Coal mine safety risk intelligent early warning method empowered by large language model

Abstract

The application relates to a coal mine safety risk intelligent early warning method based on a large language model, and relates to the technical field of coal mine safety risk early warning, and comprises the following steps: acquiring multi-source heterogeneous monitoring data, and constructing a safety event narrative flow based on the physical cause-effect relationship among gas emission, ventilation dilution and ignition source appearance; performing semantic analysis on the safety event narrative flow, extracting a risk representation and a chain reaction path, and fusing a dynamic underground environment entropy to calculate a gas explosion risk entropy value; inputting the gas explosion risk entropy value and the corresponding risk representation into a large language model to generate a risk narrative description; performing scene-based risk mode matching and deconstruction on the risk narrative description and the safety event narrative flow, identifying and associating key risk elements and composite risk structures, and dynamically generating an early warning level and a disposal strategy plan. The application solves the problem that the traditional coal mine safety risk early warning is not sensitive to unknown and composite risks due to the dependence on preset rules and thresholds, and the early warning accuracy is insufficient.

Description

Intelligent Early Warning Method for Coal Mine Safety Risks Empowered by Big Language Model Technical Field

[0001] This application relates to the field of coal mine safety risk early warning, and in particular to intelligent early warning methods for coal mine safety risks empowered by large language models. Background Technology

[0002] The underground environment of coal mines is complex and changeable. Safety risks such as gas explosions pose a great threat to the lives of workers and the safety of mine production. Therefore, accurate safety risk early warning has become a core technical requirement for safe production in coal mines.

[0003] Currently, traditional coal mine safety risk early warning methods mostly rely on preset rules and fixed thresholds for monitoring and early warning, which is difficult to adapt to the complex dynamic scenarios underground. They lack sensitivity to unknown risks not included in the preset scope and complex risks formed by multiple intertwined factors, directly resulting in insufficient early warning accuracy and creating hidden dangers for coal mine safety production. Summary of the Invention

[0004] This application provides a method for intelligent early warning of coal mine safety risks empowered by a large language model, which improves the problem of insufficient accuracy of traditional coal mine safety risk early warning due to reliance on preset rules and thresholds and insensitivity to unknown and complex risks.

[0005] The embodiments of this application disclose the following technical solutions:

[0006] This application provides a method for intelligent early warning of coal mine safety risks powered by a large language model, the method comprising:

[0007] Acquire multi-source heterogeneous monitoring data of target areas in underground coal mines, and construct a safety event narrative stream based on the physical causal relationship between gas outburst, ventilation dilution and ignition source occurrence;

[0008] Semantic parsing is performed on the safety event narrative to extract risk representations and chain reaction paths, and the underground dynamic environment entropy is fused to calculate the gas explosion risk entropy value;

[0009] The gas explosion risk entropy value and the corresponding risk characterization are input into a pre-trained large language model for risk inference, generating a risk narrative description that includes the gas diffusion path and detonation probability.

[0010] The risk narrative description and the safety event narrative flow are matched and deconstructed in a scenario-based manner to identify and associate key risk elements and complex risk structures, and to dynamically generate early warning levels and response strategy plans.

[0011] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0012] This application proposes a method for intelligent early warning of coal mine safety risks empowered by a large language model. By acquiring multi-source heterogeneous monitoring data step by step, constructing a safety event narrative flow, extracting risk characteristics and calculating the entropy value of gas explosion risk, performing risk inference using a large language model, and matching and deconstructing scenario-based risk patterns, it achieves accurate early warning and targeted handling of underground gas explosion risks in coal mines. First, time-series data from underground environmental safety sensors, main ventilation fan operation data from the production equipment monitoring system, and positioning and image data from the spatial perception system are acquired to form a multi-source heterogeneous monitoring dataset. Then, based on the physical causal relationship between gas emission, ventilation dilution, and ignition source occurrence, the data is integrated according to time sequence and correlation logic to generate a safety event narrative stream. Subsequently, the narrative stream is semantically parsed to extract three types of risk representations and connect them to form a chain reaction path. The underground dynamic environmental entropy is then fused to calculate the gas explosion risk entropy value. Next, the risk entropy value and risk representations are input into a pre-trained large language model to deduce and generate a risk narrative description containing gas diffusion paths and ignition possibilities. Finally, through scenario-based risk pattern matching, the risk narrative description and safety event narrative stream are deconstructed to identify risk primitives and composite risk structures, dynamically generating early warning levels and integrated disposal instructions.

[0013] The technical solution of this application solves the problems of data fragmentation and unclear risk correlation in traditional coal mine safety early warning by integrating multi-source heterogeneous data causally and semantically, thereby improving the intelligence level and reliability of underground gas explosion risk early warning in coal mines and providing technical support for safe coal mine production. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 is a flowchart illustrating the intelligent early warning method for coal mine safety risks empowered by a large language model provided in an embodiment of this application.

[0016] Figure 2 is a schematic diagram of the process for calculating the entropy value of gas explosion risk provided in an embodiment of this application. Detailed Implementation

[0017] This application provides a method for intelligent early warning of coal mine safety risks empowered by a large language model, which is used to solve the technical problem that the traditional coal mine safety risk early warning relies on preset rules and thresholds and is not sensitive to unknown and complex risks, resulting in insufficient early warning accuracy.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] As shown in Figure 1, this application provides a method for intelligent early warning of coal mine safety risks powered by a large language model. The method includes the following steps:

[0020] S110: Acquire multi-source heterogeneous monitoring data of the target area in the coal mine, and construct a safety event narrative stream based on the physical causal relationship between gas emission, ventilation dilution and ignition source occurrence;

[0021] In this embodiment of the application, in the scenario where the underground environment of a coal mine is dynamically changing and the risk of gas explosion is affected by multiple intertwined factors, in order to analyze the complete process of risk evolution, it is necessary to comprehensively collect multi-source monitoring data covering environmental, equipment, and spatial dimensions, and integrate them based on the physical causal relationship of key risk factors to form a coherent safety event narrative, providing a data foundation for subsequent risk analysis and deduction.

[0022] Specifically, the first step is to acquire multi-source heterogeneous monitoring data of the target area underground in the coal mine to reflect the safety status underground from different dimensions and ensure that the monitoring coverage is comprehensive.

[0023] In the method provided in this application embodiment, the multi-source heterogeneous monitoring data includes:

[0024] The environmental safety sensor collects time-series data, wherein the time-series data includes at least gas concentration, carbon monoxide concentration, and temperature;

[0025] The main ventilation fan operation data reported by the production equipment monitoring system includes at least the start / stop status, operating current, voltage, and vibration amplitude.

[0026] The spatial perception system generates positioning and image data, which includes personnel positioning coordinates, equipment location information, and video surveillance images of key areas.

[0027] First, environmental safety sensors continuously collect time-series data at preset intervals, focusing on monitoring three core environmental parameters that directly affect the risk of gas explosion: methane concentration, carbon monoxide concentration, and temperature. These sensors are deployed in areas prone to methane accumulation, densely populated areas, and critical ventilation points underground to ensure real-time capture of changes in environmental conditions.

[0028] In addition, the production equipment monitoring system and the main ventilation fan control system are linked in real time, continuously reporting the main ventilation fan's operating data, including start / stop status, operating current, voltage, and vibration amplitude. The main ventilation fan, as the core equipment for underground ventilation and gas dilution, determines whether gas can be effectively dispersed, making it a key equipment factor affecting risk development.

[0029] Furthermore, the spatial perception system simultaneously generates positioning and image data through personnel positioning terminals, equipment positioning modules, and high-definition cameras deployed underground. Personnel positioning coordinates reflect the real-time distribution of workers; for example, if five workers are detected lingering in a gas accumulation risk area for an extended period, this information will alert personnel to potential safety hazards. Equipment location information monitors whether critical equipment is in its preset working position; if a gas extraction device deviates from its preset installation coordinates by more than 5 meters, it will be considered an abnormal equipment position. Video surveillance images of key areas can visually capture the on-site operational status, such as identifying workers using open flames in prohibited areas or operating electrical equipment in violation of regulations, providing a basis for identifying ignition source risk events.

[0030] During the collection of multi-source heterogeneous monitoring data, it is also necessary to ensure that the collection frequency of various types of data remains synchronized, generally set to once per second, in order to ensure the consistency of data in the time dimension.

[0031] Furthermore, after completing the synchronous collection of multi-source heterogeneous monitoring data, based on the physical causal relationship between gas outburst, ventilation dilution and ignition source occurrence, various monitoring data are transformed into associative event information, thereby constructing a complete safety event narrative.

[0032] The method provided in this application embodiment constructs a safety event narrative stream based on the physical causal relationship between gas emission, ventilation dilution, and the appearance of an ignition source, including:

[0033] Identify abnormal gas outburst events based on time-series data;

[0034] Based on the main ventilation fan's operating data, determine the ventilation dilution capacity status and correlate it with the aforementioned abnormal gas outburst event;

[0035] Based on location and image data, ignition source risk events are identified and spatiotemporally correlated with the gas outburst anomaly events and ventilation dilution capacity status.

[0036] The gas outburst anomaly, ventilation dilution capacity status, and ignition source risk events are integrated according to time sequence and causal relationship to generate a safety event narrative flow.

[0037] First, abnormal gas outburst events are identified based on time-series data collected by environmental safety sensors.

[0038] Specifically, the gas concentration, carbon monoxide concentration, and temperature in the time-series data are all preset with corresponding safety thresholds. For example, the safety threshold for gas concentration is set to 0.5%, the safety threshold for carbon monoxide concentration is 0.0024%, and the safety threshold for temperature is 35°C.

[0039] Furthermore, by comparing the collected time-series data with preset thresholds in real time, when one or more parameters meet the abnormal judgment conditions, it is determined to be a gas outburst abnormal event.

[0040] For example, when the gas concentration remains stable at 0.6% for five consecutive collection cycles, exceeding the preset safety threshold, and the carbon monoxide concentration rises to 0.003% and the temperature remains at 33°C during the same period, the system automatically identifies this situation as an abnormal gas outburst event, records the timestamp of the event as 10:12, and marks the specific underground area where the event occurred as the No. 4 mining area working face.

[0041] Furthermore, based on the main ventilation fan operation data reported by the production equipment monitoring system, the ventilation dilution capacity status is determined, and the identified abnormal gas outburst events are correlated.

[0042] Similarly, the start / stop status, operating current, voltage, and vibration amplitude of the main ventilation fan all have corresponding normal operating ranges. For example, during normal operation, the current is 30A-40A, the voltage is 380V±5%, and the vibration amplitude does not exceed 0.08mm.

[0043] Furthermore, by analyzing the real-time collected operating data, the real-time operating status indicators of the main ventilation fan are calculated. If the operating current is below 30A or above 40A, the voltage fluctuation exceeds ±5%, or the vibration amplitude exceeds 0.08mm, it indicates that the main ventilation fan is operating abnormally, and thus the ventilation dilution capacity status is determined.

[0044] For example, when the main ventilation fan's operating current drops to 25A, the voltage is 360V, and the vibration amplitude reaches 0.1mm, it is determined that the ventilation dilution capacity is insufficient. This state is then associated with the abnormal gas outburst event identified at 10:12 in the No. 4 mining area working face. It is clear that insufficient ventilation dilution capacity will cause the gas in the area to be unable to be effectively diffused, exacerbating the risk of gas accumulation. At the same time, the timestamp of the ventilation dilution capacity status determination is recorded as 10:15.

[0045] Furthermore, based on the positioning and image data generated by the spatial perception system, ignition source risk events are identified and spatiotemporally correlated with abnormal gas outburst events and ventilation dilution capacity status.

[0046] Among these methods, personnel location coordinates are analyzed to determine whether there are any abnormal clustering situations. For example, if six workers are detected staying in the abnormal gas outburst area of ​​the No. 4 mining face for more than 10 minutes, it is determined to be an abnormal clustering of personnel location coordinates.

[0047] In addition, equipment location information is compared with the real-time collected equipment location and the preset installation location. If a critical electrical device deviates from the preset location by more than 3 meters, it is determined to be an abnormal change in equipment location information. Video surveillance images use image recognition technology to capture illegal operations and open flame characteristics. For example, if it identifies workers illegally using lighters or performing electric welding operations in the area, or directly detects open flame images, it is determined to be an ignition source risk event.

[0048] For example, at 10:20, the video monitoring image of the spatial perception system identified an open flame at the working face of mining area No. 4. At the same time, the personnel positioning coordinates showed that three workers were staying near the open flame. This was determined to be an ignition source risk event, and the timestamp was recorded as 10:20.

[0049] Further, a spatiotemporal correlation analysis was conducted, confirming that the ignition source risk event, the previous abnormal gas outburst event, and the insufficient ventilation and dilution capacity all occurred in the No. 4 mining area working face, and were sequentially linked in time, indicating a direct risk correlation.

[0050] Finally, the abnormal gas outburst events, ventilation dilution capacity status, and ignition source risk events are integrated according to time sequence and causal relationship to generate a safety event narrative flow, so as to present the complete context of risk from its inception to its development and formation.

[0051] The method provided in this application integrates the abnormal gas outburst event, ventilation dilution capacity status, and ignition source risk event based on time sequence and causal relationship to generate a safety event narrative flow, including:

[0052] Determine the timestamps corresponding to the gas outburst anomaly, ventilation dilution capacity status, and ignition source risk event, and sort the events according to their chronological order;

[0053] Identify and establish the causal relationship between the ventilation dilution capacity status and the abnormal gas emission event, as well as the temporal and spatial overlap between the abnormal gas emission event and the ignition source risk event;

[0054] The aforementioned abnormal gas outburst events, ventilation dilution capacity status, and ignition source risk events are integrated into a coherent textual description with causes, development, and potential consequences, serving as a safety incident narrative flow.

[0055] First, the timestamps corresponding to each type of event and status are determined and sorted according to their chronological order. Specifically, when gas outburst anomalies, ventilation dilution capacity status, and ignition source risk events are identified, their corresponding timestamps are recorded simultaneously, with the timestamps accurate to the second, to ensure the accuracy of the time series.

[0056] For example, the timestamp of the abnormal gas outburst event is recorded as 09:35:12, the timestamp of the ventilation dilution capacity status determination is 09:35:48, and the timestamp of the ignition source risk event is 09:36:24. They are sorted in chronological order as abnormal gas outburst event, ventilation dilution capacity status, and ignition source risk event, thus presenting the occurrence sequence of various events and statuses.

[0057] Furthermore, the causal relationship between ventilation dilution capacity status and abnormal gas outburst events, as well as the temporal and spatial overlap between abnormal gas outburst events and ignition source risk events, were identified and established.

[0058] In causal relationship identification, the impact of ventilation dilution capacity on abnormal gas emission events is determined based on the physical laws governing gas dilution through ventilation. For example, when ventilation dilution capacity is insufficient, the main ventilation fan cannot provide enough ventilation, resulting in the gas concentration failing to decrease after an abnormal gas emission, or even continuing to rise. This establishes a direct causal relationship between insufficient ventilation dilution capacity and abnormal gas emission events.

[0059] In identifying spatiotemporal correlations, the correlation is confirmed by comparing the time intervals and spatial locations of events. For example, a gas outburst anomaly occurs at the No. 5 tunneling face underground, with a time interval from 09:35:12 to the subsequent continuous period. An ignition source risk event also occurs at the No. 5 tunneling face, with a timestamp of 09:36:24. The time intervals of the gas outburst anomaly and the spatial locations of the two events are continuously connected and completely overlap, thus establishing a spatiotemporal overlap correlation between the two events.

[0060] Finally, all events and states are integrated into a coherent textual description with causes, development, and potential consequences, serving as a safety incident narrative. Specifically, during the integration process, the core information of the event must be fully preserved, including the time, location, key parameters, and related logic, presenting the risk evolution trajectory using clear textual descriptions.

[0061] For example, the integrated text description is as follows: "At 09:35:12, the gas concentration at the No. 5 tunneling face underground was detected to rise to 0.6%, exceeding the preset safety threshold of 0.5%, and the carbon monoxide concentration simultaneously rose to 0.003%, which was determined to be an abnormal gas outburst event. At 09:35:48, the operating current of the main ventilation fan in this area dropped to 28A (normal range 30A-40A), and the vibration amplitude reached 0.1mm (normal threshold 0.08mm), indicating insufficient ventilation dilution capacity, resulting in the gas not being able to diffuse effectively and the concentration remaining at a high level. At 09:36:24, an open flame was found at the No. 5 tunneling face through video monitoring, which was determined to be an ignition source risk event. Currently, due to gas accumulation, ventilation failure, and the coexistence of an ignition source, this area has formed an extremely high risk of gas explosion."

[0062] S120: Perform semantic parsing on the safety event narrative stream, extract risk representations and chain reaction paths, and integrate the underground dynamic environment entropy to calculate the gas explosion risk entropy value;

[0063] In this embodiment of the application, in order to quantify the degree of gas explosion risk, it is necessary to extract risk representations from the safety event narrative stream through semantic parsing and sort out the chain reaction path, and then combine the characteristics of the dynamic environment downhole to calculate the entropy value, so as to form a quantitative indicator that can support subsequent risk projection and early warning decision-making.

[0064] Specifically, the safety incident narrative is first semantically analyzed to extract risk representations and chain reaction paths. The safety incident narrative contains coherent textual descriptions related to downhole risks. Semantic analysis techniques are used to deeply mine the text content, identifying and extracting different types of risk representations.

[0065] Furthermore, based on the chronological order of events in the narrative of a safety incident and combined with the causal correlation words implied in the text, the extracted risk representations are linked together according to the actual risk transmission logic to form a clear chain reaction path, so as to fully present the transmission process of risk from generation to development.

[0066] Furthermore, after completing the risk characterization and chain reaction path extraction, the dynamic environmental entropy of the underground environment is integrated. By analyzing the spatiotemporal correlation between ventilation capacity and gas concentration changes, calculating the rationality compensation coefficient, and correcting the environmental entropy benchmark value, the gas explosion risk entropy value is calculated by combining the number of three types of risk characterizations and the complexity of the chain reaction path, so as to quantify the overall severity of the current underground gas explosion risk.

[0067] This step, by extracting core risk information, sorting out the risk transmission logic, and integrating environmental characteristics and risk factors for quantitative calculation, not only clarifies the key links in risk evolution but also quantifies the degree of risk, laying a data foundation for subsequent risk simulation, generation of early warning levels, and contingency plans for response strategies.

[0068] As shown in Figure 2, step S120 of the method provided in this application embodiment includes:

[0069] From the safety incident narrative stream, identify and extract statements about gas concentration, carbon monoxide concentration, and temperature exceeding preset thresholds as the first type of risk characterization;

[0070] Extract the abnormal state descriptions of the main ventilation fan's start / stop status, operating current, voltage, and vibration amplitude from the safety event narrative stream as the second type of risk characterization;

[0071] The descriptions of abnormal clustering of personnel location coordinates, abnormal changes in equipment location information, or violations of operation and open flames identified in video surveillance images are extracted from the safety event narrative stream and used as the third type of risk characterization.

[0072] Based on the chronological order of events and causal correlation words in the security incident narrative, the first type of risk representation, the second type of risk representation, and the third type of risk representation are linked together to generate a chain reaction path.

[0073] Based on the first and second types of risk characterization, the spatiotemporal correlation between ventilation capacity and gas concentration changes within the preset monitoring period is analyzed.

[0074] Based on the physical laws of ventilation dilution of gas, the matching degree between the ventilation capacity and the change in gas concentration is calculated as a reasonable compensation coefficient;

[0075] Obtain the baseline value of the downhole dynamic environment entropy calculated based on the stability of the ventilation network and the influence of temperature, and use the rationality compensation coefficient to correct the baseline value of the downhole dynamic environment entropy to obtain the corrected downhole dynamic environment entropy.

[0076] By combining the corrected downhole dynamic environment entropy, the number of three types of risk representations, and the complexity of the chain reaction path, the gas explosion risk entropy value is calculated.

[0077] In this embodiment of the application, in order to analyze key risk information from the structured safety event narrative, it is necessary to extract multi-dimensional risk representations and construct chain reaction paths through semantic parsing, and then calculate risk entropy values ​​by combining the characteristics of the downhole dynamic environment and physical laws, so as to form quantitative indicators that can support subsequent risk projection and early warning decision-making.

[0078] Specifically, three types of risk representations are first extracted from the safety incident narrative stream. The safety incident narrative stream contains coherent textual descriptions related to downhole risks. Semantic parsing technology is used to analyze the text content sentence by sentence to filter out expressions related to preset risk indicators.

[0079] For the first type of risk characterization, we focus on the numerical descriptions of gas concentration, carbon monoxide concentration and temperature in the text. When a statement exceeds a preset safety threshold, the statement is directly extracted as the first type of risk characterization.

[0080] For the second type of risk characterization, the focus is on identifying the description of the main ventilation fan's operating status in the text. If it involves abnormal start / stop status, operating current or voltage deviating from the normal range, or vibration amplitude exceeding the standard, it will be extracted as the second type of risk characterization.

[0081] For the third type of risk representation, focus on the descriptions related to personnel location, equipment location, and video surveillance in the text. When there are descriptions of abnormal clustering of personnel location coordinates, abnormal changes in equipment location information, or video surveillance images identifying violations of regulations, open flames, etc., extract the description as the third type of risk representation.

[0082] For example, if the safety incident narrative contains the description "the methane concentration in area 3 underground rose to 0.7% (preset safety threshold 0.5%), carbon monoxide concentration to 0.003% (preset safety threshold 0.0024%), and temperature to 36℃ (preset safety threshold 35℃)," this description is extracted as a first-type risk characterization. If the text contains content such as "the main ventilation fan operating current dropped to 26A (normal range 30A-40A), and vibration amplitude to 0.11mm (standard threshold 0.08mm)," this content is extracted as a second-type risk characterization. If the description "eight workers remained in the methane accumulation area for more than 15 minutes, and video surveillance captured unauthorized hot work in the area" exists, this content is extracted as a third-type risk characterization.

[0083] Furthermore, after extracting the three types of risk representations, based on the chronological order of events in the safety incident narrative and incorporating causal conjunctions such as "caused," "further," and "triggered" implicit in the text, the three types of risk representations are linked together according to the actual risk transmission logic to generate a chain reaction path. Through the chain reaction path, the complete process of risk evolution from the initial triggering factor can be clearly presented, clarifying the mutual influence relationships between various risk representations.

[0084] For example, combining the three types of risk characteristics extracted from the above example, the chain reaction path formed in series is as follows: the gas concentration, carbon monoxide concentration and temperature in area 3 underground successively exceed the preset safety threshold, forming an abnormal risk of gas outburst; at the same time, the main ventilation fan's operating current is lower than the normal range and the vibration amplitude exceeds the standard, resulting in a decrease in ventilation efficiency and insufficient ventilation dilution capacity, thus failing to effectively disperse the accumulated gas; under these circumstances, 8 workers stayed in the gas accumulation area for a long time, and there was illegal hot work in the area, causing an extremely high risk of gas explosion.

[0085] Furthermore, based on the first and second types of risk characterization, the spatiotemporal correlation between ventilation capacity and gas concentration changes within a preset monitoring period is analyzed. The preset monitoring period is determined according to the underground production rhythm and risk response requirements to ensure complete coverage of the time range during which ventilation capacity changes affect gas concentration; for example, it is set to 30 minutes.

[0086] Specifically, the analysis focuses on the synchronicity and correlation between changes in ventilation capacity and gas concentration over time within the same spatial area, in order to clarify the interaction patterns between the two.

[0087] For example, within 10 minutes after the main ventilation fan operating current increases and the ventilation efficiency index improves, the slope of the gas concentration in the corresponding area changes from positive to negative, indicating that the enhanced ventilation capacity has a significant effect on gas dilution; if the gas concentration slope remains positive and the value increases after the ventilation efficiency index decreases, it indicates that insufficient ventilation capacity exacerbates gas accumulation.

[0088] Furthermore, based on the physical laws of ventilation dilution of gas, the matching degree between ventilation capacity and gas concentration change is calculated as a rationality compensation coefficient to quantify the control effect of the current ventilation status on gas accumulation.

[0089] The method provided in this application embodiment calculates the matching degree between the ventilation capacity and the change in gas concentration based on the physical laws of ventilation dilution of gas, as a reasonable compensation coefficient, including:

[0090] Real-time power is calculated based on the operating current and voltage of the main ventilator, and the ventilation efficiency index is calculated in combination with the vibration amplitude.

[0091] Extract historical data points of the gas concentration and calculate the slope of the gas concentration within a preset monitoring period as the trend of gas concentration change;

[0092] The ventilation efficiency index is compared with the trend of gas concentration change to determine the physical consistency of the direction of change of the two and to obtain the reasonable compensation coefficient.

[0093] The reasonableness compensation coefficient is obtained through the following methods:

[0094] When the ventilation efficiency index is higher than the normal operation threshold and the gas concentration change trend is a negative slope, the rationality compensation coefficient is set to the first preset value.

[0095] When the ventilation efficiency index is lower than the normal operation threshold and the gas concentration change trend is positive, the rationality compensation coefficient is set to the second preset value.

[0096] Otherwise, the rationality compensation coefficient is set to a third preset value;

[0097] The first preset value is greater than the third preset value, and the third preset value is greater than the second preset value.

[0098] First, the real-time power is calculated based on the operating current and voltage of the main ventilator, and the ventilation efficiency index is calculated in combination with the vibration amplitude. Specifically, the real-time power is directly calculated from the collected operating current and voltage data of the main ventilator, and the calculation formula is "real-time power value = operating current value × operating voltage value". This value directly reflects the energy output status of the main ventilator.

[0099] For example, when the main ventilation fan operates at a current of 32A and a voltage of 380V, the real-time power value is 32A × 380V = 12160W.

[0100] Furthermore, the vibration amplitude adjustment coefficient is calculated in conjunction with the vibration amplitude. Specifically, the standard vibration amplitude allowed by the equipment under healthy conditions is first obtained, and then calculated using the formula "Vibration amplitude adjustment coefficient = 1 - (current vibration amplitude / standard vibration amplitude)". The closer the current vibration amplitude is to the standard vibration amplitude, the closer the adjustment coefficient is to 1, indicating that the vibration has a smaller impact on ventilation efficiency.

[0101] For example, if the standard vibration amplitude is 0.08 mm and the current vibration amplitude is 0.04 mm, the vibration amplitude adjustment factor is 1 - (0.04 mm / 0.08 mm) = 0.5; if the current vibration amplitude is 0.02 mm, the adjustment factor is 1 - (0.02 mm / 0.08 mm) = 0.75.

[0102] Finally, the final ventilation efficiency index is calculated using the formula: "Ventilation efficiency index = Real-time power value × Vibration amplitude adjustment coefficient". For example, multiplying the real-time power value of 12160W by the adjustment coefficient of 0.75 yields a ventilation efficiency index of 9120W, which comprehensively reflects the actual ventilation capacity of the main ventilator.

[0103] Furthermore, historical data points on gas concentration are extracted, and the slope of gas concentration within a preset monitoring period is calculated as the trend of gas concentration change. The preset monitoring period needs to be determined in conjunction with the response time of underground ventilation adjustments and risk monitoring requirements to ensure that the impact of ventilation capacity changes on gas concentration is fully captured.

[0104] Specifically, gas concentration data for each collection moment within a preset monitoring period is extracted from the historical monitoring database to form a continuous gas concentration time series. Based on this gas concentration time series, a linear regression method is used to calculate the slope. A positive slope indicates that the gas concentration is increasing, while a negative slope indicates that the gas concentration is decreasing. The absolute value of the slope corresponds to the rate of concentration change.

[0105] For example, within a preset monitoring period of 30 minutes, the gas concentration gradually decreased from 0.55% to 0.45%, and the slope calculated by linear regression was -0.0033% / min, indicating that the gas concentration showed a steady downward trend; if the concentration increased from 0.4% to 0.6%, the slope was 0.0067% / min, indicating that the gas concentration increased rapidly.

[0106] Furthermore, the obtained ventilation efficiency index is compared with the trend of gas concentration change to determine the physical consistency of the direction of change between the two.

[0107] Specifically, according to the basic physical laws of ventilation dilution of methane, when the ventilation efficiency index increases, it means that the ventilation capacity is enhanced, and theoretically the methane concentration should show a downward trend; when the ventilation efficiency index decreases, the ventilation capacity weakens, and the methane concentration should show an upward trend. Both of these situations are considered to have a physically consistent direction of change. If there is a situation where the ventilation efficiency index increases but the methane concentration increases, or the ventilation efficiency index decreases but the methane concentration decreases, it is considered to be physically inconsistent, and there may be other interfering factors such as abnormal methane outbursts or distorted monitoring data.

[0108] Finally, a reasonable compensation coefficient is obtained based on the judgment results to quantify the effectiveness of ventilation status in controlling gas accumulation. Specifically, by comparing the calculated ventilation efficiency index with the normal operating threshold and combining the positive and negative slope characteristics of the gas concentration change trend, the fit between ventilation capacity and gas concentration change is clarified, and a corresponding preset value is assigned as the reasonable compensation coefficient.

[0109] The normal operating threshold is determined by analyzing the historical operating data of the main ventilation fan. The distribution of the ventilation efficiency index during long-term stable operation of the equipment is statistically analyzed, and the statistical mean is taken as the normal operating threshold. For example, the normal operating threshold is calculated to be 8000W based on historical data.

[0110] In addition, the first preset value is set to a relatively high value close to 1, representing that ventilation effectively controls gas accumulation and the environmental uncertainty is low; the second preset value is set to a relatively low value close to 0, representing that ventilation fails, the risk of gas accumulation increases, and the environmental uncertainty is high; the third preset value is set to an intermediate value between the two, representing that the matching relationship between ventilation and gas concentration changes is unclear and there is a certain degree of uncertainty.

[0111] For example, when the ventilation efficiency index is higher than the normal operating threshold and the gas concentration change trend is a negative slope, such as a ventilation efficiency index of 9120W and a gas concentration slope of -0.0033% / min, the rationality compensation coefficient is set to 0.9; when the ventilation efficiency index is lower than the normal operating threshold and the gas concentration change trend is a positive slope, such as a ventilation efficiency index of 7000W and a gas concentration slope of 0.0067% / min, the rationality compensation coefficient is set to 0.1; in other cases, such as a ventilation efficiency index of 8500W and a gas concentration slope of 0.001% / min, the rationality compensation coefficient is set to 0.5.

[0112] Furthermore, after determining the reasonableness compensation coefficient, the benchmark value of the downhole dynamic environment entropy calculated based on the stability of the ventilation network and the influence of temperature is obtained, and the benchmark value of the downhole dynamic environment entropy is corrected using the reasonableness compensation coefficient to obtain the corrected downhole dynamic environment entropy.

[0113] Specifically, calculating the stability of the ventilation network requires first obtaining airflow data from key measuring points underground within a preset period. Key measuring points are selected from core areas such as branch nodes of the underground ventilation network and the intake and return airways of the working face, to ensure a comprehensive reflection of the overall operating status of the ventilation system. The ventilation stability value is calculated using the formula "Ventilation stability value = Standard deviation of airflow data / Average value of airflow data". The larger this value, the more drastic the airflow fluctuations in the ventilation network, and the worse the stability.

[0114] For example, the average air volume data of a key measuring point within a preset period is 50m³. 3 / min, standard deviation is 5m 3 If the ventilation stability value is 5 / 50 = 0.1, then the standard deviation is 10m. 3 / min, the ventilation stability value is 10 / 50=0.2, the latter has worse ventilation stability.

[0115] Further, the temperature impact value is calculated by extracting the measured average temperature within a preset period from the time-series data collected by environmental safety sensors, and combining it with a preset safe temperature threshold. The temperature impact value is calculated using the formula: "Temperature Impact Value = (Measured Average Temperature - Safe Temperature Threshold) / Safe Temperature Threshold". When the measured average temperature is not higher than the safe temperature threshold, the temperature impact value is recorded as zero. The safe temperature threshold is determined according to the "Coal Mine Safety Regulations," for example, set to 35℃.

[0116] For example, if the measured average temperature is 36℃, then the temperature influence value = (36-35) / 35≈0.0286; if the measured average temperature is 34℃, the temperature influence value is recorded as 0.

[0117] Furthermore, the benchmark value of the downhole dynamic environment entropy is calculated by weighted summation. The specific calculation formula can be expressed as "benchmark value of downhole dynamic environment entropy = ventilation weight coefficient × (ventilation stability value / maximum ventilation stability reference value) + temperature weight coefficient × (temperature influence value / maximum temperature influence reference value)".

[0118] Among them, the ventilation weight coefficient and the temperature weight coefficient are set according to the degree of impact of underground environmental risks, and the sum of the two is 1. For example, the ventilation weight coefficient is set to 0.6 and the temperature weight coefficient is set to 0.4. The maximum ventilation stability reference value and the maximum temperature influence reference value are also determined according to the "Coal Mine Safety Regulations" and historical risk data. For example, the maximum ventilation stability reference value is 0.5 and the maximum temperature influence reference value is 0.3.

[0119] For example, assuming the ventilation stability value is 0.1 and the temperature influence value is 0.0286, substituting into the formula, we can obtain the baseline value of the downhole dynamic environmental entropy = 0.6×(0.1 / 0.5)+0.4×(0.0286 / 0.3)≈0.158.

[0120] Furthermore, the previously determined rationality compensation coefficient is used to correct the baseline value of the downhole dynamic environmental entropy in order to eliminate the interference of the matching relationship between ventilation status and gas concentration changes on the environmental entropy value. The specific correction formula can be expressed as "corrected downhole dynamic environmental entropy = rationality compensation coefficient × environmental entropy baseline value".

[0121] For example, when the reasonableness compensation coefficient is 0.9, the corrected downhole dynamic environmental entropy is 0.9 × 0.158 ≈ 0.142; if the reasonableness compensation coefficient is 0.1, the corrected downhole dynamic environmental entropy is 0.1 × 0.158 ≈ 0.0158. The correction makes the downhole dynamic environmental entropy value more consistent with the actual environmental risk state under the current ventilation control effect.

[0122] Furthermore, by combining the corrected downhole dynamic environment entropy, the number of three types of risk representations, and the complexity of the chain reaction path, the gas explosion risk entropy value is calculated.

[0123] Specifically, the total number of risk representations is first calculated, which is the sum of the number of risk representations in the first category, the second category, and the third category. For example, if there are 2 risk representations in the first category, 1 in the second category, and 3 in the third category, then the total number of risk representations = 2 + 1 + 3 = 6.

[0124] Furthermore, the complexity of a chain reaction path can be quantified by using the number of nodes in the path as a quantitative indicator; the more nodes, the higher the complexity. For example, a chain reaction path of "excessive gas concentration - insufficient ventilation efficiency - gas accumulation - personnel gathering - unauthorized open flame" contains 5 nodes, and its complexity quantification value is 5; if the path is "excessive temperature - gas emission - open flame appearance", it contains 3 nodes, and the complexity quantification value is 3.

[0125] Finally, the gas explosion risk entropy value is calculated by weighted summation. The formula can be expressed as "Gas explosion risk entropy value = (environmental entropy weight × corrected downhole dynamic environmental entropy) + (risk characterization weight × total number of risk characterizations) + (path complexity weight × path complexity)".

[0126] The environmental entropy weight, risk characterization weight, and path complexity weight are set according to the priority of risk impact, and the sum of the three is 1, for example, set to 0.3, 0.4, and 0.3 respectively. Assuming the corrected downhole dynamic environmental entropy is 0.142, the total number of risk characterizations is 6, and the path complexity is 5, substituting into the formula, we get: Gas explosion risk entropy value = 0.3 × 0.142 + 0.4 × 6 + 0.3 × 5 = 3.9426. This value comprehensively quantifies the overall severity of the current downhole gas explosion risk, providing a basis for subsequent risk projection and early warning decision-making.

[0127] S130: Input the gas explosion risk entropy value and the corresponding risk characterization into the pre-trained large language model to perform risk deduction and generate a risk narrative description that includes the gas diffusion path and detonation probability;

[0128] In this embodiment of the application, in order to accurately predict the development trend of gas explosion risk, it is necessary to use a pre-trained large language model to deeply deduce the quantified risk entropy value and key risk characteristics, so as to generate a coherent narrative description containing key risk information, and provide a basis for subsequent risk pattern matching and early warning disposal.

[0129] The method provided in this application embodiment includes the following pre-training steps for the large language model:

[0130] Based on the historical safety database and simulation case library of coal mines, data on gas explosion risk scenarios containing clear causal chains are collected to form a training sample set;

[0131] For each sample in the training sample set, the causal relationships between each risk representation are labeled, and a quality score label is generated for the entire narrative description, forming a causal relationship label set and a quality score label set;

[0132] A network architecture for a large language model is constructed based on machine learning. The input layer of the network architecture is set to a dual-branch channel. The first branch channel is used to receive the gas explosion risk entropy value, and the second branch channel is used to receive three types of risk representations. The output layer of the network architecture is configured as a text generation channel.

[0133] The large language model is trained under supervision using the training sample set, causal relationship label set, and quality score label set until the risk narrative description generated by the model reaches the preset standard, thus completing the pre-training of the model.

[0134] First, based on the historical safety database and simulation case library of coal mines, data on gas explosion risk scenarios containing clear causal chains are collected to form a training sample set.

[0135] Specifically, the data collection process must ensure that the samples cover different risk evolution scenarios, including combinations of different gas outburst intensities, ventilation conditions, and ignition source types, as well as risk cases under different working areas and production rhythms underground, in order to improve the generalization ability of the model.

[0136] For example, the samples include scenario data such as "slowly exceeding the gas concentration standard + insufficient ventilation efficiency + unauthorized hot work by personnel" as well as emergency scenario data such as "sudden large-scale gas outburst + ventilation system failure + equipment friction fire". At the same time, each sample must fully record the entire process of risk evolution, including risk characterization at each stage, parameter changes, and final results, to ensure the integrity and validity of the sample data.

[0137] Furthermore, for each sample in the training sample set, the causal relationships between each risk representation are labeled, and quality score labels are generated for the entire narrative description, forming a causal relationship label set and a quality score label set.

[0138] Specifically, when labeling causal relationships, it is necessary to clarify the triggering and being triggered relationships between various risk characteristics. For example, label the logical chain of "abnormal main ventilator current - decreased ventilation efficiency - gas accumulation - explosion caused by ignition source" to ensure that the model can learn the inherent laws of risk evolution.

[0139] In addition, the quality rating labels are assigned values ​​based on dimensions such as the completeness of the narrative description, the causal logic, and the accuracy of the risk information. For example, samples with clear logic and complete information are labeled with a high score of 0.9, while samples with missing information or logical contradictions are labeled with a low score of 0.3.

[0140] Furthermore, a network architecture for building a large language model based on machine learning is constructed. Considering the needs of coal mine gas explosion risk simulation, a deep learning network adapted to text generation and logical reasoning is selected as the basic architecture to effectively capture the quantitative characteristics of risk entropy and the semantic features of risk representation, thus achieving a mapping from input data to risk narrative description.

[0141] Specifically, the input layer of the network architecture is set to a dual-branch channel. The first branch channel is used to receive the gas explosion risk entropy value. The fully connected layer converts the quantized index into a feature vector that the model can recognize. The fully connected layer has two hidden layers. The first layer has 128 neurons and uses ReLU as the activation function. The second layer has 64 neurons and uses ReLU as the activation function to ensure the deep extraction of quantized features.

[0142] In addition, the second branch channel receives three types of risk representations. The semantic information is transformed into feature vectors through the text embedding layer. The embedding dimension is set to 256. The Word2Vec algorithm is used for pre-training initialization and then dynamically optimized through training. The feature vectors of the two branch channels are concatenated and fused before being input into the subsequent network layer. The dimension of the fused feature vector is 320 (64+256).

[0143] Furthermore, the output layer is configured as a text generation channel, using an autoregressive decoding method to generate a coherent risk narrative description. The decoding layer adopts a Transformer decoder structure, with 6 decoding layers. Each decoding layer includes a multi-head attention mechanism (8 heads), a feedforward neural network (512 hidden layer dimensions), and layer normalization and residual connections are added to ensure that the output results conform to natural language expression habits. The maximum length of the generated text is limited to 512 characters.

[0144] During the architecture construction process, network parameters need to be set reasonably according to the scale and complexity of the sample data. For example, when the sample scale reaches more than 50,000 and covers 20 types of risk scenarios, the decoding layer can be increased to 8 layers and the number of multi-head attention heads can be set to 12 to improve the feature learning ability of the model. When the sample scale is less than 10,000 and covers less than 10 types of risk scenarios, the decoding layer can be simplified to 4 layers and the number of multi-head attention heads can be set to 4. At the same time, the dimension of the hidden layer of the feedforward neural network can be reduced to 256 to avoid overfitting.

[0145] Furthermore, a large language model was trained under supervised supervision using a training sample set, a causal relationship label set, and a quality score label set. During training, the training, validation, and test sets were divided in a 7:2:1 ratio, with a batch size of 32 and an initial learning rate of 0.001. A cosine annealing strategy was used to adjust the learning rate, which was reduced to 0.8 times its current value every 10 rounds. The gas explosion risk entropy value and the three types of risk representations from the samples were input into the model, which generated risk narrative descriptions. The generated results were then compared with the real narrative descriptions, causal relationship labels, and quality score labels from the samples.

[0146] Furthermore, the difference between the generated results and the true labels is quantified by calculating the cross-entropy loss function, where the text generation loss weight is set to 0.6, the causal relationship matching loss weight is set to 0.3, and the quality score loss weight is set to 0.1. The total loss is the weighted sum of the three. The network parameters are continuously adjusted by combining the Adam optimizer and the gradient descent algorithm, with the gradient clipping threshold set to 1.0 to avoid gradient explosion and optimize the model's inference and generation capabilities.

[0147] During training, the validation set is used to monitor model performance. When the risk narrative descriptions generated by the model on the validation set meet the preset standards in causal accuracy (accuracy ≥ 97%) and information completeness (recall ≥ 88%), and remain stable for 5 consecutive rounds without significant decline, training is stopped and the model pre-training is completed.

[0148] During supervised training, the training effect needs to be verified and optimized periodically. For example, after every 20 rounds of training, the model performance should be tested using validation set data. If the generated narrative description is found to have causal logic inconsistencies, the network's attention mechanism parameters need to be adjusted, such as changing the dropout rate of multi-head attention from 0.1 to 0.2.

[0149] Meanwhile, the weight of causal relationship loss is increased to 0.4 to strengthen the model's learning of causal relationships. If the model's inference accuracy for a certain type of risk scenario is low, such as the poor performance in handling the "sudden failure of the ventilation system" scenario, 5,000 sample data of this type of scenario are added, the dataset is re-divided proportionally, and training continues until the inference accuracy of this type of scenario is ≥88%.

[0150] Ultimately, the pre-trained large language model can accurately receive the gas explosion risk entropy value and the corresponding risk representation. Based on the learned risk evolution law and causal relationship, it can perform in-depth inference and generate a risk narrative description that includes the gas diffusion path and detonation probability.

[0151] S140: Perform scenario-based risk pattern matching and deconstruction on the risk narrative description and the safety event narrative flow, identify and associate key risk primitives and composite risk structures, and dynamically generate early warning levels and response strategy plans.

[0152] In this embodiment of the application, in order to accurately locate the core risk sources and evolution logic, and to ensure that the warning level matches the actual risk level and the handling strategy is targeted, it is necessary to mine key risk information through scenario matching and deconstruction, construct a risk structure and transform it into an executable warning and handling plan, so as to improve the effectiveness of safety warning and the timeliness of emergency response.

[0153] Specifically, the risk narrative is first analyzed to extract textual descriptions of gas diffusion paths and detonation probabilities, forming an abstract risk model to be matched. By sorting out the risk evolution descriptions in the text, the direction of risk propagation, scope of impact, and potential degree of harm are clarified, laying the foundation for subsequent matching with actual scenarios.

[0154] Furthermore, the abstract risk model is compared and mapped one by one with the gas outburst anomaly, ventilation dilution capacity status, and ignition source risk events in the safety incident narrative. By associating the event descriptions in the related text with the core features of the abstract model, a correspondence between the two is established to ensure that the risk model accurately reflects the actual monitored safety events.

[0155] Furthermore, based on the established mapping relationship and chain reaction path, the first, second, and third types of risk representations are identified as gas accumulation units, ventilation failure units, and ignition source units, respectively. By clarifying the core attributes of each type of risk representation, the basic units constituting the overall risk are extracted, clearly presenting the core components of the risk.

[0156] Furthermore, by combining the transmission and impact chain of the risk narrative, the interaction relationships between the gas accumulation unit, the ventilation failure unit, and the ignition source unit are established, forming a complex risk structure. By sorting out the triggering, aggravation, or correlation logic between the units, the process of risk from generation to evolution is fully presented.

[0157] After constructing the composite risk structure, the basic warning level is determined based on the number of gas accumulation units, ventilation failure units, and ignition source units within it. A higher number of units indicates a wider risk coverage and more prominent hidden dangers, resulting in a higher basic warning level.

[0158] Based on the magnitude of the gas explosion risk entropy value, the basic warning level is revised to generate the final warning level. The gas explosion risk entropy value, as a core indicator for quantifying the degree of risk, objectively reflects the severity of the risk. Fine-tuning the basic warning level using this value makes the warning results more closely reflect the actual risk situation.

[0159] This step, through scenario-based matching and risk deconstruction, sorts out the core components and evolution logic of risks, making the generated early warning levels scientific and reliable and the response strategies highly targeted, providing a basis for action for the rapid control and emergency response to underground gas explosion risks in coal mines.

[0160] Step S140 in the method provided in this application embodiment includes:

[0161] The risk narrative description is analyzed, and the textual expressions of the gas diffusion path and the detonation probability are extracted as abstract risk patterns to be matched.

[0162] The abstract risk pattern is compared and mapped one by one with the gas outburst anomaly, ventilation dilution capacity status and ignition source risk event in the safety event narrative stream;

[0163] Based on the mapping relationship and the chain reaction path, the first type of risk characterization, the second type of risk characterization and the third type of risk characterization are respectively identified as gas accumulation unit, ventilation failure unit and ignition source unit;

[0164] By combining the transmission and impact chain described in the risk narrative, the interaction relationship between the gas accumulation unit, the ventilation failure unit, and the ignition source unit is established to form a composite risk structure.

[0165] Based on the number of gas accumulation units, ventilation failure units, and ignition source units in the composite risk structure, the basic early warning level is determined.

[0166] Based on the magnitude of the gas explosion risk entropy value, the basic warning level is corrected to generate the final warning level;

[0167] For the gas accumulation unit, a handling instruction is generated, including adjusting the main ventilation fan operation data to enhance ventilation;

[0168] For the aforementioned ignition source unit, a disposal instruction is generated that includes cutting off power to the area involved in the ignition source risk event and extinguishing open flames;

[0169] For the identified ignition source risk events and their impact range, evacuation instructions for personnel within the location coordinates of the organization are generated, and all instructions are integrated to form the aforementioned emergency response strategy plan.

[0170] In this embodiment of the application, in order to ensure that the warning level matches the actual risk level and generate a targeted and feasible emergency response plan, it is necessary to improve the accuracy of safety warnings and the timeliness of emergency response by matching scenario-based risk patterns, identifying risk primitives, constructing composite risk structures, and combining quantitative indicators to correct the warning level and integrate response instructions.

[0171] Specifically, the risk narrative description is first analyzed to extract textual expressions of gas diffusion paths and detonation probabilities, which serve as abstract risk patterns to be matched. The risk narrative description encompasses the entire risk development process deduced by a large language model. Through semantic parsing techniques, core information is extracted to clarify the specific paths, diffusion speeds, and impact ranges of gas diffusion from the accumulation area to the surrounding areas. Simultaneously, key expressions such as detonation probability intervals and triggering conditions are extracted to form a structured abstract risk pattern.

[0172] For example, the core content of "gas spreads along the return airway of working face No. 3 to auxiliary roadway No. 5 at a speed of 0.28 m / s, and the affected area covers working face No. 3, auxiliary roadway No. 5 and a 10-meter radius around it; there is an illegal ignition source in the current area, with an ignition probability of 85%-95%" is extracted from the risk narrative description to construct the corresponding abstract risk model.

[0173] Furthermore, the abstract risk model is compared and mapped one by one with the gas outburst anomaly, ventilation dilution capacity status, and ignition source risk events in the safety incident narrative.

[0174] Specifically, the security incident narrative stream is constructed based on actual monitoring data and includes various real-world risk-related events. By comparing the core characteristics of the abstract risk pattern with the event descriptions in the narrative stream, a one-to-one correspondence is established.

[0175] For example, the "gas diffusion" feature in the abstract risk model matches the abnormal gas outburst event in the narrative stream where "the gas concentration in area 3 rises to 0.72% (preset threshold 0.5%)", the "high probability of explosion" feature matches the ignition source risk event in the narrative stream where "video surveillance captures illegal hot work in the southeast corner of work face 3", and the "diffusion affected by ventilation" feature matches the insufficient ventilation dilution capacity in the narrative stream where "the main ventilation fan operates at a current of 28A (normal range 30A-40A)".

[0176] Furthermore, based on the established mapping relationship and the previously identified chain reaction path, the first type of risk characterization, the second type of risk characterization, and the third type of risk characterization are identified as gas accumulation unit, ventilation failure unit, and ignition source unit, respectively.

[0177] Among them, risk primitives are the basic units that constitute the overall risk, and are classified and counted by clarifying the core attributes of various risk characteristics. For example, the three statements in the first category of risk characteristics, namely "gas concentration 0.72% exceeding the threshold, carbon monoxide concentration 0.0031% exceeding the threshold, and temperature 36.2℃ exceeding the threshold", are identified as three gas accumulation primitives; the two descriptions in the second category of risk characteristics, namely "main ventilation fan operating current 28A below the normal range and vibration amplitude 0.10mm exceeding the standard threshold", are identified as two ventilation failure primitives; and the two contents in the third category of risk characteristics, namely "eight people abnormally gathering in the gas accumulation area and illegal hot work", are identified as two ignition source primitives.

[0178] Furthermore, by combining the transmission and impact chain of the risk narrative, the interaction relationship between the gas accumulation unit, the ventilation failure unit, and the ignition source unit is established to form a composite risk structure.

[0179] The transmission chain clearly presents the logical sequence of risk from its generation to its evolution, thereby clarifying the triggering and aggravating relationships between the basic elements. For example, two ventilation failure elements lead to a 30% decrease in ventilation dilution capacity, which in turn increases the risk level of three gas accumulation elements. Since the three gas accumulation elements and the two ignition source elements are less than 5 meters apart in space, a complex risk structure of "ventilation failure - aggravated gas accumulation - ignition source triggering explosion at close range" is formed, fully presenting the overall evolution framework of the risk.

[0180] Furthermore, after constructing the composite risk structure, the basic early warning level is determined based on the number of gas accumulation units, ventilation failure units, and ignition source units within it. The basic early warning level is divided into four levels according to the combination and distribution of units, with clear quantitative judgment criteria: Level 1 is determined when only one type of unit exists (quantity ≥ 1); Level 2 is determined when two types of key units exist simultaneously (quantity ≥ 1 of each type); Level 3 is determined when all three types of units exist simultaneously, and the quantity of one type of unit is ≤ 1 or its location is ≥ 10 meters away from other units; Level 4 is determined when all three types of units exist simultaneously, and the number of gas accumulation units is ≥ 2, the number of ignition source units is ≥ 2, and the spatial distance between the two types of units is ≤ 5 meters.

[0181] For example, if there are only 3 gas accumulation units in the composite risk structure and no other types of units, the basic warning level is Level 1; if there are 2 gas accumulation units and 1 ignition source unit, the basic warning level is Level 2; if all three types of units exist, but there is only 1 ignition source unit and it is 12 meters away from the gas accumulation unit, the basic warning level is Level 3; if all three types of units exist, and there are 3 gas accumulation units and 2 ignition source units, distributed in the same work area and spaced 3 meters apart, the basic warning level is Level 4.

[0182] Furthermore, the basic warning level is adjusted based on the magnitude of the gas explosion risk entropy value to generate the final warning level. Specifically, the gas explosion risk entropy value is divided into three ranges: low (0-2.0), medium (2.1-3.5), and high (3.6 and above). The adjustment rules are clear: if the entropy value is in the low range, the basic warning level is slightly adjusted downward by 1 level (minimum is level 1); if the entropy value is in the medium range, the basic warning level is maintained; if the entropy value is in the high range, the basic warning level is adjusted upward by 1 level (maximum is level 4).

[0183] For example, if the basic warning level is level three and the gas explosion risk entropy value is 3.8 (high range), it needs to be revised upward to level four; if the basic warning level is level two and the entropy value is 1.8 (low range), it needs to be slightly adjusted downward to level one; if the basic warning level is level three and the entropy value is 2.8 (medium range), it remains level three; if the basic warning level is level four and the entropy value is 3.7 (high range), it remains level four.

[0184] Furthermore, for different types of risk elements, specific handling instructions with clear parameters are generated. For the gas accumulation element, the core objective is to reduce the gas concentration to below 0.5% within 30 minutes, generating instructions to adjust the operating parameters of the main ventilation fan, such as increasing the operating current from 28A to 35A and stabilizing the voltage at 380V, thereby increasing the ventilation volume by 40% and enhancing the ventilation dilution effect.

[0185] In addition, regarding the ignition source, the ignition source must be eliminated within 10 minutes, and an order must be generated to cut off the power to the area involved by the ignition source in the southeast corner of the No. 3 work face, and organize two full-time personnel to carry fire extinguishing equipment to extinguish the open flame on site.

[0186] In addition, based on the risk of ignition sources and their impact range, an instruction was generated to organize all personnel (a total of 23 people) within 15 minutes from work face No. 3, auxiliary roadway No. 5, and a 10-meter radius around them to evacuate to a safe area on the ground along a pre-set evacuation route. Finally, the response strategy was integrated into a logically clear and parameter-defined emergency response plan, prioritizing the response to "eliminating the ignition source - enhancing ventilation - evacuating personnel".

[0187] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects:

[0188] This application proposes a large language model-enabled intelligent early warning method for coal mine safety risks. First, it acquires multi-source heterogeneous monitoring data covering the environmental, equipment, and spatial dimensions of the target area underground in the coal mine. Based on the physical causal relationship between gas emission, ventilation dilution, and ignition source occurrence, various monitoring data are transformed into related events and integrated into a safety event narrative stream, ensuring the coherence and completeness of risk information. Next, the safety event narrative stream is semantically parsed to extract three types of risk representations and connect them to form a chain reaction path. The gas explosion risk entropy value is calculated by integrating the underground dynamic environmental entropy and the rationality compensation coefficient, achieving a quantitative representation of the risk. Subsequently, the risk entropy value and corresponding risk representations are input into a pre-trained large language model to generate a risk narrative description containing gas diffusion paths and ignition possibilities. Finally, the risk narrative description and the safety event narrative stream are matched and deconstructed using scenario-based risk patterns to identify three types of risk primitives: gas accumulation, ventilation failure, and ignition sources, and a composite risk structure is constructed. Early warning levels are dynamically generated based on the number of primitives and the risk entropy value. Targeted handling instructions are formulated for different risk primitives and integrated into a contingency plan, completing the intelligent management and control of the entire process from risk perception to early warning and handling.

[0189] The method provided in this application, through the technical solution of "multi-source data integration - risk quantification and characterization - large language model deduction - risk structure deconstruction - early warning and response generation", solves the problems of data fragmentation, unclear risk evolution logic, low early warning accuracy, and insufficient targeted response strategies in traditional coal mine safety early warning. It realizes intelligent management and control of the entire chain of underground gas explosion risks in coal mines, and improves the timeliness, accuracy and operability of safety early warning and emergency response.

Claims

1. A method for intelligent early warning of coal mine safety risks empowered by a large language model, characterized in that: The method includes: acquiring multi-source heterogeneous monitoring data of a target area in a coal mine, and constructing a safety event narrative stream based on the physical causal relationship between gas emission, ventilation dilution, and ignition source occurrence; performing semantic analysis on the safety event narrative stream to extract risk representations and chain reaction paths, and integrating the underground dynamic environmental entropy to calculate the gas explosion risk entropy value. The semantic analysis of the safety event narrative stream to extract risk representations and chain reaction paths includes: identifying and extracting statements about gas concentration, carbon monoxide concentration, and temperature exceeding preset thresholds from the safety event narrative stream as first-type risk representations; and extracting information from the safety event narrative stream regarding the start / stop status of the main ventilation fan and its operating current. The abnormal state descriptions of voltage and vibration amplitude are used as the second type of risk characterization; descriptions of abnormal clustering of personnel positioning coordinates, abnormal changes in equipment location information, or violations of operation and open flames identified in video surveillance images are extracted from the safety event narrative stream and used as the third type of risk characterization; based on the chronological order of events and causal correlation words in the safety event narrative stream, the first, second, and third types of risk characterizations are linked together to generate a chain reaction path; the gas explosion risk entropy value and the corresponding risk characterization are input into a pre-trained large language model for risk inference to generate a risk narrative description containing the gas diffusion path and detonation probability; the risk narrative description and the safety event narrative stream are then subjected to scenario-based risk assessment. Risk pattern matching and deconstruction identifies and associates key risk primitives and complex risk structures, dynamically generating early warning levels and response strategy plans. This includes: parsing the risk narrative description, extracting textual descriptions of the gas diffusion path and the ignition probability as abstract risk patterns to be matched; comparing and mapping these abstract risk patterns one by one with gas outburst anomalies, ventilation dilution capacity status, and ignition source risk events in the safety event narrative flow; based on the mapping relationship and the chain reaction path, identifying the first, second, and third types of risk representations as gas accumulation primitives, ventilation failure primitives, and ignition source primitives, respectively; and combining the transmission and influence chain of the risk narrative description to establish the gas accumulation primitives, ventilation failure primitives, and ignition source primitives. The interaction between failure elements and ignition source elements forms a composite risk structure. Based on the number of gas accumulation elements, ventilation failure elements, and ignition source elements in the composite risk structure, a basic warning level is determined. The basic warning level is modified according to the magnitude of the gas explosion risk entropy value to generate a final warning level. For the gas accumulation elements, a disposal instruction including adjusting the main ventilation fan operation data to enhance ventilation is generated. For the ignition source elements, a disposal instruction including cutting off power and extinguishing open flames in the area involved in the ignition source risk event is generated. For the identified ignition source risk event and its impact range, a disposal instruction for evacuating personnel within the personnel positioning coordinate range is generated. All instructions are integrated to form the disposal strategy plan.

2. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 1, characterized in that, The multi-source heterogeneous monitoring data includes: time-series data collected by environmental safety sensors, wherein the time-series data includes at least gas concentration, carbon monoxide concentration, and temperature; main ventilation fan operation data reported by the production equipment monitoring system, wherein the main ventilation fan operation data includes at least start / stop status, operating current, voltage, and vibration amplitude; and positioning and image data generated by the spatial perception system, wherein the positioning and image data includes personnel positioning coordinates, equipment location information, and video surveillance images of key areas.

3. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 1, characterized in that, Based on the physical causal relationship between gas outbursts, ventilation dilution, and ignition source occurrence, the multi-source heterogeneous monitoring data is constructed into a safety event narrative stream, including: identifying abnormal gas outburst events based on time-series data; determining the ventilation dilution capacity status based on main ventilator operation data and associating it with the abnormal gas outburst events; identifying ignition source risk events based on location and image data, and spatiotemporally associating them with the abnormal gas outburst events and the ventilation dilution capacity status; and integrating the abnormal gas outburst events, ventilation dilution capacity status, and ignition source risk events according to time series and causal relationships to generate a safety event narrative stream.

4. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 3, characterized in that, The gas emission anomaly, ventilation dilution capacity status, and ignition source risk event are integrated based on time sequence and causal relationship to generate a safety event narrative flow. This includes: determining the timestamps corresponding to each of the gas emission anomaly, ventilation dilution capacity status, and ignition source risk event, and sorting the events according to their chronological order; identifying and establishing the causal relationship between the ventilation dilution capacity status and the gas emission anomaly, as well as the temporal and spatial overlap between the gas emission anomaly and the ignition source risk event; and integrating the gas emission anomaly, ventilation dilution capacity status, and ignition source risk event into a coherent textual description with cause, development, and potential consequences, as the safety event narrative flow.

5. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 1, characterized in that, The calculation of the gas explosion risk entropy value by integrating the underground dynamic environmental entropy includes: analyzing the spatiotemporal correlation between ventilation capacity and gas concentration changes within a preset monitoring period based on the first and second types of risk characterization; calculating the matching degree between the ventilation capacity and gas concentration changes according to the physical law of gas dilution by ventilation, as a reasonable compensation coefficient; obtaining the underground dynamic environmental entropy benchmark value calculated based on the stability of the ventilation network and the influence of temperature, and correcting the underground dynamic environmental entropy benchmark value using the reasonable compensation coefficient to obtain the corrected underground dynamic environmental entropy; and combining the corrected underground dynamic environmental entropy, the number of the three types of risk characterizations, and the complexity of the chain reaction path to calculate the gas explosion risk entropy value.

6. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 5, characterized in that, Based on the physical laws governing the dilution of methane by ventilation, the matching degree between the ventilation capacity and the change in methane concentration is calculated as a reasonable compensation coefficient. This includes: calculating real-time power based on the operating current and voltage of the main ventilation fan, and calculating the ventilation efficiency index in conjunction with the vibration amplitude; extracting historical data points of the methane concentration, calculating the slope of the methane concentration within a preset monitoring period as the methane concentration change trend; comparing the ventilation efficiency index with the methane concentration change trend to determine the physical consistency of their change directions, and obtaining the reasonable compensation coefficient. The reasonable compensation coefficient is obtained as follows: when the ventilation efficiency index is higher than the normal operating threshold and the methane concentration change trend has a negative slope, the reasonable compensation coefficient is set to a first preset value; when the ventilation efficiency index is lower than the normal operating threshold and the methane concentration change trend has a positive slope, the reasonable compensation coefficient is set to a second preset value; otherwise, the reasonable compensation coefficient is set to a third preset value; the first preset value is greater than the third preset value, and the third preset value is greater than the second preset value.

7. The intelligent early warning method for coal mine safety risks empowered by a large language model according to claim 1, characterized in that, The pre-training steps of the large language model include: collecting gas explosion risk scenario data containing clear causal chains based on the coal mine historical safety database and simulation case library to form a training sample set; labeling the causal relationships between various risk representations for each sample in the training sample set, and generating quality score labels for the entire narrative description to form a causal relationship label set and a quality score label set; constructing the network architecture of the large language model based on machine learning, wherein the input layer of the network architecture is set to a dual-branch channel, the first branch channel is used to receive the gas explosion risk entropy value, and the second branch channel is used to receive three types of risk representations, and the output layer of the network architecture is configured as a text generation channel; using the training sample set, causal relationship label set, and quality score label set, the large language model is subjected to supervised training until the risk narrative description generated by the model reaches a preset standard, thus completing the pre-training of the model.

Images