A coal mine safety situation awareness method and system based on graph convolution

Abstract

The application discloses a coal mine safety situation awareness method and system based on graph convolution, and belongs to the technical field of mine multi-disaster safety situation awareness. The application obtains mine multi-source heterogeneous monitoring data and basic environment data, and uniformly quantizes the data, constructs a disaster heterogeneous graph, a global connection graph and a space-time graph structure containing a time sequence correlation edge, obtains a safety situation evaluation value and a node attention weight through a self-attention graph convolution model, obtains a situation prediction result and a factor risk prediction result through XGBoost and Autoformer respectively, updates the factor risk prediction result to a future node and performs graph convolution reevaluation, corrects the prediction deviation according to the reevaluation result and the situation prediction result, and determines a key situation factor according to the node attention weight change. The application can simultaneously represent the time sequence evolution relationship of multi-disaster correlation and safety situation factors, and improves the reliability of coal mine safety situation prediction and key risk factor identification.

Description

Technical Field

[0001] This invention relates to the field of mine multi-hazard safety situation awareness technology, specifically to a coal mine safety situation awareness method and system based on graph convolution. Background Technology

[0002] The underground mining environment in coal mines is complex, with risks of hazards such as gas, fire, mine pressure, and dust often coexisting, and different hazards may have direct or potential influence relationships. With the continuous deployment of intelligent equipment, sensors, and video monitoring systems in mines, a large amount of monitoring information from different sources, with different data structures and time scales is generated during the production process, providing a data foundation for conducting safety situation assessments and predictions for the entire mine.

[0003] Existing coal mine safety monitoring typically relies on monitoring systems for gas concentration, dust, CO, surrounding rock stress, extraction, and ventilation to collect data, and makes risk assessments based on monitoring indicators, early warning rules, or prediction models corresponding to single types of disasters. Some methods also extract monitoring features through statistical analysis, data mining, or deep learning models to conduct integrated assessments of safety status or predict future changes based on historical trend values.

[0004] However, multi-source monitoring data differ in data format, sampling period, and spatial objects. Single disaster early warning or single path prediction mainly focuses on the analysis of local indicators, making it difficult to simultaneously characterize the correlation between multiple disasters and the changing relationship of various safety situation elements over time in a unified spatial and temporal context. It is also difficult to further verify the overall situation prediction results by combining future element risks and identify continuously increasing key risk elements, resulting in insufficient reliability of overall coal mine safety situation prediction and key risk element identification.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a coal mine safety situation perception method and system based on graph convolution, in order to solve the problem that the existing technology is unable to simultaneously characterize the correlation between multiple disasters and the changing relationship of safety situation elements over time, resulting in insufficient reliability of overall safety situation prediction and key risk element identification.

[0007] On the one hand, to solve the above-mentioned technical problems, this invention provides a method for coal mine safety situation awareness based on graph convolution, including: acquiring multi-source heterogeneous monitoring data and basic environmental data of the mine area, and obtaining quantitative situation indicators through feature extraction, risk screening, and unified quantification; determining the influence relationship between disaster risk variables based on the quantitative situation indicators, constructing a disaster heterogeneous graph for each disaster type, and constructing a global connectivity graph based on the node associations between graphs; performing time slicing on the disaster heterogeneous graph and the global connectivity graph, establishing temporal correlation edges between corresponding nodes in adjacent time slices to form a spatiotemporal graph structure; inputting the spatiotemporal graph structure into a self-attention graph convolution model, performing attention-weighted aggregation on node features to obtain safety situation assessment values ​​and node attention weights; and converting continuous time... The security situation assessment value of each slice is input into the first prediction model to obtain the situation prediction result. The historical quantification sequence of each security situation element is input into the second prediction model to obtain the element risk prediction result. According to the correspondence between security situation elements and graph nodes, the element risk prediction result is updated to the future corresponding node features and a new graph convolutional situation assessment is performed to obtain the reassessed situation result. The prediction deviation between the reassessed situation result and the situation prediction result is calculated. Based on the comparison result of the prediction deviation and the preset deviation threshold, the situation prediction result is corrected with the reassessed situation result to obtain the security situation prediction value. The changes in node attention weights in continuous time slices are statistically analyzed. Key situation elements are determined based on the trend of weight changes, and the security situation prediction value and key situation elements are output.

[0008] Optionally, the unified quantification process includes: performing multi-timescale local weighted regression smoothing decomposition on continuous time series monitoring data, extracting trend terms, and extracting local risk features based on a sliding window; performing exponential weighted quantification on the occurrence frequency of discontinuous monitoring events and early warning information, and mapping the basic environmental data to environmental risk features according to preset risk rules to form candidate risk features; obtaining the risk rating results of the candidate risk features, and correcting the candidate risk features according to the risk rating results to obtain a unified scale quantitative situation indicator.

[0009] Optionally, the steps of constructing the disaster heterogeneous graph and the global connectivity graph include: determining the variables characterizing the target disaster risk as risk nodes, determining the variables that have a direct influence relationship with the risk nodes as first-order neighbor nodes, and determining the variables that are directly related to the first-order neighbor nodes as second-order neighbor nodes; establishing connection edges according to the influence relationship between each node to form a disaster heterogeneous graph corresponding to the disaster type, and establishing cross-disaster connections for risk nodes and related nodes in different disaster heterogeneous graphs to form the global connectivity graph.

[0010] Optionally, the step of establishing connection edges based on the influence relationship between nodes includes: establishing direct influence connection edges between nodes with direct influence relationships, and determining potential influence variables corresponding to the direct influence variables; determining whether the potential influence variables and the direct influence variables have a substitutable relationship, and if so, establishing substitution connection edges between the corresponding nodes; and fusing the direct influence connection edges, the substitution connection edges, and the corresponding nodes to generate the disaster heterogeneous graph.

[0011] Optionally, the step of establishing temporal correlation edges includes: retaining the graph nodes and spatial connection relationships corresponding to each time period according to the same time scale to obtain multiple time slice graphs; establishing the temporal correlation edges between corresponding nodes based on the temporal sequence correlation of the same security situation element in adjacent time slices, and forming the spatiotemporal graph structure with the spatial connection relationships and the temporal correlation edges.

[0012] Optionally, the attention weighted aggregation step includes: calculating the attention coefficients between the target node and its spatially adjacent nodes and the corresponding nodes of adjacent time slices; normalizing the attention coefficients using a multi-head attention mechanism; and weighting and aggregating the features of the corresponding neighboring nodes according to the normalized attention coefficients to obtain the updated node features.

[0013] The security situation assessment value is output based on the updated node characteristics, and the attention coefficient is determined as the attention weight of the corresponding node.

[0014] Optionally, the first prediction model is an XGBoost model, and the second prediction model is an Autoformer model; the security situation assessment values ​​of continuous time slices are input into the XGBoost model to obtain the situation prediction results; the historical quantitative sequences of each security situation element are input into the Autoformer model for trend and periodic decomposition, and the periodic dependency is extracted based on the autocorrelation mechanism to obtain the element risk prediction results.

[0015] Optionally, the step of re-evaluating and correcting the graph convolutional situation includes: updating the feature vector of the corresponding node in the future time period with the risk prediction result of the element according to the correspondence between security situation elements and graph nodes; performing attention-weighted aggregation on the updated node features to obtain the re-evaluated situation result; calculating the absolute value of the difference between the re-evaluated situation result and the situation prediction result as the prediction deviation; when the prediction deviation exceeds the preset deviation threshold, correcting the situation prediction result according to the re-evaluated situation result to obtain the security situation prediction value.

[0016] Optionally, the step of determining key situation elements includes: statistically analyzing the attention weight distribution of each node in continuous time slices, calculating the weight difference between corresponding nodes in adjacent time slices; when the attention weight of the target node shows an upward trend in continuous time slices, determining the security situation element corresponding to the target node as the key situation element, and outputting the key situation element and the corresponding security situation prediction value.

[0017] On the other hand, the present invention also provides a coal mine safety situation awareness system based on graph convolution, comprising: a data processing module configured to acquire multi-source heterogeneous monitoring data and basic environmental data of the mine area and obtain quantitative situation indicators; a graph structure construction module configured to construct a disaster heterogeneous graph, a global connectivity graph, and a spatiotemporal graph structure containing temporally related edges; a situation assessment module configured to perform attention-weighted aggregation of node features through a self-attention graph convolution model to obtain a safety situation assessment value and node attention weights; a situation prediction module configured to input the safety situation assessment value of continuous time slices into a first prediction model to obtain a situation prediction result, input the historical quantitative sequence of each safety situation element into a second prediction model to obtain an element risk prediction result, update the element risk prediction result to the corresponding node features in the future and re-perform graph convolution situation assessment to obtain a reassessed situation result, and correct the prediction deviation between the reassessed situation result and the situation prediction result and a preset deviation threshold to obtain a safety situation prediction value; and a key element determination module configured to determine key situation elements based on the changes in node attention weights in continuous time slices and output the safety situation prediction value and the key situation elements.

[0018] Beneficial effects:

[0019] This invention unifies and quantifies multi-source heterogeneous monitoring data and basic environmental data in mines. Based on the influence relationships between safety situation elements, it constructs a disaster heterogeneous graph and a cross-disaster global connectivity graph. Through time slices and the temporal correlation between corresponding nodes, a spatiotemporal graph structure is formed, enabling the self-attention graph convolution model to integrate spatial adjacency information and cross-time node information to characterize the temporal evolution of multi-disaster correlations and safety situation elements. Furthermore, it forms a dual-branch prediction by combining XGBoost overall situation prediction and Autoformer element risk prediction. The element risk prediction results are written back to future nodes for graph convolution re-evaluation, and corrections are implemented based on the deviation between the re-evaluated situation results and the overall situation prediction results. At the same time, key situation elements are determined based on the changes in node attention weights in continuous time slices, thereby improving the reliability of overall safety situation prediction and key risk element identification in coal mines. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the data acquisition module provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the real-time monitoring data processing flow provided in an embodiment of the present invention;

[0022] Figure 3 This is a specific disaster safety situation diagram structure network diagram provided in the embodiments of the present invention;

[0023] Figure 4 This is a flowchart illustrating the relevant disaster map structure provided in an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of the supplementary time correlation graph structure provided in an embodiment of the present invention;

[0025] Figure 6 This is a flowchart of the overall security situation assessment provided in an embodiment of the present invention;

[0026] Figure 7 This is a flowchart of security situation prediction provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0028] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0029] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0030] As mentioned earlier, the underground mining environment in coal mines involves multiple types of disaster risks, including gas, fire, mine pressure, and dust, and these risks have direct or potential relationships with each other. With the application of intelligent equipment and monitoring systems in mines, a large amount of monitoring information with different sources, data structures, and time scales is generated in the mines. When using single disaster early warning or single path prediction, it is difficult to simultaneously express the correlation between multiple disasters and the changes of various safety situation elements over time.

[0031] To address this, the present invention provides a method and system for coal mine safety situation awareness based on graph convolution. It extracts features, screens risks, and quantifies unified data from multi-source heterogeneous monitoring data and basic environmental data in the mine. Using quantified situation indicators, it constructs heterogeneous disaster maps and cross-disaster global connectivity maps corresponding to different disaster types, and further establishes temporal correlations between adjacent time slices. A self-attention graph convolution model is used to aggregate features of spatially adjacent nodes and temporally related nodes to obtain safety situation assessment values ​​and node attention weights. Based on this, overall situation prediction results and future risk prediction results for each safety situation element are obtained. The element risk prediction results are written into future graph nodes and graph convolution situation assessment is performed again. The safety situation prediction value is obtained by analyzing the deviation between the reassessed situation result and the overall situation prediction result. Simultaneously, key situation elements are determined by utilizing changes in node attention weights in continuous time slices.

[0032] The following is combined Figures 1 to 7 This invention is described in detail.

[0033] Example 1:

[0034] This embodiment provides a coal mine safety situation awareness method based on graph convolution, which mainly includes the following steps:

[0035] S100. Acquire multi-source heterogeneous monitoring data and basic environmental data of the mining area, and obtain quantitative situation indicators through feature extraction, risk screening and unified quantification.

[0036] S200. Based on quantitative situation indicators, determine the influence relationship between disaster risk variables, construct a disaster heterogeneity graph for each disaster type, and construct a global connectivity graph based on the node association between graphs;

[0037] S300. Perform time slicing on the disaster heterogeneous graph and the global connectivity graph, establish temporal correlation edges between corresponding nodes of adjacent time slices to form a spatiotemporal graph structure; input the spatiotemporal graph structure into the self-attention graph convolution model, perform attention weighted aggregation on node features to obtain the security situation assessment value and node attention weight;

[0038] S400. Input the security situation assessment value of the continuous time slice into the first prediction model to obtain the situation prediction result, and input the historical quantitative sequence of each security situation element into the second prediction model to obtain the element risk prediction result; according to the correspondence between security situation elements and graph nodes, update the element risk prediction result to the future corresponding node features and re-perform graph convolution situation assessment to obtain the re-assessed situation result.

[0039] S500: Calculate the prediction deviation between the reassessed situation result and the situation prediction result; based on the comparison between the prediction deviation and the preset deviation threshold, correct the situation prediction result with the reassessed situation result to obtain the security situation prediction value; statistically analyze the changes in node attention weights in continuous time slices, determine key situation elements based on the weight change trend, and output the security situation prediction value and key situation elements.

[0040] Based on the above steps, this invention unifies and quantifies multi-source heterogeneous monitoring data and basic environmental data in mines, and constructs a spatiotemporal graph structure based on the influence relationship and temporal correlation between safety situation elements. It uses a self-attention graph convolution model to fuse spatial adjacency information and cross-time node information to obtain safety situation assessment results. Furthermore, it forms a dual-branch prediction process that mutually verifies overall situation prediction and element risk prediction. The element risk prediction results are written back to future graph nodes for graph convolution reassessment, and the results are corrected based on the prediction deviation between the reassessed situation results and the situation prediction results. At the same time, the key situation elements are determined by combining the continuous change of node attention weights. Thus, a coal mine safety situation perception process is formed, which is interconnected from multi-source data quantification, correlation mapping, spatiotemporal situation assessment, future risk prediction, reassessment correction to key element identification.

[0041] Example 2:

[0042] To provide a more detailed explanation of the technical solutions provided in the above embodiments, the present invention also provides another preferred embodiment. This embodiment proceeds in the following order: multi-source data access and quantification in the mine, disaster map structure construction, spatiotemporal map convolutional situation assessment, dual-branch prediction and future node reassessment, prediction correction and determination of key situational elements.

[0043] First, feature extraction, risk screening, and unified quantification are performed on the multi-source heterogeneous monitoring data and basic environmental data of the mining area. Specifically:

[0044] S110. Perform multi-time-scale local weighted regression smoothing decomposition on continuous time series monitoring data, extract trend terms, and extract local risk features based on sliding windows.

[0045] S120. The frequency of occurrence of discontinuous monitoring events and early warning information is quantified by exponential weighting, and the basic environmental data is mapped into environmental risk characteristics according to preset risk rules to form candidate risk characteristics.

[0046] S130. Obtain the risk rating result of the candidate risk features, and modify the candidate risk features according to the risk rating result to obtain a quantitative situation indicator with a unified scale.

[0047] In the above steps, multi-source heterogeneous data of the mine area can be obtained through the already installed big data monitoring system and existing databases. The data processing is compatible with the existing mine monitoring systems, collecting monitoring data from various sensors in existing systems such as microseismic monitoring systems, fire monitoring systems, extraction systems, and ventilation monitoring systems via network protocols. Data sources are identified according to different monitoring systems to serve as a reference for the type of nodes in different hazard heterogeneous maps. Simultaneously, basic data such as mine geology, hydrology, coal seam gas content, and coal physical properties can be obtained by reading tabular files or existing databases, achieving the integration of basic information from the entire mine with big data monitoring information.

[0048] like Figure 1 As shown, the collected data can be categorized into sensor data, video data, real-time statistical data, and basic mine information. Sensor and real-time monitoring information may include gas extraction monitoring, gas concentration monitoring, ventilation system monitoring, coal mining machine operation monitoring, dust monitoring, dust removal equipment monitoring, temperature monitoring, fire-related gas monitoring, electromechanical equipment temperature, belt and fiber optic temperature, micro-vibration monitoring, surrounding rock stress monitoring, and roof deformation monitoring, etc. Video and other data may include mechanical equipment operation information, equipment inspection records, drilling videos, floating coal accumulation videos, outburst prediction indicators, and sensor calibration results, etc. Basic mine information may include coal body physical properties, working face structure and faults, coal seam gas pressure, coal seam gas content, and other geological and hydrological data.

[0049] Data from different sources are uniformly identified upon access. Spatial information in multi-source heterogeneous data is encoded using data coding to form a unique regional identifier, while retaining the data source system identifier and time information. Existing information from different monitoring systems, such as gas concentration trend warnings, exceedance warnings, gas emission warnings, and abnormal changes in surrounding rock stress warnings, are mapped using unified warning keywords and corresponding identification information is generated.

[0050] Based on this, a hierarchical analysis of each monitoring system and basic data is conducted to establish a unified "element-feature-indicator" library. For real-time monitoring data, MSTL and sliding windows are used for online feature extraction, which can extract features such as slope, seasonal term amplitude, seasonal term phase drift, residual volatility, and anomaly integral. At the same time, data mining methods are used to establish a cross-source candidate feature library, which can include statistical features, frequency domain features, event features, and cross-hysteresis features. Valid features are screened through correlation, importance assessment, and expert experience to form security situation elements for subsequent fusion assessment and heterogeneous graph construction.

[0051] like Figure 2 As shown, gas concentration data, CO monitoring data, wind speed monitoring data, extraction monitoring data, dust monitoring data, and other real-time monitoring data first enter the data processing process, and then undergo MSTL decomposition, data mining, and other predetermined data processing. Subsequently, feature filtering is performed based on indicators such as correlation and importance, ultimately forming different safety status indicators.

[0052] Taking real-time monitoring data of methane concentration as an example, the time series of methane concentration is denoted as... , .in, The number of sampling points in the current processing sequence (e.g., when the sampling interval is 1 minute and 1 day of data is used as a processing sequence). 1440). MSTL employs an inner and outer loop of LOESS. In the inner loop, local weighted smoothing is performed on the seasonal subsequences corresponding to different periods, while low-pass smoothing and the trend term LOESS are used to calculate the trend of the sequence after removing the seasonal components. In the outer loop, robust weights are updated based on the residuals to reduce the impact of outliers on the fitting results, and the updated robust weights are fed back to the next round of the inner loop. To reduce the confusion between the trend term and the seasonal term, zero-sum constraints are applied to each seasonal component.

[0053] The additive decomposition relationship of the gas concentration time series is as follows:

[0054]

[0055] in, This is a trend item; For the first The seasonal components corresponding to each seasonal cycle; The number of seasonal cycles involved in the decomposition (e.g., ); This represents the residual term. Through the above decomposition, while preserving the smooth changes in the long-term trend, multi-period fluctuations are removed, and the trend information, seasonal information, and residual information used for subsequent risk feature extraction are obtained.

[0056] The time series data of gas concentration are further represented as follows:

[0057]

[0058] in, For the first At a certain point in time, This represents the gas concentration value at the corresponding time. The gas concentration value uses the data units of the original mine monitoring system; for example, when the original monitoring system recorded it as a volume fraction percentage, The unit is %; The number of sampling points contained in the current data sequence (e.g., The sampling interval is denoted as (For example, ).

[0059] The gas concentration data is segmented using a sliding window, and the k-th sliding window is represented as follows:

[0060]

[0061] The window start time is:

[0062]

[0063] The window ends at:

[0064]

[0065] in, This represents the start time of the current sequence; The sliding window number; For the defined window advance interval (e.g., ); For the defined window length (e.g., ).

[0066] Construct the improved -tsfresh input matrix based on each sliding window:

[0067]

[0068] For each sliding window, statistical features, time-series dynamic features, and nonlinearity test features can be extracted. Taking skewness, kurtosis, autocorrelation, and the Hearst exponent as examples, the skewness coefficient and kurtosis feature of the k-th time window are as follows:

[0069]

[0070] in, For the first The average gas concentration within each time window, with the same units as the original gas concentration data; For the first The standard deviation of gas concentration in each time window, with the same units as the original gas concentration data; and All of them are dimensionless.

[0071] No. The autocorrelation function (ACF) in the time-series dynamic features of each time window is:

[0072]

[0073] in, For the time lag order (e.g., ); Used to describe the time-series dynamic correlation of gas concentration data in the current window.

[0074] No. The Hearst exponent for each time window is:

[0075]

[0076] in, This is a Hurst exponent feature in nonlinear characteristics.

[0077] Therefore, trend changes are first extracted from continuous time series monitoring data using MSTL, and then statistical features, time series dynamic features, and nonlinear features are obtained using a sliding window to form candidate risk features corresponding to the continuous monitoring data.

[0078] For discontinuous information such as video risk identification, non-continuous prominent prediction indicators, edge risk identification results, and early warnings generated by the original monitoring system, the frequency of occurrence within a predetermined statistical period is calculated based on the time of event occurrence, and the frequency is quantified using an exponentially weighted moving average (EWMA) method to ensure that recent risk events have a higher quantitative contribution. The exponential weighting coefficient is determined based on historical valid event sequences or on-site initialization results (e.g., an exponential weighting coefficient of 0.8). For basic environmental data such as geology, hydrology, coal seam gas content, and coal physical properties, they are mapped to environmental risk characteristics according to preset risk rules corresponding to regional identifiers. Continuous sequence risk characteristics, discontinuous event risk characteristics, and environmental risk characteristics together constitute candidate risk characteristics.

[0079] Different mines have different geological parameters and operating conditions. The risk rules corresponding to the same safety situation elements may change with the mine conditions. Therefore, we need to combine expert theoretical knowledge with theoretical risk analysis to classify the risk characteristics of candidate risks.

[0080] The Delphi process includes determining expert credibility, expressing expert opinions, and assessing the consistency of expert opinions. The number of experts participating in each round of expert surveys can be selected as 25, as configured in the briefing materials (n=25). Expert credibility can be determined based on information such as years of experience and the number of relevant projects participated in, and confidence levels can be used. Description of the The credibility of the expert's judgment (e.g.) ).

[0081] Experts used linguistic judgment terminology to express their degree of acceptance of the risk rating results for security situation elements. The assessment set results are expressed as follows:

[0082]

[0083] in, These correspond to seven levels of language judgment terms, from lowest to highest. Each language judgment term is further represented using triangular fuzzy numbers, and its corresponding set of fuzzy numbers is as follows:

[0084]

[0085] and:

[0086]

[0087] Among them, the endpoints of each triangular fuzzy number are formed according to the pre-determined language judgment terminology-triangular fuzzy number mapping relationship.

[0088] The expert evaluation results are weighted and averaged based on the confidence levels of each expert to obtain a comprehensive fuzzy evaluation of a certain risk rule by the experts:

[0089]

[0090] in, This indicates a vague result of the experts' overall opinion on a certain rule. , and They represent the first The three components of the triangular fuzzy number corresponding to each expert.

[0091] Using the centroid method (COG) to defuzzify the comprehensive fuzzy evaluation, we obtain:

[0092]

[0093] in, This is the evaluation result of the expert consensus after defuzzification. When... When the value is between 5 and 6, it indicates that the expert consensus is nearly complete; when... A score of 3-5 indicates a moderate level of consensus among experts; when... A value between 0 and 3 indicates that the expert consensus is weakly consistent.

[0094] After each round of expert evaluation, Conduct an inspection. When When the risk level is between 0 and 3, the corresponding theoretical risk analysis rules are revised before proceeding to the next round of evaluation; when... When the scores are between 3 and 5, the relevant rule functions are further explained and confirmed before proceeding to the next round of expert evaluation. When the evaluation results for each parameter rule reach 5 to 6, the revised theoretical risk analysis results for the security situation variables are obtained, and these results are used as the basis for risk rating of candidate risk characteristics. It is only used to determine the degree of consensus of expert evaluation results and is not directly used as the quantitative value of candidate risk features. Based on the finally confirmed risk rating results, the candidate risk features are uniformly scaled and corrected to obtain a quantitative situation indicator used for graph node feature representation.

[0095] For example, in one preferred embodiment, the system first follows... Figure 1 The data organization method shown collects gas concentration, wind speed, dust, CO, surrounding rock stress, extraction negative pressure, and corresponding early warning records from a certain longwall mining face. Simultaneously, it obtains basic data on the region's geology, hydrology, coal seam gas pressure, coal seam gas content, and coal physical properties. Different data sources are mapped to the same longwall face using a unified regional code. Continuous data is processed at 1-minute sampling intervals to form a one-day data sequence. Trend, seasonal, and residual terms are extracted using MSTL, and local risk features such as skewness, kurtosis, autocorrelation, and Hurst exponent are extracted using a 10-minute sliding window. Non-continuous risk events are quantified using an exponential weighting coefficient of 0.8. Basic environmental information is used to form environmental risk features according to existing risk rules for the region. Subsequently, effective candidate features are screened using correlation and importance assessments and expert experience, and 25 experts use the Delphi process to generate risk rating results. Candidate risk features are uniformly converted into quantitative situation indicators usable by the final confirmed risk level and then incorporated into the subsequent disaster map structure construction process.

[0096] Subsequently, based on the quantitative situation indicators and the influence relationships between various security situation elements, a disaster heterogeneity diagram and a global connectivity diagram are constructed, and the specific implementation is as follows:

[0097] S210. The variables characterizing the target disaster risk are identified as risk nodes, the variables that have a direct influence relationship with the risk nodes are identified as first-order neighbor nodes, and the variables that are directly related to the first-order neighbor nodes are identified as second-order neighbor nodes.

[0098] S220. Establish direct influence connection edges between nodes with direct influence relationships, and determine the potential influence variables corresponding to the direct influence variables;

[0099] S230. Determine whether there is a substitutable relationship between the potential influencing variable and the direct influencing variable. If so, establish a substitution connection edge between the corresponding nodes.

[0100] S240. Merge the directly affecting connecting edges, the alternative connecting edges, and the corresponding nodes to generate the disaster heterogeneous graph;

[0101] S250. Establish connection edges based on the influence relationship between each node to form a disaster heterogeneous graph corresponding to the disaster type, and establish cross-disaster connections for risk nodes and related nodes in different disaster heterogeneous graphs to form the global connection graph.

[0102] In the above steps, a graph structure is constructed based on the established logical influence relationships between different safety situation elements. The direct influence relationships, potential influence relationships, and substitutable relationships are all based on the risk influence relationships confirmed before mapping and the current mine monitoring configuration; cross-hazard connections are established according to the existing cross-hazard node relationships in the mine. The hazard of each type of hazard is characterized by the safety situation elements directly related to it, while variables with potential influence relationships are included in the node relationships, enabling the graph convolution process to aggregate not only directly influencing variables but also variables with confirmed potential relationships.

[0103] like Figure 3 As shown, taking the convolutional network of the security situation assessment graph as an example, variable P represents the security situation risk of the corresponding disaster. Variables that have a direct causal relationship with variable P are selected as first-order neighbor nodes of P, and variables that are directly related to the first-order neighbor nodes are selected as second-order neighbor nodes. Figure 3 C1 to C22 represent different security situation variables, and each node forms a multi-layered association around the risk node P.

[0104] For variables with a direct impact relationship, establish direct impact edges; for variables that have been determined to serve as alternative representations of the same risk relationship, establish alternative impact edges. For example, Figure 3 C9 and C21 in the model are interchangeable. Some mines do not verify both simultaneously in regional outburst prediction, but their corresponding risk impact relationships can be substituted for each other based on the actual conditions of the mine, and therefore can be expressed through substitution connections. Different mines determine the types and number of nodes based on actual geological parameters, monitoring configurations, and production conditions.

[0105] like Figure 4As shown, different types of safety situation indicators form different disaster map structures. Gas concentrations at different locations, gas concentration trends, wind speed variation characteristics, and other relevant indicators can form gas exceedance safety situation indicators, which further form gas exceedance map structures; gas concentration trend characteristics, conventional prediction indicators, and geological indicators can form gas outburst safety situation indicators and gas outburst map structures; maximum dust concentration, video monitoring anomalies, and dust composition information can form dust explosion safety situation indicators and dust explosion map structures; surrounding rock stress at different locations, roof layer information, and anchor bolt monitoring data can form roof collapse safety situation indicators and roof collapse map structures; CO data trend characteristics, video hazards, and electromechanical equipment information can also serve as safety situation elements for fire, flood, and other related map structures.

[0106] For high-risk work areas in mines, the influence relationship between different disaster safety situation elements in tunneling and mining faces can be analyzed from both spatial and temporal dimensions. A heterogeneous map of the safety situation of each disaster in the corresponding high-risk area can be established. At the same time, the overall safety situation connection relationship between tunneling and mining faces can be established to achieve safety situation assessment at the working face level.

[0107] For the safety situation assessment of major disasters in the entire mine, safety situation assessment maps can be established separately for four types of disasters: gas, fire, roof collapse, and dust. Sensitive information related to various disasters in the mine can be integrated, and a global connection map can be established based on the confirmed cross-disaster node associations, thereby forming the overall safety situation assessment map structure of the entire mine.

[0108] For example, in this preferred embodiment, the system establishes nodes for the aforementioned unified and quantified gas concentration trends, CO changes, surrounding rock stress, dust concentration, and basic environmental risks, according to disaster types. Taking gas disaster as an example, the overall gas risk is set as the risk node P, variables directly related to P are identified as first-order neighbors, and variables directly related to first-order neighbors are identified as second-order neighbors; substitution connections are established for C9 and C21, which have been confirmed to be interchangeable in the configuration. Subsequently, disaster heterogeneous diagrams corresponding to gas, fire, roof fall, and dust disasters are formed respectively, and the risk nodes and related nodes of different disaster diagrams are connected according to the existing cross-disaster relationships in the mine to form the global connection diagram used in subsequent time slices.

[0109] Based on the aforementioned disaster heterogeneity diagram and global connectivity diagram, establish time relationships and perform a current security situation assessment, specifically:

[0110] S310. Retain the graph nodes and spatial connections corresponding to each time period according to the same time scale to obtain multiple time slice graphs;

[0111] S320. Based on the temporal sequence correlation of the same security situation element in adjacent time slices, establish the temporal correlation edge between corresponding nodes, and form the spatiotemporal graph structure with the spatial connection relationship and the temporal correlation edge.

[0112] S330. Calculate the attention coefficients between the target node and its spatially adjacent nodes, as well as the corresponding nodes in adjacent time slices.

[0113] S340. The attention coefficients are normalized using a multi-head attention mechanism, and the corresponding neighbor node features are weighted and aggregated according to the normalized attention coefficients to obtain the updated node features.

[0114] S350. Output a security situation assessment value based on the updated node characteristics, and determine the attention coefficient as the attention weight of the corresponding node.

[0115] In the above steps, the temporal relationships between security situation elements are further constructed based on the disaster heterogeneity map and the global connectivity map. The data collection process can form time slices according to a unified time scale. The length of the time slice is determined jointly based on the monitoring sampling cycle, the disaster evolution speed, and the security situation prediction task (for example, using 1 day as a time slice).

[0116] like Figure 5 As shown, These represent different time slices. Each time slice retains the node characteristics of the corresponding time period and the spatial logical connections in the security situation network diagram. For adjacent time slices... and The nodes in the middle can identify the same security situation element, and their historical time series are used to determine whether they meet the configured time-related conditions. If the time-related conditions are met, a time-series correlation edge is added between the corresponding nodes. Thus, spatial connectivity describes the logical influence between security situation elements in the same time slice, while the time-series correlation edge describes the temporal influence of security situation variables in the previous time period on the next time period.

[0117] like Figure 6 As shown, the security situation map structure and the risk quantification results of security situation elements are jointly input into the security situation map convolutional neural network. After model processing, the security situation assessment results and node attention information are formed, and on this basis, the sensitivity analysis of situation elements is carried out.

[0118] A set of node variables participating in GAT processing is represented as follows:

[0119]

[0120] in, Represents the set of all nodes in GAT; For the number of nodes (e.g., in the local node structure shown in the disclosure material). ); For each node, the feature length (e.g., ); {1,..., }

[0121] With nodes For example, its set of neighboring nodes can be represented as:

[0122]

[0123] For the target node and neighboring nodes First, calculate the nodes. For nodes Degree of impact:

[0124]

[0125] in, Represents a node For nodes The degree of impact; The weight matrix in graph convolution; These are the training parameters corresponding to the attention mechanism; the symbol " "" indicates the concatenation operation of the features after the transformation of two nodes. and All parameters are determined during the offline modeling phase of the self-attention map convolutional model and remain fixed during online inference of the same model version.

[0126] Normalization is performed based on the degree of attention influence of neighboring nodes, resulting in:

[0127]

[0128] in, For nodes For neighboring nodes The attention coefficient is a dimensionless value; Represents a node The set of neighboring nodes, Index the neighboring nodes.

[0129] GAT further employs a multi-head attention mechanism to calculate the weighted aggregation of node neighbor information separately. (Number of multi-head attention points) Determined by model configuration (e.g., ).node The new feature obtained after aggregating neighbor nodes is represented as follows:

[0130]

[0131] in, For nodes New features after aggregating neighbor nodes; 1 / This represents the average bullish attention level; For the first Each attention node For neighboring nodes Attention coefficient For attention head number, ; This corresponds to the weight matrix in the attention head; This is the activation function.

[0132] Time period obtained through multi-head attention mechanism and After calculating the attention coefficients of each variable, The security situation assessment results for the time period are expressed as follows:

[0133]

[0134] in, and They represent time periods respectively. and The Middle One security situation variable, and These represent the attention coefficients of the security situation variables in the corresponding time periods. express The results of the security situation assessment formed over a period of time.

[0135] The distribution of node attention coefficients can differ under different security situation assessment states. By retaining the attention coefficients of the same node in continuous time slices, we can further analyze the direction of change of node attention weights with the security situation risk state, and use this information to determine subsequent key security situation elements.

[0136] For example, in this preferred embodiment, the continuous monitoring results of a certain longwall face are divided into time slices of 1 day each. Each time slice retains the nodes and spatial connections in the gas, fire, roof, and dust hazard maps. If the same gas concentration trend node can be matched according to the node identifier in two adjacent days, and the historical sequence meets the configured time correlation conditions, then a temporal correlation edge is established between the corresponding nodes in the two time slices. After forming the spatiotemporal map, the degree of attention influence and the normalized attention coefficient are calculated for spatially adjacent nodes and temporally correlated nodes. Then, feature aggregation is performed according to the K=3 multi-head attention mechanism to output the safety situation assessment value of the current time slice and save the attention weight of each node. The safety situation assessment value continues to enter the overall safety situation prediction process, and the node attention weights are continuously saved according to the node identifier for subsequent analysis of key situation elements.

[0137] After obtaining the security situation assessment results of continuous time slices, further overall situation prediction, element risk prediction, and future node reassessment are formed, and specific implementation is carried out as follows:

[0138] S410. The first prediction model is an XGBoost model, and the second prediction model is an Autoformer model.

[0139] S420. Input the security situation assessment values ​​of the continuous time slices into the XGBoost model to obtain the situation prediction results;

[0140] S430. Input the historical quantitative sequence of each security situation element into the Autoformer model for trend and periodic decomposition, and extract the periodic dependency based on the autocorrelation mechanism to obtain the risk prediction result of the element.

[0141] S440. According to the correspondence between security situation elements and graph nodes, update the risk prediction results of the elements to the feature vector of the corresponding node in the future time period.

[0142] S450. Perform attention-weighted aggregation on the updated node features to obtain the reassessment situation result;

[0143] S460. Calculate the absolute value of the difference between the reassessed situation result and the situation prediction result as the prediction deviation. When the prediction deviation exceeds the preset deviation threshold, correct the situation prediction result according to the reassessed situation result to obtain the security situation prediction value.

[0144] In the above steps, such as Figure 7As shown, the safety situation prediction consists of two prediction branches with different input levels. The first prediction branch takes the safety situation assessment results as input and generates safety situation prediction results at different scales through XGBoost. These different spatial scales include at least the tunneling face, the longwall face, and the entire mine level; the different time scales are determined based on the monitoring sampling cycle, the speed of disaster evolution, and the time slice scale corresponding to the prediction task. The second prediction branch takes the risk quantification values ​​of safety situation elements as input and generates safety situation element prediction results through Autoformer. The predicted safety situation element risk values ​​are then rewritten into the corresponding graph nodes and re-evaluated through safety situation graph convolution. The re-evaluation results and the overall safety situation prediction results generated by XGBoost continue to enter the result comparison and key element control process.

[0145] XGBoost uses Newton's method to find the extreme value of the loss function and performs a second-order expansion of the loss function. It also introduces a regularization term to control the complexity of the tree model. Its training objective consists of the loss term and the regularization term.

[0146]

[0147] in, For the number of training samples (e.g., using 365 valid historical samples formed daily), =365); Indicates the first The loss function for each training sample; For the model to the first Predicted values ​​for each training sample; The true label is the value of the training sample; the loss function can be the mean squared error (MSE) or the log loss. The first in the model A tree, The tree number is B, and the total number of trees B is determined based on the offline training configuration (e.g., B=100). This represents the regularization term corresponding to the b-th tree.

[0148] The regularization term is represented as:

[0149]

[0150] in, Tree complexity control parameters (e.g., =1); The number of leaf nodes in the tree (e.g., =8); For regularization strength (e.g., =1); The leaf node weight vector, This represents the norm of the leaf node weight vector. All of the above parameters are determined during the XGBoost offline training phase; the online situation prediction phase uses a model that has already been trained and its version confirmed.

[0151] Using the trained XGBoost model, the security situation assessment values ​​formed by continuous time slices are used as input to obtain the overall situation prediction result P corresponding to the future time period of the target.

[0152] The second prediction branch uses Autoformer to predict the historical risk quantification sequence of each security situation element. Autoformer is a variant of Transformer, and its structure consists of N encoder layers and... It consists of several decoder layers, where the number of encoder layers N and the number of decoder layers are... Determined by the offline model configuration.

[0153] The Autoformer decoder consists of a sequential flow of seasonal information and an accumulation of trend information. Assume the decoder has... There are several decoder layers, each including encoder-decoder series correlation processing (CSAM), as well as internal autocorrelation and encoder-decoder autocorrelation. Using the latent variable from the encoder output, the... Each decoder layer processes data according to the following relationship:

[0154]

[0155] in:

[0156]

[0157] when When =1, Corresponding to the initial embedded input formed by the seasonal input of the decoder, the trend information is provided by the trend input of the decoder; and They represent the first In the layer decoder The seasonal and trend information obtained from the subsequence decomposition, among which {1,2,3}; Indicates the model parameters of the corresponding decomposition layer; symbol " " indicates element-wise multiplication.

[0158] go through After processing by the layer decoder, the prediction result of the target future interval is expressed as follows:

[0159]

[0160] in, Predict the length for the target (e.g., =1 future time slice); To make the first The layer decoder outputs model parameters mapped to the target dimension; Indicates the future The predicted output corresponding to each time position.

[0161] Autoformer model inputs include encoder inputs and decoder inputs. The encoder input is represented as:

[0162]

[0163] in, Indicates the length of the historical time period used by the encoder (e.g., =30 historical time points); F represents the input feature dimension (e.g., for a single security situation element historical quantification sequence, F=1).

[0164] The seasonal and trend components of the decoder are represented as follows:

[0165]

[0166] First, perform sequence decomposition on the latter half of the encoder input:

[0167]

[0168] in, This represents the seasonal component obtained by decomposing the latter half of the encoder input. This indicates the corresponding trend.

[0169] The decoder input is initialized as follows:

[0170]

[0171] in, The length represents the target predicted length. Zero-filled tensors; This represents the mean-filled tensor corresponding to the encoder's historical inputs; This represents the process of concatenating two tensors.

[0172] The autocorrelation mechanism in Autoformer extracts periodic dependencies by calculating the autocorrelation of time series data. For a length of... Real discrete-time process { }, sequence length Determined based on the length of the input history (e.g., =30), and its autocorrelation is:

[0173]

[0174] in, {1,..., } indicates a time delay; Used to characterize the degree of periodic correlation under the corresponding time delay. The autocorrelation of long sequences can be obtained through FFT based on the Wiener-Khinchin relationship, thereby improving the efficiency of autocorrelation calculation.

[0175] In multi-head autocorrelation processing, the results of multiple autocorrelation heads are represented as follows:

[0176]

[0177] in, These are the query value, key value, and variable, respectively. For the number of autocorrelation heads (e.g., =4); Output the corresponding projection parameters for the multi-head autocorrelation mechanism.

[0178] For the The time-delay aggregation process for each autocorrelation head is as follows:

[0179]

[0180] in, Indicates according to time delay For variables Scrolling is performed so that elements that have exceeded the first position are reintroduced into the last position, thus completing the alignment of subsequences with the same periodic position.

[0181] Selected The autocorrelation results are normalized using SoftMax:

[0182]

[0183] in:

[0184]

[0185] The hyperparameter used to control the selected time delay ranges from 1 to 3 (e.g., =2); Indicates the period before entering time-delayed aggregation Each autocorrelation position.

[0186] Autoformer also decomposes complex time patterns using moving averages. For lengths of... Feature dimension is input sequence :

[0187]

[0188] Its trend and seasonal components are represented as follows:

[0189]

[0190] in, For the extracted long-term trend portion; This refers to the seasonal component after stripping away the trend. This indicates that the sequence boundaries are padded; This indicates a moving average processing method.

[0191] Therefore, Autoformer uses sequence decomposition, autocorrelation, and time-delay aggregation to extract long-term trends, seasonal changes, and cyclical dependencies in the historical quantitative sequences of various security situation elements, and obtains the element risk prediction results for the same future prediction time period.

[0192] The risk prediction results generated by the second prediction model are not directly output as the final security situation result. The system, based on the correspondence between security situation element identifiers, region identifiers, and graph node identifiers saved during the mapping process, locates each valid element risk prediction result to the corresponding graph node in the target future time slice and updates the feature dimension of that node corresponding to the relevant security situation element. Other feature dimensions in the future node that have not been updated by the second prediction model retain their original definitions, and the direct impact connections, substitution connections, cross-hazard connections, and temporal correlation edges between adjacent time slices in the disaster heterogeneity map remain unchanged.

[0193] After updating the features of future nodes, the updated features of future nodes and the existing spatial and temporal connections are re-input into the aforementioned self-attention graph convolutional model. Then, a second inference is performed according to the same node attention calculation and multi-head feature aggregation process as the current security situation assessment to obtain the reassessed situation result R.

[0194] The overall situation prediction result generated by the first prediction model is denoted as... The reassessment of the situation is recorded as follows: Before comparing the two methods, they should be applied to the same mining area, the same future forecast period, and the same quantitative scale of safety status. The forecast discrepancy between the two methods should be considered. for:

[0195]

[0196] in, For prediction bias; , and All are expressed according to a unified security status quantification scale. For example, when the security status value is uniformly quantified to the range of 0 to 1, , and All are dimensionless values.

[0197] Preset deviation threshold The prediction bias distribution in historical valid prediction samples, on-site initialization configuration, or manual verification results are used to determine (e.g., under a uniform quantization scale of 0 to 1). =0.10).

[0198] Security situation prediction value Determined based on the comparison between the predicted deviation and the preset deviation threshold:

[0199]

[0200] Therefore, when the difference between the first prediction branch and the future map reassessment branch does not exceed the allowable range, the overall situation prediction result formed by XGBoost is maintained; when the prediction deviation between the two exceeds the preset deviation threshold, the first prediction result is corrected by the reassessment situation result formed by the re-aggregation of the element risk prediction and the existing disaster topology.

[0201] For example, in this preferred embodiment, the system inputs the security situation assessment values ​​of consecutive time slices into the XGBoost model to obtain the overall situation prediction result for the next time slice. =0.62; Simultaneously, historical quantified sequences of gas concentration trends, CO changes, surrounding rock stress, dust, and other safety situation elements are input into Autoformer. After sequence decomposition, autocorrelation period dependency extraction, and encoder-decoder processing, the risk prediction results for each element in the next time slice are generated. The system writes the effective element risk prediction results into the feature vector of the corresponding graph node in the next time slice according to the element-node correspondence, while maintaining the original spatial and temporal connections. The updated future nodes are re-input into the same self-attention graph convolutional model to obtain the reassessed situation results. =0.78. At this time... =|0.78-0.62|=0.16, which is higher than the preset deviation threshold. =0.10, therefore the predicted value for this round of security situation is... =0.78. This forms... Figure 7 The actual data connection between the XGBoost overall situation prediction branch and the Autoformer feature prediction-future node update-graph convolution reassessment branch is shown, rather than simply outputting the results of the two prediction models side by side.

[0202] Finally, the attention weights of nodes in the continuous time slices are statistically analyzed to determine key situational factors, and the specific execution is as follows:

[0203] S510. Statistically analyze the attention weight distribution of each node in consecutive time slices, and calculate the weight difference between corresponding nodes in adjacent time slices.

[0204] S520. When the attention weight of the target node shows an upward trend in the continuous time slice, the security situation element corresponding to the target node is determined as the key situation element, and the key situation element and the corresponding security situation prediction value are output.

[0205] In the above steps, the self-attention graph convolutional model saves the attention weights of the corresponding nodes after completing the security posture assessment for each valid time slice. The key element determination process maps the attention weights of the same node in consecutive time slices according to the region identifier and node identifier, and only compares valid time slices from the same model version with consistent node identifiers.

[0206] The difference in attention weights for the same node in adjacent time slices is:

[0207]

[0208] in, Represents a node Slice of the current effective time Attention weights in This represents the attention weight of the same node in the previous valid time slice. These represent the changes in attention weights between adjacent time slices; all three are dimensionless values.

[0209] The number of time slices used for a continuous upward trend is determined based on historical effective slice statistics and mine risk analysis requirements (e.g., three consecutive effective time slices). The attention weight of the target node increases sequentially across the three consecutive effective time slices, i.e., the weight of adjacent time slices... When all values ​​are greater than 0, the node is mapped back to the corresponding security situation element according to the existing node-security situation element correspondence relationship, and is identified as a key situation element.

[0210] For example, in this preferred embodiment, the attention weights of a certain gas concentration trend node in three consecutive effective time slices are 0.18, 0.25, and 0.33, respectively. The difference in attention weights between the first two time slices is 0.07, and the difference between the last two time slices is 0.08. Both differences are greater than 0, therefore the node meets the condition of continuous increase. Based on the node identifiers and safety situation element identifiers retained during the mapping process, the system determines the gas concentration trend corresponding to this node as the key situation element for this round and compares it with the aforementioned safety situation prediction value. =0.78 (Common Output). The identified key situational elements are used to characterize risk factors whose impact continues to increase during the current escalation of security situation risks, and serve as the basis for determining key targets and control directions in subsequent security management.

[0211] In the above implementation process, for cases where input data is missing or abnormal, control is implemented according to the effectiveness of the data processing chain. Individual outliers in continuous time series can have their impact on trend extraction reduced through robust weights in the MSTL outer loop; when the number of valid data within the sliding window is insufficient, the window is not used to form local risk features; when a certain basic environment field does not have a confirmed risk mapping rule, the field and its corresponding integrity information are retained, and a risk score is not automatically generated.

[0212] When a certain security situation element cannot be associated with a unique graph node in a future time slice, the risk prediction result of that element will not be written into the future node; when the historical input length required by the first prediction model or the second prediction model is insufficient, the corresponding prediction result will be treated as an invalid result for the current period; when the updated future graph node cannot obtain a valid reassessment result, the invalid reassessment result will not be used to replace the valid prediction result formed by the first prediction model.

[0213] For judging the trend of node attention weights, when the node identifiers in adjacent time slices cannot correspond, or when the self-attention graph convolution model undergoes a version switch, the relevant weights are not directly formed into a continuous change sequence; after accumulating valid time slices of the same version that meet the continuity requirements, the judgment of key situational elements is then performed.

[0214] Accordingly, this invention unifies continuous monitoring data, discontinuous risk events, and basic environmental information into quantitative situation indicators. It organizes heterogeneous maps of different disasters based on direct impact relationships, substitution relationships, and cross-disaster correlations. Furthermore, it forms a spatiotemporal graph structure through time slices and temporal correlations between corresponding nodes, enabling the self-attention graph convolutional model to comprehensively characterize the spatial influence between different disaster elements and their evolution over time. Based on this, it performs deviation verification and selective correction on the overall situation prediction results generated by XGBoost and the graph convolutional reassessment results generated after writing back to future nodes using Autoformer element prediction results. It also identifies key situation elements by combining changes in node attention weights in continuous time slices. Simultaneously, it effectively controls for situations such as missing data, node correspondence failures, invalid prediction results, and model version changes, thereby improving the data consistency and reliability of the coal mine safety situation assessment, prediction, and key risk element identification processes.

[0215] Example 3:

[0216] Based on the same inventive concept, this invention also provides a coal mine safety situational awareness system based on graph convolution, used to implement the aforementioned method. The system includes:

[0217] The data processing module is configured to acquire multi-source heterogeneous monitoring data and basic environmental data of the mining area and obtain quantitative situation indicators;

[0218] The graph structure construction module is configured to construct a disaster heterogeneous graph, a global connectivity graph, and a spatiotemporal graph structure containing temporally related edges;

[0219] The situation assessment module is configured to perform attention-weighted aggregation of node features through a self-attention graph convolutional model to obtain the security situation assessment value and node attention weights.

[0220] The situation prediction module is configured to input the security situation assessment values ​​of continuous time slices into a first prediction model to obtain the situation prediction result, input the historical quantitative sequence of each security situation element into a second prediction model to obtain the element risk prediction result, update the element risk prediction result to the future corresponding node features and re-perform graph convolution situation assessment to obtain the reassessed situation result, and correct it according to the comparison result of the prediction deviation between the reassessed situation result and the situation prediction result and the preset deviation threshold to obtain the security situation prediction value.

[0221] The key element determination module is configured to determine key situational elements based on the changes in node attention weights in continuous time slices, and output the security situation prediction value and the key situational elements.

[0222] The data processing module is compatible with the mine's existing big data monitoring system via network protocols. It receives data from systems monitoring microseismic activity, fire, gas extraction, ventilation, and safety, and reads basic data such as mine geology, hydrology, coal seam gas pressure, coal seam gas content, and coal physical properties through database interfaces or spreadsheets. This module also receives information on equipment operation, equipment maintenance, drilling videos, loose coal accumulation videos, outburst prediction indicators, and existing early warnings from the monitoring system, ensuring consistent data time information, spatial location information, data source identifiers, and early warning keywords.

[0223] The data processing module, following the processing relationships described in Embodiment 2, performs MSTL decomposition and sliding window feature extraction on the real-time continuous monitoring sequence, quantifies the frequency of discontinuous events and early warning information, and maps risk rules to the basic environmental data. Simultaneously, it establishes candidate features such as statistical, frequency domain, event, and cross-lag features, and filters features through correlation, importance assessment, and expert experience. Finally, it combines these with Delphi risk rating results to form a quantitative situation indicator with a unified quantitative scale. The data acquisition, unified encoding, MSTL processing, data mining, feature filtering, and risk quantification functions undertaken by the data acquisition module and the security situation element understanding module in the disclosure materials are taken over by the data processing module in this embodiment.

[0224] The graph structure construction module establishes a disaster heterogeneous graph based on the risk nodes, first-order neighbor nodes, and second-order neighbor node relationships corresponding to different disasters, and establishes corresponding connecting edges based on the confirmed direct impact relationships, potential impact relationships, and substitutable relationships. For high-risk areas such as tunneling faces and mining faces, disaster safety situation heterogeneous graphs can be formed at the corresponding working face level; for the entire mine, disaster heterogeneous graphs can be established according to disaster types such as gas, fire, roof fall, and dust, and a cross-disaster global connection graph can be established based on the confirmed node relationships between different disasters.

[0225] The graph structure construction module further generates multiple consecutive time slices according to a unified time scale. Each time slice retains the corresponding nodes and spatial connections, and establishes temporal correlation edges between corresponding nodes based on the temporal correlation of the same security situation element in adjacent time slices, thereby forming a spatiotemporal graph structure that simultaneously possesses spatial and temporal relationships.

[0226] The situation assessment module receives the aforementioned spatiotemporal graph structure and node quantification features, and invokes a pre-trained self-attention graph convolutional model. During the security situation assessment of the current time slice, it calculates the degree of attention influence and normalized attention coefficients for the target node and its spatially adjacent nodes, as well as the corresponding nodes in adjacent time slices. Then, it aggregates the features of neighboring nodes through a multi-head attention mechanism to form updated node features, and outputs the security situation assessment value and node attention weights.

[0227] The node attention weights are continuously stored and passed to the key element determination module; the security situation assessment values ​​are passed to the situation prediction module in continuous time sequence.

[0228] The situation prediction module includes overall security situation prediction processing and security situation element risk prediction processing. The overall security situation prediction processing uses a trained XGBoost model, taking the security situation assessment values ​​of continuous time slices as input to generate a situation prediction result for the target future time period. The risk prediction and processing of security situation elements adopts the Autoformer model, which takes the historical quantitative sequence of each security situation element as input, and forms the element risk prediction results for the corresponding future time period through trend and seasonal decomposition, autocorrelation periodic dependence extraction and encoder-decoder processing.

[0229] The situation prediction module updates the corresponding node features in future time slices with the effective element risk prediction results based on the correspondence between security situation element identifiers, region identifiers, and graph node identifiers saved by the graph structure construction module. The updated future node features are returned to the situation assessment module, and attention aggregation is performed again using the same version of the self-attention graph convolutional model as the current security situation assessment, along with existing spatial and temporal connectivity relationships, to obtain the reassessed situation result. .

[0230] The situation assessment module will reassess the situation results. Return to the situation prediction module. The situation prediction module calculates... Overall situation prediction results Prediction deviation between and will Deviation threshold from preset Comparison. When Greater than At that time, with As the corrected security situation prediction value ;when Not greater than At that time, keep As a value for predicting the security situation .

[0231] The key element determination module receives the node attention weights in the continuous time slices and the security situation prediction values ​​output by the situation prediction module. The system assigns continuous attention weights to the same node based on region identifier, node identifier, and model version, and calculates the weight difference between adjacent time slices. When a target node shows a continuous upward trend across a specified number of consecutive valid time slices, the security situation element corresponding to the target node is identified as a key situation element and output along with the predicted security situation value.

[0232] The data processing module is used to collect, analyze, extract features, screen risks, and quantify the multi-source heterogeneous monitoring data and basic environmental data of the mine, forming quantitative situation indicators for subsequent graph structure construction. The graph structure construction module constructs a disaster heterogeneous graph and a global connectivity graph based on the direct impact, potential impact, substitutability, and cross-disaster correlation between safety situation elements, and retains the nodes and spatial connections of each time period according to a unified time scale. Temporal correlation edges are established between corresponding nodes that meet the time correlation conditions to form a spatiotemporal graph structure.

[0233] The situation assessment module inputs the spatiotemporal graph structure and quantified situation indicators into the self-attention graph convolutional model to aggregate node features, obtaining the security situation assessment value and node attention weights. The situation prediction module uses XGBoost to perform overall situation prediction on the security situation assessment values ​​of continuous time slices, and uses Autoformer to perform element risk prediction on the historical quantified sequence of each security situation element. After updating the element risk prediction results to the corresponding future nodes, the graph convolutional situation assessment is performed again to obtain the reassessed situation result. Then, the prediction deviation between the reassessed situation result and the situation prediction result and the preset deviation threshold are used for correction. The key element determination module determines the key situation elements based on the changes in node attention weights in continuous time slices.

[0234] Each module can be deployed on the same mine safety situation analysis server, or it can be deployed on edge computing nodes and the central server respectively, depending on the mine's existing computing resources. Regardless of the deployment method, each module transmits quantitative situation indicators, graph node characteristics, node connection relationships, safety situation assessment values, node attention weights, situation prediction results, element risk prediction results, reassessment situation results, and safety situation prediction values ​​through data interfaces. This enables a continuous data processing process, including mine multi-source heterogeneous data acquisition, risk quantification, disaster map construction, spatiotemporal graph convolutional assessment, dual-branch prediction, future node updates, graph convolutional reassessment, prediction bias correction, and output of key situation elements.

[0235] The modules described in this invention can be implemented by a processor calling computer programs stored in memory. Different modules can be deployed on the same computing device or distributed according to the existing information system and computing resources of the mine. During implementation, each processing step is executed according to its inherent data inheritance relationship to maintain consistent correspondence between mine area identifiers, safety situation element identifiers, graph node identifiers, time slices, prediction times, and model versions.

[0236] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present 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 coal mine safety situation awareness based on graph convolution, characterized in that, include: Multi-source heterogeneous monitoring data and basic environmental data of the mining area are acquired, and quantitative situation indicators are obtained through feature extraction, risk screening and unified quantification. Based on quantitative situation indicators, the influence relationship between disaster risk variables is determined, a disaster heterogeneity graph for each disaster type is constructed, and a global connectivity graph is constructed based on the node associations between graphs; The disaster heterogeneous graph and the global connectivity graph are sliced ​​over time, and temporal correlation edges are established between corresponding nodes in adjacent time slices to form a spatiotemporal graph structure. The spatiotemporal graph structure is input into a self-attention graph convolutional model, and attention-weighted aggregation of node features is performed to obtain the security situation assessment value and node attention weights. The security situation assessment values ​​of continuous time slices are input into the first prediction model to obtain the situation prediction results, and the historical quantitative sequences of each security situation element are input into the second prediction model to obtain the element risk prediction results. Based on the correspondence between security situation elements and graph nodes, the risk prediction results of the elements are updated to the features of the corresponding nodes in the future, and the graph convolution situation assessment is performed again to obtain the reassessed situation results. Calculate the prediction deviation between the reassessed situation result and the situation prediction result. Based on the comparison between the prediction deviation and a preset deviation threshold, correct the situation prediction result with the reassessed situation result to obtain the security situation prediction value. The changes in node attention weights in continuous time slices are statistically analyzed. Key situational elements are determined based on the trend of weight changes, and the predicted security situation value and key situational elements are output.

2. The method according to claim 1, characterized in that, The steps of the unified quantization process include: Multi-timescale local weighted regression smoothing decomposition is performed on continuous time series monitoring data to extract trend terms, and local risk features are extracted based on sliding windows; The frequency of non-continuous monitoring events and early warning information is quantified by exponential weighting, and the basic environmental data is mapped into environmental risk characteristics according to preset risk rules to form candidate risk characteristics; Obtain the risk rating results of the candidate risk features, and revise the candidate risk features based on the risk rating results to obtain a quantitative situation indicator with a unified scale.

3. The method according to claim 1, characterized in that, The steps for constructing the disaster heterogeneity graph and the global connectivity graph include: The variables characterizing the target disaster risk are identified as risk nodes, the variables that have a direct influence relationship with the risk nodes are identified as first-order neighbor nodes, and the variables that are directly related to the first-order neighbor nodes are identified as second-order neighbor nodes. Based on the influence relationships between nodes, connection edges are established to form a disaster heterogeneous graph corresponding to the disaster type. Cross-disaster connections are then established for risk nodes and associated nodes in different disaster heterogeneous graphs to form the global connection graph.

4. The method according to claim 3, characterized in that, The step of establishing connecting edges based on the influence relationships between nodes includes: Establish direct influence connection edges between nodes with direct influence relationships, and identify potential influence variables corresponding to the direct influence variables; Determine whether there is a substitutable relationship between the potential influencing variable and the direct influencing variable. If so, establish a substitution connection edge between the corresponding nodes. The disaster heterogeneous graph is generated by integrating the directly affecting connecting edges, the alternative connecting edges, and the corresponding nodes.

5. The method according to claim 1, characterized in that, The steps for establishing temporal correlation edges include: By preserving the graph nodes and spatial connections corresponding to each time period according to the same time scale, multiple time slice graphs are obtained. Based on the temporal correlation of the same security situation element in adjacent time slices, the temporal correlation edge is established between corresponding nodes, and the spatiotemporal graph structure is formed by the spatial connection relationship and the temporal correlation edge.

6. The method according to claim 5, characterized in that, The attention-weighted aggregation steps include: Calculate the attention coefficients between the target node and its spatially adjacent nodes, as well as the nodes corresponding to adjacent time slices; The attention coefficients are normalized using a multi-head attention mechanism, and the corresponding neighbor node features are weighted and aggregated according to the normalized attention coefficients to obtain the updated node features. The security situation assessment value is output based on the updated node characteristics, and the attention coefficient is determined as the attention weight of the corresponding node.

7. The method according to claim 1, characterized in that, The first prediction model is the XGBoost model, and the second prediction model is the Autoformer model; The security situation assessment values ​​of continuous time slices are input into the XGBoost model to obtain the situation prediction results; The historical quantitative sequences of each security situation element are input into the Autoformer model for trend and periodic decomposition, and the periodic dependency is extracted based on the autocorrelation mechanism to obtain the risk prediction results of the element.

8. The method according to claim 7, characterized in that, The steps for re-evaluating and correcting the graph convolution state include: Based on the correspondence between security situation elements and graph nodes, the risk prediction results of the elements are updated to the feature vectors of the corresponding nodes in future time periods. The updated node features are then aggregated using attention weighting to obtain the reassessment situation result. The absolute value of the difference between the reassessed situation result and the situation prediction result is calculated as the prediction deviation. When the prediction deviation exceeds the preset deviation threshold, the situation prediction result is corrected according to the reassessed situation result to obtain the security situation prediction value.

9. The method according to claim 1, characterized in that, The steps for determining key situational elements include: The attention weight distribution of each node in consecutive time slices is statistically analyzed, and the weight difference between corresponding nodes in adjacent time slices is calculated. When the attention weight of the target node shows an upward trend in a continuous time slice, the security situation element corresponding to the target node is determined as the key situation element, and the key situation element and the corresponding security situation prediction value are output.

10. A coal mine safety situation awareness system based on graph convolution, used to implement the method described in any one of claims 1 to 9, characterized in that, include: The data processing module is configured to acquire multi-source heterogeneous monitoring data and basic environmental data of the mining area and obtain quantitative situation indicators; The graph structure construction module is configured to construct a disaster heterogeneous graph, a global connectivity graph, and a spatiotemporal graph structure containing temporally related edges; The situation assessment module is configured to perform attention-weighted aggregation of node features through a self-attention graph convolutional model to obtain the security situation assessment value and node attention weights. The situation prediction module is configured to input the security situation assessment values ​​of continuous time slices into a first prediction model to obtain the situation prediction result, input the historical quantitative sequence of each security situation element into a second prediction model to obtain the element risk prediction result, update the element risk prediction result to the future corresponding node features and re-perform graph convolution situation assessment to obtain the reassessed situation result, and correct it according to the comparison result of the prediction deviation between the reassessed situation result and the situation prediction result and the preset deviation threshold to obtain the security situation prediction value. The key element determination module is configured to determine key situational elements based on the changes in node attention weights in continuous time slices, and output the security situation prediction value and the key situational elements.

Images