Intelligent Prediction Method and System for Outstanding Hazards Based on Spatiotemporal Evolution Analysis

Abstract

This invention discloses an intelligent prediction method and system for outburst hazards based on spatiotemporal evolution analysis, belonging to the field of artificial intelligence technology. The method includes: acquiring multi-source heterogeneous data from a mine and preprocessing the data to obtain rasterized spatiotemporal grid data; constructing a mine spatial topology map and extracting persistent homology features; using a topology set network and a temporal convolutional network to spatially and temporally encode the rasterized spatiotemporal grid data, respectively, to obtain a spatial topology-encoded feature matrix and a temporal-encoded feature matrix; weighted fusion of the spatial topology-encoded feature matrix and the temporal-encoded feature matrix to generate a spatiotemporal joint feature matrix; constructing and using a dual-structure spatiotemporal network model to perform evolutionary analysis on the spatiotemporal joint feature matrix to predict the probability distribution of outburst hazards; and performing risk classification and visualization based on the prediction results. This achieves online dynamic identification and high-precision prediction of mine-scale outburst hazards by integrating spatial topology and physical law constraints.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a method and system for intelligent prediction of prominent hazards based on spatiotemporal evolution analysis. Background Technology

[0002] During mining operations, gas outbursts and rock bursts are characterized by their suddenness and wide-ranging destructive impact, seriously threatening mine safety. With the continuous expansion of mining depth and scale, existing outburst hazard prediction technologies have the following four significant shortcomings:

[0003] Insufficient spatial topological representation capability: The spatial structure of mines has complex non-Euclidean characteristics. Traditional grid modeling methods lose the connectivity of mine tunnels and the characteristics of geological cavities, while existing graph modeling methods can only capture single-dimensional relationships and cannot fully represent multi-scale topological patterns.

[0004] Quantization of unordered topological data is difficult: the persistent graph generated by persistent homology analysis is an unordered set of points, and traditional vectorization methods cannot effectively preserve the stability information of topological features, which limits the full utilization of topological information in deep learning;

[0005] Lack of spatiotemporal coupled evolution modeling: Existing methods adopt a spatiotemporal feature stacking architecture, which severs the inherent coupling relationship between spatial topology and temporal dynamics, resulting in poor anti-interference ability in short-term prediction and difficulty in accurately tracking risk evolution trends in medium-term prediction;

[0006] Insufficient physical consistency and feature adaptability: The model lacks multi-physics prior knowledge embedding, and the prediction results are prone to violating physical laws; at the same time, the dimensional mismatch and semantic differences during spatiotemporal feature fusion lead to prominent feature confusion problems.

[0007] Existing technologies have systematic deficiencies in spatial feature representation, topological information utilization, spatiotemporal coupling modeling, and physical constraints, making it difficult to meet the urgent needs of mine safety production for high precision, long-term effectiveness, and interpretable prediction. Therefore, it is currently necessary to break through the technical bottlenecks in four aspects: spatial topological representation, disordered data quantification, spatiotemporal coupling modeling, and physical constraint fusion. Summary of the Invention

[0008] This invention aims to address the core problems in existing mine outburst risk prediction technologies, such as one-sided spatial topological representation, difficulty in quantifying disordered topological data, lack of spatiotemporal coupling modeling, and insufficient physical consistency. It provides an intelligent prediction method and system for outburst risks based on spatiotemporal evolution analysis. Through multi-source data preprocessing, spatiotemporal coding fusion, evolutionary analysis under physical constraints, and decision support integration, it achieves online dynamic identification and high-precision short / medium-term prediction of outburst risks at the mine scale, and can output risk visualization results and disposal suggestions.

[0009] In a first aspect, embodiments of the present invention provide a method for intelligent prediction of prominent hazards based on spatiotemporal evolution analysis, comprising:

[0010] Acquire multi-source heterogeneous data from the mine, and perform unified temporal sequencing, missing value imputation, and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data;

[0011] A spatial topology map is constructed based on the rasterized spatiotemporal grid data, and topological features are extracted for spatial encoding to obtain a spatial encoding feature matrix; a temporal convolutional network is used to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix.

[0012] By employing a cross-modal self-attention fusion mechanism, the spatial encoding feature matrix and the temporal encoding feature matrix are weighted and fused to generate a spatiotemporal joint feature matrix;

[0013] Based on the spatiotemporal joint feature matrix, a dual-structure spatiotemporal network model coupling physical graph and functional graph is constructed, and the spatiotemporal evolution analysis is performed by inputting the spatiotemporal joint feature matrix to predict the probability distribution of prominent hazards in multiple future time steps.

[0014] Based on the predicted probability distribution of salient hazards over multiple future time steps, risk classification and visualization are performed, and automated response recommendations are generated.

[0015] In a preferred embodiment, the multi-source heterogeneous data is subjected to unified temporal sequencing, missing value imputation, and spatial gridding to obtain rasterized spatiotemporal grid data, including:

[0016] The multi-source heterogeneous data includes static geological data, dynamic monitoring data, and imagery and topological data;

[0017] The static geological data is expanded to a time-series dimension, the dynamic monitoring data is resampled at a preset sampling interval, and the images and topological data are converted into a frame sequence time-series format to obtain a unified time-series dataset.

[0018] A spatiotemporal interpolation method based on an attention mechanism is used to calculate weights based on the inherent spatiotemporal correlation of the data, and to fill in the missing data points in the unified time series dataset.

[0019] A spatial reference grid is established under the local coordinate system of the mine, and the completed unified time series dataset is mapped to the grid cells to obtain the rasterized spatiotemporal grid data.

[0020] In a preferred embodiment, a spatial topology map is constructed based on the rasterized spatiotemporal grid data, and topological features are extracted for spatial encoding to obtain the spatial encoding feature matrix, including:

[0021] Topological map elements are extracted from the rasterized spatiotemporal grid data, wherein the topological map elements include shaft and tunnel nodes, geological units, shaft and tunnel connectivity relationships between grid units, and adjacent relationships between geological units;

[0022] A mine spatial topology map is constructed based on the aforementioned topology map elements, wherein the mine tunnel nodes and geological units are used as vertices of the topology map, and the mine tunnel connectivity and geological unit adjacency relationships are used as edges of the topology map.

[0023] Persistent homology analysis is performed on the spatial topology graph to generate a persistent graph that quantifies the stability of topological features;

[0024] The persistent graph is vectorized and encoded using the TS-Layer of the topology set network to generate spatial feature vectors.

[0025] The spatial feature vectors are mapped to grid cells, and the feature weights of key topological regions are enhanced through a residual attention mechanism, outputting a spatially encoded feature matrix consistent with the dimensions of the rasterized spatiotemporal grid data.

[0026] In a preferred embodiment, a temporal convolutional network is used to perform temporal encoding on the spatiotemporal grid data to obtain a temporal encoding feature matrix, including:

[0027] Extract multi-channel time-series monitoring data of each grid cell from the rasterized spatiotemporal grid data;

[0028] A temporal convolutional network is used to extract the corresponding temporal dynamic features from the multi-channel temporal monitoring data of each grid cell;

[0029] The time-dimensional global average pooling is performed on the time dynamic features corresponding to each grid cell to obtain a time-encoded feature matrix that corresponds one-to-one with each grid cell.

[0030] In a preferred embodiment, the spatial coding feature matrix and the temporal coding feature matrix are weighted and fused through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix, including:

[0031] The spatial coding feature matrix is ​​mapped to a grid to generate a grid-dimensional spatial feature matrix, wherein the grid-dimensional spatial feature matrix has the same dimension as the temporal coding feature matrix.

[0032] Calculate the cross-modal attention weights between the grid dimension spatial features in the grid dimension spatial feature matrix and the temporal features in the temporal coding feature matrix;

[0033] Based on the cross-modal attention weights, the spatial and temporal features of the same grid cell are weighted and fused to generate a spatiotemporal joint feature matrix that covers all grid cells and integrates spatiotemporal information.

[0034] As a preferred embodiment, a dual-structure spatiotemporal network model coupling a physical graph and a functional graph is constructed based on the spatiotemporal joint feature matrix, and the spatiotemporal evolution analysis is performed by inputting the spatiotemporal joint feature matrix to predict the probability distribution of salient hazards at multiple future time steps, including:

[0035] Based on the spatiotemporal joint feature matrix, a dual-structure spatiotemporal network model coupling the physical graph and the functional graph is constructed;

[0036] The physical graph and the functional graph are connected through the same grid cells to generate a spatiotemporal adjacency matrix;

[0037] Learnable spatial and temporal embeddings are added to the spatiotemporal adjacency matrix, and the matrix is ​​summed with the spatiotemporal joint feature matrix through a broadcast mechanism to obtain an enhanced spatiotemporal joint feature matrix.

[0038] Chebyshev graph convolution is used to extract spatiotemporal correlation features across time steps and grid cells from the enhanced spatiotemporal joint feature matrix. The extracted spatiotemporal correlation features are then concatenated using max pooling and average pooling to obtain graph-level features that fuse global information.

[0039] The graph-level features are input into the dual-structure spatiotemporal network model, which outputs the probability distribution of prominent hazards and their corresponding real risk labels for multiple future time steps.

[0040] In a preferred embodiment, the method further includes:

[0041] The classification loss is calculated based on the predicted probability distribution of salient hazards at multiple future time steps and their corresponding true risk labels.

[0042] During the training process of the dual-structure spatiotemporal network model, a physical constraint loss function is introduced, which includes gas diffusion constraint loss and pressure conservation constraint loss.

[0043] The classification loss and the physical constraint loss are fused according to preset weights to guide the training of the dual-structure spatiotemporal network model.

[0044] In a preferred embodiment, the functional diagram is constructed through the following steps:

[0045] Calculate the cosine similarity of each grid cell in the gas concentration and stress value monitoring data;

[0046] Grid cell pairs with cosine similarity less than or equal to a preset threshold are defined as functional neighbor relationships;

[0047] Construct a functional graph adjacency matrix based on the aforementioned functional neighbor relationships.

[0048] As a preferred implementation, the step of performing risk classification and visualization based on the predicted probability distribution of prominent hazards over multiple future time steps, and generating automated response recommendations, includes:

[0049] Based on the predicted probability distribution of prominent hazards over multiple future time steps and the preset risk probability threshold, each grid cell is divided into different risk levels;

[0050] Based on the spatial geographic information of the mine, the risk level of each grid unit is dynamically presented, and risk visualization results covering the entire prediction cycle are generated.

[0051] The system automatically matches the preset emergency response plan for mine outburst disasters with the risk level and generates targeted graded disposal suggestions that are adapted to the risk level.

[0052] The risk visualization results and targeted graded handling suggestions are pushed to the mine monitoring terminal, and confirmation or correction feedback is received from on-site personnel. The prediction model is iteratively optimized based on the feedback information.

[0053] Secondly, embodiments of the present invention also provide a prominent hazard intelligent prediction system based on spatiotemporal evolution analysis, comprising:

[0054] The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data from the mine, and to perform unified temporal sequencing, missing value imputation and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data.

[0055] The data encoding module is used to construct a spatial topology map based on the rasterized spatiotemporal grid data, extract topological features for spatial encoding, and obtain a spatial encoding feature matrix; and to use a temporal convolutional network to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix.

[0056] The cross-modal self-attention fusion module is used to weightedly fuse the spatial encoding feature matrix and the temporal encoding feature matrix through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix;

[0057] The hazard probability distribution prediction module is used to construct a dual-structure spatiotemporal network model that couples a physical graph and a functional graph based on the spatiotemporal joint feature matrix, and input the spatiotemporal joint feature matrix to perform spatiotemporal evolution analysis to predict the salient hazard probability distribution at multiple future time steps.

[0058] The risk grading and visualization module is used to grade and visualize risks based on the predicted probability distribution of salient hazards over multiple future time steps, and generate automated response recommendations.

[0059] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0060] One or more processors;

[0061] Storage device for storing one or more programs;

[0062] When the one or more programs are executed by the one or more processors, the one or more processors implement the intelligent prediction method for prominent hazards based on spatiotemporal evolution analysis as described in any embodiment of the present invention.

[0063] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent prediction method for prominent hazards based on spatiotemporal evolution analysis as described in any embodiment of the present invention.

[0064] The present invention achieves the following beneficial effects:

[0065] (1) In view of the non-Euclidean characteristics of the spatial structure of mines, this invention proposes a spatial coding method based on the TSNet (TopologicalSet Network) framework. By converting tunnel nodes and geological units into vertices of the topological graph and tunnel connectivity and geological unit adjacency into edges, a mine spatial topological graph is constructed. Then, topological features are extracted through persistent homology analysis, which solves the problem of losing non-Euclidean information in traditional modeling and provides an accurate spatial basis for subsequent risk positioning.

[0066] (2) In view of the problem of disorder in persistent graphs, this invention proposes a vectorization scheme based on rational hat function. The rational hat function is used to generate a probability distribution of the birth and death coordinates of each point in the persistent graph. The unordered point set is converted into a fixed-length vector through hybrid entropy aggregation, which preserves the stability information of topological features, provides structured input for deep learning models, and avoids the waste of topological information.

[0067] (3) In response to the problem of spatiotemporal feature fragmentation, this invention proposes a dual-structure spatiotemporal network model. By constructing a dual-structure graph to capture multi-dimensional spatial correlations, and then constructing a time sequence graph and graph convolution, the coupled modeling of spatial topology guiding time dynamics and time dynamics updating spatial risks is realized. Finally, the probability prediction distribution of prominent risks in multiple future time steps is output, which improves the accuracy and stability of predictions at different time points.

[0068] (4) In view of the problem that single graph modeling cannot reflect risk-driven association, this invention proposes a functional graph construction method based on the feature similarity of grid cells. Based on the cosine distance of monitoring data such as gas concentration and stress value, grid cells with similarity less than a preset threshold are set as neighbors to construct a functional graph, accurately capture grid cells that are not physically adjacent but are risk-related, supplement the physical graph that can only reflect spatial location association, and improve the representation of spatial risk propagation path.

[0069] (5) To address the issues of mismatch between spatiotemporal feature dimensions and lack of physical consistency, this invention proposes a broadcast embedding mechanism. By summing spatiotemporal joint features through the broadcast mechanism, the semantic and dimensional adaptation of spatial and temporal features is solved. In addition, physical constraints are introduced to ensure that the prediction results conform to the laws of multi-physics fields in the mine, thereby improving the reliability and interpretability of the model output. Ultimately, this invention achieves online dynamic identification and short / medium-term high-precision prediction of mine-scale outburst hazards, providing full-process support for on-site personnel from risk positioning to evolution trend to disposal suggestions, thereby reducing the probability of outburst disasters. Attached Figure Description

[0070] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0071] Figure 1 This is a flowchart of the intelligent prediction method for prominent hazards based on spatiotemporal evolution analysis provided in an embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of the overall framework of the intelligent prediction method for prominent hazards based on spatiotemporal evolution analysis provided in this embodiment of the invention.

[0073] Figure 3 This is a flowchart of data encoding and feature fusion provided in an embodiment of the present invention;

[0074] Figure 4 This is a flowchart of the BS-STN model evolution analysis provided in the embodiments of the present invention;

[0075] Figure 5 This is a schematic diagram of the structure of the intelligent prediction system for prominent hazards based on spatiotemporal evolution analysis provided in an embodiment of the present invention;

[0076] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0078] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0079] Example 1

[0080] like Figure 1 The diagram shows a flowchart of a prominent hazard intelligent prediction method 100 based on spatiotemporal evolution analysis provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0081] S110. Obtain multi-source heterogeneous data from the mine, and perform unified temporal sequencing, missing value interpolation, and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data.

[0082] The aforementioned multi-source heterogeneous data is collected through sensor networks, geological information systems (GIS), and drone inspection equipment deployed in the mine, and specifically includes the following three types of data:

[0083] Static geological data Static geological layers including fault distribution, stratigraphic dip angle, and coal seam thickness;

[0084] Dynamic monitoring data Gas concentration Gas emission volume Number of micro-seismic events Stress value ,temperature Wind pressure Equal time series data;

[0085] Imagery and topology data Fiber optic monitoring images, UAV tunnel images, and tunnel topology data.

[0086] Furthermore, in order to eliminate data heterogeneity and achieve unified representation, the aforementioned multi-source heterogeneous data needs to be preprocessed, specifically including unified temporal sequencing, missing value imputation, and spatial gridding of the multi-source heterogeneous data:

[0087] First, the above static geological data are respectively... Expanding to a time-series dimension, the aforementioned dynamic monitoring data Resample at a preset sampling interval (e.g., 10 minutes) and combine the above images with topological data. Converted to frame sequence time-series format, the final result is a unified time-series dataset. ;

[0088] Secondly, a spatiotemporal interpolation method based on an attention mechanism is adopted to calculate weights according to the spatiotemporal correlation of the data, and then applied to the aforementioned unified time-series dataset. Complete the missing data points in the data;

[0089] Finally, a spatial reference grid (using the mine's local coordinate system) is established based on the working face / mining area, and all data is mapped to the grid cells. ( (using grid coordinates) to obtain rasterized spatiotemporal grid data ( The total number of time steps. For the number of grid cells, Spatial alignment of data is achieved by using the number of features for each grid.

[0090] Among them, the unified time series dataset To fill in missing data points, use the following formula:

[0091] ;

[0092] in, : Missing data point values ​​completed by attention mechanism spatiotemporal interpolation, with scalar dimension, corresponding to the _th The time step, the first The completion results of each monitoring parameter (such as the gas concentration at the 10th time step) should be used to ensure that the completed values ​​conform to the spatiotemporal evolution trend of the data and avoid contradicting the patterns of the preceding and following data. : Two-dimensional summation operator, where Traverse all time steps From 1 to Traverse all monitoring parameters ( From 1 to (This is used to calculate the effective data point pairs with missing points across the entire spatiotemporal range.) The weighted sum of contributions; Total number of time steps, i.e., the total time series length of mine monitoring data (e.g., collecting one month's worth of data at 10-minute sampling intervals). ), defines the calculation range of the time dimension; The total number of monitoring parameters, i.e., the number of monitoring indicators deployed in the mine (such as gas concentration, stress value, temperature, etc.). ), define the calculation range of the parameter dimension; Attention weights, with scalar dimension, corresponding to the [number]th [dimension]. The first time step The parameter and the first The first time step The relevance weights of the parameters, with values ​​ranging from [value range missing]. The values ​​are calculated based on the spatiotemporal correlation of the data (such as continuity in time and correlation between parameters). The higher the correlation, the greater the weight, ensuring that the completion values ​​are preferentially referenced from valid data that are consistent with the patterns of the missing points. : No. The time step, the first The valid data point values ​​of each monitoring parameter are scalar in dimension, used to fill in missing points. The basic data source is used for weighted calculation only if the data point is complete; : The time step index of the missing data point, corresponding to the specific time position where the data is missing (e.g., (Indicates the 50th time step), used to locate the position of the missing point in the time series; : Index of monitoring parameters for missing data points, corresponding to the specific monitoring indicator where the data is missing (e.g., ... This represents the second monitoring parameter, namely gas concentration, used to locate the missing point in the parameter dimension; : The time step index of the valid data points, traversing all non-missing time steps to identify missing points. Provides reference data in the time dimension; : Index of monitoring parameters for valid data points; iterate through all non-missing monitoring parameters to identify missing points. Provide reference data for parameter dimensions (such as using valid data of other relevant parameters at the same time step to assist in completion).

[0093] S120. Construct a spatial topology map based on the rasterized spatiotemporal grid data, extract topological features for spatial encoding, and obtain a spatial encoding feature matrix; use a temporal convolutional network to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix.

[0094] In some embodiments, a spatial topology map is constructed based on rasterized spatiotemporal grid data, and topological features are extracted for spatial encoding to obtain a spatial encoding feature matrix. This process specifically includes the following steps:

[0095] Topological map elements are extracted from the rasterized spatiotemporal grid data, wherein the topological map elements include shaft and tunnel nodes, geological units, shaft and tunnel connectivity relationships between grid units, and adjacent relationships between geological units;

[0096] Construct a mine space topology map based on the aforementioned topology map elements. ,in For the vertex set (number of vertexes) ), The topology graph is defined as an edge set, with the tunnel nodes and geological units as vertices and the tunnel connectivity and geological unit adjacency relationships as edges.

[0097] The mine spatial topology map Perform persistent homology analysis to generate a persistent graph that reflects topological characteristics. ;in The moment when topological features are "born" For the moment of "extinction", Indicates the stability of topological features;

[0098] The persistent graph is obtained through the TS-Layer (containing 100 TS-Blocks) of the topology network. Perform vectorization encoding to generate fixed-length spatial feature vectors;

[0099] The spatial feature vectors are mapped to grid cells, and the feature weights of key topological regions are enhanced through a residual attention mechanism, outputting a spatially encoded feature matrix consistent with the dimensions of the rasterized spatiotemporal grid data.

[0100] The persistent graph is constructed using a TS-Layer (containing 100 TS-Blocks) of the topology set network. During vectorization encoding, each TS-Block uses a rational hat function to generate a probability distribution, which quantifies the degree of matching between a single point in the persistent graph and the TS-Block probability distribution. The output range is... The larger the value, the more closely the point matches the characteristic distribution of the current TS-Block. The formula is:

[0101] ;

[0102] in, : Probability values ​​generated based on the rational hat function; The mean parameter vector of TS-Block is derived from... and The components are used to define the probability distribution center of the rational hat function and determine the focus of TS-Block on the features of specific regions in the persistent graph; Mean parameter vector The first component corresponds to the distribution center in the dimension of the birth time in the persistent graph, and is used to determine the feature capture range of TS-Block in the time dimension. Mean parameter vector The second component corresponds to the distribution center of the extinction moment dimension in the persistent graph, and is used to determine the capture range of TS-Block in the dimension of topological feature stability. : A constant parameter used to control the probability distribution width of the rational hat function, adjusting the range of TS-Block's perception of feature points in the persistent graph. It is typically preset based on the topological complexity of the mine space (e.g., a value of...). ); : Norm parameter, used to specify computational paradigms (such as) It is an L1 norm. (Using the L2 norm) determines the weighting method for distance calculation, adapting to the distance measurement needs of different mine topological features; The coordinates of the birth time of a single topological feature point in the persistent graph reflect the scale or time at which the topological feature (such as a well-connected component or a geological unit cavity) begins to appear, and are one of the core attributes of the topological feature. The coordinates of the disappearance time of a single topological feature point in the persistent graph, reflecting the scale or time of the disappearance of that topological feature. The larger the value, the more stable the topological feature; Feature points in persistent graphs With TS-Block distribution center The Euclidean distance (default L2 norm) is used to measure the spatial distance between feature points and the distribution center. The smaller the distance, the higher the matching degree. Feature points in persistent graphs With TS-Block distribution center The q-norm distance, according to The distance calculation method is adjusted to adapt to the different measurement needs of different topological features in the mine; The spatial encoding feature matrix output by TS-Layer has dimensions of [missing information]. ( (This represents the number of vertices in the mine spatial topology map, with 100 representing the number of TS-Blocks). Each row corresponds to the spatial characteristics of a topology vertex, and each column corresponds to the output of a TS-Block, comprehensively reflecting the global characteristics of the mine spatial topology. The number of vertices in the mine's spatial topology map, with each vertex corresponding to a key spatial unit within the mine. The value is determined by the actual size of the mine (e.g., small and medium-sized mines). large mines ).

[0103] In some embodiments, a residual attention mechanism (RA-Layer) is used to enhance the feature weights of key topological regions, highlighting the feature contributions of key topological regions (such as areas near faults or high-gas regions), and outputting a spatially encoded feature matrix consistent with the dimension of the rasterized spatiotemporal grid data. The following formula is used:

[0104] ;

[0105] Among them, Attention The self-attention calculation function generates attention weights by calculating the relevance between the query (Q), key (K), and value (V), and then applies these weights to the value (K). Weighted summation, outputting a feature matrix focusing on key regions, with dimensions equal to... Consistent; Query matrix, derived from original spatial features Generated through a multilayer perceptron (MLP) mapping, used to calculate the correlation with the key matrix (K), with dimensions of [missing information]. ( This represents the number of vertices in the mine's spatial topology. (For attention head dimension) : Key matrix, derived from the original spatial features Generated via a multilayer perceptron (MLP) mapping, and matched with the query matrix (Q) to compute attention weights, with dimensions consistent with Q. ); Value matrix, derived from the original spatial features The feature is generated through a multilayer perceptron (MLP) mapping, weighted according to attention weights, and finally outputs an attention-enhanced feature with a dimension of [dimensionality missing]. MLP: Multilayer Perceptron, consisting of an input layer, hidden layers (usually with ReLU activation function), and an output layer, used to... The feature dimension is mapped from 100 to Generate an adaptive attention computation Matrix; Softmax The softmax activation function is used to evaluate the attention score matrix of the input. Normalize the output so that the weights of each row sum to 1, and the output range is within... The attention weight matrix between them has dimensions of ; The query matrix (Q) is multiplied by the transpose of the key matrix (K). The relevance of each element in Q and K is calculated to generate the original attention score matrix, with dimensions [missing information]. ; Attention head dimension The square root of the original attention score is used to calculate the original attention score. Scaling is performed to avoid [damage / losses]. An excessively large value leads to an excessively high score, which in turn affects the gradient stability of the Softmax function; Attention head dimension: The feature dimension of each attention head in a self-attention mechanism; a hyperparameter, typically based on the original spatial features. Feature dimension adjustment (e.g., taking a value) ),Decide The number of columns in a matrix.

[0106] In some embodiments, a temporal convolutional network is used to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix, specifically including the following steps:

[0107] Extract multi-channel time-series monitoring data (gas concentration, stress, microseismic activity, etc.) from each grid cell of the rasterized spatiotemporal grid data.

[0108] Temporal convolutional networks are used to extract corresponding temporal dynamic features from the multi-channel temporal monitoring data of each grid cell. (64 represents the number of TCN output channels);

[0109] The time dynamic features corresponding to each grid cell Perform global average pooling along the time dimension to obtain the time-encoded feature matrix corresponding to each grid cell. 64 represents the number of TCN output channels.

[0110] Preferably, the Temporal Convolutional Network (TCN) employs 3 layers of one-dimensional (1D) causal convolutions with a kernel size of 3, dilation coefficients of 1, 2, and 4 (to achieve multi-scale temporal receptive fields), and the activation function is ReLU.

[0111] S130. The spatial coding feature matrix and the temporal coding feature matrix are weighted and fused through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix.

[0112] In some embodiments, the spatial encoding feature matrix is ​​used. With time-coded feature matrix The steps for fusion via cross-modal self-attention layers are as follows:

[0113] First, encode the feature matrix of this space. Perform grid-based mapping to generate a grid-dimensional spatial feature matrix. The grid dimension spatial feature matrix With the time-coded feature matrix The dimensions are consistent;

[0114] Calculate the feature matrix of the grid dimension space Spatial features and temporal coding feature matrix Cross-modal attention weights between temporal features ;

[0115] Based on the cross-modal attention weights The spatial and temporal features of the same grid cell are weighted and fused to generate a spatiotemporal joint feature matrix that covers all grid cells and integrates spatiotemporal information. .

[0116] Among them, the aforementioned cross-modal attention weights The following formula is used for calculation:

[0117] ;

[0118] Spatial and temporal characteristics in grid cells The mutual attention weight matrix at point , with weight values ​​ranging from The larger the value, the more significant the contribution of the spatial feature at that location to the joint feature; Time-coded feature matrix The transpose of the matrix has dimensions of . (Only the feature dimension is transposed; the grid dimension remains unchanged), used to connect with the feature matrix in the grid dimension space. Matrix multiplication is used to calculate cross-modal correlation; the higher the score, the stronger the correlation between the spatial and temporal characteristics of the corresponding grid cell. Cross-modal feature dimensions The square root is used to scale the original attention score, avoiding bias. Excessive values ​​can cause scores to exceed a reasonable range, affecting the gradient stability of the Softmax function and ensuring the reliability of attention weight calculation. Cross-modal feature dimension, a hyperparameter fixed at 100, corresponding to the feature matrix in the grid dimension space. The feature dimension (100) is used to unify the computational dimensions of spatial and temporal features, ensuring the validity of cross-modal attention score calculation;

[0119] Preferably, through attention weights The spatial and temporal features of the same grid cell are weighted and fused to obtain a spatiotemporal joint feature matrix. The following formula is used:

[0120] ;

[0121] Dimensions are ( (where 100 represents the number of grid cells, 100 represents the spatial feature dimension, and 64 represents the temporal feature dimension). Weighting coefficients of time features, dimensions and Consistency, used for time-encoded feature matrices Weighting is performed to balance the contribution ratios of spatial and temporal features; : Time-encoded feature matrix, dimension is It records the temporal dynamic attributes of grid cells and is the source of temporal information in the joint features.

[0122] S140. Based on the spatiotemporal joint feature matrix, construct a dual-structure spatiotemporal network model that couples the physical graph and the functional graph, and input the spatiotemporal joint feature matrix to perform spatiotemporal evolution analysis to predict the probability distribution of prominent hazards at multiple future time steps.

[0123] The dual-structure spatiotemporal network model (BS-STN (Bi-Structural Spatial-Temporal Network)) adopts an encoder-decoder architecture. The encoder contains three layers of Chebyshev graph convolutions, with output dimensions of 128, 64, and 32 for each layer. The decoder contains two fully connected layers with 64 and 32 neurons respectively. The activation function is ReLU, and the dropout rate is set to 0.2.

[0124] In some embodiments, the purpose of this step is to use a pre-built BS-STN model to analyze the spatiotemporal joint feature matrix. Perform evolutionary analysis to predict the future. The salient risk indicators (probability distribution) for each time step (short-term: 1-6 hours, medium-term: 1-7 days) are determined as follows:

[0125] Construct a dual-structure graph, including a physical graph based on spatial distance and a functional graph based on feature similarity;

[0126] Connect the dual-structure graphs through the same grid cells to form a time sequence graph, and generate a spatiotemporal adjacency matrix;

[0127] Learnable spatial and temporal embeddings are added to the spatiotemporal adjacency matrix, and the matrix is ​​summed with the spatiotemporal joint feature matrix through a broadcast mechanism to obtain an enhanced spatiotemporal joint feature matrix.

[0128] Chebyshev graph convolution is used to extract spatiotemporal correlation features across time steps and grid cells from the enhanced spatiotemporal joint feature matrix. The extracted spatiotemporal correlation features are then concatenated using max pooling and average pooling to obtain graph-level features that fuse global information.

[0129] The graph-level features are input into the fully connected layer and the Softmax activation function of the BS-STN model, and the output is the probability distribution of salient risks and their corresponding true risk labels for multiple future time steps.

[0130] Preferably, constructing a dual-structure diagram includes constructing a physical diagram and constructing a functional diagram, wherein:

[0131] Physical graph construction: Construct a physical graph based on the physical location of grid cells, with adjacent grid cells (spatial distance) ( ) are set as neighbors, and the adjacency matrix of the physical graph is obtained. This reflects the spatial physical relationships within the mine.

[0132] Functional graph construction: A functional graph is constructed based on the feature similarity of mesh cells (such as gas concentration and cosine distance of stress values). If the similarity... If the threshold is used, then it is set as a neighbor, resulting in the functional graph adjacency matrix. This reflects the functional relationships between grid cells.

[0133] Preferably, the dual-structure graphs are connected using the same grid cells to form a time series graph, generating a spatiotemporal adjacency matrix. Specifically, this includes connecting the dual-structure graphs of the current time step and the previous two time steps using the same grid cells to form a time series graph, and generating a spatiotemporal adjacency matrix. ,in For a single-time-step dual-structure graph adjacency matrix ( or ), This is a temporal connection matrix (edges exist only between identical grid cells, and diagonal elements are 1).

[0134] Next, learnable spatial and temporal embeddings are added to the spatiotemporal adjacency matrix, and then summed with the spatiotemporal joint feature matrix through a broadcast mechanism to obtain an enhanced spatiotemporal joint feature matrix, specifically including:

[0135] Broadcast embedding: Adding learnable spatial embeddings With time embedding Through broadcasting mechanism and Summation is performed using the following formula:

[0136] ;

[0137] in, Enhanced spatiotemporal joint feature matrix with added spatiotemporal embedding, dimension and spatiotemporal joint feature matrix Consistent For time steps, (where 164 is the feature dimension and 164 is the grid number) integrates the original spatiotemporal features with learnable location information, avoiding the confusion of spatiotemporal information in graph convolution; : Learnable spatial embedding vectors, with dimension (where 164 is the feature dimension) is used to distinguish the spatial location attributes of different grid cells. It is extended to 3 time steps through a broadcast mechanism to ensure that the spatial location information of each grid cell is captured by graph convolution. Learnable temporal embedding vectors, with dimensions of (3 represents the number of time steps, and 164 represents the feature dimension), used to distinguish the temporal attributes of different time steps. It is extended to all grid cells through a broadcast mechanism to ensure that the temporal information of each time step is captured by graph convolution.

[0138] Next, Chebyshev graph convolution is used to extract spatiotemporal correlation features across time steps and across grid cells from the enhanced spatiotemporal joint feature matrix. The following formula is used:

[0139] ;

[0140] in, The spatiotemporal correlation matrix output by graph convolution has dimensions of . It integrates spatiotemporal correlation features across time steps and grid cells, providing highly representative feature inputs for subsequent prominent hazard prediction. The summation operator, for... From 0 to ( The order of the Chebyshev polynomial is 2), which is the default. The summation of the results of each term in 2) integrates the features captured by polynomials of different orders, thereby improving the multi-scale feature extraction capability of graph convolution. The convolution coefficients of the Chebyshev graph convolution, with dimension 1. (164 is the feature dimension), which can be learned through model training and used to weight the features of polynomial outputs of different orders, adjusting the contribution ratio of each order of features; Chebyshev polynomials in normalized Laplace matrix The calculation results are in dimension 1. It is used to transform graph convolution into polynomial computation, reducing the computational complexity of traditional graph convolution and adapting to large-scale grid data in mines; : Normalized Laplace matrix, dimension , from the original Laplace matrix The standardization yields the following formula: This is used to eliminate the influence of graph node degree differences on the convolution results and ensure the stability of feature propagation; The original Laplace matrix, with dimension 1. , by the spatiotemporal adjacency matrix Calculated for The degree matrix (the matrix of the time series graph) is used to describe the topological structure of the time series graph and is the core input of graph convolution; : Original Laplace matrix The largest eigenvalue, used for... Normalization is performed to avoid polynomial calculation overflow due to excessively large eigenvalues. This is usually achieved by pre-calculating eigenvalues ​​through eigenvalue decomposition. Identity matrix, dimensions and Consistent, with diagonal elements set to 1 and the rest to 0, used for participation. The calculation ensures that the normalized Laplace matrix meets the numerical stability requirements.

[0141] Next, the extracted spatiotemporal correlation feature matrix Graph-level features incorporating global information are obtained by concatenating max pooling and average pooling. Finally, the graph-level features Input a fully connected layer and a Softmax activation function, output the future The probability distribution of salient hazards for each grid cell at each time step (C=3 represents low risk, medium risk, and high risk, respectively), using the following formula:

[0142] ;

[0143] in, This is the weight matrix. For bias, To highlight the probability distribution of risks.

[0144] In addition, to ensure that the prediction results conform to the multiphysics laws of the mine, this step also includes:

[0145] Based on the predicted probability distribution of prominent risks Calculate the classification loss based on the corresponding real risk label;

[0146] During the training of the BS-STN model, a physical constraint loss function is introduced, which includes gas diffusion constraint loss and pressure conservation constraint loss.

[0147] The classification loss and the physical constraint loss are fused together with preset weights to guide the training of the BS-STN model.

[0148] The classification loss and the physical constraint loss are fused according to preset weights to obtain the total loss function of the BS-STN model, specifically using the following formula:

[0149] ;

[0150] in, The total loss function of the BS-STN model guides parameter updates during model training, ensuring that the prediction results are both consistent with the data labels and conform to the multi-physics field laws of the mine. Physical constraint weights are determined through grid search, with a search range of [0.3, 0.5]. They are used to adjust the contribution ratio of physical constraint loss to the total loss. The larger the value, the more the model training focuses on satisfying physical laws, balancing the optimization goal of data fitting and physical rationality. The classification loss, using the cross-entropy loss function, measures the difference between the model's predicted salient risk probability distribution and the actual risk label. It is the core loss term for the model to learn the association between data and risk. The total physical constraint loss is calculated by summing and integrating the gas diffusion constraint loss and the pressure conservation constraint loss to ensure that the model prediction results conform to the physical laws of mine gas diffusion and rock pressure, and to avoid outputting results that contradict reality.

[0151] Among them, classification loss The calculation formula is:

[0152] ;

[0153] in, Prediction time steps: The number of future time steps the model needs to predict (short-term forecasting) Corresponding to 6 hours; medium-term forecast (corresponding to 7 days), which determines the time frame for loss calculation; The number of horizontal grids in the mine spatial reference grid is determined by the area of ​​the mine mining area and the grid resolution (e.g., if the mining area is 500m horizontally and the resolution is 25m, then...). ), is the calculation range parameter for spatial dimension; The number of vertical grid cells in the mine spatial reference grid, and Together they form a spatial grid (e.g., a mining area with a vertical length of 300m and a resolution of 25m, then...). ), defines the computational boundary of spatial dimensions; The total number of calculated samples, i.e. the total number of combinations of prediction time steps and grid cells, is used to average the sum of cross-entropy loss results, eliminating the influence of sample size on the magnitude of loss value and making loss values ​​comparable. : The summation operator for the predicted time steps, iterating from 1 to For all time steps, calculate the classification loss contribution at each time step; : A summation operator for the horizontal dimensions of a spatial grid, iterating from 1 to 1. All horizontal grid cells cover all horizontal locations within the mining area; : A summation operator for the vertical dimensions of a spatial grid, iterating from 1 to 1. All vertical grid cells cover all locations along the longitudinal direction of the mining area; Summation operator for risk category dimensions. (Corresponding to low, medium, and high risk categories), iterate through all risk categories to calculate the complete cross-entropy; : for the first Time step, grid Location, Category The true risk label, : This refers to the category of the grid cell at this time step predicted by the BS-STN model. The probability of.

[0154] Total physical constraint loss The calculation formula is:

[0155] ;

[0156] Among them, gas diffusion constraint Based on the gas diffusion equation ( Where is the diffusion coefficient. (for airflow velocity), calculate and predict gas concentration. The difference between the solution and the equation is expressed by the following formula:

[0157] ;

[0158] in, The model predicts the first Time step, grid cell The gas concentration value at the location, in units of , dimension , is the core predictor variable for gas diffusion constraints; The model predicts the first Time step, grid cell The gas concentration value at that location, and The time-varying rate of gas concentration is calculated together to reflect the dynamic evolution trend of gas concentration; Time step interval, i.e., the time difference between two adjacent prediction time steps (fixed at 10 minutes, corresponding to the data sampling interval), is measured in seconds. ), used to convert the time step difference of gas concentration into a time change rate ( (discrete approximation); Approximate rate of change of gas concentration over time, in units of The corresponding gas diffusion equation (Partial derivative of concentration with respect to time), describing the trend of gas concentration change over time; B: Gas diffusion coefficient, the diffusion coefficient of mine gas in coal seams or rock masses, a known physical parameter (value range is...). The value is determined by the geological conditions of the mine (such as coal seam porosity and permeability) and is used to calculate the contribution of the diffusion term.

[0159] The Laplace operator for predicting gas saturation, in units of 1. It describes the second-order change of gas concentration in space (i.e., the gradient of the diffusion gradient), corresponds to the core of the diffusion term in the gas diffusion equation, and reflects the trend of gas diffusion from high concentration to low concentration.

[0160] : Airflow velocity, the average flow velocity of methane gas in the mine, is a known physical parameter (value range is...). The value is determined by the mine ventilation system and gas extraction parameters, and is used to calculate the contribution of the convection term. Gradient operator for predicting gas content, in units of It describes the first-order change of gas concentration in space, corresponds to the core of the convection term in the gas diffusion equation, and reflects the flow effect of gas carrying gas. Gas diffusion term, unit: The gas diffusion rate, which is driven solely by concentration difference, is a core component of the gas diffusion equation. Gas convection term, unit: It describes the gas transport rate driven by airflow and together with the diffusion term, constitutes the physical mechanism of gas density change; The absolute value operator is used to convert the difference between the predicted rate of change and the theoretical value of the equation into a non-negative number, avoiding the cancellation of positive and negative differences and ensuring that the loss accurately reflects the magnitude of the deviation. : A three-dimensional summation operator that iterates through all prediction time steps and grid cells to calculate the total gas diffusion deviation over the entire spatiotemporal range; The averaging operator averages the total temporal and spatial deviations, eliminating the influence of the size of the temporal and spatial range on the loss value and making the loss value horizontally comparable.

[0161] Pressure conservation constraint Based on the law of conservation of rock mass pressure, the predicted stress is calculated. The divergence difference is expressed by the formula:

[0162] ;

[0163] in, The model predicts the first Time step, grid cell The rock mass stress value at the location is expressed in MPa, with dimensions of [missing information]. It encompasses comprehensive stress indicators such as vertical stress and horizontal stress, and is the core predictive variable of the pressure conservation constraint. Predict the divergence of the stress field, in units of This describes the degree of divergence / convergence of stress in space. According to the law of conservation of pressure, the stress field divergence of a stable rock mass should approach 0 (i.e., The larger the deviation, the more it violates the laws of physics.

[0164] S150. Based on the predicted probability distribution of salient hazards over multiple future time steps, perform risk classification and visualization, and generate automated response recommendations.

[0165] The purpose of this step is to transform the prediction results of the spatiotemporal evolution analysis module into actionable decision-making recommendations, enabling risk visualization, tiered early warning, and manual closed-loop confirmation. Specific functions are as follows:

[0166] First, based on a preset risk probability threshold, the probability distribution of salient hazards for each grid cell over multiple future time steps is determined. The risk levels are divided into low, medium, and high, as follows:

[0167] Low risk: Medium risk: High risk: ;

[0168] Secondly, based on the mine GIS map, the risk level of grid units is overlaid (using red / yellow / green three-color marking) to dynamically display the future. It displays the risk evolution trend at each time step, supports multi-view (2D / 3D) viewing, and generates visual reports such as risk heatmaps and evolution curves.

[0169] Next, based on the risk classification results and the mine emergency plan, targeted response suggestions are generated, such as:

[0170] High-risk areas: It is recommended to immediately isolate the working face and activate the backup extraction system (increase the extraction flow rate to...). ), evacuation of personnel;

[0171] Medium-risk areas: It is recommended to increase the frequency of gas monitoring (from 10 minutes / time to 5 minutes / time) and optimize the layout of extraction pipelines;

[0172] Low-risk areas: Maintain routine monitoring and extraction parameters (extraction flow rate) ), and regularly review the geological structure.

[0173] Finally, a manual confirmation process is implemented to close the loop. The system pushes risk warnings and handling suggestions to the mine monitoring center, allowing manual review of the prediction basis (such as key influencing factors: sudden increase in gas concentration near the fault, stress value exceeding the threshold). On-site personnel confirm / adjust suggestions based on the actual situation and provide feedback to the system. The system then updates the parameters of the prediction model (such as adjusting physical constraint weights). This forms a closed loop of prediction, decision-making, feedback, and optimization.

[0174] According to embodiments of the present invention, the following beneficial effects are achieved:

[0175] (1) This invention proposes a spatial coding method based on the TSNet framework, which solves the problem of losing non-Euclidean information in traditional modeling and provides a precise spatial basis for subsequent risk positioning.

[0176] (2) To address the problem of disorder in persistent graphs, this invention uses a rational hat function to generate a probability distribution of the birth and death coordinates of each point in the persistent graph. By using hybrid entropy aggregation, the disordered point set is transformed into a fixed-length vector, preserving the stability information of topological features, providing structured input for deep learning models, and avoiding the waste of topological information.

[0177] (3) In response to the problem of spatiotemporal feature fragmentation, this invention proposes the BS-STN model, which captures multi-dimensional spatial correlation by constructing a dual-structure graph, and then constructs a time series graph and graph convolution to realize the coupled modeling of spatial topology guiding temporal dynamics and temporal dynamics updating spatial risks. Finally, it outputs the probability distribution of prominent risks in multiple future time steps, improving the accuracy and stability of predictions with different time effects.

[0178] (4) In view of the problem that single graph modeling cannot reflect risk-driven association, this invention proposes a functional graph construction method based on the feature similarity of grid cells. Based on the cosine distance of monitoring data such as gas concentration and stress value, grid cells with similarity less than a preset threshold are set as neighbors to construct a functional graph, accurately capture grid cells that are not physically adjacent but are risk-related, supplement the physical graph that can only reflect spatial location association, and improve the representation of spatial risk propagation path.

[0179] (5) To address the issues of mismatched spatiotemporal feature dimensions and lack of physical consistency, this invention proposes a broadcast embedding mechanism, adding learnable spatial and temporal embeddings. By summing the spatiotemporal joint features through the broadcast mechanism, the semantic and dimensional adaptation of spatial and temporal features is resolved. In addition, physical constraints such as gas diffusion and pressure conservation are incorporated to ensure that the prediction results conform to the laws of multi-physics fields in the mine, thereby improving the reliability and interpretability of the model output. Ultimately, this invention achieves online dynamic identification and short / medium-term high-precision prediction of mine-scale outburst hazards, providing full-process support for on-site personnel from risk positioning to evolution trend to disposal suggestions, thereby reducing the probability of outburst disasters.

[0180] Example 2

[0181] Figure 5 This is a schematic diagram of the structure of a prominent hazard intelligent prediction system based on spatiotemporal evolution analysis provided in Embodiment 2 of the present invention, as shown below. Figure 5 As shown, the system includes:

[0182] The data acquisition and preprocessing module 510 is used to acquire multi-source heterogeneous data from the mine, and to perform unified temporal serialization, missing value imputation, and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data; wherein, the multi-source heterogeneous data includes static geological data, dynamic monitoring data, and image and topology data;

[0183] The data encoding module 520 is used to construct a spatial topology map based on the rasterized spatiotemporal grid data, extract topological features for spatial encoding, and obtain a spatial encoding feature matrix; and to use a temporal convolutional network to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix.

[0184] The cross-modal self-attention fusion module 530 is used to weightedly fuse the spatial coding feature matrix and the temporal coding feature matrix through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix;

[0185] The hazard probability distribution prediction module 540 is used to construct a dual-structure spatiotemporal network model that couples a physical graph and a functional graph based on the spatiotemporal joint feature matrix, and input the spatiotemporal joint feature matrix to perform spatiotemporal evolution analysis to predict the salient hazard probability distribution at multiple future time steps.

[0186] The risk grading and visualization module 550 is used to grade and visualize risks based on the predicted probability distribution of salient hazards over multiple future time steps, and generate automated response recommendations.

[0187] The prominent hazard intelligent prediction system based on spatiotemporal evolution analysis provided in this embodiment of the invention can execute the prominent hazard intelligent prediction method based on spatiotemporal evolution analysis provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the prominent hazard intelligent prediction method based on spatiotemporal evolution analysis. For detailed process, please refer to the relevant operations of the prominent hazard intelligent prediction method based on spatiotemporal evolution analysis in the foregoing embodiments.

[0188] Example 3

[0189] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0190] like Figure 6As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0191] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0192] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the salient hazard intelligent prediction method based on spatiotemporal evolution analysis described above.

[0193] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0194] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for intelligent prediction of outburst danger based on spatio-temporal evolution analysis, characterized in that, include: Acquire multi-source heterogeneous data from the mine, and perform unified temporal sequencing, missing value imputation, and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data; A spatial topology map is constructed based on the rasterized spatiotemporal grid data, and topological features are extracted for spatial encoding to obtain a spatial encoding feature matrix. A temporal convolutional network is used to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix, including: Topological map elements are extracted from the rasterized spatiotemporal grid data, wherein the topological map elements include shaft and tunnel nodes, geological units, shaft and tunnel connectivity relationships between grid units, and adjacent relationships between geological units; A mine spatial topology map is constructed based on the aforementioned topology map elements, wherein the mine tunnel nodes and geological units are used as vertices of the topology map, and the mine tunnel connectivity and geological unit adjacency relationships are used as edges of the topology map. Persistent homology analysis is performed on the spatial topology graph to generate a persistent graph for quantifying the stability of topological features; The persistent graph is vectorized and encoded using a TS-Layer of the topology set network to generate spatial feature vectors. Specifically, during the TS-Layer vectorization, each topology-sensitive module in the TS-Layer uses a rational hat function to generate a probability distribution, which quantifies the degree of matching between a single point in the persistent graph and the probability distribution of the topology-sensitive module. The output range is... ; The spatial feature vector is mapped to grid cells, and the feature weights of key topological regions are enhanced through a residual attention mechanism to highlight the feature contribution of key topological regions, and a spatial coding feature matrix with the same dimension as the rasterized spatiotemporal grid data is output. The spatial coding feature matrix and the temporal coding feature matrix are weighted and fused through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix. Based on the aforementioned spatiotemporal joint feature matrix, a dual-structure spatiotemporal network model coupling a physical graph and a functional graph is constructed. The spatiotemporal evolution analysis is then performed using the spatiotemporal joint feature matrix to predict the probability distribution of salient hazards at multiple future time steps, including: Based on the spatiotemporal joint feature matrix, a dual-structure spatiotemporal network model coupling the physical graph and the functional graph is constructed; The physical graph and the functional graph are connected through the same grid cells to generate a spatiotemporal adjacency matrix; Learnable spatial and temporal embeddings are added to the spatiotemporal adjacency matrix, and the matrix is ​​summed with the spatiotemporal joint feature matrix through a broadcast mechanism to obtain an enhanced spatiotemporal joint feature matrix. Chebyshev graph convolution is used to extract spatiotemporal correlation features across time steps and grid cells from the enhanced spatiotemporal joint feature matrix. The extracted spatiotemporal correlation features are then concatenated using max pooling and average pooling to obtain graph-level features that fuse global information. The graph-level features are input into the dual-structure spatiotemporal network model, which outputs the probability distribution of prominent hazards at multiple future time steps and their corresponding real risk labels. Based on the predicted probability distribution of prominent hazards at multiple future time steps, risk classification and visualization are performed, and automated response suggestions are generated.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data is subjected to unified temporal sequencing, missing value imputation, and spatial gridding to obtain rasterized spatiotemporal grid data, including: The multi-source heterogeneous data includes static geological data, dynamic monitoring data, and imagery and topological data; The static geological data is expanded to a time-series dimension, the dynamic monitoring data is resampled at a preset sampling interval, and the images and topological data are converted into a frame sequence time-series format to obtain a unified time-series dataset. A spatiotemporal interpolation method based on an attention mechanism is used to calculate weights based on the inherent spatiotemporal correlation of the data, and to fill in the missing data points in the unified time series dataset. A spatial reference grid is established under the local coordinate system of the mine, and the completed unified time series dataset is mapped to the grid cells to obtain the rasterized spatiotemporal grid data.

3. The method according to claim 1, characterized in that, A temporal convolutional network is used to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix, including: Extract multi-channel time-series monitoring data of each grid cell from the rasterized spatiotemporal grid data; A temporal convolutional network is used to extract the corresponding temporal dynamic features from the multi-channel temporal monitoring data of each grid cell; The time-dimensional global average pooling is performed on the time dynamic features corresponding to each grid cell to obtain a time-encoded feature matrix that corresponds one-to-one with each grid cell.

4. The method according to claim 3, characterized in that, By employing a cross-modal self-attention fusion mechanism, the spatial encoding feature matrix and the temporal encoding feature matrix are weighted and fused to generate a spatiotemporal joint feature matrix, including: The spatial coding feature matrix is ​​mapped to a grid to generate a grid-dimensional spatial feature matrix, wherein the grid-dimensional spatial feature matrix has the same dimension as the temporal coding feature matrix. Calculate the cross-modal attention weights between the grid dimension spatial features in the grid dimension spatial feature matrix and the temporal features in the temporal coding feature matrix; Based on the cross-modal attention weights, the spatial and temporal features of the same grid cell are weighted and fused to generate a spatiotemporal joint feature matrix that covers all grid cells and integrates spatiotemporal information.

5. The method according to claim 1, characterized in that, The method further includes: The classification loss is calculated based on the predicted probability distribution of salient hazards at multiple future time steps and their corresponding true risk labels. During the training process of the dual-structure spatiotemporal network model, a physical constraint loss function is introduced, which includes gas diffusion constraint loss and pressure conservation constraint loss. The classification loss and the physical constraint loss are fused according to preset weights to guide the training of the dual-structure spatiotemporal network model.

6. The method according to claim 1, characterized in that, The functional diagram is constructed through the following steps: Calculate the cosine similarity of each grid cell in the gas concentration and stress value monitoring data; Grid cell pairs with cosine similarity less than or equal to a preset threshold are defined as functional neighbor relationships; Construct a functional graph adjacency matrix based on the aforementioned functional neighbor relationships.

7. The method according to claim 1, characterized in that, The process involves risk classification and visualization based on the predicted probability distribution of salient hazards over multiple future time steps, generating automated response recommendations, including: Based on the predicted probability distribution of prominent hazards over multiple future time steps and the preset risk probability threshold, each grid cell is divided into different risk levels; Based on the spatial geographic information of the mine, the risk level of each grid unit is dynamically presented, and risk visualization results covering the entire prediction cycle are generated. The system automatically matches the preset emergency response plan for mine outburst disasters with the risk level and generates targeted graded disposal suggestions that are adapted to the risk level. The risk visualization results and targeted graded handling suggestions are pushed to the mine monitoring terminal, and confirmation or correction feedback is received from on-site personnel so as to iteratively optimize the prediction model based on the feedback information.

8. A prominent hazard intelligent prediction system based on spatiotemporal evolution analysis, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data from the mine, and to perform unified temporal sequencing, missing value imputation and spatial gridding on the multi-source heterogeneous data to obtain rasterized spatiotemporal grid data. The data encoding module is used to construct a spatial topology map based on the rasterized spatiotemporal grid data, extract topological features for spatial encoding, and obtain a spatial encoding feature matrix. A temporal convolutional network is used to perform temporal encoding on the rasterized spatiotemporal grid data to obtain a temporal encoding feature matrix, including: Topological graph elements are extracted from the rasterized spatiotemporal grid data. These elements include shaft / tunnel nodes, geological units, shaft / tunnel connectivity relationships between grid units, and geological unit adjacency relationships. A mine spatial topological graph is constructed based on these elements, where shaft / tunnel nodes and geological units are used as vertices, and shaft / tunnel connectivity relationships and geological unit adjacency relationships are used as edges. Persistent homology analysis is performed on the spatial topological graph to generate a persistent graph for quantifying the stability of topological features. The persistent graph is vectorized using a TS-Layer of the topological set network to generate spatial feature vectors. When the persistent graph is vectorized using the TS-Layer, each topology-sensitive module in the TS-Layer uses a rational hat function to generate a probability distribution, which quantifies the matching degree between a single point in the persistent graph and the probability distribution of the topology-sensitive module. The output range is... ; The spatial feature vector is mapped to the grid cell, and the feature weight of the key topological region is enhanced by the residual attention mechanism to highlight the feature contribution of the key topological region, and the spatial coding feature matrix with the same dimension as the rasterized spatiotemporal grid data is output. The cross-modal self-attention fusion module is used to weightedly fuse the spatial encoding feature matrix and the temporal encoding feature matrix through a cross-modal self-attention fusion mechanism to generate a spatiotemporal joint feature matrix; The hazard probability distribution prediction module is used to construct a dual-structure spatiotemporal network model coupling a physical graph and a functional graph based on the spatiotemporal joint feature matrix, and to input the spatiotemporal joint feature matrix for spatiotemporal evolution analysis to predict the salient hazard probability distribution at multiple future time steps, including: Based on the spatiotemporal joint feature matrix, a dual-structure spatiotemporal network model coupling physical graph and function graph is constructed; the physical graph and function graph are connected through the same grid cells to generate a spatiotemporal adjacency matrix; learnable spatial embedding and temporal embedding are added to the spatiotemporal adjacency matrix, and the matrix is ​​summed with the spatiotemporal joint feature matrix through a broadcast mechanism to obtain an enhanced spatiotemporal joint feature matrix; Chebyshev graph convolution is used to extract spatiotemporal correlation features across time steps and grid cells from the enhanced spatiotemporal joint feature matrix. The extracted spatiotemporal correlation features are then concatenated using max pooling and average pooling to obtain graph-level features that fuse global information. The graph-level features are then input into the dual-structure spatiotemporal network model to output the probability distribution of prominent hazards in multiple future time steps and their corresponding real risk labels. The risk grading and visualization module is used to grade and visualize risks based on the predicted probability distribution of salient hazards over multiple future time steps, and generate automated response recommendations.

Images