Mine ecological restoration management method and system based on multiple modes

Abstract

The invention relates to the technical field of ecological restoration management, in particular to a mine ecological restoration management method and system based on multiple modalities, and the method comprises the steps: obtaining boundary disturbance intensity distribution through extracting the spatial change gradient of multi-source variables of vegetation, hydrology, terrain and soil, and achieving the spatial deconstruction of an edge disturbance region; the path structure is in a spatial connection relation; frequency statistics and serial number aggregation are carried out on operation execution sequences, low-frequency behavior combinations are screened out, the expression sparsity and generalization ability of operation data are improved, sensitive behaviors in process evolution are monitored, and the expression sparsity and generalization ability of the operation data are improved; an action distribution diagram is formed by marking adjacent node frequency offset and extreme value recognition through tail end action, a key control position under disturbance influence is visually reflected, synchronous improvement of quantitative analysis capacity and intervention path screening efficiency is achieved, and the system supporting capacity for accurate regulation and control of mine ecological restoration operation is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological restoration management, and in particular to a multimodal-based mine ecological restoration management method and system. Background Art

[0002] The field of ecological restoration and management technology mainly targets ecosystems damaged by human activities and affected by natural disasters, especially ecological degradation problems caused by the development of mineral resources, and carries out organized and planned ecological function restoration and environmental governance. The core goal is to restore ecological service functions, soil and water conservation capacity, biodiversity and surface vegetation systems to ensure the stability and sustainability of the ecosystem. It covers multiple aspects such as environmental information acquisition, ecological status assessment, restoration plan design, construction supervision and restoration effect monitoring. Through remote sensing images, geographic information systems and environmental model technologies, it realizes systematic restoration and management of damaged areas.

[0003] The main purpose of the multimodal mine ecological restoration management method and system is to build a digital management mechanism that can take into account the accuracy of ecological restoration and the feasibility of engineering implementation. Through a systematic approach, the relationship between various elements in mine restoration is coordinated, and multi-source heterogeneous data such as terrain, remote sensing, climate, soil, and vegetation are integrated to establish an ecological degradation evaluation model and a restoration path decision-making mechanism. The aim is to improve the scientific nature, execution efficiency and ecological restoration quality of the restoration plan, achieve effective reconstruction and continuous maintenance of the mine ecosystem, improve operational response efficiency, shorten the average intervention time in the restoration path identification and execution scheduling process by no less than 30%, and increase the vegetation recovery index in the restoration area by no less than 0.2 within 180 days after the intervention.

[0004] Although existing technologies have covered the links of remote sensing data extraction, ecological status assessment and restoration plan formulation, they have obvious limitations in the precise identification of damaged areas, the formation mechanism of intervention paths and the monitoring of dynamic behavioral anomalies. It is difficult to characterize the differences in response structures within the spatial disturbance boundary, resulting in discontinuous distribution of restoration plans at the edge. Existing plans usually adopt fixed rules and experience-based operating logic, and fail to fully consider the structural dependencies and superimposed effects between multi-factor interventions. The path distribution is prone to imbalance, and there is a lack of dynamic analysis between continuous behavioral states. It is difficult to timely identify high-frequency fluctuation segments and potential abnormal operations. At the system level, there is a lack of a chain modeling mechanism from disturbance identification to behavioral sequence evolution, which leads to a disconnect between plan formulation and process execution, increasing regulation costs and response delays. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a multimodal mine ecological restoration management method and system.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solution: a multimodal mine ecological restoration management method, comprising the following steps: Step 1: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, extract unit differences and calculate gradients. After locating boundaries, superimpose structures to generate a boundary disturbance intensity distribution map. Step 2: Based on the boundary disturbance intensity distribution map, the strength of the spoil dump edge segment is grouped, the change path is extracted, the continuous peaks are screened and a node sequence is constructed. The disturbance relationship between nodes is weighted modeled using a graph neural network, the interaction intensity weight is obtained, and an intervention priority structure path set is generated; Step 3: Based on the intervention priority structure path set, the collapse area operation sequences are numbered and the frequencies are counted. The low-frequency operations are eliminated and the high-frequency operations are combined and recombined to generate a stable intervention behavior sequence cluster; Step 4: Based on the stable intervention behavior sequence cluster, the state is transferred and windowed, and the Markov model is used to identify the stage change of the state sequence. The ratio is calculated and the increase segment is suddenly divided. The density cluster and difference are determined to generate the path state abnormality segmentation structure; Step 5: Based on the path state abnormality segmentation structure, traverse the adjacent nodes of the abnormal segment and calculate the operation frequency offset, identify the extreme value position and operation distribution range, and generate a key action group distribution map under the influence of disturbance.

[0007] As a further solution of the present invention, the specific steps of generating the boundary disturbance intensity distribution map are: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, the surface unit grid is divided and five numerical parameters corresponding to each unit cell are extracted. The corresponding values ​​between adjacent units are subtracted and the change slope is calculated according to the direction. The gradient values ​​are synthesized and superimposed to generate a unit gradient distribution matrix. Based on the unit gradient distribution matrix, high-value edges in continuous areas are extracted and segments with fluctuation value differences less than a threshold in the boundary segments are screened out. The edges of the remaining areas are traced and the contours are extracted. The consistency of the structural connection and orientation between multiple edge segments is checked and the sidebands are merged to generate a boundary disturbance intensity distribution map.

[0008] As a further solution of the present invention, the specific steps of generating the intervention priority structure path set are: Based on the boundary disturbance intensity distribution map, the disturbance value within the edge of the spoil dump is extracted and the peak difference value of adjacent grids is calculated. The intensity level boundary is set according to the numerical interval. The continuity of each intensity segment is determined and classified by labeling to generate a disturbance intensity jump structure map. Based on the disturbance intensity jump structure diagram, equal-length steps in the jump segment area are extracted and it is determined whether the local peak value exceeds the mean of the adjacent two-side segments. The low-intensity isolated segments are eliminated and the continuous high-value path is retained. The node identifiers are constructed and numbered in sequence according to the position to generate an edge segment path node sequence diagram; Based on the edge path node sequence diagram, a graph neural network is used to perform weighted modeling of the disturbance relationship between nodes and extract significant response features between structures. The distance value and the disturbance value difference between each node are calculated as the product to form an influence matrix, and the matrix is ​​normalized into an interaction ratio sequence according to the rows and columns. The sequence is used to construct a connection weight table to generate an intervention priority structural path set.

[0009] As a further solution of the present invention, the graph neural network is according to the formula:

[0010] in: Representation node For adjacent nodes The attention weight coefficient, represents the exponential function, Represents the activation function LeakyRectifiedLinearUnit, Represents the attention weight vector, which is transformed with the perturbation feature matrix during network training Synchronous iterative update, using the node feature splicing perturbation difference and spatial position difference to calculate the attention coefficient, and then converge after minimizing the loss function through back propagation, represents the transpose of the attention weight vector, represents the perturbation feature transformation matrix, Representation node The initial feature vector of Representation node The initial feature vector of represents the vector concatenation operation, Representation node With node The disturbance value difference, Representation node With node The spatial position difference, Representation node The set of adjacent nodes of Indicates traversal of nodes The subscript variables of all adjacent nodes, Representation node With node The disturbance value difference, Representation node With node The spatial position difference.

[0011] As a further solution of the present invention, the execution process of the graph neural network includes the following operations: constructing the connection relationship between nodes based on the edge path node sequence graph, extracting the disturbance attribute value and spatial index information of each node, initializing the disturbance response representation vector of each node, traversing the set of adjacent nodes, calculating the disturbance value difference and spatial distance difference between the target node and the adjacent node, and normalizing the above two differences as the attention weighting factor, performing aggregation on the adjacent node disturbance representation vector according to the weighting factor, updating the disturbance response representation vector of the current node, repeating the above process until all nodes complete the embedding update, and outputting the disturbance correlation distribution value between each node and the adjacent node.

[0012] As a further solution of the present invention, the specific steps of generating the stable intervention behavior sequence cluster are: Based on the intervention priority structure path set, all action types in the intervention nodes in the collapse area are numbered according to their sequence positions and an action comparison table is established. The continuous action combinations in each path sequence are read and the number of combination occurrences is counted. The numbered combinations below the specified frequency threshold are eliminated to generate an operation sequence number frequency table. Based on the operation sequence number frequency table, the position index difference between high-frequency combination numbers is calculated and the number groups with the difference within the set range are screened. The connection and reorganization are performed by judging the similarity of the action types within the group, and the number sequence is rearranged according to the path structure to generate a stable intervention behavior sequence cluster.

[0013] As a further solution of the present invention, the specific steps of generating the abnormal path status segmentation structure are: Based on the stable intervention behavior sequence cluster, the states in the sequence are numbered in execution order and the time window length is set in sequence. The Markov model is used to model the stability of the state transition process and identify stage changes. The ratio value of the number of state changes and the number of state stays in each window is calculated, and the sequence segments with a change ratio greater than the set value are extracted to generate a state jump segment structure set. Based on the state transition segment structure set, the state position index and frequency statistics in each transition segment are extracted and a quantity matrix is ​​constructed. The dense segments are extracted and the index intervals are marked according to the proximity degree and node concentration degree between consecutive states to generate a state clustered high-density segment mark set. Based on the state-aggregated high-density segment marker set, the difference between the inter-segment state transition value sequence and the change frequency sequence is calculated according to the same start and end point positions. The boundaries of the segments with continuous differences greater than the threshold are intercepted and the center of gravity of the segmentation line is adjusted to generate a path state abnormality segmentation structure.

[0014] As a further solution of the present invention, the Markov model is according to the formula:

[0015] in: represents the improved state stability index, An index variable representing the state number, Indicates the total number of states involved in the calculation in the current time window, represents the weighted coefficient of the adjustment state stability difference factor, Represents the weighted coefficient for adjusting the state access frequency factor, Represents the weighted coefficient for adjusting the state dwell step length factor, represents the weighting coefficient for adjusting the transfer frequency difference factor, Indicates the The self-transition probability of a number, Indicates the The mean of all transition probabilities of a number, Indicates the The total number of visits to a number in the current sliding window, Indicates the peak value of all state access times in the current sliding window, Indicates the The number stays in the window continuously for the cumulative value of the step length, represents the average of the step lengths within the window, Indicates the The standard deviation of the numbered transfer frequencies, Indicates the average reference value of the standard deviation of the transition frequencies within the window.

[0016] As a further solution of the present invention, the specific steps of generating the key action group distribution map under the influence of the disturbance are: Based on the path state anomaly segmentation structure, the start and end nodes of each segment are numbered in sequence and a connection index table is established. The difference between the action set in each node and the operation of the previous and next nodes is calculated. The frequency of operation type changes in each adjacent node group is counted to generate an adjacent node frequency offset set. Based on the frequency offset set of adjacent nodes, the frequency offset sequence is extracted by sliding the window range and the peak position is calculated. The operation index corresponding to each extreme value is used to find the associated number and establish an operation coverage matrix. The operation number is aggregated to form a continuous block number set, and a distribution map of the key action group under the influence of disturbance is generated.

[0017] A multimodal mine ecological restoration management system, comprising: Disturbance Perception Module: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, it performs gridded numerical comparison and gradient intensity superposition of multiple ecological factors to generate a boundary disturbance intensity distribution map; Path identification module: Based on the boundary disturbance intensity distribution map, the intervention intensity edge segments in the spoil dump area are classified and the peak connection paths are extracted. The graph neural network is used to weight the relationship between nodes and extract the structural path set to generate the intervention priority structural path set; Behavior aggregation module: Based on the intervention priority structure path set, it constructs the intervention operation number sequence in the collapse area and screens and rearranges the high-frequency combinations, removes the low-frequency segments and reconstructs the continuous segment combination structure to generate a stable intervention behavior sequence cluster; Phase division module: Based on the stable intervention behavior sequence cluster, the operation state is segmented by sliding the time window, and the Markov model is used to determine the stability of the state transition and separate the mutation window segments to generate the path state abnormality segmentation structure; Action labeling module: Based on the path state anomaly segmentation structure, the operation frequency change amplitude in the abnormal segment is extracted and the extreme value position is determined centrally. The operation area partition information is generated by combining the adjacent node distribution analysis, and the key action group distribution map under the influence of disturbance is generated.

[0018] Compared with the prior art, the advantages and positive effects of the present invention are: 1. In this method, the spatial variation gradients of multiple source variables, including vegetation, hydrology, topography, and soil, are extracted to obtain the boundary disturbance intensity distribution. This method then constructs a highly variable path along the edge of the dump and forms a node topology sequence. This allows for spatial deconstruction of the edge disturbance zone, enabling the path structure to not only demonstrate spatial connectivity but also express the intensity of the disturbance impact. 2. In this invention, by counting the frequency and aggregating the numbers of the operation execution sequence, low-frequency behavior combinations are screened out, and the expression sparsity and generalization ability of the operation data are improved. This enables monitoring of sensitive behaviors in process evolution, screens out low-frequency behavior combinations, and improves the expression sparsity and generalization ability of the operation data. 3. In the present invention, an action distribution diagram is formed through terminal action labeling, adjacent node frequency offset and extreme value identification, which intuitively reflects the key control positions under the influence of disturbance, realizes the simultaneous improvement of quantitative analysis capability and intervention path screening efficiency, and enhances the system support capability for precise regulation of mine ecological restoration operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the workflow of the present invention; Figure 2 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0021] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0022] Example 1 See also Figure 1 The present invention provides a technical solution: a multimodal mine ecological restoration management method, comprising the following steps: Step 1: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, extract unit differences and calculate gradients. After locating boundaries, superimpose structures to generate a boundary disturbance intensity distribution map. Step 2: Based on the boundary disturbance intensity distribution map, the intensity of the spoil dump edge segments is grouped, the change path is extracted, the continuous peaks are screened and a node sequence is constructed. The disturbance relationship between nodes is weighted modeled using a graph neural network, the interaction intensity weight is obtained, and a set of intervention priority structure paths is generated. Step 3: Based on the intervention priority structure path set, the collapse area operation sequences are numbered and the frequencies are counted. The low-frequency operations are eliminated and the high-frequency operations are combined and recombined to generate a stable intervention behavior sequence cluster. Step 4: Based on the stable intervention behavior sequence cluster, the state is transferred and windowed, and the Markov model is used to identify the stage change of the state sequence. The ratio is calculated and the increase segment is suddenly divided. The density cluster and difference are determined to generate the path state abnormal segmentation structure; Step 5: Based on the path state anomaly segmentation structure, traverse the adjacent nodes of the abnormal segment and calculate the operation frequency offset, identify the extreme value position and operation distribution range, and generate the key action group distribution map under the influence of disturbance.

[0023] The specific steps to generate the boundary disturbance intensity distribution map are: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, the surface unit grid is divided and five numerical parameters corresponding to each unit cell are extracted. The corresponding values ​​between adjacent units are subtracted and the change slope is calculated according to the direction. The gradient values ​​are synthesized and superimposed to generate a unit gradient distribution matrix. Based on the unit gradient distribution matrix, high-value edges in continuous regions are extracted and segments with fluctuation values ​​less than a threshold are filtered out. The edges of the remaining regions are traced and contours are extracted. The consistency of the structural connection and orientation between multiple edge segments is checked and the edge bands are merged to generate a boundary disturbance intensity distribution map. Based on the vegetation coverage, slope aggregation degree, number of hydrological aggregation points, surface soil acid-base difference and sediment flow trajectory in the rainy season, the surface units are divided into grids and the five numerical parameters corresponding to each unit are extracted. The difference between the upper, lower, left and right adjacent units of each unit is calculated. The numerical changes corresponding to each group of directions are traversed in sequence according to the surface unit number and the size of the numerical changes are recorded. The horizontal change data table and the vertical change data table are constructed respectively. The values ​​in the above two tables are processed by absolute value and then normalized and coded. Then, the gradient sequence matrix is ​​generated according to the five variables. Five groups of matrices are superimposed in sequence in a fixed order. The participation proportional coefficient is set before each group is superimposed. The proportional coefficients are 0.25, 0.2, 0.2, 0.15 and 0.2 respectively. After superposition, the five-value merge result matrix is ​​output to generate the unit gradient distribution matrix. Based on the unit gradient distribution matrix, high-value edges in continuous areas are extracted and segments with fluctuation value differences less than the threshold in the boundary segments are screened out. The high threshold of edge detection is set to the 60th percentile value in the regional gradient sample, and the low threshold is set to the 30th percentile value. The non-maximum suppression option is enabled, and the edge connection mode is set to sequential tracking. After performing the double threshold detection operation, the initial edge image is output. Subsequently, the edge extraction tool is called to extract the point sequence of the closed edge band contour. The simplified point option is enabled when calling the contour fitting function, and the point compression coefficient is set to 1% of the total number of original vertices. The deviation angles of the direction vectors between the edge segments are compared in sequence, and the edge segment sequence with an angle difference less than 30 degrees is retained. The edge segment point set is integrated into a continuous segment set to generate a boundary disturbance intensity distribution map.

[0024] The specific steps to generate the intervention priority structure path set are: Based on the boundary disturbance intensity distribution map, the disturbance value within the edge of the dump is extracted and the peak difference between adjacent grids is calculated. The intensity level boundary is set according to the numerical interval. The continuity of each intensity segment is determined and classified by number to generate the disturbance intensity jump structure map. Based on the disturbance intensity jump structure diagram, the equal-length steps in the jump segment area are extracted and the local peak value is judged to be higher than the mean of the adjacent two-side segments. The low-intensity isolated segments are eliminated and the continuous high-value path is retained. The node identifiers are constructed and numbered in sequence according to the position to generate the edge segment path node sequence diagram. Based on the edge path node sequence graph, a graph neural network is used to perform weighted modeling of the disturbance relationship between nodes and extract significant response features between structures. The influence matrix is ​​calculated by multiplying the distance value and the disturbance value difference between each node. The matrix is ​​normalized into an interaction ratio sequence according to the rows and columns. The connection weight table is constructed using the sequence to generate an intervention priority structural path set. Based on the boundary disturbance intensity distribution map, the disturbance value within the edge of the dump was extracted and the peak difference between adjacent grids was calculated. The disturbance values ​​of the cells were traversed, and the interval of the numerical interval was set to 10. The segments were mapped in the interval order, and the identifiers of the grid cells in each segment were assigned. Then, the numbered cells were aggregated into four-way connected regions. During the aggregation process, the regional continuity judgment rule was set as that the four adjacent relationships all met the same number and the area exceeded three times the cell area threshold. After the execution, the regional boundaries were marked with numbers and uniformly recoded according to the connected domain numbers to generate a disturbance intensity jump structure diagram. Based on the disturbance intensity jump structure diagram, the equal-length steps in the jump segment area are extracted and it is judged whether the local peak value exceeds the mean of the adjacent two-side segments. The window length is set to 5 and the step length is set to 1. The jump segment value sequence is traversed and the value of the center position in each window is compared with the mean of the two adjacent positions on both sides. If it is established, the center position is retained as the local peak mark point. The judged mark points are aggregated and the isolated segment judgment rule is set to eliminate the isolated segment when there is no other mark point between the left and right peaks and the valley length in the segment is less than five position units. Then the retained peak positions are numbered in sequence and a coordinate index table is established based on the row and column positions to generate an edge segment path node sequence diagram. Based on the edge path node sequence graph, a graph neural network is used to perform weighted modeling of the disturbance relationship between nodes and extract significant response features between structures. The input graph structure includes a node feature matrix and an adjacency matrix. The node feature items are set as disturbance value, coordinate number, and edge label. The adjacency matrix establishes first-order connections based on adjacent numbers. The number of attention heads is set to 4, corresponding to different initialization weights. The weighted activation function is enabled and set to LeakyReLU, with a negative coefficient set to 0.2. The nodes of each node pair are traversed and weighted adjacent feature aggregation is performed. The aggregated multi-head features are spliced ​​and merged and then mapped to the output layer dimension. The weighted value sorting results are extracted for each connection pair between nodes. The distance value and the disturbance value difference between each node are calculated as the product to form an influence matrix, and the matrix is ​​normalized into an interaction ratio sequence according to the rows and columns. The connection weight table is constructed using the sequence to generate an intervention priority structural path set.

[0025] The execution process of the graph neural network includes the following operations: constructing the connection relationship between nodes based on the edge path node sequence graph, extracting the perturbation attribute value and spatial index information of each node, initializing the perturbation response representation vector of each node, traversing the set of adjacent nodes, calculating the perturbation value difference and spatial distance difference between the target node and the adjacent node, and normalizing the above two differences as the attention weighting factor, performing aggregation on the adjacent node perturbation representation vector according to the weighting factor, updating the perturbation response representation vector of the current node, repeating the above process until all nodes are embedded and updated, and outputting the perturbation correlation distribution value between each node and the adjacent nodes; Graph neural network, according to the formula:

[0026] in: Representation node For adjacent nodes The attention weight coefficient, represents the exponential function, Represents the activation function LeakyRectifiedLinearUnit, Represents the attention weight vector, which is transformed with the perturbation feature matrix during network training Synchronous iterative update, using the node feature splicing perturbation difference and spatial position difference to calculate the attention coefficient, and then converge after minimizing the loss function through back propagation, represents the transpose of the attention weight vector, represents the perturbation feature transformation matrix, Representation node The initial feature vector of Representation node The initial feature vector of represents the vector concatenation operation, Representation node With node The disturbance value difference, Representation node With node The spatial position difference, Representation node The set of adjacent nodes of Indicates traversal of nodes The subscript variables of all adjacent nodes, Representation node With node The disturbance value difference, Representation node With node The spatial position difference.

[0027] Execution process: First, the system constructs a path node sequence diagram of the dumping site edge section based on multimodal input information of remote sensing images, topographic data, vegetation index, soil pH value and hydrological aggregation points, extracts the disturbance value, spatial coordinate index and landform type label of each node, and combines them to form a feature vector , using the perturbation feature transformation matrix Perform linear transformation on node features to enhance expression ability and traverse each node Adjacent nodes , calculate the disturbance value difference Difference in spatial position , and 、 、 、 Perform vector splicing and then use the attention parameter vector Perform the inner product operation, pass it into the LeakyReLU activation function, take the exponent, and normalize it to get the attention coefficient , repeat the process to complete the modeling and expression of the disturbance correlation between nodes in the graph structure.

[0028] The specific steps to generate a stable intervention behavior sequence cluster are: Based on the intervention priority structure path set, all action types in the intervention nodes in the collapse area are numbered according to their sequence positions and an action comparison table is established. The continuous action combinations in each path sequence are read and the number of combination occurrences is counted. The number combinations below the specified frequency threshold are eliminated to generate an operation sequence number frequency table. Based on the frequency table of operation sequence numbers, the position index differences between high-frequency combination numbers are calculated and number groups with differences within a set range are screened. The similarity of the action types within the group is judged and connected and reorganized, and the number sequence is rearranged according to the path structure to generate a stable intervention behavior sequence cluster. Based on the intervention priority structure path set, all action types in the intervention nodes in the collapse area are numbered according to the sequence position and an action comparison table is established. The combined sliding window length is set to 3 and the step size is 1. The continuous action combinations in the path sequence are extracted and an action combination key-value pair table is generated in the window. A pair of combination keys corresponds to a unique index number. The number of times each number appears in the entire path sequence is counted. The frequency threshold is set to 5. The number index is traversed in sequence and the key values ​​with a number of occurrences less than 5 are removed from the combination key-value pair table. The original combination corresponding sequence position and node number index are retained before removal. After removal, the number index mapping relationship is re-established, the number frequency record table is updated, and the operation sequence number frequency table is generated. Based on the operation sequence number frequency table, the position index difference between high-frequency combination numbers is calculated and the number groups with the difference within the set range are screened. The difference range is set to 1 to 4, and the number groups that meet the conditions are grouped into number pairing sets. The character sequence similarity of the corresponding action type in each pair of number combinations is compared. The comparison threshold is set to 0.8, and the literal difference of the action names is compared using character matching. The action group numbers that meet the conditions are recorded, and each number combination is connected and rearranged according to the number sorting rules in the original path structure to generate a complete number index list. The rearranged sequence numbers are written into the new action sequence structure table according to the path index to generate a stable intervention behavior sequence cluster.

[0029] The specific steps for generating the path status abnormal segmentation structure are: Based on the stable intervention behavior sequence cluster, the states in the sequence are numbered according to the execution order and the time window length is set in sequence. The Markov model is used to model the stability of the state transition process and identify the stage changes. The ratio value of the number of state changes and the number of state stays in each window is calculated, and the sequence segments with a change ratio greater than the set value are extracted to generate a state jump segment structure set. Based on the state transition segment structure set, the state position index and frequency statistics in each transition segment are extracted and a quantity matrix is ​​constructed. The dense segments are extracted and the index intervals are marked according to the proximity degree and node concentration degree between consecutive states, thus generating a state clustered high-density segment marking set. Based on the state-aggregated high-density segment marker set, the difference between the state transition value sequence and the change frequency sequence between segments with the same start and end points is calculated. The boundaries of segments with continuous differences greater than the threshold are intercepted and the center of gravity of the segmentation line is adjusted to generate a path state anomaly segmentation structure. Based on the stable intervention behavior sequence cluster, the states in the sequence are numbered in execution order and the time window length is set in sequence. The Markov model is used to model the state transition process and identify stage changes. Each state number is constructed into a finite state set. The window length is set to 5 and the step size is 1. The state changes in each sliding window are traversed and counted. A two-dimensional state transition count matrix is ​​constructed. The diagonal represents the number of state stays and the non-diagonal elements represent the state change counts. The transition probability distribution of each state is obtained by row normalization. The stability difference coefficient of each window is calculated. The stability change ratio threshold is set to 0.25 for judgment. The execution boundary of each window sequence that continuously exceeds the threshold is extracted, and the corresponding start and end state position index intervals are recorded to generate a state jump segment structure set. Based on the state jump segment structure set, the state position index and frequency statistics in each jump segment are extracted and a quantity matrix is ​​constructed. The index position of each state is counted, and the state point with an index difference less than 3 is marked as a candidate aggregation segment. The threshold of the number of valley states in the segment is set to 4. The state occurrence position is sorted in ascending order to form a segment candidate set. The candidate segments are traversed in order and the number of occurrences is accumulated. The average position difference between consecutive states is calculated and the density limit is set to 1.5. The paragraphs below the limit value are filtered out. The position segment index range is output from the remaining paragraph set as the aggregation paragraph boundary. The start and end points of the index interval and the state number are recorded to generate a state aggregation high-density segment marker set. Based on the state-aggregated high-density segment marker set, the difference between the state transition value sequence and the change frequency sequence between segments is calculated according to the same start and end point positions. Two corresponding sequences are constructed for each segment according to the state number. Each state change value in the sequence is counted to form a difference sequence. The sliding window mean of each difference sequence is calculated. The window width is set to 3 and the threshold is set to 0.3. The segments with the difference mean greater than the threshold in the continuous window are set as jump candidate regions. The center of gravity value of the state number in the candidate region is calculated. The relative center of gravity offset method is used with the right boundary of the previous interval and the left boundary of the next interval as reference to redefine the segmentation line position and generate a path state anomaly segmentation structure.

[0030] The execution process of the Markov model includes the following operations: based on the numbered state sequence in the stable intervention behavior sequence cluster, a finite state set is constructed and a state transition record matrix is ​​generated. The state sequence is segmented according to the set time window length. The number of transitions between state pairs and the number of stays in each state in each window are counted. A local state transition probability matrix is ​​constructed based on the number of transitions. The state stability index curve is calculated based on the stay probability. The trend of the stability index in adjacent time windows is compared. The time period where the index change amplitude exceeds the preset threshold is identified, and the state sequence corresponding to the time period is marked as a stage mutation segment. Markov model, according to the formula:

[0031] in: represents the improved state stability index, An index variable representing the state number, Indicates the total number of states involved in the calculation in the current time window, represents the weighted coefficient of the adjustment state stability difference factor, Represents the weighted coefficient for adjusting the state access frequency factor, Represents the weighted coefficient for adjusting the state dwell step length factor, represents the weighting coefficient for adjusting the transfer frequency difference factor, Indicates the The self-transition probability of a number, Indicates the The mean of all transition probabilities of a number, Indicates the The total number of visits to a number in the current sliding window, Indicates the peak value of all state access times in the current sliding window, Indicates the The number stays in the window continuously for the cumulative value of the step length, represents the average of the step lengths within the window, Indicates the The standard deviation of the numbered transfer frequencies, represents the average reference value of the standard deviation of the transfer frequency within the window; Execution process: First, the stable intervention behavior sequence cluster constructed based on multi-source input data is segmented into multiple time windows. For the state number sequence in each time window, a finite state set is extracted and the number of state transitions is counted to construct a local state transition matrix. , extract the diagonal elements Indicates status The self-transition probability of , calculate the mean of the state transition probability , then record the status Number of visits , extract the peak access frequency from each state , and then count the status Continuous dwell step cumulative value , and the average stay value Compare the relative aggregation factors and determine the status The standard deviation of the transfer frequency , and the global reference standard deviation After comparison, the fluctuation level of the transfer rule is expressed by introducing the weight coefficient 、 、 、 The impact intensity of the four indicators is regulated, among which The normalized variance of the state stability difference is set to suppress the impact of high fluctuation segments. The ratio of the current state access times to the global peak times is dynamically calculated through the Sigmoid function to balance the contribution of high-frequency states. According to the difference between the current state continuous stay step length and the average stay value, it is distributed in a linear proportion to strengthen the continuous stay segment detection. The ratio of the current state transition frequency standard deviation to the reference standard deviation is determined using the exponential amplification method, and the results are accumulated in each state to obtain the stability index of the current time window. , and then for the continuous time window The sequence performs difference sliding to identify stability mutation areas. The entire process integrates the state information expression of multimodal input, frequency statistics, sequence behavior structure analysis and probability modeling.

[0032] The specific steps to generate the key action group distribution map under the influence of disturbance are: Based on the path state anomaly segmentation structure, the start and end nodes of each segment are numbered in sequence and a connection index table is established. The difference between the action set in each node and the operation of the previous and next nodes is calculated. The frequency of operation type changes in each adjacent node group is counted to generate the adjacent node frequency offset set. Based on the frequency offset set of adjacent nodes, the frequency offset sequence is extracted by sliding the window range and the peak position is calculated. The operation index corresponding to each extreme value is used to find the associated number and establish an operation coverage matrix. The operation number is aggregated to form a continuous block number set, and a distribution map of key action groups under the influence of disturbance is generated. Based on the path state anomaly segmentation structure, the start and end nodes of each segment are numbered in sequence and a connection index table is established. The node action set difference is calculated, and each node action set is traversed. The set element type is mapped to a number sequence. The execution position of the corresponding number sequence of the previous and next nodes is compared. The number of non-repeated elements is extracted and counted as the difference frequency. A two-dimensional matrix is ​​constructed for adjacent node pairs to record the difference frequency. The change in the number of operation types in each node group is counted. The node window width is set to 3, and the average offset number of each group is set as the benchmark value to calculate the upper and lower floating range. The node pairs with floating values ​​greater than the set threshold of 0.4 are extracted to construct the result set, and the adjacent node frequency offset set is generated. Based on the frequency offset set of adjacent nodes, the frequency offset sequence is extracted by sliding the window range and the peak position is calculated to locate the offset sequence. The sliding window length is set to 5 and the sliding step is set to 1. The five values ​​in each window are compared to perform the median and the upper and lower boundaries are calibrated. It is determined whether the current center value is simultaneously greater than the two previous and next points and greater than the mean of the entire window. Those that meet the conditions are recorded as local peaks. The index number is extracted, and the corresponding action index in the operation sequence is found. It is then mapped back to the action combination identifier to construct a numbered operation table. Based on the numbered operation table, a covering matrix is ​​constructed. Each column of the covering matrix is ​​summed and the numbered block is extracted. The consecutive numbered segments are merged in sequence to construct a numbered set segment, and a distribution map of key action groups under the influence of disturbance is generated.

[0033] See also Figure 2 , a multimodal mine ecological restoration management system, the system includes: Disturbance Perception Module: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, it performs gridded numerical comparison and gradient intensity superposition of multiple ecological factors to generate a boundary disturbance intensity distribution map; Path identification module: Based on the boundary disturbance intensity distribution map, the intervention intensity edges in the dump area are classified and the peak connection paths are extracted. The graph neural network is used to weight the relationship between nodes and extract the structural path set to generate the intervention priority structural path set. Behavior aggregation module: Based on the intervention priority structure path set, it constructs the intervention operation number sequence in the collapse area and screens and rearranges the high-frequency combinations, eliminates the low-frequency segments and reconstructs the continuous segment combination structure to generate a stable intervention behavior sequence cluster; Phase division module: Based on the stable intervention behavior sequence cluster, the operation state is segmented by sliding the time window. The Markov model is used to determine the stability of state transition and separate the mutation window segments to generate the path state abnormality segmentation structure. Action labeling module: Based on the path state anomaly segmentation structure, it extracts the amplitude of the operation frequency change in the abnormal segment and centrally determines the extreme value position. Combined with the adjacent node distribution analysis, it generates the operation area partition information and generates the key action group distribution map under the influence of disturbance.

[0034] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A multimodal mine ecological restoration management method, characterized by: The following steps are involved: Step 1: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, extract unit differences and calculate gradients. After locating boundaries, superimpose structures to generate a boundary disturbance intensity distribution map. Step 2: Based on the boundary disturbance intensity distribution map, the strength of the spoil dump edge segment is grouped, the change path is extracted, the continuous peaks are screened and a node sequence is constructed. The disturbance relationship between nodes is weighted modeled using a graph neural network, the interaction intensity weight is obtained, and an intervention priority structure path set is generated; Step 3: Based on the intervention priority structure path set, the collapse area operation sequences are numbered and the frequencies are counted. The low-frequency operations are eliminated and the high-frequency operations are combined and recombined to generate a stable intervention behavior sequence cluster; Step 4: Based on the stable intervention behavior sequence cluster, the state is transferred and windowed, and the Markov model is used to identify the stage change of the state sequence. The ratio is calculated and the increase segment is suddenly divided. The density cluster and difference are determined to generate the path state abnormality segmentation structure; Step 5: Based on the path state abnormality segmentation structure, traverse the adjacent nodes of the abnormal segment and calculate the operation frequency offset, identify the extreme value position and operation distribution range, and generate a key action group distribution map under the influence of disturbance.

2. The multimodal mine ecological restoration management method according to claim 1 is characterized in that: The specific steps of generating the boundary disturbance intensity distribution map are: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, the surface unit grid is divided and five numerical parameters corresponding to each unit cell are extracted. The corresponding values ​​between adjacent units are subtracted and the change slope is calculated according to the direction. The gradient values ​​are synthesized and superimposed to generate a unit gradient distribution matrix. Based on the unit gradient distribution matrix, high-value edges in continuous areas are extracted and segments with fluctuation value differences less than a threshold in the boundary segments are screened out. The edges of the remaining areas are traced and the contours are extracted. The consistency of the structural connection and orientation between multiple edge segments is checked and the sidebands are merged to generate a boundary disturbance intensity distribution map.

3. The multimodal mine ecological restoration management method according to claim 1 is characterized in that: The specific steps for generating the intervention priority structure path set are: Based on the boundary disturbance intensity distribution map, the disturbance value within the edge of the spoil dump is extracted and the peak difference value of adjacent grids is calculated. The intensity level boundary is set according to the numerical interval. The continuity of each intensity segment is determined and classified by labeling to generate a disturbance intensity jump structure map. Based on the disturbance intensity jump structure diagram, equal-length steps in the jump segment area are extracted and it is determined whether the local peak value exceeds the mean of the adjacent two-side segments. The low-intensity isolated segments are eliminated and the continuous high-value path is retained. The node identifiers are constructed and numbered in sequence according to the position to generate an edge segment path node sequence diagram; Based on the edge path node sequence diagram, a graph neural network is used to perform weighted modeling of the disturbance relationship between nodes and extract significant response features between structures. The distance value and the disturbance value difference between each node are calculated as the product to form an influence matrix, and the matrix is ​​normalized into an interaction ratio sequence according to the rows and columns. The sequence is used to construct a connection weight table to generate an intervention priority structural path set.

4. The multimodal mine ecological restoration management method according to claim 3 is characterized in that: The graph neural network is based on the formula: in: Representation node For adjacent nodes The attention weight coefficient, represents the exponential function, Represents the activation function LeakyRectifiedLinearUnit, represents the attention weight vector, represents the transpose of the attention weight vector, represents the perturbation feature transformation matrix, Representation node The initial feature vector of Representation node The initial feature vector of represents the vector concatenation operation, Representation node With node The disturbance value difference, Representation node With node The spatial position difference, Representation node The set of adjacent nodes of Indicates traversal of nodes The subscript variables of all adjacent nodes, Representation node With node The disturbance value difference, Representation node With node The spatial position difference.

5. The multimodal mine ecological restoration management method according to claim 4 is characterized in that: The execution process of the graph neural network includes the following operations: constructing the connection relationship between nodes based on the edge path node sequence diagram, extracting the disturbance attribute value and spatial index information of each node, initializing the disturbance response representation vector of each node, traversing the set of adjacent nodes, calculating the disturbance value difference and spatial distance difference between the target node and the adjacent node, and normalizing the above two differences as the attention weighting factor, performing aggregation on the adjacent node disturbance representation vector according to the weighting factor, updating the disturbance response representation vector of the current node, repeating the above process until all nodes complete the embedding update, and outputting the disturbance correlation distribution value between each node and the adjacent node.

6. The multimodal mine ecological restoration management method according to claim 1, characterized in that: The specific steps for generating the stable intervention behavior sequence cluster are: Based on the intervention priority structure path set, all action types in the intervention nodes in the collapse area are numbered according to their sequence positions and an action comparison table is established. The continuous action combinations in each path sequence are read and the number of combination occurrences is counted. The numbered combinations below the specified frequency threshold are eliminated to generate an operation sequence number frequency table. Based on the operation sequence number frequency table, the position index difference between high-frequency combination numbers is calculated and the number groups with the difference within the set range are screened. The connection and reorganization are performed by judging the similarity of the action types within the group, and the number sequence is rearranged according to the path structure to generate a stable intervention behavior sequence cluster.

7. The multimodal mine ecological restoration management method according to claim 1, characterized in that: The specific steps of generating the path status abnormal segmentation structure are: Based on the stable intervention behavior sequence cluster, the states in the sequence are numbered in execution order and the time window length is set in sequence. The Markov model is used to model the stability of the state transition process and identify stage changes. The ratio value of the number of state changes and the number of state stays in each window is calculated, and the sequence segments with a change ratio greater than the set value are extracted to generate a state jump segment structure set. Based on the state transition segment structure set, the state position index and frequency statistics in each transition segment are extracted and a quantity matrix is ​​constructed. The dense segments are extracted and the index intervals are marked according to the proximity degree and node concentration degree between consecutive states to generate a state clustered high-density segment mark set. Based on the state-aggregated high-density segment marker set, the difference between the inter-segment state transition value sequence and the change frequency sequence is calculated according to the same start and end point positions. The boundaries of the segments with continuous differences greater than the threshold are intercepted and the center of gravity of the segmentation line is adjusted to generate a path state abnormality segmentation structure.

8. The multimodal mine ecological restoration management method according to claim 7, characterized in that: The Markov model is based on the formula: in: represents the improved state stability index, An index variable representing the state number, Indicates the total number of states involved in the calculation in the current time window, represents the weighted coefficient of the adjustment state stability difference factor, Represents the weighted coefficient for adjusting the state access frequency factor, Represents the weighted coefficient for adjusting the state dwell step length factor, represents the weighting coefficient for adjusting the transfer frequency difference factor, Indicates the The self-transition probability of a number, Indicates the The mean of all transition probabilities of a number, Indicates the The total number of visits to a number in the current sliding window, Indicates the peak value of all state access times in the current sliding window, Indicates the The number stays in the window continuously for the cumulative value of the step length, represents the average of the step lengths within the window, Indicates the The standard deviation of the numbered transfer frequencies, Indicates the average reference value of the standard deviation of the transition frequencies within the window.

9. The multimodal mine ecological restoration management method according to claim 1, characterized in that: The specific steps of generating the key action group distribution map under the influence of the disturbance are: Based on the path state anomaly segmentation structure, the start and end nodes of each segment are numbered in sequence and a connection index table is established. The difference between the action set in each node and the operation of the previous and next nodes is calculated. The frequency of operation type changes in each adjacent node group is counted to generate an adjacent node frequency offset set. Based on the frequency offset set of adjacent nodes, the frequency offset sequence is extracted by sliding the window range and the peak position is calculated. The operation index corresponding to each extreme value is used to find the associated number and establish an operation coverage matrix. The operation number is aggregated to form a continuous block number set, and a distribution map of the key action group under the influence of disturbance is generated.

10. The multimodal mine ecological restoration management system is characterized by: According to any one of claims 1 to 9, the multimodal mine ecological restoration management method comprises: Disturbance Perception Module: Based on vegetation coverage, slope aggregation, number of hydrological aggregation points, surface soil pH difference, and rainy season sediment flow trajectory, it performs gridded numerical comparison and gradient intensity superposition of multiple ecological factors to generate a boundary disturbance intensity distribution map; Path identification module: Based on the boundary disturbance intensity distribution map, the intervention intensity edge segments in the spoil dump area are classified and the peak connection paths are extracted. The graph neural network is used to weight the relationship between nodes and extract the structural path set to generate the intervention priority structural path set; Behavior aggregation module: Based on the intervention priority structure path set, it constructs the intervention operation number sequence in the collapse area and screens and rearranges the high-frequency combinations, removes the low-frequency segments and reconstructs the continuous segment combination structure to generate a stable intervention behavior sequence cluster; Phase division module: Based on the stable intervention behavior sequence cluster, the operation state is segmented by sliding the time window, and the Markov model is used to determine the stability of the state transition and separate the mutation window segments to generate the path state abnormality segmentation structure; Action labeling module: Based on the path state anomaly segmentation structure, the operation frequency change amplitude in the abnormal segment is extracted and the extreme value position is determined centrally. The operation area partition information is generated by combining the adjacent node distribution analysis, and the key action group distribution map under the influence of disturbance is generated.

Images