Construction and dynamic early warning method and system for micro-seismic risk field

Abstract

The invention discloses a micro-seismic risk field construction and dynamic early warning method and system, and belongs to the technical field of risk field construction and dynamic early warning. The method specifically comprises the steps of collecting micro-seismic data of a target monitoring area, writing the micro-seismic data into a storage area, constructing a micro-seismic risk field based on the micro-seismic data of the target monitoring area, and obtaining a micro-seismic risk field sequence signal set; based on the microseismic risk field sequence signal set, dividing an early warning section, constructing a time-space multi-relation tensor diagram of the target monitoring area, and constructing a time-space propagation kernel function to obtain a time sequence risk representation set; based on the time sequence risk representation set, constructing a distorted hyperbolic space-time risk manifold space of the target monitoring area and performing geometric embedding to obtain a hyperbolic evolution representation set; based on the hyperbolic evolution representation set, hyperbolic risk distances of all the early warning sections are calculated, the hyperbolic risk distances are graded, and dynamic early warning is carried out on risks; according to the invention, the accuracy of dynamic early warning of the micro-seismic risk field is improved, and the stability of dynamic early warning of the micro-seismic risk field is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of risk field construction and dynamic early warning, specifically relating to a method and system for constructing and dynamically warning of microseismic risk fields. Background Technology

[0002] In the process of risk monitoring and early warning in the target monitoring area, facing an environment with frequent disturbances and unstable spatiotemporal evolution of microseismic activity, micro-ruptures or stress releases may trigger microseismic events with spatial location and energy characteristics. Relying solely on the experience and judgment of on-site monitoring personnel, scattered point observations, or post-event analysis and summaries makes it difficult to grasp the dynamic changes in microseismic risk within the target monitoring area in a timely manner, easily leading to biased early warning judgments and delayed early warning feedback. Constructing a microseismic risk field based on microseismic data from the target monitoring area, dividing the target monitoring area into multiple early warning segments, and generating microseismic risk field sequences for each early warning segment can provide a data foundation for regional risk early warning. Therefore, how to conduct objective and stable dynamic early warning for different spatial segments after constructing the microseismic risk field is a crucial problem that urgently needs to be solved.

[0003] In existing technologies, methods for constructing and dynamically warning of microseismic risk fields typically utilize data such as the time, spatial coordinates, and energy of microseismic events to build a sequence of microseismic risk fields. These sequences are then predicted using spatiotemporal prediction models, and indicators such as peak energy, average energy, and event frequency are extracted from the prediction results to classify and warn of risks in different warning zones. However, existing methods for constructing and dynamically warning of microseismic risk fields suffer from the following problems: they adjust warning levels based solely on minor fluctuations in risk indicators over a short period, leading to premature downgrading or cancellation of warnings before the actual risk has diminished. This results in poor accuracy and stability of the dynamic warning results for microseismic risk fields. Summary of the Invention:

[0004] The purpose of this invention is to provide a method and system for constructing and dynamically warning of microseismic risk fields.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for constructing and dynamically warning of microseismic risk fields, comprising: Collect microseismic data of the target monitoring area and write it into the storage area. Based on the microseismic data of the target monitoring area, construct a microseismic risk field and obtain a set of microseismic risk field sequence signals. Based on the microseismic risk field sequence signal set, the early warning zone is divided, and a spatiotemporal multi-relation tensor map of the target monitoring area is constructed to obtain a time series risk representation set. Based on the temporal risk representation set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and geometric embedding is performed in the distorted hyperbolic spatiotemporal risk manifold space to obtain a hyperbolic evolution representation set; Based on the hyperbolic evolution representation set, the hyperbolic risk distance of each warning segment is calculated, and the hyperbolic risk distance is classified according to the multi-level dynamic warning judgment rule to provide dynamic warning of the risk.

[0006] Specifically, based on microseismic data from the target monitoring area, a microseismic risk field is constructed, resulting in a set of microseismic risk field sequence signals, including: Based on the occurrence time of each microseismic event and a preset window capacity threshold, the microseismic data of the target monitoring area are sorted and divided into sliding windows to obtain a sliding window sequence. Based on the sliding window sequence and the preset grid resolution, the regular grid corresponding to each sliding window is determined, and the maximum value is filtered to obtain the microseismic risk field image frame corresponding to each sliding window. The image frames of the microseismic risk field are subjected to structural similarity interpolation and combined to obtain a set of microseismic risk field sequence signals.

[0007] Specifically, based on the sliding window sequence and the preset grid resolution, the regular grid corresponding to each sliding window is determined, and the maximum value is filtered to obtain the microseismic risk field image frame corresponding to each sliding window, including: Based on the sliding window sequence and the preset grid resolution, the coordinate range of each microseismic event within each sliding window is determined, and the grid is divided to obtain a regular grid. Based on the spatial coordinates and event energy of each microseismic event, the energy diffusion value of each microseismic event at each grid node in the regular grid is calculated. The maximum value of energy diffusion for each grid node is selected to obtain the microseismic risk field image frame corresponding to each sliding window.

[0008] Specifically, based on the microseismic risk field sequence signal set, warning zones are divided, and a spatiotemporal multi-relationship tensor map of the target monitoring area is constructed to obtain a time-series risk representation set, including: Based on the microseismic risk field sequence signal set, the warning sections are divided, and the energy diffusion values ​​of the corresponding grid nodes in each warning section are time-series aggregated to obtain the section risk characteristic sequence of each warning section at each time. Based on the segment risk feature sequence and the preset multi-time scale time window, the segment risk features are weighted and accumulated and extreme value statistics are performed to obtain the multi-time scale initial node features of each warning segment at each time. Based on the initial node features at multiple time scales, each warning segment is mapped to a segment time node at each time, and a continuous-time neural evolution model is constructed to obtain a spatiotemporal multi-relationship tensor map of the target monitoring area. Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, a graph Laplacian operator is constructed, and a spatiotemporal propagation kernel function is built to obtain the spatiotemporal propagation weight set; Based on the spatiotemporal propagation weight set, time decay weighted aggregation is performed on the time nodes of each segment to generate the spatiotemporal memory risk representation of each warning segment at each time, and then combined to obtain the time-series risk representation set of each warning segment.

[0009] Specifically, based on the initial node features at multiple time scales, each warning segment is mapped to a segment time node at each moment, and a continuous-time neural evolution model is constructed to obtain a spatiotemporal multi-relationship tensor map of the target monitoring area, including: Based on the initial node features at multiple time scales, each warning segment is mapped to segment time nodes at each time, and a candidate adjacency matrix is ​​constructed to obtain the set of segment time nodes and the set of candidate adjacency matrices. Based on the candidate adjacency matrix set, a continuous-time neural evolution model is constructed to obtain the continuous-time state equation of the nodes at each time segment. Based on the continuous-time state equation, the time series of segment time nodes composed of initial node features of multiple time scales is fitted to obtain the continuous-time influence coefficient matrix. Based on the continuous time influence coefficient matrix, the time segment node pairs with continuous time influence coefficients greater than a preset threshold are selected, and a set of causal directed edges is constructed to obtain the causal directed graph of time segment nodes. Based on the causal directed graph of nodes in the segment, the candidate adjacency matrix set is updated by the continuous time influence coefficient matrix to obtain the spatiotemporal multi-relationship tensor graph of the target monitoring area.

[0010] Specifically, based on the candidate adjacency matrix set, a continuous-time neural evolution model is constructed to obtain the continuous-time state equations of nodes at each time segment, including: Based on the candidate adjacency matrix set, the historical memory kernel of the node at each time segment is calculated, and a graph structure operator is constructed. The historical memory kernel and the initial node features at multiple time scales are combined to obtain the enhanced node state, the graph structure operator set, and the memory kernel weight set. Based on the enhanced node state, graph structure operator set, and memory kernel weight set, a continuous-time state evolution equation is constructed to obtain the continuous-time neural evolution model to be identified. Based on the continuous-time neural evolution model to be identified, the continuous-time state trajectory of each time node in each segment is calculated to obtain the predicted state evolution trajectory. The initial node features at multiple time scales are processed to obtain the time series of nodes at different time segments; By minimizing the error between the predicted state evolution trajectory and the time series of time nodes in each segment, the coefficients and memory kernel parameters in the continuous-time neural evolution model are determined, and the continuous-time state equations of each time node in each segment are obtained.

[0011] Specifically, based on the spatiotemporal multi-relation tensor graph of the target monitoring area, a graph Laplacian operator is constructed, and a spatiotemporal propagation kernel function is built to obtain the spatiotemporal propagation weight set, including: Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, the risk intensity of each segment time node is calculated, the risk weighted graph Laplacian operator of each segment time node is constructed, and spectral decomposition is performed to obtain the feature spectral values ​​and feature vector set. Each characteristic spectrum value and its corresponding risk intensity are multiplied by a preset frequency adjustment coefficient and risk adjustment coefficient, respectively, to generate a first intermediate quantity and a second intermediate quantity. The first intermediate quantity is divided by the second intermediate quantity to obtain a set of ratio-type independent variables. Based on the set of ratio-type independent variables, each ratio-type independent variable is multiplied by a preset spatial decay coefficient, and the negative value of the exponential function is taken to generate an exponential decay subset. Each ratio-type independent variable is multiplied by a preset focusing coefficient and subjected to a hyperbolic tangent function transformation to obtain a focusing subset. Based on the ratio-type independent variable set, the exponential decay sub-item set, and the focusing sub-item set, the symbol modulation coefficient is determined, the symbol modulation sub-item set is generated, and the exponential decay sub-item set and the focusing sub-item set are linearly combined according to the preset modulation weight and the symbol modulation sub-item set to obtain the spatiotemporal propagation kernel function set. Based on the set of spatiotemporal propagation kernel functions and the set of feature vectors, the spatiotemporal propagation kernel functions are mapped from the spectral domain to the spatiotemporal multi-relation tensor graph of the target monitoring area, and the spatiotemporal propagation weights are calculated to obtain the set of spatiotemporal propagation weights.

[0012] Specifically, based on the temporal risk representation set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and geometrically embedded in the distorted hyperbolic spatiotemporal risk manifold space to obtain a hyperbolic evolutionary representation set, including: Based on the time-series risk representation set, the time-series risk representations corresponding to each warning segment at each time are reorganized to obtain the set of time nodes of the reorganized segment and the corresponding time-series risk vector set. Based on the set of time nodes of the recombined segment and the corresponding temporal risk vector set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and tangent space projection is performed to obtain the tangent space risk vector set; Based on the tangent space risk vector set, the tangent space risk vector is mapped to fiber sphere coordinate points on the twisted hyperbolic spatiotemporal risk manifold space through exponential mapping, thus obtaining the initial fiber sphere spatiotemporal embedding coordinate set; Based on the initial spatiotemporal embedding coordinate set of the fiber spheres, the fiber sphere embedding coordinate sequence of each warning zone is determined, and the fiber sphere embedding coordinate sequence is interpolated and smoothed to obtain the spatiotemporal evolution trajectory set of the fiber spheres in each warning zone. Based on the spatiotemporal evolution trajectory set of the fiber sphere, the coordinates of the distorted hyperbolic spatiotemporal risk manifold corresponding to each warning segment at each time are extracted, and the hyperbolic evolution representation set of each warning segment at each time is obtained.

[0013] Specifically, based on the set of time nodes in the reconstructed segment and the corresponding temporal risk vector set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and tangent spatial projection is performed to obtain the tangent spatial risk vector set, including: Based on the set of time nodes in the recombined segment and the corresponding temporal risk vector set, the temporal base space is determined, and fibers are constructed to obtain the spatiotemporal risk fiber bundle of the target monitoring area. Based on the spatiotemporal risk fiber bundle of the target monitoring area, the corresponding fiber ball risk state space is constructed at each time node of the fiber bundle, and the fiber ball risk state space set is obtained. Based on the fiber ball risk state space set and the preset distortion function, the metric between the time base space and the fiber ball risk state space is distorted to obtain the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. Based on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, a tangent space is constructed at a preset base point. The temporal risk vector set is projected onto the tangent space through a linear transformation to obtain the tangent space risk vector set.

[0014] A microseismic risk field construction and dynamic early warning system, comprising: The data acquisition module is used to collect microseismic data of the target monitoring area and write it into the storage area. Based on the microseismic data of the target monitoring area, a microseismic risk field is constructed to obtain a set of microseismic risk field sequence signals. The early warning segmentation module, based on the microseismic risk field sequence signal set, divides the early warning sections, constructs a spatiotemporal multi-relation tensor map of the target monitoring area, and obtains a time-series risk representation set; The hyperbolic embedding module constructs a spatiotemporal risk fiber bundle of the target monitoring area based on the temporal risk representation set, and performs geometric embedding in the distorted hyperbolic spatiotemporal risk manifold space to obtain the hyperbolic evolution representation set; The early warning generation module calculates the hyperbolic risk distance for each early warning segment based on the hyperbolic evolution representation set, and classifies the hyperbolic risk distance according to the multi-level dynamic early warning judgment rules to provide dynamic early warning for the risk.

[0015] Compared with existing technologies, the beneficial effects of this invention include: constructing a time-evolving microseismic risk field based on microseismic data of the target monitoring area, and dividing the target monitoring area into warning segments according to space to obtain a microseismic risk field sequence characterizing the risk distribution of each segment; constructing a spatiotemporal multi-relation tensor map of the target monitoring area based on the risk characteristics of each segment, combining it with a continuous-time neural evolution model and performing causal directed updates to obtain a temporal risk representation; constructing a spatiotemporal risk fiber bundle of the target monitoring area based on the temporal risk representation, performing geometric embedding and hyperbolic risk distance measurement in a distorted hyperbolic spatiotemporal risk manifold space, and combining it with multi-level dynamic warning judgment rules to dynamically warn of the risk of each warning segment; solving the problem in existing technologies that adjust the warning level only based on small fluctuations in risk indicators within a short period of time, resulting in the warning level being prematurely reduced or the warning being lifted before the risk is substantially alleviated, and improving the accuracy and stability of the dynamic warning results of the microseismic risk field. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for constructing and dynamically warning of a microseismic risk field, provided by this invention. Figure 2 This invention provides a schematic diagram of the target monitoring area grid and high-risk microseismic zones. Figure 3 This is a schematic diagram of hyperbolic risk distance and early warning level provided by the present invention; Figure 4 The present invention provides a structural diagram of a microseismic risk field construction and dynamic early warning system. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0018] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0019] Example 1: Please see Figures 1-3 The present invention provides an embodiment of a method for constructing and dynamically warning of a microseismic risk field, comprising the following specific steps: Step S1: Collect microseismic data of the target monitoring area and write it into the storage area. Based on the microseismic data of the target monitoring area, construct the microseismic risk field and obtain the microseismic risk field sequence signal set.

[0020] The specific steps of step S1 are as follows: Step S101: Collect microseismic data of the target monitoring area and write it into the storage area. Based on the occurrence time of each microseismic event and the preset window capacity threshold, sort the microseismic data of the target monitoring area and divide it into sliding windows to obtain a sliding window sequence.

[0021] In this embodiment, microseismic data of the target monitoring area is collected and written to the storage area. The collected microseismic data of the target monitoring area includes microseismic events and the occurrence time, spatial coordinates, and event energy of each microseismic event. The collected microseismic data is encapsulated according to a preset data field format, and the generation time index and segment index of each microseismic event are recorded. The encapsulated microseismic event records are written to the original data table in the storage area. An index structure is established for the original data table according to the time index, and the key fields of the microseismic event records are written to the cache area to support subsequent sliding window reading. It should be noted that after collecting the microseismic data of the target monitoring area, the data needs to be preprocessed. The preprocessing specifically includes: outlier handling, data standardization, missing value imputation, time alignment, and feature smoothing. The data field format is set by those skilled in the art according to the actual situation. The target monitoring area can be coal mine roadways, urban roads, bridge subgrades, public buildings and venues of all sizes, and open-air storage yards, etc.

[0022] The chronological order of the occurrence of each microseismic event is used as the time sequence. Based on the time sequence, each microseismic event in the microseismic data of the target monitoring area is sorted to obtain a time-sorted sequence of microseismic events.

[0023] Starting from the first microseismic event in the microseismic event sequence, consecutive microseismic events not exceeding a preset window capacity threshold are sequentially selected to obtain the first sliding window. The window capacity threshold is set by those skilled in the art according to the actual situation and is used to limit the upper limit of the number of microseismic events contained in each sliding window.

[0024] Based on the time sequence, the starting microseismic event of the first sliding window is moved back one microseismic event position, and the operation of filtering continuous microseismic events and forming sliding windows is repeated until the microseismic event sequence is completely traversed, resulting in a sliding window sequence arranged in time sequence.

[0025] Step S102: Based on the sliding window sequence and the preset grid resolution, determine the regular grid corresponding to each sliding window, and perform maximum value filtering to obtain the microseismic risk field image frame corresponding to each sliding window.

[0026] The specific steps of step S102 are as follows: Step S1021: Based on the sliding window sequence and the preset grid resolution, determine the coordinate range of each microseismic event within each sliding window, and perform grid division to obtain a regular grid.

[0027] In this embodiment, each sliding window in the sliding window sequence is traversed to obtain the spatial coordinates of all microseismic events within each sliding window. The minimum and maximum coordinate values ​​of each microseismic event within each sliding window on the horizontal and vertical axes of the preset regional plane coordinate axis are calculated respectively to obtain the horizontal coordinate range and vertical coordinate range of each sliding window. The regional plane coordinate axis is set by those skilled in the art according to the actual situation.

[0028] Based on the preset grid resolution, grid horizontal coordinate sequences and grid vertical coordinate sequences are generated along the horizontal and vertical axes, respectively, within the horizontal and vertical coordinate ranges, to construct a two-dimensional regular grid, which serves as the regular grid corresponding to each sliding window. The grid resolution is set by those skilled in the art according to the actual situation.

[0029] Step S1022: Based on the spatial coordinates and event energy corresponding to each microseismic event, calculate the energy diffusion value of each microseismic event at each grid node in the regular grid.

[0030] In this embodiment, all microseismic events within each sliding window are traversed, and each grid node in the regular grid corresponding to each microseismic event is traversed. The difference between the horizontal coordinate of each grid node and the horizontal coordinate of the corresponding microseismic event is calculated to obtain the horizontal distance difference. The difference between the vertical coordinate of each grid node and the vertical coordinate of the corresponding microseismic event is calculated to obtain the vertical distance difference.

[0031] Based on the lateral and longitudinal distance differences, the squared distance between each grid node and the corresponding microseismic event is calculated. The squared distance and the event energy of the corresponding microseismic event are input into a preset Gaussian function to obtain the energy diffusion value of the corresponding microseismic event at each grid node. The Gaussian function is set by those skilled in the art according to the actual situation.

[0032] Step S1023: Filter out the maximum value of energy diffusion value of each grid node to obtain the microseismic risk field image frame corresponding to each sliding window.

[0033] In this embodiment, all energy diffusion values ​​of each grid node in the regular grid corresponding to each sliding window are obtained, the maximum value of the energy diffusion value of each grid node is selected, and the maximum value of the energy diffusion value corresponding to each grid node is arranged according to the grid position to obtain the microseismic risk field image frame corresponding to each sliding window.

[0034] Step S103: Perform structural similarity interpolation on the microseismic risk field image frames and combine them according to the time order to obtain the microseismic risk field sequence signal set.

[0035] In this embodiment, the microseismic risk field image frames corresponding to each sliding window arranged in chronological order are used as the initial endpoint frame sequence. An image frame sequence for storing the interpolation results is preset, the image frame sequence is initialized to empty, and the initial value of the interpolation trigger flag of the image frame sequence is set to no. The image frame sequence for storing the interpolation results is set by those skilled in the art according to the actual situation.

[0036] Based on the time order, each adjacent frame pair in the initial endpoint frame sequence is traversed sequentially to determine the current frame and the next frame of each adjacent frame pair. The structural similarity between the current frame and the next frame is calculated according to a preset structural similarity index, which is set by those skilled in the art based on the actual situation.

[0037] When the structural similarity of adjacent frame pairs is less than a preset structural similarity threshold, a weighted average of the current frame and the next frame is generated by pixel-level linear interpolation. The current frame and the intermediate transition frame are then added sequentially to the image frame sequence used to store the interpolation results. The interpolation trigger flag is set to "yes". The structural similarity threshold is set by those skilled in the art according to the actual situation.

[0038] When the structural similarity is greater than or equal to the structural similarity threshold, only the current frame is added to the image frame sequence used to store the interpolation results.

[0039] After traversing all adjacent frame pairs, the last frame in the initial endpoint frame sequence is added to the image frame sequence used to store the interpolation results, resulting in a microseismic risk field image frame sequence after one round of interpolation processing.

[0040] When the final interpolation trigger is marked as yes, the microseismic risk field image frame sequence after one round of interpolation is used as the new initial endpoint frame sequence. The above adjacent frame traversal and interpolation processing operations are repeated until the structural similarity of all adjacent frame pairs is greater than or equal to the structural similarity threshold.

[0041] When the final frame interpolation trigger flag is no, the frame interpolation process is terminated. After the frame interpolation process is terminated, the final microseismic risk field image frame sequence is combined in chronological order to obtain the microseismic risk field sequence signal set.

[0042] Step S2: Based on the microseismic risk field sequence signal set, divide the early warning section, construct the spatiotemporal multi-relation tensor map of the target monitoring area, and obtain the temporal risk representation set.

[0043] The specific steps of step S2 are as follows: Step S201: Based on the microseismic risk field sequence signal set, the warning sections are divided, and the energy diffusion values ​​of the corresponding grid nodes in each warning section are time-series aggregated to obtain the section risk characteristic sequence of each warning section at each time.

[0044] In this embodiment, based on the microseismic risk field image frames and regular grids corresponding to each moment in the microseismic risk field sequence signal set, the spatial range of the target monitoring area within the regional plane is determined by the regional plane coordinate axis.

[0045] The spatial layout information of the target monitoring area is collected. Based on the spatial layout information of the target monitoring area, the main axis direction of the target monitoring area is determined. A coordinate axis parallel to the main axis direction is selected in the regional plane coordinate axis as the main extension direction of the area. According to the coordinate range in the main extension direction of the area, the target monitoring area is divided into n1 continuous spatial segments along this direction. Each spatial segment is determined as a warning segment, and the warning segment division result is obtained.

[0046] Based on the results of the warning zone division and the microseismic risk field image frames at each time, the regular grid in the microseismic risk field image frames at each time is traversed. According to the coordinate position of the grid node in each regular grid on the regional plane coordinate axis, each grid node is divided into the corresponding warning zone, thus obtaining the set of grid nodes corresponding to each warning zone at each time. The energy diffusion value corresponding to each grid node at each time is read from the microseismic risk field image frames.

[0047] Based on the set of grid nodes corresponding to each warning section at each time, the energy diffusion value of the same warning section at each time is statistically processed. The sum, average and maximum values ​​of the energy diffusion values ​​of all grid nodes in each warning section at each time are calculated. The sum, average and maximum values ​​are combined to obtain the section risk feature vector of each warning section at that time.

[0048] The risk feature vectors of each time period are arranged in chronological order to form the risk feature sequence of each warning zone. The risk feature sequences of all warning zones are then summarized to obtain the risk feature sequence of each warning zone at each time period.

[0049] exist Figure 2 In the diagram, the black rectangular border represents the projection range of the target monitoring area onto the regional plane. The thick horizontal line below the black rectangular border and the thick vertical line on the left represent coordinate references without specific scales, i.e., the regional plane coordinate axes.

[0050] The regular fine grid of colored lines inside the black rectangular border represents the grid discretization of the target monitoring area according to the preset grid resolution. Each small rectangle represents a grid node or grid unit in the regular grid. The colored grid vertical lines also represent p longitudinal strips divided along the main extension direction of the area. That is, according to the main axis direction and spatial location of the target monitoring area, the target monitoring area is divided into p warning segments, and each segment consists of q grid units.

[0051] The yellow crosses distributed on the grid represent the spatial locations of the monitored microseismic events projected onto the regional plane. The denser the crosses, the more active the microseismic activity in the region. The three sets of colored dashed ellipses on the left, middle, and right represent high-risk clusters formed by the diffusion and superposition of microseismic energy in different warning sections. Different colors represent different risk levels in different warning sections. The pink dashed ellipse on the left represents the lowest risk, the yellow dashed ellipse on the right represents the moderate risk, and the gray dashed ellipse in the middle represents the highest risk.

[0052] Step S202: Based on the segment risk feature sequence and the preset multi-time scale time window, perform weighted accumulation and extreme value statistics on the segment risk features to obtain the multi-time scale initial node features of each warning segment at each time.

[0053] In this embodiment, based on the segment risk feature sequence of each warning segment at each time and the preset multi-time scale time window, the segment risk feature sequence of each warning segment at each time is traversed. For each time of each warning segment, the segment risk feature vector set corresponding to the warning segment is extracted in the short time window, medium time window and long time window respectively. The multi-time scale time window includes a short time window, a medium time window and a long time window, which are set by those skilled in the art according to the actual situation.

[0054] Based on a preset time decay weighting coefficient, risk feature vectors of segments within a multi-timescale time window that are at a time distance greater than or equal to a preset time threshold are assigned a preset large weight, while risk feature vectors of segments at a time distance less than the time threshold are assigned a preset small weight. The risk feature vectors of each segment within the multi-timescale time window are then weighted and accumulated to generate a weighted accumulated risk feature vector corresponding to each time scale, thus obtaining a set of risk feature vectors of each segment within the multi-timescale time window. The time decay weighting coefficient, time threshold, large weight, and small weight are set by those skilled in the art according to the actual situation.

[0055] Based on the risk feature vector set of each segment within a multi-timescale time window, extreme value statistics are performed on the segment risk feature vector of each warning segment at each time. The maximum value of the risk indicator value corresponding to the segment risk feature vector of each warning segment within each multi-timescale time window is extracted to obtain the extreme value risk feature vector at each time scale. The risk indicator value includes the sum of energy diffusion values, the average value of energy diffusion values, and the maximum value of energy diffusion values.

[0056] The identifiers of each warning segment are extracted. Based on a preset splicing order, the segment risk feature vector, the weighted cumulative risk feature vector corresponding to each time scale, and the extreme value risk feature vector corresponding to each time scale are spliced ​​together to form the multi-time scale initial node feature vector corresponding to each warning segment at each time. According to the time order and the identifiers of each warning segment, the multi-time scale initial node feature vectors are sorted to obtain the multi-time scale initial node features of each warning segment at each time. The splicing order is set by those skilled in the art according to the actual situation.

[0057] Step S203: Based on the initial node features of multiple time scales, each warning segment is mapped to segment time nodes at each time, and a continuous time neural evolution model is constructed to obtain a spatiotemporal multi-relationship tensor map of the target monitoring area.

[0058] The specific steps of step S203 are as follows: Step S2031: Based on the initial node features of multiple time scales, map each warning segment to segment time nodes at each time, and construct a candidate adjacency matrix to obtain the segment time node set and the candidate adjacency matrix set.

[0059] In this embodiment, the time index of each warning segment is extracted. Based on the initial node features of multiple time scales, the identifier and time index of each warning segment, a set of initial node features of multiple time scales corresponding to each warning segment at each time is associated with the corresponding warning segment identifier and time index. The association result is defined as a segment time node. A unique number is assigned to each segment time node to obtain a set of segment time nodes, and the correspondence between segment time nodes, warning segments and time indexes is established.

[0060] Based on the set of time nodes in the segment, a candidate adjacency matrix is ​​constructed to represent the connectivity between time nodes in the segment. According to the arrangement order of the warning segments in the main extension direction of the region, adjacent warning segments are determined. The time nodes of the segments belonging to adjacent warning segments at the same time are combined to form pairs of time nodes in the segment. Spatial candidate edges are established between pairs of time nodes in the segment to obtain a set of spatial candidate edges.

[0061] Based on the set of spatial candidate edges, a spatial candidate adjacency matrix is ​​constructed. In the spatial candidate adjacency matrix, the time nodes of each segment are used as the row index and column index of the matrix. For the pair of time nodes belonging to the set of spatial candidate edges, the corresponding matrix elements are set to non-zero values. The spatial candidate adjacency matrix is ​​used to represent the spatial adjacency relationship of time nodes in segments.

[0062] Based on the spatial candidate edge set, temporal candidate edges are established between pairs of time nodes in a segment, generating a temporal candidate edge set and constructing a temporal candidate adjacency matrix. In the temporal candidate adjacency matrix, each time node in a segment is used as the row index and column index of the matrix. For pairs of time nodes in a segment that belong to the temporal candidate edge set, their corresponding matrix elements are set to non-zero values. The temporal candidate adjacency matrix is ​​used to represent the temporal order of time nodes in a segment.

[0063] Based on the initial node features at multiple time scales, the similarity of the initial node feature vectors of each segment time node pair is calculated. Segment time node pairs with similarity greater than a preset similarity threshold are regarded as node pairs with similar risk states, and a risk similarity candidate adjacency matrix is ​​constructed. In the risk similarity candidate adjacency matrix, each node pair with similar risk states is used as the row index and column index of the matrix. For segment time node pairs belonging to the time candidate edge set, their corresponding matrix elements are set to values ​​related to similarity. The risk similarity candidate adjacency matrix is ​​used to represent the similarity relationship of segment time nodes in terms of risk features. The similarity threshold is set by those skilled in the art according to the actual situation.

[0064] Based on the spatial candidate adjacency matrix, the temporal candidate adjacency matrix, and the risk-similar candidate adjacency matrix, the candidate adjacency matrices are organized and classified to form a candidate adjacency matrix set composed of multiple candidate adjacency matrices, thus obtaining the segment time node set and the candidate adjacency matrix set.

[0065] Step S2032: Based on the candidate adjacency matrix set, construct a continuous-time neural evolution model to obtain the continuous-time state equations of nodes at each time segment.

[0066] The specific steps of step S2032 are as follows: Step S20321: Based on the candidate adjacency matrix set, calculate the historical memory kernel of the nodes at each time segment, construct the graph structure operator, and combine the historical memory kernel with the initial node features at multiple time scales to obtain the enhanced node state, the graph structure operator set, and the memory kernel weight set.

[0067] In this embodiment, based on the candidate adjacency matrix set and the initial node features of multiple time scales, the set of time nodes in each segment is traversed to determine the index order of each time node on the preset time axis and its row and column position in the candidate adjacency matrix. The time axis is set by those skilled in the art according to the actual situation.

[0068] Based on the preset memory duration and time step, historical segment time nodes that have candidate edges with the current segment time node are selected. The time interval between the historical segment time node and the current segment time node is determined. The time interval and the matrix elements of the corresponding positions of the two nodes in the candidate adjacency matrix are used as edge weights. The preset memory kernel function is input to calculate the memory weights corresponding to each historical segment time node. All memory weights corresponding to each segment time node are sorted to obtain the historical memory kernels of each segment time node. The memory kernel function, memory duration, and time step are set by those skilled in the art according to the actual situation.

[0069] Based on the historical memory kernel and the initial node features of multiple time scales, the initial node features of multiple time scales of each segment time node within the memory time length are weighted and accumulated. The initial node features of multiple time scales of the historical segment time node are multiplied by the corresponding memory weights and then summed to obtain the memory state vector of each segment time node.

[0070] The feature vectors of the initial nodes at multiple time scales at each time point and the corresponding memory state vectors are concatenated to generate the enhanced node states of the nodes at each time segment. The memory weights corresponding to the nodes at each time segment are then uniformly organized to obtain the memory kernel weight set.

[0071] Based on the candidate adjacency matrix set, each candidate adjacency matrix is ​​normalized, the number of edges connecting nodes in the candidate adjacency matrix at each time segment is counted, the node degree is determined, and the rows and columns of the candidate adjacency matrix are scaled according to the node degree to construct the initial matrix of the graph structure operator.

[0072] The initial graph structure operator matrix and the preset identity matrix are weighted and added together to generate a graph structure operator matrix for describing spatial relationships, temporal relationships and risk similarity relationships. The graph structure operator matrix is ​​then organized according to different relationship types to generate a graph structure operator set, resulting in the enhanced node state, the graph structure operator set and the memory kernel weight set. The identity matrix is ​​set by those skilled in the art according to the actual situation.

[0073] Step S20322: Based on the enhanced node states, the set of graph structure operators, and the set of memory kernel weights, construct the continuous-time state evolution equation to obtain the continuous-time neural evolution model to be identified.

[0074] In this embodiment, based on the enhanced node state, the set of graph structure operators and the set of memory kernel weights, the state change relationship of nodes in each segment over continuous time is modeled. The enhanced node state corresponding to each segment's time node is used as the state vector, and the time variable is used as a continuous variable to construct a first-order differential equation to describe the change of the state vector over time.

[0075] Based on the set of graph structure operators, the set of time nodes in each segment is traversed. The currently traversed time node in the segment is taken as the target time node in the segment. The enhanced node states of adjacent time nodes that are connected to the target time node in the segment are weighted and linearly combined. The spatial propagation increment is calculated based on the graph structure operators constructed from the spatial candidate adjacency matrix. The temporal propagation increment is calculated based on the graph structure operators constructed from the temporal candidate adjacency matrix. The risk similarity propagation increment is calculated based on the graph structure operators constructed from the risk similarity candidate adjacency matrix. The propagation increments are then weighted and synthesized to obtain the graph propagation state increment of the target time node in the segment.

[0076] Based on the memory kernel weight set, the enhanced node states of historical segment time nodes that have memory relationships with the target segment time nodes are accumulated by memory kernel weighting. According to the memory weights corresponding to the historical memory kernels, the weighted sum of each historical enhanced node state is obtained to obtain the memory state increment that represents the historical memory effect.

[0077] Based on the graph propagation state increment, memory state increment, and a preset family of nonlinear basis functions, the graph propagation state increment and memory state increment are nonlinearly transformed and linearly weighted. The first derivative of the enhanced node state with respect to time of each segment time node obtained from the first-order differential equation is expressed as a nonlinear combination of the graph propagation state increment and memory state increment, forming the continuous-time state evolution equation of each segment time node. The continuous-time state evolution equations of all segment time nodes are sorted out to obtain the continuous-time neural evolution model to be identified. The family of nonlinear basis functions is set by those skilled in the art according to the actual situation.

[0078] Step S20323: Based on the continuous-time neural evolution model to be identified, calculate the continuous-time state trajectory of each time node in each segment to obtain the predicted state evolution trajectory.

[0079] In this embodiment, based on the continuous-time neural evolution model and enhanced node states, the state evolution process of each segment's time node is calculated. According to the microseismic risk field sequence signal set, the start time and end time corresponding to each segment's time node are determined. A discrete time sequence is constructed between the start time and end time according to a preset discrete time step. The discrete time sequence is used as the time grid for continuous-time state evolution. The discrete time step is set by those skilled in the art according to the actual situation.

[0080] Based on discrete time series, at the corresponding start time of each segment time node, the enhanced node state is used as the initial state vector of the continuous time state equation and input into the continuous time neural evolution model to determine the time derivative of each segment time node at the start time.

[0081] Based on the time derivative and discrete time step, continuous-time numerical integration is performed to iteratively calculate the update value of the initial state vector of the continuous-time state equation between adjacent time points in the time grid.

[0082] Based on the state vector update results obtained by continuous-time numerical integration, the predicted state vectors of each segment time node are extracted at each time point of the discrete-time series. The predicted state vectors of each segment time node corresponding to each warning segment are arranged according to the time order to form the state evolution trajectory of each warning segment in continuous time. The state evolution trajectories corresponding to each warning segment are summarized to obtain the predicted state evolution trajectory.

[0083] Step S20324: Organize the initial node features of multiple time scales to obtain the time series of nodes in the segment.

[0084] In this embodiment, based on the initial node features of multiple time scales, the identifiers corresponding to each warning segment and each time, and according to the time order of the microseismic risk field sequence signal set, the initial node features of multiple time scales and the time index corresponding to each warning segment at each time are associated. Combined with the segment time nodes, the initial node features of multiple time scales corresponding to each warning segment at each time are bound to the corresponding segment time nodes respectively, resulting in a set of node status records with warning segment identifiers and time indexes.

[0085] Based on the node status record set, the time nodes of each segment corresponding to the same warning segment are arranged from early to late according to the time index. The initial node features of multiple time scales corresponding to each time are sequentially connected to form the node status time series of the warning segment. The node status time series corresponding to all warning segments are summarized to obtain the segment time node time series.

[0086] Step S20325: Minimize the error between the predicted state evolution trajectory and the time series of segment time nodes, determine the coefficients and memory kernel parameters in the continuous-time neural evolution model, and obtain the continuous-time state equation for each segment time node.

[0087] In this embodiment, based on the predicted state evolution trajectory, the time series of segment time nodes, and the continuous-time neural evolution model, the model parameters of each segment time node are identified. According to the time index and warning segment identifier of each segment time node in the segment time node time series, the corresponding predicted state vector is found in the predicted state evolution trajectory. The predicted state vector corresponding to the same time of the same warning segment and the feature vector of the initial node at multiple time scales are aligned to obtain the correspondence between the predicted state vector and the target state vector indexed by the warning segment identifier and the time index.

[0088] Based on the correspondence between the predicted state vector and the target state vector, the difference vector between the predicted state vector and the target state vector at each time point in each warning segment is calculated one by one. The components of the difference vector are squared and summed to obtain the error scalar of each segment's time node at the corresponding time point.

[0089] Based on the error scalars of all warning segments and all time points, the error scalars are accumulated to construct a total error metric function. In the total error metric function, the coefficients and memory kernel parameters in the continuous-time neural evolution model are used as parameters to be determined. The coefficients are linear weight parameters that appear in the continuous-time state evolution equation, and the memory kernel parameters are parameters that control the decay pattern and intensity of historical memory weights.

[0090] Based on the total error metric function, the numerical range of coefficients and memory kernel parameters in the continuous-time neural evolution model is constrained according to the preset actual requirements. By adding a parameter amplitude penalty term, the constraints are incorporated into the total error metric function, resulting in an error minimization objective function with parameter constraints. The actual requirements are set by those skilled in the art according to the actual situation.

[0091] Based on the objective function of minimizing error with parameter constraints, the coefficients and memory kernel parameters in the continuous-time neural evolution model are used as optimization variables. The least squares method is used to solve for the set of coefficient and memory kernel parameter values ​​that minimize the objective function of minimizing error.

[0092] The coefficients and memory kernel parameters obtained by solving are substituted into the continuous-time neural evolution model to update the continuous-time state evolution equations corresponding to the time nodes of each segment, thereby obtaining the continuous-time state equations of each time node.

[0093] Step S2033: Based on the continuous-time state equation, fit the segment time node time series composed of initial node features of multiple time scales to obtain the continuous-time influence coefficient matrix.

[0094] In this embodiment, based on the continuous-time state equation and the time series of each segment time node, the influence coefficient in the continuous-time state equation is fitted. According to the time interval between adjacent time points in the time series of the segment time nodes and the corresponding multi-timescale initial node characteristics, the state change rate of each segment time node at each time point is calculated, and the state change rate is used as the first derivative sample on the left side of the continuous-time state equation.

[0095] Based on the continuous-time state equation, the first derivative samples of each segment time node at each time point are represented as a linear combination of the enhanced node state vectors of all segment time nodes at the current time and a weighted sum of the historical memory terms.

[0096] Based on the first derivative samples at all time points and the corresponding enhanced node state samples, a system of linear equations with the influence coefficients as unknowns is constructed and solved to obtain the values ​​of each influence coefficient in the continuous-time state equation.

[0097] Based on the values ​​of each influence coefficient in the continuous-time state equation, the corresponding influence coefficients are filled into the corresponding matrix element positions by using the index of the target segment time node in the row direction and the index of the segment time node that affects the target segment time node in the column direction. This constructs a continuous-time influence coefficient matrix. The continuous-time influence coefficient matrices corresponding to different warning segments and different times are then sorted to obtain the continuous-time influence coefficient matrix.

[0098] Step S2034: Based on the continuous time influence coefficient matrix, select the time segment node pairs whose continuous time influence coefficient is greater than the preset threshold, construct the causal directed edge set, and obtain the causal directed graph of the time segment nodes.

[0099] In this embodiment, based on the continuous time influence coefficient matrix and the set of segment time nodes, the continuous time influence coefficient matrix is ​​structurally organized. Each continuous time influence coefficient matrix is ​​indexed by the number of the target segment time node in the row direction and by the number of the segment time node that affects the target segment time node in the column direction. Each matrix element is defined as the influence coefficient of the segment time node corresponding to the column index on the segment time node corresponding to the row index in continuous time.

[0100] Based on the continuous-time influence coefficient matrix and a preset continuous-time influence coefficient threshold, each matrix element is traversed. When the continuous-time influence coefficient corresponding to a certain matrix element is greater than the continuous-time influence coefficient threshold, the segment time node corresponding to the matrix column index is determined as the cause node, and the segment time node corresponding to the matrix row index is determined as the result node. A causal directed edge is determined between the cause node and the result node, pointing from the cause node to the result node, and this causal directed edge is added to the causal directed edge set. After traversing all continuous-time influence coefficient matrices, the causal directed edge set is obtained. The continuous-time influence coefficient threshold is set by those skilled in the art according to the actual situation.

[0101] Based on the set of time nodes in the segment and the set of causal directed edges, each time node in the set of time nodes in the segment is used as a node in the directed graph, and each causal directed edge in the set of causal directed edges is used as a directed edge in the directed graph. Directed connections are established between nodes according to the start and end points of the causal directed edges to construct a causal directed graph of time nodes in the segment.

[0102] Step S2035: Based on the causal directed graph of nodes in the segment time, update the candidate adjacency matrix set through the continuous time influence coefficient matrix to obtain the spatiotemporal multi-relationship tensor graph of the target monitoring area.

[0103] In this embodiment, based on the causal directed graph of segment time nodes, the continuous time influence coefficient matrix, and the candidate adjacency matrix set, the candidate adjacency matrix set is updated. According to the starting segment time node and ending segment time node of each causal directed edge in the segment time node causal directed graph, the position of the matrix element corresponding to each causal directed edge in the continuous time influence coefficient matrix is ​​determined, and the value of the matrix element is determined as the continuous time influence coefficient of the corresponding causal directed edge.

[0104] Based on the candidate adjacency matrix set, each candidate adjacency matrix is ​​traversed. The positions of non-zero matrix elements in each candidate adjacency matrix are compared with the causal directed edges in the causal directed graph of the segment time nodes. When the segment time node pair corresponding to a certain matrix element in the candidate adjacency matrix has a causal directed edge in the causal directed graph of the segment time nodes, the continuous time influence coefficient corresponding to the segment time node pair in the continuous time influence coefficient matrix is ​​extracted. The original weight value of the matrix element in the candidate adjacency matrix is ​​multiplied by the corresponding continuous time influence coefficient to obtain the updated edge weight value. The updated edge weight value replaces the original weight value in the candidate adjacency matrix.

[0105] Based on the updated set of candidate adjacency matrices and the preset stacking order, the candidate adjacency matrices in the set are arranged and stacked to form a three-dimensional array with the segment time node number as the row index, the segment time node number as the column index, and the candidate adjacency matrix number as the stacking index. The three-dimensional array is determined as the spatiotemporal multi-relationship tensor graph of the target monitoring area. The stacking order is set by those skilled in the art according to the actual situation.

[0106] Step S204: Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, construct the graph Laplacian operator and build the spatiotemporal propagation kernel function to obtain the spatiotemporal propagation weight set.

[0107] The specific steps of step S204 are as follows: Step S2041: Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, calculate the risk intensity of each segment time node, construct the risk weighted graph Laplacian operator for each segment time node, and perform spectral decomposition to obtain the feature spectral values ​​and feature vector set.

[0108] In this embodiment, based on the spatiotemporal multi-relation tensor graph of the target monitoring area, the risk intensity of each segment time node is calculated. The spatiotemporal multi-relation tensor graph of the target monitoring area is expanded in the relation stacking dimension. For each relation type, a corresponding candidate adjacency matrix is ​​extracted. According to the edge weights of the edges connecting each segment time node in each candidate adjacency matrix, the weights of all edges connected to the target segment time node are weighted and summed. The weighted summation result is used as the local risk measure of the target segment time node under the current relation type. According to the preset relation importance weights, the local risk measures under each relation type are weighted and superimposed to obtain the risk intensity of each segment time node. The relation importance weights are set by those skilled in the art according to the actual situation.

[0109] Based on the risk intensity of each segment time node and the candidate adjacency matrix corresponding to each relation type in the spatiotemporal multi-relation tensor graph of the target monitoring area, risk-weighted graph Laplacian operators for each relation type are constructed respectively. For each relation type, the degree value of each segment time node is obtained by summing the edge weights of each segment time node in the row direction according to the candidate adjacency matrix. The degree value of each segment time node and the corresponding risk intensity are weighted and combined to obtain the risk-weighted degree matrix.

[0110] Based on the risk-weighted degree matrix and the candidate adjacency matrix, a risk-weighted graph Laplacian matrix is ​​constructed by subtracting the candidate adjacency matrix from the risk-weighted degree matrix. The risk-weighted graph Laplacian matrix is ​​then normalized to obtain the risk-weighted graph Laplacian operator for each relation type.

[0111] Based on the risk-weighted graph Laplacian operator under each relation type, spectral decomposition is performed on the risk-weighted graph Laplacian operator to determine the spectral domain. For the risk-weighted graph Laplacian operator corresponding to each relation type, the eigenvalues ​​and eigenvectors of the matrix are calculated. All eigenvalues ​​are arranged in ascending order to obtain the eigenvalue set. The eigenvectors corresponding to each eigenvalue are arranged in column vector form to obtain the eigenvector set. The decomposition results of each relation type are then organized to obtain the eigenvalue and eigenvector sets.

[0112] Step S2042: Multiply each characteristic spectrum value and the corresponding risk intensity by the preset frequency adjustment coefficient and risk adjustment coefficient respectively to generate the first intermediate quantity and the second intermediate quantity. Divide the first intermediate quantity by the second intermediate quantity to obtain the ratio-type independent variable set.

[0113] In this embodiment, based on the feature spectrum value, the risk intensity of each segment time node, the preset frequency adjustment coefficient and risk adjustment coefficient, the feature spectrum value and the risk intensity of the corresponding segment time node under each relation type are traversed. The feature spectrum value is multiplied by the frequency adjustment coefficient to obtain the first intermediate quantity. The risk intensity is multiplied by the risk adjustment coefficient and a small constant is added to prevent the denominator from being zero to obtain the second intermediate quantity. The frequency adjustment coefficient and the risk adjustment coefficient are set by those skilled in the art according to the actual situation.

[0114] Based on the first and second intermediate values, for each feature spectrum value and the corresponding segment time node under each relation type, the first intermediate value is divided by the second intermediate value to obtain the corresponding ratio-type independent variables. According to the relation type index, feature spectrum value index and segment time node index, the ratio-type independent variables corresponding to each feature spectrum value and each segment time node are organized to obtain the set of ratio-type independent variables.

[0115] Step S2043: Based on the set of ratio-type independent variables, multiply each ratio-type independent variable by a preset spatial decay coefficient, and take the negative value of the exponential function to generate an exponential decay subset. Multiply each ratio-type independent variable by a preset focusing coefficient and perform a hyperbolic tangent function transformation to obtain a focusing subset.

[0116] In this embodiment, based on the set of ratio-type independent variables, the preset spatial decay coefficient, and the focusing coefficient, the ratio-type independent variables under each relation type are traversed, and each ratio-type independent variable is multiplied by the spatial decay coefficient to obtain the spatial decay intermediate quantity. The negative value of the exponential function corresponding to the spatial decay intermediate quantity is calculated to generate the exponential decay sub-item corresponding to each ratio-type independent variable. According to the relation type and the ratio-type independent variables, the exponential decay sub-item is sorted to obtain the exponential decay sub-item set. The spatial decay coefficient and the focusing coefficient are set by those skilled in the art according to the actual situation.

[0117] Based on the set of ratio-type independent variables and the focusing coefficient, each ratio-type independent variable is multiplied by the focusing coefficient to obtain the focusing intermediate quantity. The focusing intermediate quantity is then transformed by the hyperbolic tangent function to generate the focusing sub-items corresponding to each ratio-type independent variable. Based on the relationship type and the ratio-type independent variables, the focusing sub-items are organized to obtain the focusing sub-item set.

[0118] Step S2044: Based on the ratio-type independent variable set, the exponential decay sub-item set, and the focusing sub-item set, determine the symbol modulation coefficients, generate the symbol modulation sub-item set, and linearly combine the exponential decay sub-item set and the focusing sub-item set according to the preset modulation weight and the symbol modulation sub-item set to obtain the spatiotemporal propagation kernel function set.

[0119] In this embodiment, the symbol modulation coefficient is determined based on the ratio-type independent variable set, the exponential decay sub-item set, and the focusing sub-item set. Each ratio-type independent variable under each relation type is traversed. Based on the comparison between the sign and the corresponding absolute value of the ratio-type independent variable and a preset zero-neighbor threshold, when the ratio-type independent variable is positive and its absolute value is greater than the zero-neighbor threshold, the symbol modulation coefficient corresponding to that ratio-type independent variable is set to a preset first constant. When the ratio-type independent variable is negative and its absolute value is greater than the zero-neighbor threshold, the symbol modulation coefficient corresponding to that ratio-type independent variable is set to a preset second constant. When the absolute value of the ratio-type independent variable is less than or equal to the zero-neighbor threshold, the symbol modulation coefficient corresponding to that ratio-type independent variable is set to a preset third constant. Based on the symbol modulation coefficients corresponding to each ratio-type independent variable, the set is organized according to the relation type index and the ratio-type independent variable index to generate a symbol modulation sub-item set. The zero-neighbor threshold, the first constant, the second constant, and the third constant are set by those skilled in the art according to the actual situation.

[0120] Based on the exponential decay sub-items set, the focused sub-items set, the symbolic modulation sub-items set, and preset first and second modulation weights, the exponential decay sub-items and the focused sub-items are linearly combined. Each sub-item in the exponential decay sub-items set is multiplied by the first modulation weight, and each sub-item in the focused sub-items set is multiplied by its corresponding symbolic modulation sub-item and then multiplied by the second modulation weight. The two weighted results are summed to obtain the spatiotemporal propagation kernel function value corresponding to each ratio-type independent variable. The spatiotemporal propagation kernel function values ​​corresponding to all relation types and all ratio-type independent variables are organized to form a spatiotemporal propagation kernel function set. The first and second modulation weights are set by those skilled in the art according to the actual situation.

[0121] Step S2045: Based on the set of spatiotemporal propagation kernel functions and the set of feature vectors, map the spatiotemporal propagation kernel functions from the spectral domain to the spatiotemporal multi-relation tensor graph of the target monitoring area, calculate the spatiotemporal propagation weights, and obtain the set of spatiotemporal propagation weights.

[0122] In this embodiment, based on the set of spatiotemporal propagation kernel functions and the set of eigenvectors, the values ​​of the spatiotemporal propagation kernel functions in the spectral domain are organized. According to the risk-weighted graph Laplace operator corresponding to each relation type, the eigenvectors under the same relation type are stacked column by column to form an eigenvector matrix. According to the sorting order of the eigenspectral values, the values ​​of the spatiotemporal propagation kernel functions corresponding to each ratio-type independent variable under the relation type are sequentially filled into the diagonal positions to construct the spectral domain propagation kernel matrix of the relation type.

[0123] Based on the spectral domain propagation kernel matrix and the corresponding eigenvector matrix of each relation type, the spatiotemporal propagation kernel function of each relation type is mapped from the spectral domain to the graph domain. For each relation type, the graph domain propagation operator matrix under that relation type is calculated by matrix multiplication of the eigenvector matrix, the spectral domain propagation kernel matrix, and the transpose of the eigenvector matrix. Each matrix element in the graph domain propagation operator matrix is ​​used as the spatiotemporal propagation weight of the corresponding segment time node pair in the spatiotemporal multi-relation tensor graph of the target monitoring area, and the rows or columns of the graph domain propagation operator matrix are normalized.

[0124] Based on the graph propagation operator matrices obtained under each relation type, the graph propagation operator matrices corresponding to different relation types are organized according to the relation type. The values ​​of each matrix element in each graph propagation operator matrix are used as the spatiotemporal propagation weights under the corresponding relation type, thus forming a spatiotemporal propagation weight set with segment time node pairs as indexes and relation type as the distinguishing dimension.

[0125] Step S205: Based on the spatiotemporal propagation weight set, perform time decay weighted aggregation on the time nodes of each segment to generate the spatiotemporal memory risk representation of each warning segment at each time, and combine them to obtain the time-series risk representation set of each warning segment.

[0126] In this embodiment, based on the spatiotemporal propagation weight set, the segment time node set, and the initial node features of multiple time scales, the risk features of each segment time node are weighted and fused. According to the segment time node set, each segment time node is traversed, and the currently traversed segment time node is taken as the target segment time node. Segment time nodes that have edges connected to the target segment time node in the spatiotemporal multi-relation tensor graph of the target monitoring area are retrieved. The retrieved segment time nodes are determined as the set of adjacent segment time nodes of the target segment time node. The propagation weights of each adjacent segment time node pointing to the target segment time node are extracted from the spatiotemporal propagation weight set.

[0127] Based on the microseismic risk field sequence signal set, the time interval between each adjacent segment time node and the target segment time node is calculated. The time interval is input into a preset time decay function to obtain the time decay coefficient corresponding to each adjacent segment time node. The propagation weight of each adjacent segment time node is attenuated and corrected according to the time decay coefficient to obtain the time decay propagation weight of each adjacent segment time node. The time decay function is set by those skilled in the art according to the actual situation.

[0128] Based on time decay propagation weights and multi-timescale initial node features, the multi-timescale initial node features of each adjacent segment time node are weighted and summed. Then, according to a preset self-influence coefficient, the weighted summation result is linearly combined with the multi-timescale initial node features of the target segment time node itself to obtain the spatiotemporal memory risk representation of the target segment time node. The self-influence coefficient is set by those skilled in the art according to the actual situation.

[0129] Based on the set of time nodes in each segment and the spatiotemporal memory risk representation of each time node in each segment, the time nodes of the same warning segment are sorted from early to late according to the time index. The sorted spatiotemporal memory risk representations are arranged in sequence to form the temporal risk representation vector sequence of the warning segment. The temporal risk representation vector sequences of all warning segments are sorted to obtain the temporal risk representation set of each warning segment.

[0130] Step S3: Based on the temporal risk representation set, construct the spatiotemporal risk fiber bundle of the target monitoring area, and perform geometric embedding in the distorted hyperbolic spatiotemporal risk manifold space to obtain the hyperbolic evolution representation set.

[0131] The specific steps of step S3 are as follows: Step S301: Based on the time-series risk representation set, reorganize the time-series risk representation corresponding to each warning segment at each time point to obtain the set of time nodes of the reorganized segment and the corresponding time-series risk vector set.

[0132] In this embodiment, based on the time-series risk representation set and the segment time node set, the time-series risk representation corresponding to each warning segment at each time is reorganized. According to the warning segment identifier and time index recorded in the time-series risk representation set, the segment time node with the same warning segment identifier and time index is searched in the segment time node set. The found segment time node is recombined with the corresponding time-series risk representation, and the combined segment time node is renumbered to form a reorganized segment time node set.

[0133] Based on the set of time nodes of the reorganization segment, the temporal risk representation corresponding to each time node of the reorganization segment is used as the temporal risk vector of that time node. The temporal risk vectors are then sorted and arranged according to the numbering order of the time nodes of the reorganization segment to obtain a set of temporal risk vectors that correspond one-to-one with the set of time nodes of the reorganization segment.

[0134] Step S302: Based on the set of time nodes of the recombined segment and the corresponding temporal risk vector set, construct the spatiotemporal risk fiber bundle of the target monitoring area, and perform tangent space projection to obtain the tangent space risk vector set.

[0135] The specific steps of step S302 are as follows: Step S3021: Based on the set of time nodes in the recombined segment and the corresponding temporal risk vector set, determine the time base space and construct fibers to obtain the spatiotemporal risk fiber bundle of the target monitoring area.

[0136] In this embodiment, based on the set of reorganization segment time nodes and the corresponding time-series risk vector set, the time structure of the time-series risk vector is organized. According to the time index of each reorganization segment time node in the set of reorganization segment time nodes, all time indices are deduplicated and sorted in ascending order to obtain a time index sequence. Each time index in this time index sequence is used as a set of discrete time points. The discrete time point set is arranged according to the chronological order to form a time base space.

[0137] Based on the time base space and the time-series risk vector set, each time point in the time base space is traversed. At the current time point, all time points with time indices equal to the current time point are selected from the set of time points of the reorganization segment. The corresponding time-series risk vectors are extracted, and the vector space where these time-series risk vectors are located is determined as the risk state space at the current time point. This risk state space is then used as the fiber corresponding to the current time point.

[0138] Based on the time base space and the risk state space corresponding to each time point, each time point and the corresponding risk state space are associated through a mapping relationship, so that a risk state fiber is suspended at each time point in the time base space. The risk state fibers in the time base space and all time points are combined according to the mapping relationship to form a spatiotemporal risk fiber bundle of the target monitoring area.

[0139] Step S3022: Based on the spatiotemporal risk fiber bundle of the target monitoring area, construct the corresponding fiber ball risk state space at each time node of the fiber bundle to obtain the fiber ball risk state space set.

[0140] In this embodiment, based on the spatiotemporal risk fiber bundle of the target monitoring area, the set of time nodes of the recombined segment, and the corresponding temporal risk vector set, a spherical model is performed on the local risk state at each time node of the fiber bundle.

[0141] Based on the time base space in the spatiotemporal risk fiber bundle of the target monitoring area and the risk state space corresponding to each time point, the risk state space corresponding to each time point is taken as the real vector space, and the Euclidean norm is used as the length measure of the temporal risk vector in it.

[0142] Based on the set of time nodes of the reorganization segment and the set of temporal risk vectors, the time nodes of each reorganization segment are traversed. According to the time index of the time node of the reorganization segment, the time point to which it belongs and the corresponding risk state space are located in the spatiotemporal risk fiber bundle of the target monitoring area. The all-zero vector in the risk state space is determined as the reference center of the time node of the reorganization segment.

[0143] Based on the Euclidean norm of the time-series risk vector corresponding to the reorganization segment time node, the preset proportional coefficient, and the minimum radius threshold, the risk radius corresponding to each reorganization segment time node is calculated. The proportional coefficient and the minimum radius threshold are set by those skilled in the art according to the actual situation.

[0144] Based on the reference center and the risk radius, within the corresponding risk state space, with the reference center as the sphere center and the risk radius as the radius, determine the set of all risk vectors whose Euclidean norm is less than or equal to the risk radius. Each set of risk vectors is used as the fiber ball risk state space of the corresponding reorganization segment time node. According to the number of the reorganization segment time node, organize the fiber ball risk state spaces corresponding to all reorganization segment time nodes to obtain the fiber ball risk state space set.

[0145] Step S3023: Based on the fiber ball risk state space set and the preset distortion function, the metric between the time base space and the fiber ball risk state space is distorted to obtain the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area.

[0146] In this embodiment, based on the time base space, the set of fiber ball risk state spaces, and a preset distortion function, the metric relationship between the time base space and the fiber ball risk state space is distorted and constructed. Each time index in the time base space is used as a scalar coordinate value arranged along the time direction. The Euclidean norm of the risk vector in each fiber ball risk state space is used as a scalar value representing the magnitude of the risk. The normalized components of the risk vector in each characteristic direction are used as vector values ​​representing the direction of the risk. The distortion function is set by those skilled in the art according to the actual situation.

[0147] Based on the Euclidean norm of time index and risk vector, each time index and the corresponding risk vector in the risk state space of the fiber ball are combined, and the Euclidean norm of the time index and the risk vector are paired to obtain a spatiotemporal risk scalar pair used to describe the time position and risk magnitude of each segment's time node.

[0148] Based on the normalized components of each risk vector, each normalized component is associated with its corresponding spatiotemporal risk scalar pair to obtain a local spatiotemporal risk coordinate representation jointly defined in the time base space and the fiber sphere risk state space.

[0149] Based on the distortion function, the time index and Euclidean norm of the risk vector in each spatiotemporal risk scalar pair are used as inputs to the distortion function. The time direction distortion coefficient and the risk direction distortion coefficient are calculated. According to the time direction distortion coefficient, the distance between adjacent time indices in the time base space is scaled. According to the risk direction distortion coefficient, the distance corresponding to the Euclidean norm of the risk vector in the fiber sphere risk state space is scaled.

[0150] Based on the distorted time-direction distance metric and the risk-direction distance metric, the distorted risk vector coordinates in the risk state space of each time index and the corresponding fiber sphere are combined to form a local hyperbolic spatiotemporal risk metric structure at each time index.

[0151] Based on the chronological order of the time index, all local hyperbolic spatiotemporal risk measurement structures are spliced ​​and smoothly transitioned to construct a globally continuous negative curvature spatiotemporal risk geometric structure on the time base space and the fiber sphere risk state space. This geometric structure is used as the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area.

[0152] Step S3024: Based on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, construct a tangent space at a preset base point, and project the temporal risk vector set onto the tangent space through a linear transformation to obtain the tangent space risk vector set.

[0153] In this embodiment, based on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, the set of time nodes of the recombined segment, and the corresponding temporal risk vector set, a linear approximation is made on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. According to the distance metric of the time direction and the risk direction in the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, a vector set is formed by the linear combination of the spatiotemporal risk change directions that pass through a preset base point and have finite first-order changes in the time direction and the risk direction. This vector set is determined as the tangent space at the base point. The base point is set by those skilled in the art according to the actual situation.

[0154] Based on the distance metric in the tangent space at the base point and the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, a linear transformation is performed on the temporal risk vector set. According to the set of time nodes in the reorganization segment, the time nodes of each reorganization segment are traversed to obtain the time index and temporal risk vector corresponding to each time node in the reorganization segment. The Euclidean norm of each temporal risk vector is used as the risk magnitude. The time index and risk magnitude corresponding to the base point are subtracted from the time index and risk magnitude of each time node in the reorganization segment to construct a change vector composed of time difference and risk difference.

[0155] According to the preset linear transformation matrix, each transformation vector is mapped to the tangent space at the base point to obtain the tangent space risk vector corresponding to each reorganization segment time node. All tangent space risk vectors are sorted according to the index order of the reorganization segment time node set to obtain the tangent space risk vector set. The linear transformation matrix is ​​set by those skilled in the art according to the actual situation.

[0156] Step S303: Based on the tangent space risk vector set, the tangent space risk vector is mapped to the fiber sphere coordinate points on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area through exponential mapping, thereby obtaining the initial fiber sphere spatiotemporal embedding coordinate set.

[0157] In this embodiment, based on the distorted hyperbolic spatiotemporal risk manifold space and the tangent space risk vector set of the target monitoring area, an exponential mapping from the tangent space to the manifold is defined at the base point of the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. According to the distance metric in the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, each tangent space risk vector in the tangent space at the base point is used as the initial change direction and change length at the base point, which is used to generate the shortest path curve passing through the base point on the manifold. The shortest path curve is the geodesic curve.

[0158] Based on the tangent space risk vector set, each tangent space risk vector is traversed, and the norm of each tangent space risk vector is calculated. The norm is used as the path length along the geodesic curve on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. The tangent space risk vectors are normalized to obtain the unit direction vector corresponding to each tangent space risk vector.

[0159] Based on the unit direction vector and path length, starting from the base point, the corresponding geodesic curve is moved along the path length in the twisted hyperbolic spatiotemporal risk manifold space of the target monitoring area to determine the endpoint of each geodesic curve. The coordinates of each endpoint in the twisted hyperbolic spatiotemporal risk manifold space of the target monitoring area are represented as the coordinate points of the fiber sphere.

[0160] Based on the correspondence between each reorganization segment time node and the tangent space risk vector in the reorganization segment time node set, the fiber ball coordinate points corresponding to each reorganization segment time node are organized according to the index order of the reorganization segment time nodes. The fiber ball coordinate points of all reorganization segment time nodes are arranged and combined in the time direction and the warning segment direction to form a fiber ball spatiotemporal coordinate set organized by the warning segment and time index, thus obtaining the initial fiber ball spatiotemporal embedding coordinate set.

[0161] like Figure 3 As shown, step S304: Based on the initial spatiotemporal embedding coordinate set of fiber spheres, determine the fiber sphere embedding coordinate sequence of each warning section, and interpolate and smooth the fiber sphere embedding coordinate sequence to obtain the spatiotemporal evolution trajectory set of fiber spheres for each warning section.

[0162] In this embodiment, based on the initial set of spatiotemporal embedding coordinates of the fiber spheres and the set of time nodes of the recombined segments, the fiber sphere embedding coordinates of each warning segment are recombined. According to the warning segment identifier and time index in the set of time nodes of the recombined segments, the segment time nodes corresponding to the same warning segment are filtered and sorted from early to late according to the time index. The fiber sphere coordinate points corresponding to the sorted segment time nodes in the initial set of spatiotemporal embedding coordinates are arranged in sequence to obtain the fiber sphere embedding coordinate sequence of the warning segment.

[0163] Based on the embedded coordinate sequence of fiber balls and the corresponding time index sequence of each warning zone, interpolation processing is performed on adjacent coordinate point pairs whose time index difference is greater than a preset first interpolation threshold or whose distance between fiber ball coordinate points is greater than a preset second interpolation threshold, according to the preset interpolation step size and interpolation judgment rules. The interpolation step size, interpolation judgment rules, first interpolation threshold and second interpolation threshold are set by those skilled in the art according to the actual situation.

[0164] Based on the time index and fiber ball coordinates of each adjacent coordinate point pair, the time index difference is divided into m1 equal steps, and linear interpolation or spline interpolation is performed on the adjacent coordinate points in the fiber ball coordinate space to generate interpolated coordinate points. The interpolated coordinate points are then inserted into the corresponding fiber ball embedding coordinate sequence to obtain the interpolated fiber ball embedding coordinate sequence.

[0165] Based on the interpolated fiber ball embedding coordinate sequence, the fiber ball embedding coordinates of each warning segment on the time axis are smoothed according to the preset smoothing window length. The smoothed fiber ball coordinates are calculated based on the fiber ball coordinates of each coordinate point within the smoothing window. The smoothed fiber ball coordinates are arranged sequentially according to the time index to obtain the spatiotemporal evolution trajectory of the fiber ball on the time axis for each warning segment. The spatiotemporal evolution trajectories of the fiber balls in all warning segments are organized to form a set of spatiotemporal evolution trajectories of the fiber balls in each warning segment. The smoothing window length is set by those skilled in the art according to the actual situation.

[0166] Step S305: Based on the spatiotemporal evolution trajectory set of the fiber sphere, extract the coordinates of the distorted hyperbolic spatiotemporal risk manifold of the target monitoring area corresponding to each early warning segment at each time, and obtain the hyperbolic evolution representation set of each early warning segment at each time.

[0167] In this embodiment, based on the set of spatiotemporal evolution trajectories of the fiber sphere and the set of time nodes of the reorganization segment, the risk position of each warning segment at each time is extracted. According to the warning segment identifier and time index recorded in the set of time nodes of the reorganization segment, the same warning segment is traversed at each time index. The trajectory point with the same or closest time index to the current time index is found in the spatiotemporal evolution trajectory of the fiber sphere corresponding to the warning segment. The coordinates of the trajectory point on the fiber sphere are determined as the spatiotemporal coordinates of the fiber sphere of the warning segment under the current time index.

[0168] Based on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, the spatiotemporal coordinates of the fiber sphere under each time index of each early warning segment are used as the coordinate representation on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. According to the components of the corresponding time direction and risk direction in the spatiotemporal coordinates of the fiber sphere, the time direction component and the risk direction component are combined into a distorted hyperbolic spatiotemporal risk manifold coordinate vector. This coordinate vector is determined as the hyperbolic evolution representation of the early warning segment under the time index.

[0169] Based on the warning segment identifier and time index, the hyperbolic evolution representations obtained by all warning segments under all time indices are grouped according to the warning segment, and sorted from early to late according to the time index within each warning segment. The sorted hyperbolic evolution representations are arranged in sequence to form a set of hyperbolic evolution representations indexed by the warning segment and the time index, thus obtaining the hyperbolic evolution representation set of each warning segment at each time.

[0170] Step S4: Based on the hyperbolic evolution representation set, calculate the hyperbolic risk distance of each warning segment, and classify the hyperbolic risk distance according to the multi-level dynamic warning judgment rules to provide dynamic warnings for the risks.

[0171] In this embodiment, the risk intensity of each warning segment is measured based on the hyperbolic evolution representation set and the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. According to the warning segment identifier and time index recorded in the hyperbolic evolution representation set, the hyperbolic evolution representation corresponding to each warning segment at each time is traversed. Reference coordinates representing the reference safety state are selected as hyperbolic risk benchmarks in the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. Based on the hyperbolic risk benchmarks and the distance measurement of each hyperbolic evolution representation in the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, the distance from each hyperbolic evolution representation to the hyperbolic risk benchmark is calculated. This distance is determined as the instantaneous hyperbolic distance, and the instantaneous hyperbolic distance sequence of each warning segment at each time is obtained.

[0172] Based on the instantaneous hyperbolic distance sequence of each warning segment, the instantaneous hyperbolic distances at the current moment and the previous m2 moments are weighted and accumulated according to the preset time window length and time weight. The time-weighted hyperbolic distance used to describe the persistence of risk and the degree of cumulative exposure is calculated. According to the preset combination coefficient, the time-weighted hyperbolic distance and the instantaneous hyperbolic distance at the current moment are linearly combined to obtain the hyperbolic risk distance of each warning segment at the current moment. The time window length, time weight and combination coefficient are set by those skilled in the art according to the actual situation.

[0173] Based on the preset preliminary warning rules, a rule template for multi-level dynamic warning judgment rules is determined, and a list of parameter items for the rule template is determined. Then, parameters to be optimized are extracted according to the list of parameter items. The extraction of parameters to be optimized includes reading the parameter items related to the warning level boundaries item by item in the rule template, and using the set of rising thresholds and the set of falling thresholds corresponding to each warning level as two sets of parameters to be optimized; reading the parameter items related to time-series judgment, and using the continuous over-threshold duration n2, the continuous under-threshold duration n3, and the trend judgment window m3 as parameters to be optimized; and reading the parameter items related to distance fusion, and using the weighting coefficients used to combine instantaneous hyperbolic distance and time-weighted hyperbolic distance as parameters to be optimized. The preliminary warning rules are set by those skilled in the art according to the actual situation.

[0174] Based on the range of values ​​of the parameters to be optimized, the initial configuration of the particle swarm optimization algorithm is determined. The range of values ​​of the parameters to be optimized is determined by the default parameter values ​​and the corresponding allowable adjustment range in the parameter item list of the preliminary warning rules. The allowable adjustment range is set by those skilled in the art based on the statistical range of historical hyperbolic risk distance sequences and the requirements for the classification of warning levels.

[0175] Historical early warning results are collected, and a fitness function is constructed based on the historical hyperbolic risk distance sequence and historical early warning results. Penalty weights are set for missed and false alarms. The fitness function includes an early warning accuracy cost term and an early warning level fluctuation penalty term, so that the fitness evaluation reflects both the effectiveness and stability of early warning.

[0176] Based on the initialization configuration, the particle swarm size, maximum number of iterations, inertia weight, individual learning factor, and swarm learning factor are set. Within the range of these values, an initial particle position vector is randomly generated for each particle, and its velocity vector is initialized. The initial particle position vector is taken as the historical best position of that particle. The initial fitness value of each particle is calculated based on the fitness function, and the particle position vector with the best initial fitness is selected as the global best position. The particle swarm size, maximum number of iterations, inertia weight, individual learning factor, and swarm learning factor are hyperparameters of the particle swarm optimization algorithm, which are preset by those skilled in the art based on actual computing resources and convergence requirements.

[0177] The particle swarm is iteratively optimized. During the iterative optimization, for each particle in each iteration, the velocity vector of the particle is updated based on the inertia weight, the individual learning factor and the group learning factor, and the particle position vector is updated based on the updated velocity vector. When the updated particle position vector has an out-of-bounds dimension, the out-of-bounds dimension is truncated or bounced back so that the particle position vector falls within the value range of the parameter to be optimized.

[0178] Based on the updated particle position vector, the parameter values ​​of each dimension in the particle position vector are sequentially backfilled into the corresponding parameter items in the rule template to obtain the candidate multi-level dynamic early warning judgment rule corresponding to the particle; based on the candidate multi-level dynamic early warning judgment rule, the historical hyperbolic risk distance sequence is maintained, upgraded, downgraded or deactivated to obtain the candidate early warning level sequence, and the candidate fitness value is calculated according to the candidate early warning level sequence and the historical early warning results; when the candidate fitness value is better than the historical best fitness value of the particle, the historical best position vector of the particle is updated; when the candidate fitness value is better than the current global best fitness value, the global best position vector is updated.

[0179] Repeatedly perform particle velocity updates, particle position updates, out-of-bounds handling, rule backfilling, and fitness evaluation until the global optimal fitness meets the preset particle swarm convergence condition or reaches the maximum number of iterations. Output the global optimal position vector as the optimal parameter vector, and backfill the optimal parameter vector into the rule template of the preliminary warning rule to obtain the multi-level dynamic warning judgment rule. The particle swarm convergence condition is set by those skilled in the art according to the actual situation.

[0180] Based on the multi-level dynamic early warning judgment rules and the hyperbolic risk distance of each early warning segment at the current moment, combined with the dynamic early warning level of each early warning segment at the previous moment, the early warning level of each early warning segment is determined to be maintained, upgraded, downgraded or lifted.

[0181] Based on the above judgment results, the dynamic warning level corresponding to each warning section at the current time is determined, the dynamic warning levels of all warning sections at all times are sorted out, and dynamic warnings are issued for risks.

[0182] exist Figure 3 In the diagram, the outermost thick circular boundary represents the overall boundary of the distorted hyperbolic spacetime risk manifold space. All points inside the disk represent the embedding position of a certain warning segment at a certain moment. The light-colored solid circle in the center represents the safe zone or low-risk zone, that is, the trajectory points near this area are basically safe and do not require warning.

[0183] The concentric rings of colored dashed lines from the inside out represent the contour lines of risk levels that increase sequentially from the inside out. The intervals between the rings represent the increasing hyperbolic risk distance, indicating that the risk increases as it moves outward. The concentric rings are slightly flattened longitudinally to represent the distorted hyperbolic spacetime geometry. The concentric rings and the central light-colored solid circle have the same center, which represents the risk benchmark or reference origin.

[0184] The outermost ring of the long arc solid line trajectory, composed of orange arc solid lines and cross-shaped dots, represents the hyperbolic spatiotemporal evolution trajectory of a certain warning segment over a longer time scale. The trajectory is close to the outer dashed ring, indicating that the segment has always been at a high risk level over a long period of time, that is, in a situation of continuous risk accumulation. The dense cross-shaped dots on the arc represent the embedding positions of each discrete moment. The points are closer to the outer side as they move towards the later part of the trajectory, indicating that the risk is accumulating or slowly intensifying. The slightly larger green dots on the arc represent the selected key moments, including the beginning, the middle, and the current moment, representing the trend of gradually moving outward over time.

[0185] The short arc-shaped dashed line trajectory, consisting of pink dashed lines and blue cross-shaped dots, represents the hyperbolic spatiotemporal evolution trajectory of another warning zone in a short period of time. The trajectory briefly bulges outward near the central dashed ring and then retracts inward, indicating that this zone only experienced short-term risk fluctuations and was not in a high-risk state for a long time.

[0186] The black ray pointing from the center of the circle to the end of the long arc represents the calculation process of the hyperbolic risk distance at the end of the long arc trajectory at the current moment. The starting point of the ray is at the risk benchmark point, and the ending point is at the current position of the long arc trajectory. The length of the ray corresponds to the current comprehensive risk distance.

[0187] The orange vertical bar on the right represents the dynamic warning level axis. The top of the bar represents the highest warning level, the bottom represents the lowest warning level, and the middle position corresponds to the medium and lower warning levels. The multiple short horizontal lines on the vertical bar represent the boundaries between different warning levels. The relative height of each scale on the vertical bar corresponds to the threshold of the hyperbolic risk distance, indicating the segment from safe to the highest warning level. The thick horizontal line at the top represents the upper boundary of the highest warning level, and the thick horizontal line at the bottom represents the lower boundary of the lowest warning level or no warning state.

[0188] Example 2: Please see Figure 4 The present invention provides an embodiment of a microseismic risk field construction and dynamic early warning system, comprising: a data acquisition module, an early warning division module, a hyperbolic embedding module, and an early warning generation module.

[0189] The data acquisition module is used to collect microseismic data of the target monitoring area and write it into the storage area. Based on the microseismic data of the target monitoring area, a microseismic risk field is constructed to obtain a set of microseismic risk field sequence signals.

[0190] The early warning segmentation module divides early warning sections based on the microseismic risk field sequence signal set, constructs a spatiotemporal multi-relationship tensor map of the target monitoring area, and obtains a time-series risk representation set.

[0191] The hyperbolic embedding module constructs a spatiotemporal risk fiber bundle for the target monitoring region based on the temporal risk representation set, and performs geometric embedding in the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring region to obtain a hyperbolic evolution representation set.

[0192] The early warning generation module calculates the hyperbolic risk distance for each early warning segment based on the hyperbolic evolution representation set, and classifies the hyperbolic risk distance according to the preset multi-level dynamic early warning judgment rules to provide dynamic early warning for the risk.

[0193] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0194] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing and dynamically warning of a microseismic risk field, characterized in that, include: Microseismic data of the target monitoring area are collected and written to the storage area. Based on the microseismic data of the target monitoring area, a microseismic risk field is constructed to obtain a set of microseismic risk field sequence signals. The microseismic data of the target monitoring area includes microseismic events and the occurrence time, spatial coordinates and event energy of each microseismic event. Based on the microseismic risk field sequence signal set, warning sections are divided, and a spatiotemporal multi-relation tensor map of the target monitoring area is constructed to obtain a time-series risk representation set. The time-series risk representation set is obtained by weighted accumulation and extreme value statistics of the section risk characteristics of each warning section, combined with a continuous-time neural evolution model and spatiotemporal propagation kernel function, and time decay weighted aggregation of the time nodes of each section. Based on the temporal risk representation set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and geometric embedding is performed in the distorted hyperbolic spatiotemporal risk manifold space to obtain a hyperbolic evolution representation set; Based on the hyperbolic evolutionary representation set, the hyperbolic risk distance of each warning segment is calculated, and the hyperbolic risk distance is classified according to the multi-level dynamic warning judgment rule to provide dynamic warning for the risk. The multi-level dynamic warning judgment rule is obtained by backfilling the parameters involved in the preset preliminary warning rule after optimizing the rise threshold, release threshold, continuous duration and weight coefficient through the particle swarm algorithm.

2. The method for constructing and dynamically warning of a microseismic risk field according to claim 1, characterized in that, The process involves constructing a microseismic risk field based on microseismic data from the target monitoring area, resulting in a microseismic risk field sequence signal set, including: Based on the occurrence time of each microseismic event and a preset window capacity threshold, the microseismic data of the target monitoring area are sorted and divided into sliding windows to obtain a sliding window sequence. Based on the sliding window sequence and the preset grid resolution, the regular grid corresponding to each sliding window is determined, and the maximum value is filtered to obtain the microseismic risk field image frame corresponding to each sliding window. The image frames of the microseismic risk field are subjected to structural similarity interpolation and combined to obtain a set of microseismic risk field sequence signals.

3. The method for constructing and dynamically warning of a microseismic risk field according to claim 2, characterized in that, The process involves determining the regular grid corresponding to each sliding window based on the sliding window sequence and a preset grid resolution, and then filtering for the maximum value to obtain the microseismic risk field image frame corresponding to each sliding window, including: Based on the sliding window sequence and the preset grid resolution, the coordinate range of each microseismic event within each sliding window is determined, and the grid is divided to obtain a regular grid. Based on the spatial coordinates and event energy of each microseismic event, the energy diffusion value of each microseismic event at each grid node in the regular grid is calculated. The maximum value of energy diffusion for each grid node is selected to obtain the microseismic risk field image frame corresponding to each sliding window.

4. The method for constructing and dynamically warning of a microseismic risk field according to claim 3, characterized in that, The method involves dividing the early warning zones based on the microseismic risk field sequence signal set, constructing a spatiotemporal multi-relation tensor map of the target monitoring area, and obtaining a time-series risk representation set, including: Based on the microseismic risk field sequence signal set, the warning sections are divided, and the energy diffusion values ​​of the corresponding grid nodes in each warning section are time-series aggregated to obtain the section risk characteristic sequence of each warning section at each time. Based on the segment risk feature sequence and the preset multi-time scale time window, the segment risk features are weighted and accumulated and extreme value statistics are performed to obtain the multi-time scale initial node features of each warning segment at each time. Based on the initial node features at multiple time scales, each warning segment is mapped to a segment time node at each time, and a continuous-time neural evolution model is constructed to obtain a spatiotemporal multi-relationship tensor map of the target monitoring area. Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, a graph Laplacian operator is constructed, and a spatiotemporal propagation kernel function is built to obtain the spatiotemporal propagation weight set; Based on the spatiotemporal propagation weight set, time decay weighted aggregation is performed on the time nodes of each segment to generate the spatiotemporal memory risk representation of each warning segment at each time, and then combined to obtain the time-series risk representation set of each warning segment.

5. The method for constructing and dynamically warning of a microseismic risk field according to claim 4, characterized in that, The method, based on the initial node features at multiple time scales, maps each warning segment to a segment time node at each moment, constructs a continuous-time neural evolution model, and obtains a spatiotemporal multi-relationship tensor map of the target monitoring area, including: Based on the initial node features at multiple time scales, each warning segment is mapped to segment time nodes at each time, and a candidate adjacency matrix is ​​constructed to obtain the set of segment time nodes and the set of candidate adjacency matrices. Based on the candidate adjacency matrix set, a continuous-time neural evolution model is constructed to obtain the continuous-time state equation of the nodes at each time segment. Based on the continuous-time state equation, the time series of segment time nodes composed of initial node features of multiple time scales is fitted to obtain the continuous-time influence coefficient matrix. Based on the continuous time influence coefficient matrix, the time segment node pairs with continuous time influence coefficients greater than a preset threshold are selected, and a set of causal directed edges is constructed to obtain the causal directed graph of time segment nodes. Based on the causal directed graph of nodes in the segment, the candidate adjacency matrix set is updated by the continuous time influence coefficient matrix to obtain the spatiotemporal multi-relationship tensor graph of the target monitoring area.

6. The method for constructing and dynamically warning of a microseismic risk field according to claim 5, characterized in that, The continuous-time neural evolution model, constructed based on the candidate adjacency matrix set, yields the continuous-time state equations for nodes at each time segment, including: Based on the candidate adjacency matrix set, the historical memory kernel of the node at each time segment is calculated, and a graph structure operator is constructed. The historical memory kernel and the initial node features at multiple time scales are combined to obtain the enhanced node state, the graph structure operator set, and the memory kernel weight set. Based on the enhanced node state, graph structure operator set, and memory kernel weight set, a continuous-time state evolution equation is constructed to obtain the continuous-time neural evolution model to be identified. Based on the continuous-time neural evolution model to be identified, the continuous-time state trajectory of each time node in each segment is calculated to obtain the predicted state evolution trajectory. The initial node features at multiple time scales are processed to obtain the time series of nodes at different time segments; By minimizing the error between the predicted state evolution trajectory and the time series of time nodes in each segment, the coefficients and memory kernel parameters in the continuous-time neural evolution model are determined, and the continuous-time state equations of each time node in each segment are obtained.

7. The method for constructing and dynamically warning of a microseismic risk field according to claim 6, characterized in that, The method involves constructing a graph Laplacian operator based on the spatiotemporal multi-relation tensor graph of the target monitoring area, and building a spatiotemporal propagation kernel function to obtain a spatiotemporal propagation weight set, including: Based on the spatiotemporal multi-relation tensor graph of the target monitoring area, the risk intensity of each segment time node is calculated, the risk weighted graph Laplacian operator of each segment time node is constructed, and spectral decomposition is performed to obtain the feature spectral values ​​and feature vector set. Each characteristic spectrum value and its corresponding risk intensity are multiplied by a preset frequency adjustment coefficient and risk adjustment coefficient, respectively, to generate a first intermediate quantity and a second intermediate quantity. The first intermediate quantity is divided by the second intermediate quantity to obtain a set of ratio-type independent variables. Based on the set of ratio-type independent variables, each ratio-type independent variable is multiplied by a preset spatial decay coefficient, and the negative value of the exponential function is taken to generate an exponential decay subset. Each ratio-type independent variable is multiplied by a preset focusing coefficient and subjected to a hyperbolic tangent function transformation to obtain a focusing subset. Based on the ratio-type independent variable set, the exponential decay sub-item set, and the focusing sub-item set, the symbol modulation coefficient is determined, the symbol modulation sub-item set is generated, and the exponential decay sub-item set and the focusing sub-item set are linearly combined according to the preset modulation weight and the symbol modulation sub-item set to obtain the spatiotemporal propagation kernel function set. Based on the set of spatiotemporal propagation kernel functions and the set of feature vectors, the spatiotemporal propagation kernel functions are mapped from the spectral domain to the spatiotemporal multi-relation tensor graph of the target monitoring area, and the spatiotemporal propagation weights are calculated to obtain the set of spatiotemporal propagation weights.

8. The method for constructing and dynamically warning of a microseismic risk field according to claim 7, characterized in that, The method involves constructing a spatiotemporal risk fiber bundle for the target monitoring area based on the temporal risk representation set, and then geometrically embedding it into the distorted hyperbolic spatiotemporal risk manifold space to obtain a hyperbolic evolutionary representation set, including: Based on the time-series risk representation set, the time-series risk representations corresponding to each warning segment at each time are reorganized to obtain the set of time nodes of the reorganized segment and the corresponding time-series risk vector set. Based on the set of time nodes of the recombined segment and the corresponding temporal risk vector set, a spatiotemporal risk fiber bundle of the target monitoring area is constructed, and tangent space projection is performed to obtain the tangent space risk vector set; Based on the tangent space risk vector set, the tangent space risk vector is mapped to fiber sphere coordinate points on the twisted hyperbolic spatiotemporal risk manifold space through exponential mapping, thus obtaining the initial fiber sphere spatiotemporal embedding coordinate set; Based on the initial spatiotemporal embedding coordinate set of the fiber spheres, the fiber sphere embedding coordinate sequence of each warning zone is determined, and the fiber sphere embedding coordinate sequence is interpolated and smoothed to obtain the spatiotemporal evolution trajectory set of the fiber spheres in each warning zone. Based on the spatiotemporal evolution trajectory set of the fiber sphere, the coordinates of the distorted hyperbolic spatiotemporal risk manifold corresponding to each warning segment at each time are extracted, and the hyperbolic evolution representation set of each warning segment at each time is obtained.

9. The method for constructing and dynamically warning of a microseismic risk field according to claim 8, characterized in that, The method involves constructing a spatiotemporal risk fiber bundle for the target monitoring area based on the set of time nodes in the reconstructed segment and the corresponding temporal risk vector set, and then performing tangent spatial projection to obtain the tangent spatial risk vector set, including: Based on the set of time nodes in the recombined segment and the corresponding temporal risk vector set, the temporal base space is determined, and fibers are constructed to obtain the spatiotemporal risk fiber bundle of the target monitoring area. Based on the spatiotemporal risk fiber bundle of the target monitoring area, the corresponding fiber ball risk state space is constructed at each time node of the fiber bundle, and the fiber ball risk state space set is obtained. Based on the fiber ball risk state space set and the preset distortion function, the metric between the time base space and the fiber ball risk state space is distorted to obtain the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area. Based on the distorted hyperbolic spatiotemporal risk manifold space of the target monitoring area, a tangent space is constructed at a preset base point. The temporal risk vector set is projected onto the tangent space through a linear transformation to obtain the tangent space risk vector set.

10. A system for constructing and dynamically warning of a microseismic risk field, used to implement the method for constructing and dynamically warning of a microseismic risk field according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect microseismic data of the target monitoring area and write it into the storage area. Based on the microseismic data of the target monitoring area, a microseismic risk field is constructed to obtain a set of microseismic risk field sequence signals. The early warning segmentation module, based on the microseismic risk field sequence signal set, divides the early warning sections, constructs a spatiotemporal multi-relation tensor map of the target monitoring area, and obtains a time-series risk representation set; The hyperbolic embedding module constructs a spatiotemporal risk fiber bundle of the target monitoring area based on the temporal risk representation set, and performs geometric embedding in the distorted hyperbolic spatiotemporal risk manifold space to obtain the hyperbolic evolution representation set; The early warning generation module calculates the hyperbolic risk distance for each early warning segment based on the hyperbolic evolution representation set, and classifies the hyperbolic risk distance according to the multi-level dynamic early warning judgment rules to provide dynamic early warning for the risk.

Images