnormal nodes. The low bearing filling body adjacent connection items are deleted. Top plate stable nodes, side wall stable nodes, and middle section associated nodes are retained. The node number, node attention weight, abnormal node mark, and retained node mark are written into the abnormal adjacent node table.

[0011] As a further aspect of the present invention, the specific steps for generating the rule-closed path cluster are as follows: Based on the aforementioned attention displacement constraint sequence, candidate connection point numbers are called according to the first endpoint of the gap and the second endpoint of the gap, respectively. Using the A* algorithm, the forward candidate point sequence and the reverse candidate point sequence are arranged, and the endpoint spacing, path segment length, adjacent turning angle, and node displacement component difference are calculated to generate a candidate endpoint path table. Based on the candidate endpoint path table, the bidirectional closure relationship is verified according to the forward path direction, reverse path direction, cumulative path segment length, cumulative angle, and endpoint proximity value. Paths with opposite directions, abrupt angle changes, and endpoint deviations are deleted, and a bidirectional closure verification table is generated. Based on the bidirectional closure check table, expansion, contraction, translation, turning, closure, and stable states are marked according to the boundary length change rate, boundary direction change angle, endpoint proximity value, adjacent turning angle change amount, and preset state judgment table. Node sequences that connect to the reverse contraction state after expansion state, the expansion state after closure state, and the turning state after stable state are filtered out to generate a regular closed path cluster.

[0012] As a further aspect of the present invention, the A* algorithm first sets the first endpoint of the gap as the forward starting point and the second endpoint of the gap as the reverse starting point. The candidate connection point number, node coordinates, directional displacement, dip displacement, and elevation displacement are written into the candidate node table. The endpoint spacing, path segment length, adjacent turning angles, and node displacement component differences are written into the path cost table. The A* algorithm is then used to read the adjacent candidate connection point numbers from the forward and reverse starting points, calculate the path segment length from the current node to the adjacent candidate connection point, calculate the remaining distance from the adjacent candidate connection point to the opposite gap endpoint, and record the cumulative path segment length, remaining distance, adjacent turning angle penalty, and node displacement component difference penalty. The values ​​are normalized by dividing each value by the corresponding cost benchmark value. The values ​​are then weighted and summed according to the first cost weight, the second cost weight, the third cost weight, and the fourth cost weight to obtain the candidate node ranking value. Candidate connection points with smaller ranking values ​​are selected and written into the forward open sequence and the reverse open sequence. Read nodes are written into the forward closed sequence and the reverse closed sequence. Expansion stops when the distance between the endpoints of the forward open sequence and the reverse open sequence is less than the closing distance threshold. The forward node chain and the reverse node chain are backtracked according to the parent node number. The forward candidate point sequence and the reverse candidate point sequence are arranged. The endpoint spacing, path segment length, adjacent turning angle, and node displacement component difference are calculated to generate a candidate endpoint path table.

[0013] As a further aspect of the present invention, the specific steps for generating the constrained backtracking contour chain are as follows: Based on the aforementioned closed path cluster, the span of the closed boundary segment is measured segment by segment, and the enclosed area, hydraulic radius, and elevation interval between the top plate node and the top surface node of the filling body are calculated. The excess boundary segment is marked with the span limit, area limit, hydraulic radius limit, and top plate gap limit, and a closed boundary constraint table is generated. Based on the closed boundary constraint table, the closed path crossing status is verified according to the boundary of the isolated pillar, the boundary of the triangular pillar, and the boundary of the low-bearing filling body. The crossing path number is deleted, the node to be retreated is located according to the number of the over-limit boundary segment, the node displacement is reduced according to the coordinates of the adjacent non-over-limit nodes, and a constraint retreat contour chain is established.

[0014] As a further aspect of the present invention, the specific steps for generating the mining area contour projection map are as follows: Based on the constrained retreat contour chain, the contour nodes are positioned according to the current layer, recovery layer and filling layer of the goaf, the boundaries to be recovered are arranged according to the recovery sequence number, the filled boundaries are arranged according to the filling sequence number, and the nodes that cannot be crossed are locked according to the isolation boundary status, thus generating a stage contour node sequence. Based on the constrained retreat contour chain, the contour nodes are positioned according to the current layer, recovery layer, and filling layer of the goaf. The boundaries to be recovered are arranged according to the recovery sequence number, the recovered boundaries are marked according to the executed recovery sequence number, the filled boundaries are arranged according to the filling sequence number, and the nodes that cannot be crossed are locked according to the isolation boundary status, thus generating a stage contour node sequence.

[0015] A system for estimating the contour of a mining area based on graph attention networks is provided. This system is used to execute the aforementioned method for estimating the contour of a mining area based on graph attention networks. The system includes: Topology chain building module: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, middle section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. Attention weighting module: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use graph attention network to calculate node attention weights, lock abnormal nodes of the top plate and side wall, remove adjacent connection items of low bearing capacity infill, and generate attention displacement constraint sequence. Gap Closure Module: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster; Constraint backoff module: Based on the rule-based closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing-capacity filling bodies, and establish a constraint backoff contour chain; Map output module: Based on the constrained back contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the planar boundary points and longitudinal boundary points, filter the boundary state paths, and generate a mining area contour projection map.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the boundary nodes are no longer connected according to a single planar outline, but form a hierarchical boundary topology chain with constraints on strata, sidewalls, roof and floor, and boundary status, so that the mining outline deduction has both mid-level hierarchical relationship and segmented spatial relationship. In this invention, based on the hierarchical boundary topology chain, coordinate difference, projection offset, inter-layer edge weight of the middle section, top plate state value, side wall state value, and infill bearing capacity value jointly participate in the weight allocation of the graph attention network. Abnormal top plate nodes, abnormal side wall nodes, and low bearing capacity infill adjacent connection items are locked or eliminated before displacement constraints, so that the strike displacement, dip displacement, and elevation displacement are jointly restricted by the boundary state and bearing capacity. In this invention, based on the displacement constraint sequence, the A* algorithm arranges candidate connection points from both ends of the gap and deletes reverse jump boundary segments by endpoint distance, path length, adjacent turning angle, and bidirectional path direction difference, so that the closed path is subject to bidirectional consistency constraints. In this invention, the regular closed path clusters are then regressed by constraints of span, enclosed area, hydraulic radius, roof gap, isolation pillar, triangular pillar, and low-bearing backfill body boundary, outputting a stope contour projection map with current layer, recovery layer, backfill layer, and boundary state paths. The gains are reflected in the contour connection relationship being more suitable for irregular goaf areas, fewer gap closure error paths, more controllable interference of roof and sidewall risk nodes on the projection results, and a more continuous contour evolution relationship between the recovery and backfill stages. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a sampling field contour inference method based on graph attention networks, comprising the following steps: S1: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, mid-section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. S2: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use a graph attention network to lock the abnormal nodes of the top plate and side slope, remove the adjacent connection terms of low-bearing infill bodies, and generate an attention displacement constraint sequence. S3: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster. S4: Based on the rule-based closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing filling bodies, and establish a constrained retreat contour chain; S5: Based on the constraint backtracking contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the plane boundary points and longitudinal boundary points, filter the boundary state paths, and generate the mining area contour projection map.

[0020] The hierarchical boundary topology chain includes contour node number, node layer number, boundary type label, and node connection order. Note that the displacement constraint sequence includes node attention weight, strike displacement limit, dip displacement limit, and elevation displacement limit. The regular closure path cluster includes candidate path number, endpoint pairing number, path status label, and closure deviation value. The constrained retreat contour chain includes retreat node number, retreat direction label, displacement reduction amplitude, and constraint boundary number. The stope contour projection map includes the recovery contour layer, the filling contour layer, and the constraint contour layer.

[0021] The specific steps for generating hierarchical boundary topology chains are as follows: Based on the planar boundaries, cross-sectional boundaries, roof boundaries, floor boundaries, sidewall boundaries, intermediate elevations, segmental elevations, stope advance sequence, stope outline gap endpoints, ore-filled boundary points, and backfill boundary points of the goaf, a three-dimensional layered topological sorting method is adopted. The intermediate elevation field is set as the primary sorting key, the segmental elevation field as the secondary sorting key, and the stope advance sequence number as the tertiary sorting key. Key-value binding is performed on boundary point numbers, planar coordinates, cross-sectional coordinates, roof elevation, floor elevation, and sidewall endpoint numbers. The layers are arranged from highest to lowest intermediate elevation and numbered from nearest to farthest segmental elevation. The boundary points within a single layer are numbered in ascending order according to the mining advance sequence. The minimum elevation difference pairing method is used to calculate the difference between the elevation of the roof boundary point and the elevation of the floor boundary point. Boundary points with a difference less than the roof-floor pairing threshold are written into the roof-floor pairing field. The endpoint spacing matching method is used to calculate the Euclidean distance between the coordinates of the endpoints of the mining contour gap and the coordinates of the boundary points in the same layer. Boundary points with a distance less than the gap identification threshold are written into the disconnection position field. The sequence of the numbering of the ore boundary point, the filling body boundary point, the layer number, the roof-floor pairing number, and the disconnection position number is registered to generate the layer boundary point sequence. Based on the stratigraphic boundary point sequence, a directed adjacency linked list construction method is adopted. The stratigraphic number, boundary point number, sidewall endpoint number, top plate point number, bottom plate point number, and disconnection position number are set as index fields. For boundary points in the same stratigraphic level, predecessor pointers and successor pointers are established according to the adjacency relationship of the boundary point numbers. A stratigraphic connection record is established between the planar boundary points and the sidewall endpoints. The cross-sectional projection matching method is used to calculate the projection distance between the planar projection coordinates of the top plate point and the planar projection coordinates of the bottom plate point. Top plate points and bottom plate points whose projection distance is less than the cross-sectional connection threshold are written into the cross-sectional connection field. A boundary status marking table is used to record the status value corresponding to the ore boundary point of the storage hole as the boundary to be recovered, the status value corresponding to the filling body boundary point as the filled boundary, and the status value corresponding to the sidewall endpoint as the sidewall boundary. The connection number of the gap endpoint is written into the disconnection segment pointer field. The stratigraphic connection record, cross-sectional connection field, boundary status value, and disconnection segment pointer field are chained and registered to establish a hierarchical boundary topology chain.

[0022] The specific steps for generating the attention displacement constraint sequence are as follows: Based on a hierarchical boundary topology chain, a three-dimensional coordinate difference weighting method is used. The adjacent node numbers at the same level are set as the search keys. The starting coordinate fields are set as Xs, Ys, and Zs, and the ending coordinate fields are set as Xe, Ye, and Ze. The endpoint numbers of the contour edges at the same stage, the endpoint numbers of the segmented associated edges, and the endpoint numbers of the middle segment associated edges are read. The directional coordinate difference is calculated by subtracting Xs from Xe, the dip coordinate difference by subtracting Ys from Ye, and the elevation coordinate difference by subtracting Zs from Ze. A segmented projection offset calculation method is used, reading the segmented projection node numbers and the projection starting point coordinates. Given Xp1, Yp1, and the projection endpoint coordinates Xp2 and Yp2, calculate the sum of the squared values ​​of Xp2 minus Xp1 and Yp2 minus Yp1, and then take the square root to obtain the plane projection offset. Using the inter-layer edge weight assignment table, assign a value of 1.00 to the contour edge of the same stage, a value of 0.75 to the segmented associated edge, and a value of 0.50 to the middle segment associated edge. Write the middle segment layer number, edge type number, strike coordinate difference, dip coordinate difference, elevation coordinate difference, plane projection offset, and inter-layer edge weight into the same node record to generate a node differential weight table. Based on a node differential weight table, a graph attention network is employed. The top slab state value, sidewall state value, infill bearing capacity, and node boundary type are defined as node state inputs. The strike coordinate difference, dip coordinate difference, elevation coordinate difference, plane projection offset, and inter-layer edge weight are defined as edge weight inputs. The number of attention heads is set to 4, and the weight normalization range is set to 0 to 1. The set of adjacent nodes is read according to node number. The current node state input is concatenated with the adjacent node state inputs to form a node state pair. Multiplying the node state pair by the edge weight input yields the attention of the adjacent node. The scoring process involves normalizing the attention scores of multiple adjacent nodes under the same current node to obtain the node attention weight. Nodes with a top plate status value greater than the top plate threshold and a node attention weight greater than the abnormal threshold are marked as top plate abnormal nodes. Nodes with a side wall status value greater than the side wall threshold and a node attention weight greater than the abnormal threshold are marked as side wall abnormal nodes. Nodes with a filling body bearing capacity value less than the bearing capacity threshold are marked as low-bearing filling body adjacent nodes. Low-bearing filling body adjacent connection items are deleted, and top plate stable nodes, side wall stable nodes, and mid-section associated nodes are retained to generate an abnormal adjacent node table. Based on an abnormal adjacency node table, a displacement component amplitude limiting calculation method is adopted. Node coordinate difference, segmented projection offset, inter-layer edge weight of the middle segment, node attention weight, and abnormal node marking are set as displacement filtering inputs. The strike coordinate difference, dip coordinate difference, and elevation coordinate difference are read according to the node number. The strike coordinate difference is multiplied by the node attention weight to obtain the strike candidate displacement, the dip coordinate difference is multiplied by the node attention weight to obtain the dip candidate displacement, and the elevation coordinate difference is multiplied by the node attention weight to obtain the elevation candidate displacement. Nodes with segmented projection offsets greater than the projection threshold are marked as abnormal nodes. Move nodes, mark nodes with inter-layer edge weights less than the inter-layer weight threshold as weak inter-layer connection nodes, multiply the candidate displacements corresponding to roof anomalous nodes, sidewall anomalous nodes, projection offset nodes, and weak inter-layer connection nodes by a reduction factor of 0.60, limit the reduced strike candidate displacements within the strike limit range, limit the reduced dip candidate displacements within the dip limit range, and limit the reduced elevation candidate displacements within the elevation limit range. Write the strike displacement, dip displacement, elevation displacement, and constraint mark according to the node number to generate a sequence of attention displacement constraints.

[0023] The graph attention network first writes the top slab state value, sidewall state value, infill bearing capacity value, and node boundary type label into the node state matrix. Then, it writes the contour edges of the same stage, segmented associated edges, and mid-segment associated edges into the adjacency relation matrix. Finally, it writes the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight into the edge weight matrix. The node state matrix, adjacency relation matrix, and edge weight matrix are then input into the graph attention network. The adjacency relation matrix is ​​used to read the adjacent node numbers of the current contour node. The node state matrix is ​​used to read the top slab state value, sidewall state value, infill bearing capacity value, and node boundary type label of adjacent nodes. The edge weight matrix is ​​used to read the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight between the current contour node and its adjacent nodes. The current contour node state value and the adjacent node state values ​​are then compared. Values ​​are concatenated into node state pairs. The node state pairs are multiplied by the corresponding edge weights in the edge weight matrix to obtain the attention scores of adjacent nodes. The attention scores of multiple adjacent nodes corresponding to the same current contour node are normalized into node attention weights. Nodes whose node attention weights are less than the abnormal weight threshold and whose filling load value is less than the load threshold are marked as low-load filling adjacent nodes. Nodes whose node attention weights are greater than the abnormal state threshold and whose top plate state value is greater than the top plate state threshold are marked as top plate abnormal nodes. Nodes whose node attention weights are greater than the abnormal state threshold and whose side wall state value is greater than the side wall state threshold are marked as side wall abnormal nodes. Low-load filling adjacent connection items are deleted. Top plate stable nodes, side wall stable nodes, and middle section associated nodes are retained. Node numbers, node attention weights, abnormal node marks, and retained node marks are written into the abnormal adjacent node table. The top plate condition value is obtained by weighted summation of the top plate crack density ratio, the top plate settlement ratio, and the top plate exposed span ratio. The side wall condition value is obtained by weighted summation of the side wall convergence ratio, the side wall crack number ratio, and the side wall inclination deviation ratio. The filling body bearing capacity value is obtained by weighted calculation of the filling body strength ratio, the filling body age ratio, the resilient modulus ratio, and the top surface elevation difference ratio.

[0024] The specific steps for generating a family of closed regular paths are as follows: Based on a hierarchical boundary topology chain, a three-dimensional coordinate difference weighting method is used. The adjacent node numbers at the same level are set as the search keys. The starting coordinate fields are set as Xs, Ys, and Zs, and the ending coordinate fields are set as Xe, Ye, and Ze. The endpoint numbers of the contour edges at the same stage, the endpoint numbers of the segmented associated edges, and the endpoint numbers of the middle segment associated edges are read. The directional coordinate difference is calculated by subtracting Xs from Xe, the dip coordinate difference by subtracting Ys from Ye, and the elevation coordinate difference by subtracting Zs from Ze. A segmented projection offset calculation method is used, reading the segmented projection node numbers and the projection starting point coordinates. Given Xp1, Yp1, and the projection endpoint coordinates Xp2 and Yp2, calculate the sum of the squared values ​​of Xp2 minus Xp1 and Yp2 minus Yp1, and then take the square root to obtain the plane projection offset. Using the inter-layer edge weight assignment table, assign a value of 1.00 to the contour edge of the same stage, a value of 0.75 to the segmented associated edge, and a value of 0.50 to the middle segment associated edge. Write the middle segment layer number, edge type number, strike coordinate difference, dip coordinate difference, elevation coordinate difference, plane projection offset, and inter-layer edge weight into the same node record to generate a node differential weight table. Based on a node differential weight table, a graph attention network is employed. The top slab state value, sidewall state value, infill bearing capacity, and node boundary type are defined as node state inputs. The strike coordinate difference, dip coordinate difference, elevation coordinate difference, plane projection offset, and inter-layer edge weight are defined as edge weight inputs. The number of attention heads is set to 4, and the weight normalization range is set to 0 to 1. The set of adjacent nodes is read according to node number. The current node state input is concatenated with the adjacent node state inputs to form a node state pair. Multiplying the node state pair by the edge weight input yields the attention of the adjacent node. The scoring process involves normalizing the attention scores of multiple adjacent nodes under the same current node to obtain the node attention weight. Nodes with a top plate status value greater than the top plate threshold and a node attention weight greater than the abnormal threshold are marked as top plate abnormal nodes. Nodes with a side wall status value greater than the side wall threshold and a node attention weight greater than the abnormal threshold are marked as side wall abnormal nodes. Nodes with a filling body bearing capacity value less than the bearing capacity threshold are marked as low-bearing filling body adjacent nodes. Low-bearing filling body adjacent connection items are deleted, and top plate stable nodes, side wall stable nodes, and mid-section associated nodes are retained to generate an abnormal adjacent node table. Based on an abnormal adjacency node table, a displacement component amplitude limiting calculation method is adopted. Node coordinate difference, segmented projection offset, inter-layer edge weight of the middle segment, node attention weight, and abnormal node marking are set as displacement filtering inputs. The strike coordinate difference, dip coordinate difference, and elevation coordinate difference are read according to the node number. The strike coordinate difference is multiplied by the node attention weight to obtain the strike candidate displacement, the dip coordinate difference is multiplied by the node attention weight to obtain the dip candidate displacement, and the elevation coordinate difference is multiplied by the node attention weight to obtain the elevation candidate displacement. Nodes with segmented projection offsets greater than the projection threshold are marked as abnormal nodes. Move nodes, mark nodes with inter-layer edge weights less than the inter-layer weight threshold as weak inter-layer connection nodes, multiply the candidate displacements corresponding to roof anomalous nodes, sidewall anomalous nodes, projection offset nodes, and weak inter-layer connection nodes by a reduction factor of 0.60, limit the reduced strike candidate displacements within the strike limit range, limit the reduced dip candidate displacements within the dip limit range, and limit the reduced elevation candidate displacements within the elevation limit range. Write the strike displacement, dip displacement, elevation displacement, and constraint mark according to the node number to generate a sequence of attention displacement constraints.

[0025] The A* algorithm first sets the first endpoint of the gap as the forward starting point and the second endpoint as the reverse starting point. It writes the candidate connection point number, node coordinates, directional displacement, dip displacement, and elevation displacement into the candidate node table. It writes the endpoint spacing, path segment length, adjacent turning angles, and node displacement component differences into the path cost table. Using the A* algorithm, it reads the adjacent candidate connection point numbers from both the forward and reverse starting points, calculates the path segment length from the current node to the adjacent candidate connection point, and calculates the remaining distance from the adjacent candidate connection point to the opposite gap endpoint. The cumulative path segment length, remaining distance, adjacent turning angle penalty, and node displacement component difference penalty are then divided by the corresponding... The cost baseline value is normalized, and the candidate node ranking value is obtained by weighted summation according to the first cost weight, the second cost weight, the third cost weight, and the fourth cost weight. The candidate connection point with the smaller ranking value is selected and written into the forward open sequence and the reverse open sequence. The read nodes are written into the forward closed sequence and the reverse closed sequence. The expansion stops when the distance between the endpoints of the forward open sequence and the endpoints of the reverse open sequence is less than the closing distance threshold. The forward node chain and the reverse node chain are backtracked according to the parent node number, the forward candidate point sequence and the reverse candidate point sequence are arranged, the endpoint spacing, path segment length, adjacent turning angle, and node displacement component difference are calculated, and the candidate endpoint path table is generated.

[0026] The specific steps for generating a constraint backtracking contour chain are as follows: Based on a cluster of rule-based closed paths, a segmented measurement method for boundary span is adopted. This involves setting the sequence of closed path nodes, node coordinate fields, and closed boundary segment numbers. Adjacent node numbers and coordinates are read, and the span of the closed boundary segment is measured according to the spatial distance between adjacent nodes. A closed contour area measurement method is used to enclose the boundary area according to the sequence of closed path nodes and read the enclosed area. A hydraulic radius determination method is used to read the hydraulic radius based on the enclosed area and the perimeter of the closed path. The elevations of the top plate nodes and the top surface nodes of the filling body are read, and the elevation interval between the top plate nodes and the top surface nodes of the filling body is read according to the two types of elevation differences. Span limits, area limits, hydraulic radius limits, and top plate gap limits are set. Boundary segments exceeding the span limit are written to the span exceedance field, closed paths exceeding the area limit are written to the area exceedance field, and closed paths exceeding the hydraulic radius limit are written to the hydraulic radius exceedance field. Nodes exceeding the top plate gap limit are written to the top plate gap exceedance field. A closed boundary constraint table is generated by registering the closed path number, closed boundary segment number, exceedance type number, and exceedance value. Based on the closed boundary constraint table, the line segment intersection judgment method is adopted. The closed path boundary segment, the isolation pillar boundary, the triangular pillar boundary, and the low bearing capacity filling body boundary are set. The coordinates of the two ends of the closed path boundary segment and the coordinates of the two ends of the constraint boundary are read. It is determined whether the two ends of the closed path boundary segment are located on both sides of the constraint boundary. If the distribution conditions on both sides are met, the closed path number is written into the crossing path field and the crossing path number is deleted. The neighbor point interpolation backtracking method is adopted. The node to be backtracked is read according to the boundary segment number. The previous non-exceeding node and the next non-exceeding node are read. The median value of the coordinates of the two non-exceeding nodes is taken as the backtracking coordinate. The node displacement reduction is read according to the difference between the original node coordinate and the backtracking coordinate. The backtracking coordinate and the node displacement reduction are written into the backtracking node record. The non-crossing path nodes and the backtracking nodes are connected in the closed path order to establish a constraint backtracking contour chain.

[0027] The specific steps for generating the mining area contour projection map are as follows: Based on the constrained retreat contour chain, a stage-state ordered arrangement method is adopted. The arrangement fields are set as follows: current goaf layer number, recovery layer number, filling layer number, contour node number, recovery sequence number, filling sequence number, and isolation boundary status code. The method reads the nodes of the uncrossed path, retreat nodes, boundary status markers, and node connection order in the constrained retreat contour chain. Current contour nodes are assigned according to the current goaf layer number, boundary nodes to be recovered are assigned according to the recovery layer number, and filled boundary nodes are assigned according to the filling layer number. The recovery sequence number is sorted in ascending order using an ascending sequence number sorting method, and the sorting result is written to the boundary node to be recovered field. The filling sequence number is also sorted in ascending order, and the sorting result is written to the filled boundary field. The isolation boundary status code is read using an isolation status locking method, and the node corresponding to the isolation boundary status code is written to the uncrossable node field. The current contour nodes, boundary nodes to be recovered, filled boundary nodes, and uncrossable nodes are registered according to the node connection order to generate a stage contour node sequence. Based on the stage contour node sequence, the horizontal and vertical coordinate mapping method is adopted. The horizontal and vertical coordinates of the plane boundary points, the vertical elevation of the plane boundary points, the layer number, the boundary status code, and the contour layer field are set as mapping fields. The existing contour nodes, boundary nodes to be recovered, filled boundary nodes, and nodes that cannot be crossed in the stage contour node sequence are read. The nodes of the same layer are located according to the horizontal and vertical coordinates of the plane boundary points, and the roof boundary nodes, floor boundary nodes, and sidewall boundary nodes are located according to the vertical elevation of the vertical boundary points. The status path filtering method is used to read the unprocessed status code, the status code to be recovered, the status code recovered, the status code filled, and the isolation status code. The node paths with the same status code sequence are retained, and the node paths with status code jumps are deleted. The retained node paths are written into the contour layer field according to the existing layer, the recovered layer, and the filled layer of the goaf. The plane boundary points, the vertical boundary points, the boundary status paths, and the layer fields are registered according to the node numbers to generate the mining area contour projection map.

[0028] Please see Figure 2 A graph attention network-based stope contour extrapolation system is used to execute the graph attention network-based stope contour extrapolation method described above. The system includes: Topology chain building module: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, middle section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. Attention weighting module: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use graph attention network to calculate node attention weights, lock abnormal nodes of the top plate and side slope, remove adjacent connection items of low bearing capacity infill, and generate attention displacement constraint sequence. Gap Closure Module: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster; Constraint backoff module: Based on the rule-based closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing-capacity filling bodies, and establish a constraint backoff contour chain; Map output module: Based on the constraint backtracking contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the plane boundary points and longitudinal boundary points, filter the boundary state paths, and generate the mining area contour projection map.

[0029] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for inferring the contour of a sampling field based on graph attention networks, characterized in that, Includes the following steps: S1: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, mid-section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. S2: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use a graph attention network to lock the abnormal nodes of the top plate and side wall, remove the adjacent connection items of low-bearing infill bodies, and generate an attention displacement constraint sequence. S3: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster. S4: Based on the rule-closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing filling bodies, and establish a constrained retreat contour chain; S5: Based on the constraint backtracking contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the plane boundary points and longitudinal boundary points, filter the boundary state paths, and generate the mining area contour projection map.

2. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The hierarchical boundary topology chain includes contour node number, node layer number, boundary type label, and node connection order. The attention displacement constraint sequence includes node attention weight, directional displacement amount and directional displacement limit, dip displacement amount and dip displacement limit, and elevation displacement amount and elevation limit. The regular closed path cluster includes candidate path number, endpoint pairing number, path status label, and closure deviation value. The constrained backtracking contour chain includes backtracking node number, backtracking direction label, displacement reduction amplitude, and constraint boundary number. The mining area contour projection map includes a recovery contour layer, a filling contour layer, and a constrained contour layer.

3. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The specific steps for generating the hierarchical boundary topology chain are as follows: Based on the planar boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, mid-level elevation, sub-level elevation, mining advance sequence, mining outline gap endpoint, ore boundary point of the goaf, and filling body boundary point, the stratum to which the boundary point belongs is determined according to the mid-level elevation and sub-level elevation, the boundary points of the same stratum are arranged according to the mining advance sequence, the roof and floor points are paired according to the elevation correspondence between the roof and floor boundaries, and the disconnection position is marked according to the mining outline gap endpoint to generate a stratum boundary point sequence; Based on the stratigraphic boundary point sequence, connect the planar boundary points and sidewall endpoints in the adjacent order of the same stratigraphic level, connect the top plate points and bottom plate points according to the cross-sectional boundary positions, set boundary status markers according to the ore boundary points and the filling body boundary points, and supplement the endpoints of the broken segments according to the connection numbers of the gap endpoints to establish a hierarchical boundary topology chain.

4. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The specific steps for generating the attention displacement constraint sequence are as follows: Based on the hierarchical boundary topology chain, the starting point coordinates and ending point coordinates are extracted according to the adjacent node numbers of the same layer, the direction coordinate difference, dip coordinate difference, and elevation coordinate difference are calculated, the plane projection offset is extracted according to the segmented projection node number, and the contour edges, segmented related edges, and middle-segment related edges of the same stage are collected according to the middle segment layer number to generate a node differential weight table. Based on the node differential weight table, a graph attention network is used to locate abnormal adjacent nodes by marking the top plate state value, side wall state value, filling body bearing capacity value, and node boundary type. Low bearing capacity filling body adjacent connection items are deleted, and top plate stable nodes, side wall stable nodes, and middle section associated nodes are retained to generate an abnormal adjacent node table. Based on the abnormal adjacent node table, node displacement components are selected according to node coordinate difference, segmented projection offset, inter-layer edge weight of middle section, and abnormal node mark. The strike displacement, dip displacement, and elevation displacement are respectively limited. The displacement components and constraint marks are arranged according to the node number to generate a attention displacement constraint sequence.

5. The sampling field contour extrapolation method based on graph attention network according to claim 4, characterized in that, The graph attention network first writes the top plate state value, side wall state value, infill bearing capacity value, and node boundary type label into the node state matrix. It then writes the contour edges of the same stage, segmented associated edges, and mid-segment associated edges into the adjacency relation matrix. Finally, it writes the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight into the edge weight matrix. The node state matrix, adjacency relation matrix, and edge weight matrix are then input into the graph attention network. The adjacency relation matrix reads the adjacent node numbers of the current contour node. The node state matrix reads the top plate state value, side wall state value, infill bearing capacity value, and node boundary type label of adjacent nodes. The edge weight matrix reads the strike coordinate difference, dip coordinate difference, elevation coordinate difference, planar projection offset, and mid-segment inter-layer edge weight between the current contour node and its adjacent nodes. The current contour node state value and the adjacent node state value are concatenated into a node state pair. The node state pair is then compared with the corresponding edge in the edge weight matrix. The weights are mapped to the first and second attention input vectors of the same dimension, respectively. The Hadamard product is performed on the first and second attention input vectors to obtain the attention scores of adjacent nodes. The attention scores of multiple adjacent nodes corresponding to the same current contour node are normalized into node attention weights. Nodes whose node attention weights are less than the abnormal weight threshold and whose filling load value is less than the load threshold are marked as low-load filling adjacent nodes. Nodes whose node attention weights are greater than the state abnormal threshold and whose top plate state value is greater than the top plate state threshold are marked as top plate abnormal nodes. Nodes whose node attention weights are greater than the state abnormal threshold and whose sidewall state value is greater than the sidewall state threshold are marked as sidewall abnormal nodes. Low-load filling adjacent connection items are deleted. Top plate stable nodes, sidewall stable nodes, and mid-section associated nodes are retained. The node number, node attention weight, abnormal node mark, and retained node mark are written into the abnormal adjacent node table.

6. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The specific steps for generating the rule-closed path cluster are as follows: Based on the aforementioned attention displacement constraint sequence, candidate connection point numbers are called according to the first endpoint of the gap and the second endpoint of the gap, respectively. Using the A* algorithm, the forward candidate point sequence and the reverse candidate point sequence are arranged, and the endpoint spacing, path segment length, adjacent turning angle, and node displacement component difference are calculated to generate a candidate endpoint path table. Based on the candidate endpoint path table, the bidirectional closure relationship is verified according to the forward path direction, reverse path direction, cumulative path segment length, cumulative angle, and endpoint proximity value. Paths with opposite directions, abrupt angle changes, and endpoint deviations are deleted, and a bidirectional closure verification table is generated. Based on the bidirectional closure check table, expansion, contraction, translation, turning, closure, and stable states are marked according to the boundary length change rate, boundary direction change angle, endpoint proximity value, adjacent turning angle change amount, and preset state judgment table. Node sequences that connect to the reverse contraction state after expansion state, the expansion state after closure state, and the turning state after stable state are filtered out to generate a regular closed path cluster.

7. The sampling field contour extrapolation method based on graph attention network according to claim 6, characterized in that, The A* algorithm first sets the first endpoint of the gap as the forward starting point and the second endpoint of the gap as the reverse starting point. It writes the candidate connection point number, node coordinates, directional displacement, dip displacement, and elevation displacement into a candidate node table. It writes the endpoint spacing, path segment length, adjacent turning angles, and node displacement component differences into a path cost table. The A* algorithm then reads the adjacent candidate connection point numbers from both the forward and reverse starting points, calculates the path segment length from the current node to the adjacent candidate connection point, and calculates the remaining distance from the adjacent candidate connection point to the opposite gap endpoint. The cumulative path segment length, remaining distance, adjacent turning angle penalty, and node displacement component difference penalty are then divided by the corresponding values. The baseline cost value is normalized, and the candidate node ranking value is obtained by weighted summation according to the first cost weight, the second cost weight, the third cost weight, and the fourth cost weight. The candidate connection point with the smaller ranking value is selected and written into the forward open sequence and the reverse open sequence. The read nodes are written into the forward closed sequence and the reverse closed sequence. The expansion stops when the distance between the endpoints of the forward open sequence and the endpoints of the reverse open sequence is less than the closing distance threshold. The forward node chain and the reverse node chain are backtracked according to the parent node number, the forward candidate point sequence and the reverse candidate point sequence are arranged, the endpoint spacing, path segment length, adjacent turning angle, and node displacement component difference are calculated, and a candidate endpoint path table is generated.

8. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The specific steps for generating the constrained backtracking contour chain are as follows: Based on the aforementioned closed path cluster, the span of the closed boundary segment is measured segment by segment, and the enclosed area, hydraulic radius, and elevation interval between the top plate node and the top surface node of the filling body are calculated. The excess boundary segment is marked with the span limit, area limit, hydraulic radius limit, and top plate gap limit, and a closed boundary constraint table is generated. Based on the closed boundary constraint table, the closed path crossing status is verified according to the boundary of the isolated pillar, the boundary of the triangular pillar, and the boundary of the low-bearing filling body. The crossing path number is deleted, the node to be retreated is located according to the number of the over-limit boundary segment, the node displacement is reduced according to the coordinates of the adjacent non-over-limit nodes, and a constraint retreat contour chain is established.

9. The sampling field contour extrapolation method based on graph attention network according to claim 1, characterized in that, The specific steps for generating the mining area contour projection map are as follows: Based on the constrained retreat contour chain, the contour nodes are positioned according to the current layer, recovery layer and filling layer of the goaf, the boundaries to be recovered are arranged according to the recovery sequence number, the filled boundaries are arranged according to the filling sequence number, and the nodes that cannot be crossed are locked according to the isolation boundary status, thus generating a stage contour node sequence. Based on the constrained retreat contour chain, the contour nodes are positioned according to the current layer, recovery layer, and filling layer of the goaf. The boundaries to be recovered are arranged according to the recovery sequence number, the recovered boundaries are marked according to the executed recovery sequence number, the filled boundaries are arranged according to the filling sequence number, and the nodes that cannot be crossed are locked according to the isolation boundary status, thus generating a stage contour node sequence.

10. A mining field contour extrapolation system based on graph attention networks, characterized in that, The sampling contour extrapolation method based on graph attention networks according to any one of claims 1-9, wherein the system comprises: Topology chain building module: Based on the goaf plane boundary, profile boundary, roof boundary, floor boundary, sidewall boundary, middle section elevation, segment elevation, mining advance sequence, mining outline gap endpoint, ore boundary point, and filling body boundary point, arrange the boundary point strata, connect the sidewall endpoints, mark the boundary point order, and establish a hierarchical boundary topology chain. Attention weighting module: Based on the hierarchical boundary topology chain, calculate the coordinate difference between adjacent contour points, collect the segmented projection offset, assign inter-layer edge weights to the middle section, use graph attention network to calculate node attention weights, lock abnormal nodes of the top plate and side wall, remove adjacent connection items of low bearing capacity infill, and generate attention displacement constraint sequence. Gap Closure Module: Based on the attention displacement constraint sequence, the A* algorithm is used to arrange the candidate connection points at both ends of the gap, calculate the endpoint distance, path length, and adjacent turning angle, verify the bidirectional path direction difference, delete the reverse jump boundary segment, and generate a regular closed path cluster; Constraint backoff module: Based on the rule-based closed path cluster, measure the span of the closed boundary segment, calculate the enclosed area, hydraulic radius, and roof gap, screen out paths that cross isolation pillars, triangular pillars, and low-bearing-capacity filling bodies, and establish a constraint backoff contour chain; Map output module: Based on the constrained back contour chain, arrange the contour nodes of the current layer, recycling layer, and filling layer, verify the recycling and filling order, map the planar boundary points and longitudinal boundary points, filter the boundary state paths, and generate a mining area contour projection map.

Images