Goaf coal spontaneous combustion multi-source partition and risk prediction method

Abstract

According to the goaf coal spontaneous combustion multi-source partitioning and risk prediction method provided by the invention, aiming at the defects that traditional three-zone partitioning is insensitive to random air leakage, lacks dynamic nature, is easy to distort in measuring point extrapolation and the like, a verifiable potential danger zone is newly added, and a hidden air leakage channel is identified through an air leakage significance index; channel opening and closing parameterization updating is achieved through the air leakage migration rate; multi-dimensional indexes are fused to generate continuous risk indexes, and robust judgment of dangerous partitions is achieved; a gating multi-mode network is constructed to output a future partition probability and risk map, disaster early warning is achieved in combination with the risk growth rate, a complete disaster prevention closed loop is formed, and the spontaneous combustion disaster recognition and prediction precision is greatly improved. According to the invention, blind areas can be effectively reduced, hidden cracks can be positioned, and recognition precision, prediction accuracy and reliability of spontaneous combustion disasters in the goaf can be significantly improved.

Description

Technical Field

[0001] This invention relates to the field of mine safety and intelligent monitoring and early warning technology, and in particular to a method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas. Background Technology

[0002] Spontaneous combustion of coal in goaf areas is a typical highly hazardous and strongly coupled disaster in coal mines. Traditional methods, based on a three-zone model (heat dissipation zone, oxidation-heating zone, and asphyxiation zone), monitor and provide early warning through fixed measuring points and numerical simulations, offering some guidance under normal circumstances. However, with increasingly complex mining conditions, the development of random fractures within the goaf leads to variable air leakage channels, and localized oxygen enrichment and high-temperature points exhibit new characteristics of random occurrence and rapid migration, breaking the traditional understanding of the relatively stable spatial distribution of the three zones. Field practice shows that high-temperature points randomly appear in the roof delamination zone, fracture enrichment zone, old goaf connection points, coal pile accumulation areas, and working face transition areas. The spatiotemporal uncertainties in these areas result in significant errors in the three-zone extrapolation method based on a limited number of measuring points, leading to insufficient reliability of early warning.

[0003] Despite the widespread application of technologies such as the Internet of Things, fiber optic temperature measurement, infrared thermal imaging, and microseismic monitoring, existing systems still face challenges such as difficulties in multi-source data fusion, lagging model updates, and insufficient ability to identify random air leakage channels, making it difficult to achieve dynamic zoning and accurate prediction. Traditional three-zone systems—heat dissipation zone, oxidation-heating zone, and asphyxiation zone—are relatively stable by default, making it difficult to characterize the rapid coupling effect of random air leakage channels leading to localized oxygen enrichment and early temperature rise, resulting in the random appearance and migration of high-temperature points. Point-to-surface extrapolation relying on a limited number of measurement points leads to large errors; model parameters cannot be corrected online with mining activities and tectonic evolution, resulting in delayed early warnings, insufficient interpretability, and difficulty in supporting closed-loop governance.

[0004] Therefore, there is an urgent need for a new method that can dynamically identify potential hazardous areas, integrate multi-source data, and achieve closed-loop management. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas.

[0006] To achieve the above objectives, the technical solution of this invention is as follows: In a first aspect, the present invention provides a method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas, the method comprising: The raw multi-source data of the goaf are acquired and preprocessed to obtain standardized multi-source data with a unified spatiotemporal reference; the standardized multi-source data includes at least gas volume fraction time series, temperature change rate and infrared thermal image data. Based on the standardized multi-source data, the air leakage significance index is calculated; if the air leakage significance index continues to exceed the preset index threshold, the air leakage mobility rate is updated online and spatially smoothed based on the air leakage significance index to obtain the updated air leakage mobility rate. Based on the standardized multi-source data, the leakage significance index, and the updated leakage mobility, cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations are calculated and fused to generate a continuous risk index. Based on the continuous risk index, the gas volume fraction time series, the temperature change rate, the air leakage significance index, and the thermal image significance, combined with the smoothing and hysteresis threshold rules of risk inertia, a mutually exclusive three-zone and one-region distribution map is dynamically divided and output; the three-zone and one-region includes the potential danger zone, the oxidation heating zone, the asphyxiation zone, and the heat dissipation zone. The standardized multi-source data is input into the FOUR-Net model, which uses the leakage significance index and the cross-modal uncertainty as gate signals, and is then weighted, fused, and risk predicted in sequence to obtain a three-zone one-region probability map and a risk prediction map. The continuous risk index, the three-zone one-district probability map, and the risk prediction map trigger tiered early warnings and implement governance measures. After the governance measures are implemented, iterative monitoring continues, and the governance effect is verified by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, and the continuous risk index.

[0007] In some embodiments, calculating the air leakage significance index based on the standardized multi-source data includes: Based on the rate of change of the ratio of oxygen to nitrogen gas volume fractions in the time series, combined with the rate of temperature change, a preliminary index of air leakage intensity is calculated. The air leakage intensity preliminary index is smoothed over a time window and spatially interpolated to obtain the air leakage significance index; the formula for calculating the air leakage significance index is as follows: , In the formula, , These are the measured predicted wind speed and the model predicted wind speed, respectively. For wind speed reference, For microseismic energy density, This serves as a reference for microseismic energy density. For pressure gradient mode, As a reference value for the pressure gradient modulus, This is the measured oxygen volume fraction. To predict the oxygen volume fraction, The oxygen fraction is the reference value, and α1, α2, α3, and α4 are the weights, respectively.i ≥0, ∑α i =1 represents the weight, and Logi(⋅)=1 / (1+e−z) represents the logistic mapping.

[0008] In some embodiments, when performing residual-driven online updates and spatial smoothing of the leakage mobility, the online update formula is: , In the formula, For a moment In position The update did not smooth out the leakage migration rate. For spatial location, For time, For discrete time step index, for Always in position The updated air leakage migration rate, This is a truncation function. The trigger threshold for the air leakage significance index, To update the step size coefficient, For a moment The air leakage significance index, These are the lower and upper limits of the mobility, respectively; The formula for spatial smoothing is: , In the formula, for Always in position The final leakage migration rate after smoothing To smooth out the weights, For spatial neighborhood set, The number of grid points in the neighborhood. For the location index within the neighborhood, For the previous moment Neighborhood Location The air leakage migration rate.

[0009] In some embodiments, the formula for calculating the cross-modal uncertainty is: , In the formula, For cross-modal uncertainty, For spatial location, For time, For the amplitude limiting operator, For the multimodal normalized residual vector, for The 2-norm, The residual metric; The formula for calculating the significance of the thermal image is: , In the formula, For thermal image saliency, This represents the equivalent temperature field of the infrared thermal image in the downhole physical coordinate domain. Let be the spatial gradient vector of the temperature field. Let L be the L2 norm of the spatial gradient vector of the temperature field. This is the gradient smoothing / stability constant; The formula for calculating the significance of the potential hazard is as follows: , In the formula, For the significance of potential hazards, , , , For different weighting coefficients, and ≥0、 , , These represent the measured oxygen volume fraction and CO volume fraction, respectively. , These are the threshold values ​​for oxygen volume fraction and CO volume fraction, respectively. For the rate of temperature rise, For reference temperature rise rate, For positive part operators, The significance index of air leakage; The formula for calculating the short-window gas fluctuation is as follows: , In the formula, Here, s represents the short-window gas fluctuation index, and s is the index of the gas type. For the set of gases participating in the statistics, For gas weighting coefficients, Let be the average volume fraction of gas s within a short time window. Let S be the standard deviation of the volume fraction of gas s within the same short time window. To prevent zero constant / stability constant.

[0010] In some embodiments, the continuous risk index is: , In the formula, It is a continuous risk index. These are the fusion weighting coefficients for cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations, respectively.

[0011] In some embodiments, the exponential smoothing formula for risk inertia is: , In the formula, Due to risk inertia, The risk inertia of the previous moment, For smoothing coefficients, For spatial location, It is a continuous risk index. For discrete time step index; Based on the risk inertia and the entry / exit dual threshold hysteresis rule, the mutual exclusion determination of the three zones and one area is performed to obtain the distribution map of the three zones and one area.

[0012] In some embodiments, a spatial location is determined to be the asphyxiation zone if and only if all of the following conditions are met simultaneously: Physical conditions: The oxygen volume fraction is less than or equal to the asphyxiation threshold; and the absolute value of the temperature change rate is less than or equal to a small threshold close to zero; and the carbon monoxide volume fraction is low; Spatial constraints: The distance of this spatial location along the working face advancement direction is greater than the distance of all units in that direction that have been identified as heat dissipation zones or oxidation heating zones. The location closer to the working surface shall be preferentially identified as the heat dissipation zone if and only if a spatial location simultaneously satisfies all of the following physical conditions, and if it conflicts with the determination of other zones in the same direction: Physical conditions: oxygen volume fraction is higher than the oxygen-enriched threshold; and the rate of temperature change is less than or equal to the temperature rise slope threshold; and carbon monoxide volume fraction is less than or equal to the low CO threshold. A spatial location is identified as a potential danger zone when it simultaneously meets all of the following conditions: If the leakage significance index is greater than or equal to the entry threshold, the oxygen volume fraction is higher than the oxygen enrichment threshold, the temperature change rate is greater than the temperature rise slope threshold, or the infrared thermal image significance is greater than or equal to the thermal image significance threshold, and the carbon monoxide volume fraction is less than the oxidation heating threshold; if the leakage significance index subsequently drops below the exit threshold, or the risk inertia drops below the exit threshold, then the current spatial location exits the potential danger zone. A spatial location is identified as the oxidation heating zone when all of the following conditions are met simultaneously: Physical conditions: The rate of temperature change is greater than the temperature rise slope threshold, and the carbon monoxide volume fraction is greater than or equal to the oxidation temperature rise CO threshold, and the risk inertia is greater than or equal to the entry threshold of the air leakage significance index; Spatial distribution characteristics: The oxidation temperature rise zone is mainly distributed in the area between the heat dissipation zone and the asphyxiation zone.

[0013] In some embodiments, the FOUR-Net model is trained in the following manner: Based on the three-zone-one-region judgment rule, a partition label is generated for each spatiotemporal grid, and a continuous risk index label is generated based on the continuous risk formula; samples are constructed in a sliding time window manner, where the input of each sample is a spatiotemporal tensor composed of multi-source data and infrared thermal image frames stacked within the time window; the sample set is divided into a training set, a validation set, and a test set. The obtained sample standardized multi-source data is input into the feature extraction module of the FOUR-Net model. The spatial branch extraction module performs two-dimensional convolution processing on the infrared thermal image sequence to extract the thermal image spatial features. The temporal branch extraction module performs one-dimensional convolution and cyclic modeling on the multi-sensor temporal data to extract multi-sensor temporal features. Using the air leakage significance index and the cross-modal uncertainty as gating signals, the gating signals are concatenated with the channel-pooled representations of the thermal spatial features and the multi-sensor temporal features. A gating map is then generated using a 1×1 convolution and a sigmoid activation function. The gating map is as follows: , In the formula, For gating graphs, For logical mapping functions, For 1×1 convolution, For thermal imaging spatial features, For multi-sensor timing characteristics, The significance index of air leakage, For channel dimension splicing operators, For uncertainty plot, Here, H is the height of the network feature map, and W is the width of the network feature map; The gated image is fused with the thermal spatial features and the multi-sensor temporal features to obtain a fused feature vector; the fused feature vector is: , In the formula, To fuse feature vectors, For point-by-point multiplication, 1−M is a gated complementary graph; The fused feature vector is input into subsequent convolutional and fully connected layers to output the three-band, one-division partitioning probability map and the risk prediction map. Construct a total loss function; the total loss function introduces a gated consistency regularization constraint; the gated consistency regularization term is: , In the formula, For gating consistency regularization, These are the gating regularization weight coefficients. For the average operator, For logical mapping functions, The linear coefficient of the air leakage significance index is . The linear coefficients of the cross-modal uncertainty are... For bias terms; The optimal model after training is evaluated using the test set. The accuracy, recall, and F1 score of the three-band-one-zone classification, as well as the mean square error and mean absolute error of risk prediction, are calculated to verify the model's ability to identify dangerous areas and the accuracy of risk prediction.

[0014] In some embodiments, the tiered early warning and remediation measures are triggered by the continuous risk index, the three-zone-one-region probability map, and the risk prediction map; after the remediation measures are implemented, iterative monitoring continues, and the remediation effect is verified by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, and the continuous risk index, including: Based on the continuous risk index and the risk prediction map, and combined with the three-zone one-district zoning probability map, the risk areas are spatially located and classified. When the continuous risk index exceeds the preset classification threshold or the risk growth rate in the risk prediction map exceeds the set threshold, a classification warning signal is triggered, and a location management instruction is generated by combining the potential danger zone and the high probability area of ​​the oxidation heating zone. After implementing inertia injection, sealing, or ventilation optimization measures, continue to collect multi-source data and iterate. Quantitatively evaluate the treatment effect by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, the continuous risk index, and the proportion of dangerous zones in the zoning probability map. If the updated air leakage migration rate decreases, the air leakage significance index falls, the continuous risk index decreases, and the proportion of the potential danger zone and the oxidation heating zone is significantly reduced, then the treatment is deemed effective; otherwise, the treatment strategy is readjusted.

[0015] The multi-source zoning and risk prediction method for spontaneous combustion of coal in goaf provided by this invention innovatively incorporates potential hazard zones as independent zones into a dynamic three-zone-one-zone zoning framework. By defining the PHZ calculable, using mutually exclusive judgment logic with the other three zones, and combining risk inertia and entry / exit dual-threshold hysteresis mechanism, this invention fills the gap in the traditional three-zone zoning for identifying the high-risk incubation stage of spontaneous combustion of coal. At the same time, it ensures the stability and operability of the zoning results at the legal and engineering implementation levels, and achieves full-cycle coverage of spontaneous combustion risk from the explicit stage to the high-risk incubation stage. Secondly, the leakage significance index L constructed in this invention, through dimensionless fusion of multi-source evidence and Sigmoid mapping, condenses multi-dimensional monitoring clues of hidden leakage channels into a unified quantitative indicator. Furthermore, L can simultaneously serve PHZ determination, D4Z-Risk calculation, and FOUR-Net gated fusion, achieving multi-purpose coupling of a single indicator across the entire chain of determination, prediction, and control. The online update scheme for leakage mobility λ based on L enables continuous tracking and denoising of leakage channel states without complex filters, significantly reducing the technical threshold and cost for engineering implementation. Thirdly, the cross-modal uncertainty and thermal image significance defined in this invention provide support for risk assessment from the dimensions of evidence consistency and thermal boundary location, respectively. Together with L, they constitute a three-dimensional feature characterization system, effectively improving the reliability of multi-source data fusion and the accuracy of thermal anomaly location. The hierarchical S / Vargas / D4Z-Risk quantitative system achieves a progressive characterization of early-stage coupling characteristics of coal spontaneous combustion, the degree of gas disturbance, and comprehensive risk, providing precise quantitative basis for graded early warning and governance priority ranking. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of another multi-source zoning and risk prediction system for spontaneous combustion of coal in goaf areas provided by an embodiment of the present invention; Figure 3 This is a flowchart illustrating another method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas provided by an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic three-band-one-zone concept and typical PHZ location provided in the embodiments of the present invention; Figure 5 This is a timing diagram of risk fusion and stability assessment provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of the FOUR-Net model provided in the embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the following description, references to "some embodiments" refer to a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of the invention pertain. The terminology used in the embodiments of the invention is for the purpose of describing the embodiments of the invention only and is not intended to limit the invention.

[0019] The following describes an exemplary application of the multi-source zoning and risk prediction device for spontaneous combustion of coal in goaf areas according to embodiments of the present invention. This device can be implemented as a terminal or a server. In one implementation, it can be implemented as a laptop, tablet, desktop computer, mobile device, or other types of terminal. In another implementation, it can also be implemented as a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention. The following will describe an exemplary application of the multi-source zoning and risk prediction device for spontaneous combustion of coal in goaf areas when it is implemented as a server.

[0020] This invention provides a method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas. (See also...) Figure 1 , Figure 1 This is a flowchart illustrating a multi-source zoning and risk prediction method for spontaneous combustion of coal in goaf areas provided by an embodiment of the present invention, which will be combined with... Figure 1 The steps shown are explained.

[0021] Step S110: Obtain the original multi-source data of the goaf area and preprocess the original multi-source data to obtain standardized multi-source data with a unified spatiotemporal reference; the standardized multi-source data includes at least gas volume fraction time series, temperature change rate and infrared thermal image data.

[0022] It should be noted that spontaneous combustion of coal in goaf areas refers to the physicochemical process by which coal resources left over from underground coal mining react with oxygen in the air in the goaf (the vacant area formed after coal is extracted), releasing heat and gradually accumulating. When the rate of heat accumulation exceeds the rate of heat dissipation, the coal temperature continues to rise, eventually triggering spontaneous combustion. Its core characteristics are the spontaneity of the oxidation reaction, the persistence of heat accumulation, and the concealed nature of the combustion, which may lead to safety and environmental problems such as coal resource waste, underground fires, and leaks of harmful gases.

[0023] In some implementations, raw multi-source data refers to a set of raw data containing different physical or chemical properties of the goaf and its surrounding environment, acquired through various types of monitoring equipment, sensors, or data acquisition equipment in order to achieve natural monitoring and risk assessment of coal in goaf areas.

[0024] In some implementations, preprocessing refers to a series of standardization operations performed on the collected raw multi-source data. The purpose is to eliminate data noise, fill in missing data, unify data format and spatiotemporal reference, and ensure data reliability and consistency. Specific processing steps include, but are not limited to: data cleaning: removing outliers such as extreme values ​​caused by sensor malfunctions and correcting data deviations; data completion: filling in missing data using interpolation methods, such as linear interpolation and Kriging interpolation; spatiotemporal registration: unifying data acquired from different monitoring devices and at different acquisition times to the same spatial coordinate system and time sampling frequency; data standardization: normalizing or standardizing data with different dimensions, such as converting temperature values ​​to standardized values ​​in the [0,1] interval to facilitate subsequent fusion calculations.

[0025] In this invention, the standardized multi-source data includes at least gas volume fraction time series, temperature change rate and infrared thermographic data.

[0026] In some implementations, gas volume fraction time series refers to the sequence data of the continuous change over time of the volume percentage of a specific gas in the goaf and surrounding areas. Examples include CO and CH4 produced during the natural oxidation of coal, and O2 consumed.

[0027] In some implementations, the temperature change rate refers to the change in temperature of the remaining coal or surrounding rock in the goaf over a unit of time, calculated as: Temperature change rate = (Current temperature - Previous temperature) / Time interval. In practical applications, the temperature change rate is used to characterize the intensity of heat generation from coal oxidation. Specifically, when the temperature change rate is consistently positive and gradually increases, it indicates that the coal oxidation reaction is intensifying and the rate of heat accumulation is accelerating.

[0028] In some implementations, infrared thermal image data refers to the distribution map of infrared radiation intensity in the target area of ​​the goaf captured by an infrared thermal imager, with the temperature difference of the target area represented by the grayscale value or color coding of the image pixels. Since heat is released during the spontaneous combustion of coal, the temperature of the spontaneous combustion area and its surroundings is higher than the ambient temperature, which is represented as a high-temperature anomaly area (such as red or yellow areas) in the infrared thermal image, which can intuitively reflect the spatial distribution range of coal spontaneous combustion in the goaf.

[0029] Step S120: Calculate the air leakage significance index based on the standardized multi-source data; if the air leakage significance index continues to exceed a preset index threshold, perform residual-driven online updating and spatial smoothing of the air leakage mobility rate based on the air leakage significance index to obtain an updated air leakage mobility rate.

[0030] In some implementations, the air leakage significance index is a dimensionless index used to quantify the severity of air leakage in goaf areas. Air leakage refers to the process by which external air enters the goaf area through fissures, roadway interfaces, and other channels, and is a key oxygen supply condition for coal spontaneous combustion. The higher the value of the air leakage significance index, the greater the air leakage intensity and the more sufficient the oxygen supply in the goaf area, and the higher the risk of coal spontaneous combustion.

[0031] In some implementations, the preset index threshold is a critical value set in advance based on factors such as geological conditions of the goaf, spontaneous combustion characteristics of coal, and safety management standards, used to determine whether air leakage is significant and whether the risk exceeds the standard.

[0032] In some implementations, the air leakage mobility rate refers to the spatial migration speed of the leaking airflow in the goaf per unit time, used to describe the flow direction and diffusion intensity of the leaking air within the goaf. The air leakage migration field refers to the distribution of the air leakage mobility rate in the three-dimensional space of the goaf, i.e., each spatial coordinate point corresponds to a leaking air mobility rate value, which can intuitively reflect the spatial flow path of the leaking air.

[0033] In some implementations, residual-driven online updates refer to an update method that dynamically adjusts the leakage mobility rate in real time based on the residual between the model's predicted values ​​and the actual monitored values. The specific implementation logic is as follows: A predicted leakage mobility rate is calculated based on the initial model; then, this predicted value is compared with actual monitored leakage-related data (such as O2 volume fraction time series and temperature change rate) to obtain the residual; if the residual exceeds the allowable range, the leakage mobility rate is corrected according to the magnitude and direction of the residual to ensure consistency between the leakage migration field and the actual leakage state. Online updates mean that this correction process is performed in real time without manual intervention.

[0034] In some implementations, spatial smoothing refers to spatial filtering of the updated leakage mobility rate. The aim is to eliminate abrupt noise in the spatial distribution of the leakage mobility rate, making the spatial distribution of the leakage migration field more consistent with the continuity of airflow. For example, algorithms such as Gaussian filtering and mean filtering are used to perform neighborhood-weighted averaging on the leakage mobility rate value at a given spatial coordinate point. This avoids local abnormal peaks in the leakage migration field caused by errors in single-point monitoring data, ensuring the smoothness and rationality of its spatial distribution.

[0035] In some implementations, the updated air leakage migration rate refers to a three-dimensional spatial distribution field reflecting the current actual air leakage flow state in the goaf, obtained after residual-driven online updating and spatial smoothing. Compared with the initial air leakage migration field, it is closer to the actual site conditions, which can improve the accuracy of subsequent risk assessment and zoning.

[0036] Step S130: Based on the standardized multi-source data, the leakage significance index, and the updated leakage mobility, calculate the cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuation, and then fuse them to generate a continuous risk index.

[0037] In some implementations, cross-modal uncertainty refers to a quantitative indicator of the degree of uncertainty in the data fusion result caused by differences in the acquisition accuracy, physical meaning, and spatiotemporal resolution of different modal data during the standardized multi-source data fusion process. In this invention, a smaller cross-modal uncertainty indicates better consistency of the standardized multi-source data and a more reliable fusion result; conversely, a larger cross-modal uncertainty indicates data conflicts or errors, requiring weight adjustment during the fusion process. For example, when the spatial resolution of infrared thermal imaging data is low, its corresponding cross-modal uncertainty increases, and the fusion weight decreases.

[0038] In some implementations, thermal image saliency refers to a dimensionless index extracted from infrared thermal image data to quantify the degree of temperature difference between high-temperature anomaly areas and background areas. Its purpose is to highlight the high-temperature anomaly characteristics of spontaneous combustion of coal in infrared thermal images and provide a spatial positioning basis for zoning.

[0039] In some implementations, the potential hazard significance is a dimensionless index used to quantify the likelihood of a region transforming into a potential hazard. A higher potential hazard significance value indicates a higher probability that the region possesses the conditions for spontaneous coal combustion, and thus a greater potential risk. For example, regions with a large residual coal layer thickness and a moderate air leakage significance index have a higher potential hazard significance index.

[0040] In some implementations, short-window gas fluctuations refer to the amplitude and frequency of fluctuations in the gas volume fraction in a certain area of ​​the goaf within a short time window, used to characterize the degree of dynamic change in gas concentration.

[0041] In some implementations, the continuous risk index is a dimensionless index used to comprehensively quantify the degree of spontaneous combustion risk of coal in goaf areas. It is obtained by integrating multiple individual indicators such as cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations through a weighted fusion algorithm. In practical applications, a higher continuous risk index indicates a higher risk of spontaneous combustion of coal in goaf areas.

[0042] Step S140: Based on the continuous risk index, the gas volume fraction time series, the temperature change rate, the air leakage significance index, and the thermal image significance, and combined with the smoothing and hysteresis threshold rules of risk inertia, dynamically divide and output mutually exclusive three-zone and one-region distribution maps; the three-zone and one-region include the potential danger zone, the oxidation heating zone, the asphyxiation zone, and the heat dissipation zone.

[0043] In some implementations, the smoothing and hysteresis threshold rules for risk relationships refer to the smoothing and threshold judgment rules set when dynamically dividing the three zones and one area, taking into account the inertia of coal spontaneous combustion risk (i.e., the risk state will not change abruptly) and the hysteresis of monitoring data (i.e., there is a time difference between the actual risk change and the monitoring data).

[0044] In this invention, risk inertia smoothing refers to performing a time-dimensional moving average on the continuous risk index to avoid frequent switching of partition results due to fluctuations in single monitoring data.

[0045] In some implementations, the mutually exclusive three-zone-one-district zoning map refers to dividing the goaf into four non-overlapping and completely covered areas based on the three-dimensional spatial coordinates of the goaf: a potential hazard zone, an oxidation and heating zone, an asphyxiation zone, and a heat dissipation zone. This means that the same spatial coordinate point belongs to only one zone, and all space within the goaf is assigned to a specific zone. The result is output as a visual image. Different zones are identified by different colors, such as yellow for the potential hazard zone, orange for the oxidation and heating zone, blue for the asphyxiation zone, and green for the heat dissipation zone, which visually reflects the spatial distribution pattern of coal spontaneous combustion risk in the goaf.

[0046] In this invention, the potential danger zone refers to the area in the goaf where coal has not undergone significant oxidation and temperature rise, but conditions for spontaneous combustion of coal exist (such as sufficient oxygen supply due to air leakage and concentrated residual coal). It is a potential source of spontaneous combustion risk. The oxidation and temperature rise zone refers to the area in the goaf where coal has undergone significant oxidation and the temperature continues to rise, but has not reached the critical temperature for spontaneous combustion; it is the active development stage of spontaneous combustion. The asphyxiation zone refers to the area in the goaf where the oxygen content is extremely low (or there is no oxygen), and the coal oxidation reaction has stopped or is in a stagnant state; there is no risk of spontaneous combustion. The heat dissipation zone refers to the area in the goaf where the heat generated by the coal oxidation reaction can be dissipated in a timely manner, and the temperature does not rise significantly; the risk of spontaneous combustion is low.

[0047] This invention breaks through the limitations of traditional three-zone zoning by incorporating the Potential Hazard Zone (PHZ) as an independent zone into a dynamic three-zone-one-zone zoning system, and providing a quantifiable definition and robust assessment of it. The PHZ is used to characterize high-risk coal spontaneous combustion incubation areas where random air leakage has created a locally oxygen-rich environment, early signs of temperature rise have appeared, but the sustained exothermic stage has not yet been reached. Its assessment logic is mutually exclusive with the other three zones. To ensure the stability and operability of the zoning results at the legal compliance and engineering implementation levels, this invention introduces a risk inertia mechanism and a dual-threshold hysteresis rule for entry / exit, avoiding frequent zoning switching due to instantaneous fluctuations in monitoring data.

[0048] Step S150: Input the standardized multi-source data into the FOUR-Net model, which uses the leakage significance index and the cross-modal uncertainty as gate signals, and perform weighted fusion and risk prediction in sequence to obtain the three-zone one-region probability map and risk prediction map.

[0049] In some implementations, the FOUR-Net model is a deep learning model specifically designed for this invention, using the wind significance index and cross-modal uncertainty as gating signals, for weighted fusion of multi-source data and prediction of coal spontaneous combustion risk.

[0050] In some implementations, the three-zone-one-district zoning probability map refers to the visualization image output by the FOUR-Net model, which is based on the three-dimensional spatial coordinates of the goaf, and each spatial coordinate point corresponds to four probability values ​​(the sum of the probabilities is 1) for belonging to the potential danger zone, oxidation and heating zone, asphyxiation zone, and heat dissipation zone.

[0051] In some implementations, the risk prediction map refers to the output of the FOUR-Net model, which uses the three-dimensional space of the goaf as coordinates, and each spatial coordinate point corresponds to the risk level of coal spontaneous combustion within a preset time period in the future. The risk level can be divided into low risk, medium risk, high risk, and extremely high risk, and is represented by different colors (such as extremely high risk being red). It can predict the development trend of coal spontaneous combustion in the goaf in advance. For example, the risk level of a certain area will rise from medium risk to high risk in the next 24 hours.

[0052] Step S160: Trial warning and governance measures are triggered and implemented through the continuous risk index, the three-zone one-district probability map and the risk prediction map; after the governance measures are implemented, iterative monitoring continues, and the governance effect is verified by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index and the continuous risk index.

[0053] In some implementations, tiered early warning refers to triggering a warning signal of the corresponding level based on the results of a continuous risk index, a probability map of the three-zone and one-region division, and a risk prediction map, according to preset warning level standards. Different warning levels correspond to different response mechanisms; for example, a blue warning only requires increasing the monitoring frequency, while a red warning requires the immediate activation of the emergency response plan.

[0054] In some implementations, mitigation measures refer to engineering and management methods or measures taken to suppress coal oxidation reactions and reduce the risk level in response to the risk of spontaneous combustion of coal in goaf areas. These measures include, but are not limited to: air leakage blocking measures, such as injecting sealing materials into the fissures in the goaf to block air leakage channels, such as polyurethane foam, cement mortar, etc.; cooling measures, such as injecting coolant (such as water, liquid nitrogen) into the spontaneous combustion area or high temperature abnormal area to reduce the coal body temperature; inerting measures, such as injecting inert gas into the goaf to reduce the oxygen concentration and inhibit oxidation reactions; and enhanced monitoring measures, such as increasing the density of monitoring points and increasing the data collection frequency in high-risk areas.

[0055] In some implementations, iterative monitoring refers to a cyclical monitoring process that involves continuously collecting, preprocessing, and calculating indices from multiple sources in the goaf area after the implementation of remediation measures. The purpose is to track changes in risk in real time after the implementation of remediation measures and provide data support for the adjustment of subsequent remediation measures.

[0056] In some implementations, the verification of the treatment effect refers to the process of determining whether the treatment measures are effective by comparing the changing trends of the updated leakage migration field, leakage significance index, and continuous risk index before and after the implementation of the treatment measures. The verification criteria are as follows: if, after treatment, the leakage intensity in the updated leakage migration rate is significantly reduced, the leakage significance index continues to fall below the preset threshold, and the continuous risk index drops to the medium-low risk range, and all three are simultaneously stabilized within a safe range, then the treatment measures are considered effective; otherwise, the treatment plan needs to be optimized.

[0057] The multi-source zoning and risk prediction method for spontaneous combustion of coal in goaf provided by this invention innovatively incorporates potential hazard zones as independent zones into a dynamic three-zone-one-zone zoning framework. By defining the PHZ calculable, using mutually exclusive judgment logic with the other three zones, and combining risk inertia and entry / exit dual-threshold hysteresis mechanism, this invention fills the gap in the traditional three-zone zoning for identifying the high-risk incubation stage of spontaneous combustion of coal. At the same time, it ensures the stability and operability of the zoning results at the legal and engineering implementation levels, and achieves full-cycle coverage of spontaneous combustion risk from the explicit stage to the high-risk incubation stage. Secondly, the leakage significance index L constructed in this invention, through dimensionless fusion of multi-source evidence and Sigmoid mapping, condenses multi-dimensional monitoring clues of hidden leakage channels into a unified quantitative indicator. Furthermore, L can simultaneously serve PHZ determination, D4Z-Risk calculation, and FOUR-Net gated fusion, achieving multi-purpose coupling of a single indicator across the entire chain of determination, prediction, and control. The online update scheme for leakage mobility λ based on L enables continuous tracking and denoising of leakage channel states without complex filters, significantly reducing the technical threshold and cost for engineering implementation. Thirdly, the cross-modal uncertainty and thermal image significance defined in this invention provide support for risk assessment from the dimensions of evidence consistency and thermal boundary location, respectively. Together with L, they constitute a three-dimensional feature characterization system, effectively improving the reliability of multi-source data fusion and the accuracy of thermal anomaly location. The hierarchical S / Vargas / D4Z-Risk quantitative system achieves a progressive characterization of early-stage coupling characteristics of coal spontaneous combustion, the degree of gas disturbance, and comprehensive risk, providing precise quantitative basis for graded early warning and governance priority ranking.

[0058] In some embodiments, step S120 above can also be implemented by steps 121 to 122: Step 121: Based on the rate of change of the ratio of oxygen to nitrogen gas volume fraction in the time series, and in combination with the rate of change of temperature, calculate the preliminary index of air leakage intensity.

[0059] Step 122: Perform time window smoothing and spatial interpolation on the preliminary air leakage intensity index to obtain the air leakage significance index.

[0060] The formula for calculating the air leakage significance index is as follows: , In the formula, , These are the measured predicted wind speed and the model predicted wind speed, respectively. For wind speed reference, For microseismic energy density, This serves as a reference for microseismic energy density. For pressure gradient mode, As a reference value for the pressure gradient modulus, This is the measured oxygen volume fraction. To predict the oxygen volume fraction, The oxygen fraction is the reference value, and α1, α2, α3, and α4 are the weights, respectively. i ≥0, ∑α i =1 represents the weight, and Logi(⋅)=1 / (1+e−z) represents the logistic mapping.

[0061] In some implementations, the preliminary air leakage intensity index refers to an intermediate index used to preliminarily quantify the air leakage intensity in the goaf, based on a coupled calculation of the rate of change of the ratio of oxygen to nitrogen gas integrals and the rate of change of temperature.

[0062] In some implementations, time window smoothing refers to applying a sliding window filter over a time dimension to the preliminary air leakage intensity index. The aim is to eliminate random noise in single monitoring data, reduce instantaneous fluctuations, and highlight the long-term trend of air leakage intensity. Here, the time window refers to a preset continuous time interval, the specific length of which can be adjusted according to the data sampling frequency and air leakage variation characteristics.

[0063] To address the difficulty in quantifying and characterizing hidden air leakage channels in goaf areas, this invention constructs a multi-evidence fusion calculation method for the air leakage significance index L. The air leakage significance index combines multi-source monitoring data such as wind speed, model prediction residuals, microseismic energy density, pressure gradient intensity, and oxygen concentration deviation. These data are normalized based on their respective reference values, weighted, and then mapped to the dimensionless interval [0,1] using a Sigmoid function, achieving single-index condensation of multi-source clues for hidden air leakage channels. L is not only the core input for PHZ determination and the calculation of the dynamic three-zone-one-region continuous risk index D4Z-Risk, but also serves as the gating signal for the FOUR-Net spatiotemporal fusion neural network model (i.e., the FOUR-Net model), directly regulating the fusion weights of thermal spatial features and multi-sensor temporal features.

[0064] In some embodiments, when performing residual-driven online updates and spatial smoothing of the leakage mobility, the online update formula is: , In the formula, For a moment In position The update did not smooth out the leakage migration rate. For spatial location, For time, For discrete time step index, for Always in position The updated air leakage migration rate, This is a truncation function. The trigger threshold for the air leakage significance index, To update the step size coefficient, For a moment The air leakage significance index, These represent the lower and upper limits of the mobility, respectively.

[0065] In some embodiments, the spatial smoothing formula is: , In the formula, for Always in position The final leakage migration rate after smoothing To smooth out the weights, For spatial neighborhood set, The number of grid points in the neighborhood. For the location index within the neighborhood, For the previous moment Neighborhood Location The air leakage migration rate.

[0066] This invention proposes a lightweight leakage mobility λ update scheme that can be stably implemented in engineering without relying on complex algorithms such as Kalman filtering. Specifically, by mapping the difference between the leakage significance index L and a preset threshold to the update increment of λ, and combining upper and lower boundary constraints of λ with neighborhood mean smoothing, continuous tracking of the opening and closing state of the leakage channel and noise elimination are achieved, forming a regularized inversion update path.

[0067] In some embodiments, the formula for calculating the cross-modal uncertainty is: , In the formula, For cross-modal uncertainty, For spatial location, For time, For the amplitude limiting operator, For the multimodal normalized residual vector, for The 2-norm, This is the residual metric.

[0068] The formula for calculating the significance of the thermal image is: , In the formula, For thermal image saliency, This represents the equivalent temperature field of the infrared thermal image in the downhole physical coordinate domain. Let be the spatial gradient vector of the temperature field. Let L be the L2 norm of the spatial gradient vector of the temperature field. This is the gradient smoothing / stability constant.

[0069] To improve the reliability of multi-source data fusion and the accuracy of thermal anomaly localization, this invention defines cross-modal uncertainty Unc and thermal image significance Sal, respectively. IR Two core auxiliary indicators. Unc uses the normalized L2 norm of the cross-modal residual vector to measure the consistency between observed data and model predictions. This serves as a reference for prudence in risk assessment and can also be used in conjunction with L to adjust gating weights; Sal IR Based on the normalization of the infrared temperature gradient, monitoring biases introduced by factors such as emissivity differences and environmental obstruction are effectively suppressed, enabling precise positioning of the coal spontaneous combustion thermal boundary. These two indicators, together with L, construct the evidence strength (L), evidence consistency (Unc), and spatial boundary (Sal). IR A three-dimensional feature representation system.

[0070] The formula for calculating the significance of the potential hazard is as follows: , In the formula, For the significance of potential hazards, , , , For different weighting coefficients, and ≥0、 , , These represent the measured oxygen volume fraction and CO volume fraction, respectively. , These are the threshold values ​​for oxygen volume fraction and CO volume fraction, respectively. For the rate of temperature rise, For reference temperature rise rate, For positive part operators, This is the air leakage significance index.

[0071] The formula for calculating the short-window gas fluctuation is as follows: , In the formula, Here, s represents the short-window gas fluctuation index, s is the index of the gas type, and G is the set of gases included in the statistics. For gas weighting coefficients, Let be the average volume fraction of gas s within a short time window. Let S be the standard deviation of the volume fraction of gas s within the same short time window. To prevent zero constant / stability constant.

[0072] In some embodiments, the continuous risk index is: , In the formula, It is a continuous risk index. These are the fusion weighting coefficients for cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations, respectively.

[0073] In some embodiments, the exponential smoothing formula for risk inertia is: , In the formula, Due to risk inertia, The risk inertia of the previous moment, For smoothing coefficients, For spatial location, It is a continuous risk index. For discrete time step index; Based on the risk inertia and the entry / exit dual threshold hysteresis rule, the three-band-one-zone mutual exclusion determination is performed to obtain the three-band-one-zone partition map.

[0074] This invention designs a hierarchical risk quantification index system to characterize the entire stage of coal spontaneous combustion risk from its early incubation to its manifest development. Specifically, the potential hazard significance S is identified by weighted synthesis of oxygen enrichment exceeding a threshold, normalized temperature rise rate, early product concentrations such as CO exceeding a threshold, and L, thus recognizing early coupled characteristics of impending coal spontaneous combustion danger; short-window gas fluctuation Var... gas The normalized ratio of mean to variance within a short time window for multiple gases is used to quantify the severity of gas concentration disturbances; D4Z-Risk, on the other hand, quantifies the severity of S and Sal... IR Var gas Based on Unc, multiple indicators are integrated to form a continuous risk score in the 0-1 range, providing a core basis for graded early warning, governance priority ranking, and governance effectiveness evaluation.

[0075] In some embodiments, a spatial location is determined to be the asphyxiation zone if and only if all of the following conditions are met simultaneously: Physical conditions: The oxygen volume fraction is less than or equal to the asphyxiation threshold; and the absolute value of the temperature change rate is less than or equal to a small threshold close to zero; and the carbon monoxide volume fraction is low; Spatial constraints: The distance of this spatial location along the working face advancement direction is greater than the distance of all units in that direction that have been identified as heat dissipation zones or oxidation heating zones. The location closer to the working surface shall be preferentially identified as the heat dissipation zone if and only if a spatial location simultaneously satisfies all of the following physical conditions, and if it conflicts with the determination of other zones in the same direction: Physical conditions: oxygen volume fraction is higher than the oxygen-enriched threshold; and the rate of temperature change is less than or equal to the temperature rise slope threshold; and carbon monoxide volume fraction is less than or equal to the low CO threshold. A spatial location is identified as a potential danger zone when it simultaneously meets all of the following conditions: If the leakage significance index is greater than or equal to the entry threshold, the oxygen volume fraction is higher than the oxygen enrichment threshold, the temperature change rate is greater than the temperature rise slope threshold, or the infrared thermal image significance is greater than or equal to the thermal image significance threshold, and the carbon monoxide volume fraction is less than the oxidation heating threshold; if the leakage significance index subsequently drops below the exit threshold, or the risk inertia drops below the exit threshold, then the current spatial location exits the potential danger zone. A spatial location is identified as the oxidation heating zone when all of the following conditions are met simultaneously: Physical conditions: The rate of temperature change is greater than the temperature rise slope threshold, and the carbon monoxide volume fraction is greater than or equal to the oxidation temperature rise CO threshold, and the risk inertia is greater than or equal to the entry threshold of the air leakage significance index; Spatial distribution characteristics: The oxidation temperature rise zone is mainly distributed in the area between the heat dissipation zone and the asphyxiation zone.

[0076] In some embodiments, the FOUR-Net model is trained in the following manner: Based on the three-zone-one-region determination rule, a three-zone-one-region zoning label is generated for each spatiotemporal grid, and a continuous risk index label is generated based on the continuous risk formula; samples are constructed in a sliding time window manner, where the input of each sample is a spatiotemporal tensor composed of multi-source data and infrared thermal image frames stacked within the time window; the sample set is divided into a training set, a validation set, and a test set. The obtained sample standardized multi-source data is input into the feature extraction module of the FOUR-Net model. The spatial branch extraction module performs two-dimensional convolution processing on the infrared thermal image sequence to extract the thermal image spatial features. The temporal branch extraction module performs one-dimensional convolution and cyclic modeling on the multi-sensor temporal data to extract multi-sensor temporal features. Using the air leakage significance index and the cross-modal uncertainty as gating signals, the gating signals are concatenated with the channel-pooled representations of the thermal spatial features and the multi-sensor temporal features. A gating map is then generated using a 1×1 convolution and a sigmoid activation function. The gating map is as follows: , In the formula, For gating graphs, For logical mapping functions, For 1×1 convolution, For thermal imaging spatial features, For multi-sensor timing characteristics, The significance index of air leakage, For channel dimension splicing operators, For uncertainty plot, Here, H is the height of the network feature map, and W is the width of the network feature map; The gated image is fused with the thermal spatial features and the multi-sensor temporal features to obtain a fused feature vector; the fused feature vector is: , In the formula, To fuse feature vectors, For point-by-point multiplication, 1-M is a gated complementary graph; The fused feature vector is input into subsequent convolutional and fully connected layers to output the three-band, one-division partitioning probability map and the risk prediction map. Construct a total loss function; the total loss function introduces a gated consistency regularization constraint; the gated consistency regularization term is: , In the formula, For gating consistency regularization, These are the gating regularization weight coefficients. For the average operator, For logical mapping functions, The linear coefficient of the air leakage significance index is . The linear coefficients of the cross-modal uncertainty are... For bias terms; The optimal model after training is evaluated using the test set. The accuracy, recall, and F1 score of the three-band-one-zone classification, as well as the mean square error and mean absolute error of risk prediction, are calculated to verify the model's ability to identify dangerous areas and the accuracy of risk prediction.

[0077] In some embodiments, step S160 above can also be implemented through steps 161 to 164: Step 161: Based on the continuous risk index and the risk prediction map, and combined with the three-zone one-district zoning probability map, the risk area is spatially located and classified into levels.

[0078] In some implementations, a risk area refers to a spatial region within a goaf that has a continuous risk index reaching a preset risk level standard or whose future risk level in a risk prediction map exceeds a safe threshold. It represents the concentrated distribution range of natural coal risks. Its definition requires a combination of spatial location and risk level classification. Spatial location clarifies the three-dimensional coordinate range of the risk area, while risk level classification clarifies the severity of the risk in that area, providing a target basis for subsequent precise management.

[0079] In some implementations, spatial positioning is the process of determining the specific location, boundary range, and spatial morphology of a risk area in the three-dimensional geographic coordinate system of the goaf based on the spatial coordinate information of the three-zone-one-district zoning probability map and the spatial distribution data of the continuous risk index.

[0080] Step 162: When the continuous risk index exceeds the preset grading threshold or the risk growth rate in the risk prediction map exceeds the set threshold, a graded early warning signal is triggered, and a location management instruction is generated by combining the potential danger zone and the high probability area of ​​the oxidation heating zone.

[0081] In some implementations, the preset grading threshold is a critical value of the continuous risk index corresponding to the risk level, set for grading early warning, and is the core judgment criterion for triggering different levels of early warning.

[0082] In some implementations, the risk growth rate is the ratio of the difference between the predicted risk index and the current continuous risk index within a predetermined future time period in the risk prediction chart to the current risk index, used to quantify the speed at which risk develops.

[0083] In some implementations, tiered early warning signals are warning messages with clear level identifiers triggered based on risk level classifications, used to notify relevant personnel to take corresponding response measures. The forms of early warning signals include, but are not limited to: audible and visual alarms (such as underground audible and visual alarms, and alarm lights in the surface control room), information notifications (such as SMS messages, app push notifications, and platform pop-ups), and written warning letters; different levels of early warning signals correspond to different response priorities. For example, a blue warning signal only requires monitoring personnel to pay attention, while a red warning signal requires the immediate activation of the emergency command center to organize personnel evacuation and rescue operations.

[0084] In some implementations, the location governance instruction refers to a specific execution instruction generated based on the spatial location results of the risk area and the distribution of danger zones in the three-zone-one-district zoning probability map, which includes the governance target area, governance measure type, and governance parameter requirements.

[0085] Step 163: After implementing inertia injection, sealing, or ventilation optimization measures, continue to collect multi-source data and iteratively execute the measures. Quantitatively evaluate the treatment effect by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, the continuous risk index, and the proportion of dangerous zones in the zoning probability map.

[0086] Step 164: If the updated air leakage migration rate decreases, the air leakage significance index falls back, the continuous risk index decreases, and the proportion of the potential danger zone and the oxidation heating zone is significantly reduced, then the treatment is deemed effective; otherwise, the treatment strategy is readjusted.

[0087] In some implementation methods, effective remediation refers to a quantitative assessment confirming that the risk of spontaneous combustion of coal in the goaf has been significantly suppressed and the risk status has continued to improve after the implementation of remediation measures. The core criteria for this determination are a significant decrease in the updated air leakage migration rate, a drop in the air leakage significance index below a preset threshold, a decrease in the continuous risk index to the low-to-medium risk range, a significant reduction in the proportion of hazardous zones, and all four indicators remaining stable for more than 36 hours without rebound. For example, if all indicators meet the above requirements after remediation and remain stable for 48 hours without rebound, then the remediation is deemed effective.

[0088] In some implementation methods, a governance strategy refers to a comprehensive plan developed to address the risk of spontaneous combustion of coal in goaf areas, including governance objectives, combinations of governance measures, governance parameters, and implementation steps. Governance strategies must be formulated based on factors such as the geological conditions of the goaf area, risk level, and distribution of hazardous zones. When governance is deemed ineffective, the reasons for the unmitigated risk need to be re-analyzed, and the combinations or parameters of governance measures adjusted to form a new governance strategy.

[0089] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0090] This invention proposes a multi-source coupled dynamic "three zones and one area" partitioning method for safety monitoring and early warning of goaf areas. This method adds a potential danger zone (PHZ) to the traditional three zones, constructing a dynamic partitioning system that includes the asphyxiation zone (HDZ), spontaneous combustion zone (AZ), potential danger zone (PHZ), and oxidation zone (OZ). It achieves closed-loop control of "monitoring-parameter update-partitioning-prediction-early warning-treatment-re-assimilation" through sequential data flow.

[0091] like Figure 2 As shown, the specific process is as follows: First, the gas volume fractions (O2, CO, CH4), temperature, wind speed, pressure time series, infrared thermography, and microseismic information in the goaf are spatiotemporally aligned and normalized to form standardized multi-source data with a unified spatiotemporal benchmark. Then, based on the standardized data, the leakage significance index L is calculated to verify the existence of hidden leakage channels. If the evidence continues to hold, residual-driven online updates and spatial smoothing of the leakage mobility λ are used to display the opening and closing status of the tracking channel in a parametric field. Based on this, the following multi-dimensional indicators are calculated: uncertainty Unc (cross-modal residual consistency), thermographic significance Sal... IR Gradient normalized thermal edge), potential hazard significance S (integrating oxygen enrichment, temperature rise, early products, and air leakage significance L), and short-window gas fluctuation Var gas The risk index D4Z-Risk is formed by combining risk inertia P and entry / exit dual threshold hysteresis rules to complete the mutual exclusion stability judgment of HDZ / AZ / PHZ / OZ, and realize the dynamic update and visualization of partitions.

[0092] To achieve short-term forward-looking prediction, a FOUR-Net spatiotemporal fusion network model with L and Unc as gating signals is proposed. This model adaptively weights and fuses the spatial texture features of infrared thermal imaging data with multi-sensor time-series data, outputting the partition probability distribution and risk map for future time periods. In the early warning stage, in addition to fixed grading thresholds, a risk growth rate (RGR) is introduced to trigger coordinated governance (inertia, airtightness, and ventilation optimization) in advance for accelerated deterioration states. After governance is implemented, data is continuously written back and parameters are updated. The synchronous decline of λ, L, and D4Z-Risk is observed to quantitatively verify the governance effect and continuously iterate and optimize.

[0093] Through the above process, the present invention significantly improves the detection sensitivity and spatial positioning accuracy of the local oxygen enrichment-heating process induced by random air leakage, reduces the extrapolation error from point to surface, and enhances the interpretability and closed-loop control capability of the prediction model.

[0094] This method specifically includes the following contents, as follows: Figure 3 As shown: S1, Data Standardization: The original multi-source time-series data (gas volume fraction O2 / CO / CH4, temperature, wind speed, pressure, infrared thermography, microseismic information) are denoised, interpolated, clock-aligned and registered to obtain standardized multi-source data with a unified spatiotemporal reference.

[0095] In this embodiment, the gas volume fraction timing ,Y CO , (dimensionless); temperature T (K); wind speed v (m·s) -1 Pressure p (Pa); Infrared thermogram I (u,v,t) (pixel coordinates); Microseismic events and energy records (converted to energy density time series Ems, J·m) -3 Here, the processing mainly includes clock alignment and resampling: unifying to t k =kΔt. Coordinate registration: The thermal image is mapped to downhole coordinates via camera calibration to obtain T. IR (x, y, tk) is used, and the dimensions of each quantity are normalized / standardized for direct use in subsequent formulas. Standardized multi-source dataset. .

[0096] S2, Channel Verification and λ Update: Calculate the leakage significance index L using standardized multi-source data; if L continuously exceeds the preset index threshold, perform residual-driven online update and spatial smoothing of the leakage mobility λ to explicitly track channel opening and closing.

[0097] In this embodiment, step S2-1 calculates the air leakage significance index L. This index integrates multi-source observation data such as wind speed residuals, microseismic information, pressure gradients, and oxygen enrichment deviations to form a single, interpretable dimensionless index, aiming to achieve quantitative identification and location of hidden air leakage channels. The L index is used for both the verification and tracking of hidden channels, driving the online update of the air leakage mobility λ; and as a key feature in the determination of potential hazard zones (PHZs), it provides gating signals for the subsequent FOUR-Net spatiotemporal fusion network. By integrating multi-dimensional information into a unified index, the L index significantly improves the detection efficiency and interpretability of hidden channels, providing a reliable quantitative basis for subsequent zoning, prediction, and remediation.

[0098] Define the air leakage significance index L∈[0,1], and calculate it using the following formula: , In the formula, , These are the measured predicted wind speed and the model predicted wind speed (m·s). -1 ), Wind speed reference quantity (m·s) -1 ), Microseismic energy density (J·m -3 ), Reference quantity for microseismic energy density (J·m) -3 ), Pressure gradient modulus (Pa·m) -1 ), The pressure gradient modulus reference value (Pa·m) -1 ), This represents the measured oxygen volume fraction (dimensionless). To predict the oxygen volume fraction (dimensionless). The oxygen fraction is a reference value (dimensionless), α1, α2, α3, and α4 are the weights, and α is the value of the oxygen fraction. i ≥0, ∑α i =1 represents the weight, σ(⋅) represents the Sigmoid function, and Logi(⋅)=1 / (1+e−z) represents the logistic mapping.

[0099] S2-2, residual-driven online update and spatial smoothing of λ are performed only when L continuously exceeds a preset exponential threshold. To characterize the opening and closing state of the channel, an online update of the parameter field based on L is proposed.

[0100] , In the formula, For a moment In position Updated but not smoothed air leakage migration rate (m 2 ·Pa -1 ·s-1 ), Let m be the spatial location (underground coordinates). For time, This is an index for discrete time steps (dimensionless). for Always in position The updated air leakage migration rate (m 2 ·Pa -1 ·s -1 ), This is a truncation function. The trigger threshold (dimensionless) for the air leakage significance index. To update the step size coefficient (m) 2 ·Pa -1 ·s -1 ), used to convert dimensionless increments The mapping is to a change of the same dimension as λ. For a moment The air leakage significance index (dimensionless, value 0-1). These are the lower and upper limits of the mobility, respectively; Without relying on complex filters / Kalman formulas, leakage is mapped to an upregulation of λ, and weak evidence is mapped to a convergence of λ, protected by upper and lower bounds and possessing spatial smoothness. Neighborhood smoothing is introduced to suppress noise artifacts. , In the formula, for Always in position The final air leakage migration rate after smoothing (m 2 ·Pa -1 ·s -1 ), For smoothing weights (dimensionless). It is a spatial neighborhood set (a dimensionless set). The number of grid points in the neighborhood (dimensionless). For the location index within the neighborhood, For the previous moment Neighborhood Location air leakage migration rate (m 2 ·Pa -1 ·s -1 ).

[0101] Ultimately, the strength index L of the concealed air leakage channel and the air leakage migration rate λ were obtained.

[0102] S3, Indicator Calculation and Risk Integration: Based on the results of S1 and S2, calculate the uncertainty Unc and the thermal image significance Sal. IRPotential hazard significance S and short-window gas fluctuation Var gas They are then integrated into the continuous risk index D4Z-Risk.

[0103] In this embodiment, S3-1, the cross-modal residual consistency measure, is defined as normalized uncertainty to reflect evidence discrepancies in early warning and attention gating. A larger value indicates greater evidence discrepancy. The formula for calculating cross-modal uncertainty is: , In the formula, This represents the cross-modal uncertainty (dimensionless). For spatial location, For time (continuous or discrete moments). For the amplitude limiting operator, For the multimodal normalized residual vector, for The 2-norm, , The residual scale is dimensionless; where , Measured wind speed (m·s) -1 ), To predict wind speed (m·s) -1 ), Wind speed reference quantity (m·s) -1 ), This is the measured oxygen volume fraction. To predict the oxygen volume fraction, This is a reference value for oxygen volume fraction. Pressure gradient modulus (Pa·m) -1 ), Reference value for pressure gradient (Pa·m) -1 ), Microseismic energy density (J·m -3 ), Microseismic reference quantity (J·m) -3 Each of the above components is divided by a reference quantity of the same dimension, so the components of r are dimensionless.

[0104] thermal imaging significance Sal IR (Based on temperature gradient normalization), to reduce the impact of emissivity / occlusion, gradient normalization significance is adopted. The formula for calculating thermal image significance is: , In the formula, For thermal image significance (dimensionless). This represents the equivalent temperature field of the infrared thermal image in the downhole physical coordinate domain. Let be the spatial gradient vector of the temperature field. Let L be the L2 norm of the spatial gradient vector of the temperature field. This is the gradient smoothing / stability constant.

[0105] It should be noted that in this embodiment, the thermal boundary is identified by gradient rather than absolute temperature to avoid amplifying calibration errors.

[0106] S3-2, The formula for calculating the significance of potential hazards is: , In the formula, The significance of potential hazards (dimensionless). , , , For different weighting coefficients, and ≥0、 This is used to adjust the contribution ratio of each symptom to S. , These are the measured volume fractions of oxygen and CO, respectively (dimensionless). , These are the threshold values ​​for oxygen volume fraction and CO volume fraction, respectively (dimensionless). The rate of temperature rise (K·s) -1 ), Reference temperature rise rate (K·s) -1 ), For positive part operators, This is the air leakage significance index.

[0107] The formula for calculating short-window gas fluctuations is: , In the formula, is the short-window gas fluctuation index, and s is the index of the gas type (dimensionless). The set of gases (dimensionless set) participating in the statistics. This is the gas weighting coefficient (dimensionless). Let be the average volume fraction of gas s within a short time window (dimensionless). Let S be the standard deviation of the volume fraction of gas s within the same short time window (dimensionless). To prevent zero constant / stability constant (dimensionless).

[0108] S3-3 synthesizes the significance of potential hazards, thermal, gaseous, and uncertainties into a continuous risk index of 0-1, used for tiered early warning and priority resource allocation. The formula for calculating the continuous risk index is: , In the formula, It is a continuous risk index. These are the fusion weighting coefficients for cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations, respectively.

[0109] S4, Stability Judgment and Zoning: Based on the continuous risk index, gas volume fraction time series, temperature change rate, air leakage significance index and thermal image significance, the risk inertia P and entry / exit dual threshold hysteresis are used to dynamically divide and output a three-zone one-region mutually exclusive zoning map.

[0110] Thermal spatial features are fused with multi-sensor temporal features using air leakage-uncertainty guided gating: , In the formula, This is a gating graph (dimensionless), with dimensions H×W, and values ​​ranging from [0,1]. Larger values ​​indicate a greater tendency towards temporal branching features. The smaller the value, the more it leans towards spatial branching features. , For logical mapping functions, For 1×1 convolution, For thermal imaging spatial features, For multi-sensor timing characteristics, The significance index of air leakage, For channel dimension splicing operators, For cross-modal uncertainty, For channel pooling operators, the size is H×W, where H is the height of the network feature map and W is the width of the network feature map; , In the formula, To fuse feature vectors, For point-by-point multiplication, 1−M is a gated complementary graph; When there is strong evidence of air leakage or high uncertainty, the temporal evidence is automatically weighted up; when the thermal texture is stronger, the spatial evidence is automatically weighted up to improve the capture rate and interpretability of pH / OZ.

[0111] To ensure that gating is governed by evidence, a gating consistency regularity is introduced (an innovative training term): , In the formula, This is a gating consistency regularization term (dimensionless). , which is the gating regularization weight coefficient (dimensionless). It is the average operator (dimensionless). It is a logical mapping function (dimensionless). The linear coefficient (dimensionless) represents the air leakage significance index. represents the linear coefficient (dimensionless) of the cross-modal uncertainty. This is the bias term (dimensionless); S5, Gated Fusion Prediction: The FOUR-Net model, using L and Unc as gating signals, fuses thermal spatial texture and multi-sensor temporal sequence to output a three-band, one-zone zoning probability map and risk prediction map for future moments.

[0112] In this embodiment, as Figure 4 As shown, 1) Asphyxiation zone AZ: when the oxygen volume fraction Y O2 (x,t k Less than or equal to the asphyxiation threshold At this point, the location is identified as the asphyxiation zone; this is generally accompanied by a temperature change rate dT / dt approaching zero and a carbon monoxide volume fraction Y. CO Lower.

[0113] Under the premise of satisfying the above physical conditions, a suffocation zone is finally determined only when the distance d(x) at position x is greater than the maximum distance of all units that have been determined to be heat dissipation zones or oxidation heating zones along that direction; that is, suffocation zones are only allowed to appear in the area farthest from the working surface and must not appear in the area closer to the working surface than oxidation heating zones or heat dissipation zones.

[0114] 2) Heat dissipation zone HDZ: When the oxygen volume fraction Y O2 (x,t k (Above the oxygen enrichment threshold) And the rate of temperature change dT / dt is less than or equal to the temperature rise slope threshold. And the volume fraction of carbon monoxide Y CO (x,t k Less than or equal to the low CO threshold At that time, the location was identified as a heat dissipation zone.

[0115] Heat dissipation zones should be assigned preferentially to areas close to the working surface; that is, heat dissipation zones should not appear in deeper areas beyond the suffocation zone.

[0116] 3) Potential Hazard Zone PHZ: When the air leakage significance index L(x,t) k (greater than or equal to the entry threshold) And oxygen volume fraction Y O2 (x,t k (Above the oxygen enrichment threshold) And satisfy the condition that dT / dt > 0. or Sal IR (x,t k )≥θ IR Meanwhile, the volume fraction of carbon monoxide Y CO (x,t k Less than the oxidation heating threshold At that time, the location was identified as a potential danger zone; if the subsequent air leakage significance index L(x,t) k(Decrease to exit threshold) The following, or risk inertia P k (x) drops to the exit threshold The location is now out of the potential danger zone.

[0117] 4) Oxidation zone (OZ): When the rate of temperature change dT / dt is greater than the temperature rise slope threshold. And the volume fraction of carbon monoxide Y CO (x,t k Greater than or equal to the oxidation heating threshold And risk inertia P k (x) is greater than or equal to the entry threshold At that time, the location was determined to be the oxidation heating zone.

[0118] The oxidation heating zone is mainly distributed in the middle area between the heat dissipation zone and the suffocation zone. That is, in a certain direction, the distance d(x) between the unit that meets the oxidation heating condition is between the heat dissipation zone unit and the suffocation zone unit. If a certain position meets both the oxidation heating and heat dissipation conditions, the area farther away from the working surface is preferentially identified as the oxidation heating zone, while the area closer is still identified as the heat dissipation zone.

[0119] By applying the above rules, each spatial unit is assigned only one zone, thereby achieving mutually exclusive partitioning of the three zones and one area—the asphyxiation zone, the heat dissipation zone, the potential danger zone, and the oxidation and heating zone—resulting in the partition label Z(x,t). k )∈{AZ,HDZ,PHZ,OHZ} and partition map.

[0120] S6, Early Warning Triggered Governance: Combine D4Z-Risk and Risk Growth Rate (RGR) to trigger governance (inertia, closure, ventilation); after governance, data is written back, and S1-S3 iterations are continued to verify the synchronous decline of λ, L, and D4Z-Risk.

[0121] In this embodiment, based on continuous scoring, a risk growth rate is defined for advance triggering, and the formula for calculating the risk growth rate is: , In the formula, x represents the spatial location (downhole coordinates), in meters (m) (used for location and not involved in dimensional calculations); t represents time (continuous or discrete moments), in seconds (s); and RGR(x,t) is the risk growth rate, used to measure the rate of increase of the current risk relative to its smoothed baseline per unit time, in seconds (s). -1D4Z-Risk(x,t) is the dynamic continuous risk index, the risk score at the current moment (dimensionless); EMAτ{D4Z-Risk}(x,t) is the exponential moving average with time constant τ as the parameter, the baseline obtained by smoothing D4Z-Risk over time (dimensionless); Τ is the time constant / evaluation window of EMA, which determines the memory length of the baseline, in seconds.

[0122] It should be noted that when D4Z-Risk exceeds the preset grading threshold, or when D4Z-Risk exceeds the risk growth rate grading threshold, the system will trigger the corresponding grading warning and governance instructions. After the governance measures are implemented, the system will continue to iterate through data collection, preprocessing, and index calculation steps (i.e., steps S1-S3 above). If the leakage significance index L, the leakage migration field characteristic parameter λ, and the dynamic continuous risk index D4Z-Risk all show a downward trend and remain stable within the safe range, then the governance measures are deemed effective. In this embodiment, when the risk score itself does not exceed the threshold but the rate of increase is abnormal, a yellow / orange warning or inspection is triggered in advance.

[0123] This invention defines the Risk Growth Rate (RGR) as the normalized rate of increase of a risk index relative to the moving average of its time window index, specifically designed to capture situations where risks are rapidly deteriorating. When the absolute value of the risk index has not yet exceeded the classification threshold, but the RGR shows an abnormal increase, the system can trigger a yellow / orange warning in advance and initiate targeted inspections, significantly improving the sensitivity to early-stage, rapidly developing risks.

[0124] This invention constructs an integrated closed-loop system encompassing prior delineation, monitoring and parameter updates, zoning and risk assessment, prediction and early warning triggering, and remediation and effect re-assimilation, ensuring that every decision can be verified through subsequent observation data and parameter rewriting. Simultaneously, the synchronous decline of leakage mobility λ, leakage significance index L, and D4Z-Risk is used as a quantitative criterion for remediation effectiveness, forming a complete technical closed loop from monitoring to remediation.

[0125] According to the appendix Figure 5 It can be seen that, Figure 5 The paper describes the temporal visualization relationship between risk index evolution, inertial smoothing, and potential hazard zone (PHZ) triggering in the multi-source coupled dynamic "three-band-one-zone" method. The figure deeply depicts the dynamic correlation between continuous risk (D4Z-Risk), risk inertia (P), and risk growth rate (RGR), and clarifies the key time nodes for PHZ entry judgment, thus intuitively demonstrating the filtering effect of "hysteresis stability judgment" on risk signals and the accuracy of PHZ identification.

[0126] Figure 5The horizontal axis represents the continuous sampling time of downhole monitoring. The entire process can be divided into three core stages according to time nodes: The first stage is the risk accumulation period, in which the D4Z-Risk curve rises slowly from a low level, the risk inertia P lags behind D4Z-Risk, and RGR gradually increases as the slope of D4Z-Risk increases; at time t3, D4Z-Risk and RGR simultaneously exceed the set threshold, indicating "entering PHZ"; the t3-t6 stage is the PHZ duration period, during which D4Z-Risk continues to rise, P tends to match it, and RGR remains at a high level. If RGR exceeds the threshold, even if D4Z-Risk has not yet reached a higher threshold, it will trigger inert injection, sealing and other mitigation measures in advance, realizing the proactive identification and prevention of risk deterioration.

[0127] According to the appendix Figure 6 It can be seen that the FOUR-Net network used in short-term forward prediction achieves high-precision fusion of thermal spatial texture of mining subsidence area and multi-sensor time series data through a full-process architecture of "hierarchical input - feature dual-branch extraction - gating adaptive fusion - physical constraint regularization - multi-dimensional output", and finally outputs a reliable partition and risk prediction map.

[0128] The network input layer accurately inputs three types of heterogeneous data in three channels: the spatial branch FI receives infrared thermal imaging data, focusing on the spatial distribution of local thermal anomalies in the goaf; the temporal branch FS receives multi-sensor time-series data, capturing the temporal evolution of each monitoring indicator; and the gating input separately receives the air leakage mobility L and the cross-modal uncertainty Unc, providing a basis for subsequent physical constraint gating fusion and ensuring that the fusion logic conforms to the actual working conditions of air leakage and temperature rise in the goaf.

[0129] In the feature extraction stage, the spatial branch extracts thermal texture features (FT) from infrared thermal images through a process of convolutional layers (Conv), ReLU activation function, and max pooling layer (MaxPool), focusing on the spatial morphology of thermal anomalies. The temporal branch extracts temporal evolution features from multi-sensor time-series data through a process of long short-term memory network (LSTM) and fully connected layers (FC), capturing the dynamic changing trends of indicators. After concatenating the air leakage mobility L and the cross-modal uncertainty Unc, a gating map M matching the feature size is generated through convolutional layers and activation functions. Temporal features FS' and Sal are weighted according to the strength of L. IR The weaker feature reduction spatial feature FT rule completes feature fusion; the gated consistency regularization layer constrains the monotonic relationship between M and L, Unc through the loss function, and the final output layer generates a three-band one-division partition map and a risk prediction map.

[0130] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may also be stored as part of a file containing other programs or data, for example, in one or more scripts within a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., files storing one or more modules, subroutines, or code sections). As an example, executable instructions may be deployed to execute on a single electronic device, or on multiple electronic devices located in one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0131] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

[0132] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0133] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, 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, 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 that element. In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multi-source zoning and risk prediction of spontaneous combustion of coal in goaf areas, characterized in that, The method includes: The raw multi-source data of the goaf are acquired and preprocessed to obtain standardized multi-source data with a unified spatiotemporal reference; the standardized multi-source data includes at least gas volume fraction time series, temperature change rate and infrared thermal image data. Based on the standardized multi-source data, the air leakage significance index is calculated; if the air leakage significance index continues to exceed the preset index threshold, the air leakage mobility rate is updated online and spatially smoothed based on the air leakage significance index to obtain the updated air leakage mobility rate. Based on the standardized multi-source data, the leakage significance index, and the updated leakage mobility, cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations are calculated and fused to generate a continuous risk index. Based on the continuous risk index, the gas volume fraction time series, the temperature change rate, the air leakage significance index, and the thermal image significance, combined with the smoothing and hysteresis threshold rules of risk inertia, a mutually exclusive three-zone and one-region distribution map is dynamically divided and output; the three-zone and one-region includes the potential danger zone, the oxidation heating zone, the asphyxiation zone, and the heat dissipation zone. The standardized multi-source data is input into the FOUR-Net model, which uses the leakage significance index and the cross-modal uncertainty as gate signals, and is then weighted, fused, and risk predicted in sequence to obtain a three-zone one-region probability map and a risk prediction map. The continuous risk index, the three-zone one-district probability map, and the risk prediction map trigger tiered early warnings and implement governance measures. After the governance measures are implemented, iterative monitoring continues, and the governance effect is verified by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, and the continuous risk index.

2. The method according to claim 1, characterized in that, The calculation of the air leakage significance index based on the standardized multi-source data includes: Based on the rate of change of the ratio of oxygen to nitrogen gas volume fractions in the time series, combined with the rate of temperature change, a preliminary index of air leakage intensity is calculated. The air leakage intensity preliminary index is smoothed over a time window and spatially interpolated to obtain the air leakage significance index; the formula for calculating the air leakage significance index is as follows: ， In the formula, , These are the measured predicted wind speed and the model predicted wind speed, respectively. For wind speed reference, For microseismic energy density, This serves as a reference for microseismic energy density. For pressure gradient mode, As a reference value for the pressure gradient modulus, This is the measured oxygen volume fraction. To predict the oxygen volume fraction, The oxygen fraction is the reference value, and α1, α2, α3, and α4 are the weights, respectively. i ≥0, ∑α i =1 represents the weight, and Logi(⋅)=1 / (1+e−z) represents the logistic mapping.

3. The method according to claim 1, characterized in that, When performing residual-driven online updates and spatial smoothing of the aforementioned leakage mobility, the online update formula is: ， In the formula, For a moment In position The update did not smooth out the leakage migration rate. For spatial location, For time, For discrete time step index, for Always in position The updated air leakage migration rate, This is a truncation function. The trigger threshold for the air leakage significance index, To update the step size coefficient, For a moment The air leakage significance index, These are the lower and upper limits of the mobility, respectively. The formula for spatial smoothing is: ， In the formula, for Always in position The final leakage migration rate after smoothing To smooth out the weights, For spatial neighborhood set, The number of grid points in the neighborhood. For the location index within the neighborhood, For the previous moment Neighborhood Location The air leakage migration rate.

4. The method according to claim 1, characterized in that, The formula for calculating the cross-modal uncertainty is as follows: ， In the formula, For cross-modal uncertainty, For spatial location, For time, For the amplitude limiting operator, For the multimodal normalized residual vector, for The 2-norm, The residual metric; The formula for calculating the significance of the thermal image is: ， In the formula, For thermal image saliency, This represents the equivalent temperature field of the infrared thermal image in the downhole physical coordinate domain. Let be the spatial gradient vector of the temperature field. Let L be the L2 norm of the spatial gradient vector of the temperature field. This is the gradient smoothing / stability constant; The formula for calculating the significance of the potential hazard is as follows: ， In the formula, For the significance of potential hazards, , , , For different weighting coefficients, and ≥0、 , , These represent the measured oxygen volume fraction and CO volume fraction, respectively. , These are the threshold values ​​for oxygen volume fraction and CO volume fraction, respectively. For the rate of temperature rise, For reference temperature rise rate, For positive part operators, The significance index of air leakage; The formula for calculating the short-window gas fluctuation is as follows: ， In the formula, Here, s represents the short-window gas fluctuation index, and s is the index of the gas type. For the set of gases participating in the statistics, For gas weighting coefficients, Let be the average volume fraction of gas s within a short time window. Let S be the standard deviation of the volume fraction of gas s within the same short time window. To prevent zero constant / stability constant.

5. The method according to claim 4, characterized in that, The continuous risk index is: ， In the formula, It is a continuous risk index. These are the fusion weighting coefficients for cross-modal uncertainty, thermal image significance, potential hazard significance, and short-window gas fluctuations, respectively.

6. The method according to claim 1, characterized in that, The exponential smoothing formula for the risk inertia is: ， In the formula, Due to risk inertia, The risk inertia of the previous moment, For smoothing coefficients, For spatial location, It is a continuous risk index. For discrete time step index; Based on the risk inertia and the entry / exit dual threshold hysteresis rule, the mutual exclusion determination of the three zones and one area is performed to obtain the distribution map of the three zones and one area.

7. The method according to claim 6, characterized in that, A spatial location is identified as the asphyxiation zone if and only if all of the following conditions are met simultaneously: Physical conditions: The oxygen volume fraction is less than or equal to the asphyxiation threshold; and the absolute value of the temperature change rate is less than or equal to a small threshold close to zero; and the carbon monoxide volume fraction is low; Spatial constraints: The distance of this spatial location along the working face advancement direction is greater than the distance of all units in that direction that have been identified as heat dissipation zones or oxidation heating zones. The location closer to the working surface shall be preferentially identified as the heat dissipation zone if and only if a spatial location simultaneously satisfies all of the following physical conditions, and if it conflicts with the determination of other zones in the same direction: Physical conditions: oxygen volume fraction is higher than the oxygen-enriched threshold; and the rate of temperature change is less than or equal to the temperature rise slope threshold; and carbon monoxide volume fraction is less than or equal to the low CO threshold. A spatial location is identified as a potential danger zone when it simultaneously meets all of the following conditions: If the leakage significance index is greater than or equal to the entry threshold, the oxygen volume fraction is higher than the oxygen enrichment threshold, the temperature change rate is greater than the temperature rise slope threshold, or the infrared thermal image significance is greater than or equal to the thermal image significance threshold, and the carbon monoxide volume fraction is less than the oxidation heating threshold; if the leakage significance index subsequently drops below the exit threshold, or the risk inertia drops below the exit threshold, then the current spatial location exits the potential danger zone. A spatial location is identified as the oxidation heating zone when all of the following conditions are met simultaneously: Physical conditions: The rate of temperature change is greater than the temperature rise slope threshold, and the carbon monoxide volume fraction is greater than or equal to the oxidation temperature rise CO threshold, and the risk inertia is greater than or equal to the entry threshold of the air leakage significance index; Spatial distribution characteristics: The oxidation temperature rise zone is mainly distributed in the area between the heat dissipation zone and the asphyxiation zone.

8. The method according to claim 1, characterized in that, The FOUR-Net model is trained in the following way: Based on the three-zone-one-region judgment rule, a partition label is generated for each spatiotemporal grid, and a continuous risk index label is generated based on the continuous risk formula; samples are constructed in a sliding time window manner, where the input of each sample is a spatiotemporal tensor composed of multi-source data and infrared thermal image frames stacked within the time window; the sample set is divided into a training set, a validation set, and a test set. The obtained sample standardized multi-source data is input into the feature extraction module of the FOUR-Net model. The spatial branch extraction module performs two-dimensional convolution processing on the infrared thermal image sequence to extract the thermal image spatial features. The temporal branch extraction module performs one-dimensional convolution and cyclic modeling on the multi-sensor temporal data to extract multi-sensor temporal features. Using the air leakage significance index and the cross-modal uncertainty as gating signals, the gating signals are concatenated with the channel-pooled representations of the thermal spatial features and the multi-sensor temporal features. A gating map is then generated using a 1×1 convolution and a sigmoid activation function. The gating map is as follows: ， In the formula, For gating graphs, For logical mapping functions, For 1×1 convolution, For thermal imaging spatial features, For multi-sensor timing characteristics, The significance index of air leakage, For channel dimension splicing operators, For uncertainty plot, Here, H is the height of the network feature map, and W is the width of the network feature map; The gated image is fused with the thermal spatial features and the multi-sensor temporal features to obtain a fused feature vector; the fused feature vector is: ， In the formula, To fuse feature vectors, For point-by-point multiplication, 1−M is a gated complementary graph; The fused feature vector is input into subsequent convolutional and fully connected layers to output the three-band, one-division partitioning probability map and the risk prediction map. Construct a total loss function; the total loss function introduces a gated consistency regularization constraint; the gated consistency regularization term is: ， In the formula, For gating consistency regularization, These are the gating regularization weight coefficients. For the average operator, For logical mapping functions, The linear coefficient of the air leakage significance index is . The linear coefficients of the cross-modal uncertainty are... For bias terms; The optimal model after training is evaluated using the test set. The accuracy, recall, and F1 score of the three-band-one-zone classification, as well as the mean square error and mean absolute error of risk prediction, are calculated to verify the model's ability to identify dangerous areas and the accuracy of risk prediction.

9. The method according to claim 1, characterized in that, The system triggers tiered early warnings and implements remedial measures using the continuous risk index, the three-zone-one-district probability map, and the risk prediction map. After implementing the remedial measures, iterative monitoring continues, and the remedial effect is verified by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, and the continuous risk index. Based on the continuous risk index and the risk prediction map, and combined with the three-zone one-district zoning probability map, the risk areas are spatially located and classified. When the continuous risk index exceeds the preset classification threshold or the risk growth rate in the risk prediction map exceeds the set threshold, a classification warning signal is triggered, and a location management instruction is generated by combining the potential danger zone and the high probability area of ​​the oxidation heating zone. After implementing inertia injection, sealing, or ventilation optimization measures, continue to collect multi-source data and iterate. Quantitatively evaluate the treatment effect by observing the synchronous decline of the updated air leakage migration rate, the air leakage significance index, the continuous risk index, and the proportion of dangerous zones in the zoning probability map. If the updated air leakage migration rate decreases, the air leakage significance index falls, the continuous risk index decreases, and the proportion of the potential danger zone and the oxidation heating zone is significantly reduced, then the treatment is deemed effective; otherwise, the treatment strategy is readjusted.

Images