A method for grading and early warning of miner state by fusing regional risk and cumulative exposure

Abstract

The application discloses a miner state grading early warning method fusing regional risk and cumulative exposure, and relates to the technical field of intelligent safety monitoring of underground mines.The application discloses a miner state grading early warning method fusing regional risk and cumulative exposure: first, a fusion sample under a unified space-time grid is acquired, multi-modal features are extracted, and cumulative exposure features and regional risk codes are calculated; then, the features are input into a recognition model combining one-dimensional and two-dimensional convolution and LSTM, the attention layer is adaptively weighted by the regional risk, and the miner state and confidence are output; finally, three-level early warnings are generated and output according to time window voting, threshold, event rules and hysteresis mechanism.The application solves the problems of difficulty in characterizing cumulative risk with instantaneous features, non-participation of regional risk in judgment, insufficient multi-modal space-time fusion, unbalanced abnormal samples and early warning jitter, and improves the miner state recognition accuracy and the stability of the early warning system.

Description

Technical Field

[0001] This invention relates to the field of intelligent safety monitoring technology for underground mines, specifically to a method for classifying and warning miners' status by integrating regional risk and cumulative exposure. Background Technology

[0002] Underground mines operate in enclosed, confined spaces with significantly varying ventilation conditions. Workers are exposed to complex environments characterized by high temperatures, high humidity, and the potential accumulation of dust and harmful gases. A miner's safety depends not only on individual physiological fluctuations and behavioral changes but also on factors such as the effectiveness of ventilation in the area, the historical frequency of abnormal harmful gas emissions, heat load levels, escape distances, and the accessibility of emergency shelters. Accurately identifying the working conditions of miners and providing timely, tiered early warnings has become a core technical challenge in underground safety monitoring.

[0003] Existing monitoring methods mainly include three categories: single-point over-limit alarms for ambient gases, personnel boundary crossing alarms, and threshold-triggered alarms from wearable devices. While these methods have some ability to identify obvious abrupt events, they lack sufficient characterization power for slowly evolving states such as fatigue accumulation, increased heat stress, and early signs of poisoning, leading to delayed identification or missed detections. Under conditions of continuous exposure to low doses of harmful gases, long-term effects of hot and humid environments, and high-intensity physical labor, personnel risk exhibits significant cumulative characteristics. Instantaneous thresholds alone cannot characterize the evolution of risk as exposure time and activity intensity increase.

[0004] Existing algorithms generally treat personnel status samples from different areas as equal, failing to incorporate regional risk attributes into the judgment system. In return airways, ventilation blind spots, areas with a high incidence of historical anomalies, and areas far from disaster refuge chambers or restricted areas, the risks associated with the same physiological fluctuations or abnormal movements are far higher than in general areas. Conversely, in low-risk areas, some transient anomalies may stem from strenuous exercise or interference from sensor wearers. This lack of differentiated regional guidance leads to frequent false alarms in low-risk areas and significant underreporting in high-risk areas, weakening the reliability of the early warning system and reducing the willingness to respond on-site.

[0005] Existing methods for utilizing personnel trajectories are limited to boundary crossing detection or dwell time statistics, failing to effectively integrate spatial distribution features such as regional dwelling clusters, path deviations, abnormal detours, and prohibited approach into the identification model. Judgment strategies are mostly based on isolated time windows, neglecting the temporal evolution of modal features between consecutive windows, making it difficult to capture the gradual transition from normal to abnormal states.

[0006] Meanwhile, the scarcity of abnormal state samples in mines leads to severe class imbalance during model training, resulting in insufficient sensitivity in recognizing abnormal states such as poisoning symptoms, heat stress, and falls. Regarding early warning output, single-threshold or single-window strategies are prone to frequent alarm jumps at class boundaries, and lack a reasonable distinction between slow-moving variable states and sudden events, making it difficult to achieve stable and reliable hierarchical early warning systems.

[0007] Therefore, addressing the technical challenges of how instantaneous features can not adequately represent cumulative risks, how regional risk attributes are not included in model judgment, how multimodal spatiotemporal feature fusion is insufficient, how the scarcity of abnormal samples leads to identification bias, and how early warning outputs lack a graded hysteresis mechanism are key areas that urgently need to be tackled in this field.

[0008] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a method for classifying and issuing early warnings of miner status by integrating regional risk and cumulative exposure, thereby resolving the issues raised in the background section. The method includes the following steps: A method for classifying and warning miners' status by integrating regional risk and cumulative exposure includes the following steps: S1. Obtain personnel-environment-region fusion samples of workers under a unified time axis and a unified spatial grid. The fusion samples include workers' location information, physiological information, action information, regional environmental information, and regional attribute information. S2. Construct a sliding time window based on the fused samples, and extract physiological time sequence features, action behavior features, and personnel trajectory features within each sliding time window; S3. Calculate the cumulative exposure characteristics of workers within the sliding time window based on regional environmental information, worker activity intensity, and time spent in the area; S4. Construct regional risk codes based on regional attribute information. Regional risk codes include ventilation risk codes, gas risk codes, heat load risk codes, escape difficulty codes, and restricted access attribute codes. S5. Input physiological time sequence characteristics, action and behavior characteristics, personnel trajectory characteristics, cumulative exposure characteristics, and regional risk codes into the status recognition model; S6. The state recognition model includes a one-dimensional convolutional sub-network, a two-dimensional convolutional sub-network, a risk embedding layer, a feature concatenation and normalization module, a long short-term memory network, a region risk-guided attention layer, and a classifier; the outputs of the one-dimensional convolutional network, the two-dimensional convolutional network, and the risk embedding layer are concatenated by the feature concatenation and normalization module and then input into the long short-term memory network. S7. The regional risk-guided attention layer uses regional risk coding to generate attention weights and performs weighted fusion of features from each modality to obtain multimodal fusion features. S8. The classifier outputs the worker's status category and confidence level based on the multimodal fusion features. The status categories include normal work, fatigue, heat stress, signs of poisoning, and abnormal behavior. S9. Generate graded early warning results based on status category, confidence level, duration, regional risk code, and event priority rules, and send the graded early warning results to wearable terminals, mining lamps, voice broadcasting systems, light strip indicator systems, or dispatch platforms.

[0010] Preferably, in step S2, the physiological time-series characteristics include mean heart rate, heart rate change rate, blood oxygen saturation rate, and skin temperature rise rate. Motor behavior characteristics include cadence, number of posture changes, fall impact intensity, duration of prolonged stillness, and path deviation. Personnel trajectory characteristics include area stay time, number of area switching, trajectory heatmap, and area occupancy matrix.

[0011] Preferably, in step S3, the cumulative exposure features are calculated according to E=Σ[R env(t) Calculate [×A(t)×Δt]; Where E represents the cumulative exposure characteristic, R env(t) Let A(t) be the environmental risk intensity at time t, A(t) be the activity intensity of the workers at time t, and Δt be the sampling time interval.

[0012] Preferably, the environmental risk intensity R env(t) It is obtained by weighting the risk items of harmful gases, oxygen deficiency, and thermal environment. The hazardous gas risk item is determined by the normalized degree of exceedance of the concentration of at least one of the gases, namely carbon monoxide, hydrogen sulfide, methane, and nitrogen dioxide, relative to the corresponding safety threshold; the thermal environment risk item is determined by temperature, humidity, wind speed, and the intensity of personnel activity.

[0013] Preferably, in step S4, the regional risk code is generated based on the ventilation conditions, historical frequency of gas anomalies, temperature baseline, escape distance, accessibility of disaster refuge chambers, no-entry signs, and communication coverage level of the work area, and is input into the state recognition model in any form of discrete level code, continuous normalized code, or embedded vector.

[0014] Preferably, in step S6, the one-dimensional convolutional subnetwork includes at least one one-dimensional convolutional layer and pooling layer, used to extract local change features of physiological and environmental sequences, including heart rate, blood oxygen, temperature, gas concentration and cumulative exposure sequence. The two-dimensional convolutional subnetwork includes at least one two-dimensional convolutional layer and a pooling layer, used to extract the spatial distribution features of personnel trajectory heatmaps or region occupancy matrices. The spatial distribution features include spatial clustering, path deviation, and proximity to restricted areas. The risk embedding layer is used to convert cumulative exposure features and regional risk codes into risk semantic features; the long short-term memory network is used to extract state evolution features within a continuous time window.

[0015] Preferably, in step S7, the regional risk-guided attention layer generates modal attention weights based on the regional risk code, and increases the weights of environmental features in high ventilation risk areas, physiological temperature rise features in high heat load areas, and abnormal behavior features in areas with difficult escape.

[0016] Preferably, in step S9, the graded early warning results include a warning level, a medium-level warning, and a high-level warning; A high-level warning is triggered when the status category is a fall, signs of poisoning, trespassing into a restricted area, or prolonged stillness and the confidence level is higher than the first threshold. When the state category is fatigue or heat stress and occurs within multiple consecutive time windows with a confidence level higher than the second threshold, a medium-level warning is triggered; when the state category is transient anomaly and the confidence level is lower than the second threshold, a prompt-level warning is triggered or observation continues.

[0017] Preferably, the training samples of the state recognition model are generated based on on-site annotations, team inspection records, dispatch and disposal records, wearable terminal event tags, and environmental over-limit records. Class resampling, loss function weighting, or hard example mining are used to reduce the class bias caused by the scarcity of abnormal state samples.

[0018] The present invention provides a miner status classification and early warning method that integrates regional risk and cumulative exposure, which has the following beneficial effects: (1) By constructing cumulative exposure features, this invention couples and characterizes the concentration of harmful gases, thermal environment indicators, intensity of human activity and time spent in the area, enabling the model to capture the increasing evolution of risk under low-dose long-term exposure and continuous hot and humid environment. This effectively solves the problem of the lag in the identification of slow evolution states by instantaneous features, and significantly improves the timeliness of early warning of cumulative risks such as fatigue, heat stress and early signs of poisoning.

[0019] (2) Through the collaborative design of regional risk coding and attention mechanism, this invention enables the state recognition model to adaptively adjust the attention weights of physiological, action, trajectory, environment and cumulative exposure characteristics based on the wind attributes, gas anomaly history, heat load level, escape difficulty and restricted access attributes of the work area. This realizes differentiated guidance of regional risk for multimodal feature judgment, effectively reducing the false alarm rate of low-risk areas and the false alarm rate of high-risk areas.

[0020] (3) This invention achieves the joint extraction of local temporal change features, trajectory spatial distribution features and continuous time window state evolution features through the cascaded architecture of one-dimensional convolutional sub-network, two-dimensional convolutional sub-network and long short-term memory network. It makes up for the shortcomings of existing methods in making insufficient use of trajectory spatial information and isolated time window decision-making, and enhances the model's ability to identify the gradual process of normal state to abnormal state.

[0021] (4) To address the problem of scarce abnormal state samples, this invention employs a training strategy that combines class resampling, loss function weighting, and hard example mining. This effectively suppresses the identification bias caused by class imbalance and improves the classification sensitivity for abnormal states such as poisoning symptoms, heat stress, and falls. Simultaneously, the hierarchical early warning mechanism, combined with event priority rules and hysteresis control, implements differentiated responses to sudden high-risk events and slow variable accumulation states, and eliminates frequent jumps in early warning levels at class boundaries, ensuring the stability and reliability of early warning output. Attached Figure Description

[0022] Figure 1 This is a structural diagram of the worker status recognition and hierarchical early warning system of the present invention.

[0023] Figure 2 This is a flowchart illustrating the construction process of cumulative exposure features and regional risk codes for this invention.

[0024] Figure 3 This is a structural diagram of the two-stage state recognition model guided by regional risk in this invention.

[0025] Figure 4 This is a flowchart of the continuous time window hierarchical early warning determination process of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of 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.

[0027] Example 1: System Composition and Work Process of the Invention

[0028] Figure 1As shown, the system in this embodiment includes a fusion sample input 1, a regional attribute input 2, a physiological action feature construction module 3, a cumulative exposure feature calculation module 4, a regional risk coding module 5, a state recognition model 6, a state category and confidence output module 7, a graded early warning module 8, and early warning output channels 9 to 11. The fusion sample input 1 can come from any data processing system that has completed spatiotemporal registration and spatial registration, including personnel location, physiological parameters, action parameters, environmental parameters, regional grid identifiers, and sample confidence.

[0029] The physiological movement feature construction module 3 extracts the mean heart rate, heart rate change rate, blood oxygen saturation decrease, skin temperature rise rate, cadence, number of posture changes, fall impact intensity, duration of prolonged stillness, and path deviation within each sliding time window. Personnel trajectories can be converted into trajectory heatmaps or area occupancy matrices, where matrix elements represent the frequency, duration, or number of times a worker stays within a corresponding grid area.

[0030] Figure 2 As shown, the cumulative exposure characteristic calculation module 4 receives the gas concentration sequence 12, thermal environment sequence 13, personnel activity intensity 14, and residence time 15. The system calculates the environmental risk intensity R for each sampling time. env(t) Then, multiplying the exposure increment by the activity intensity A(t) and the sampling interval Δt yields the exposure increment 16, which is then accumulated within a sliding window to obtain the cumulative exposure feature 17.

[0031] In one specific implementation, the environmental risk intensity R env(t) R can be calculated as follows: env(t) =β1R gas(t) +β2R o2(t) +β3R heat(t) , where R gas(t) For hazardous gas risk items, R o2(t) For the risk item of hypoxia, R heat(t) For the thermal environment risk term, β1 to β3 are weighting coefficients. R gas(t) R can be determined by the normalized degree of exceedance of the concentration of carbon monoxide, hydrogen sulfide, methane, or nitrogen dioxide relative to the corresponding safety threshold, and R can be determined by temperature, humidity, wind speed, and intensity of personnel activity.

[0032] The regional risk coding module 5 receives ventilation risk 18, gas risk 19, heat load risk 20, and escape difficulty 21, and outputs a risk coding vector 22. Ventilation risk can be determined based on regional air volume, wind speed, return air attributes, and local ventilation status; gas risk can be determined based on historical gas anomaly frequency, sensor deployment density, and recent gas fluctuations; heat load risk can be determined based on regional temperature baseline, humidity, and heat dissipation conditions; and escape difficulty can be determined based on distance to the exit or refuge chamber, tunnel slope, passage status, and restricted access attributes.

[0033] Figure 3 As shown, the state recognition model 6 includes a one-dimensional convolutional sub-network 27, a two-dimensional convolutional sub-network 28, a risk embedding layer 29, a feature concatenation and normalization module 30, a long short-term memory network 31, a region risk-guided attention layer 32, and a classifier 33. The one-dimensional convolutional sub-network 27 processes the one-dimensional sequence 24 of the physiological environment, extracting features such as local peaks, downward trends, continuous increases, periodic fluctuations, and short-term disturbances. The two-dimensional convolutional network 28 processes trajectory heatmaps or region occupancy matrices 25, extracting features such as personnel dwelling clusters, path deviations, abnormal detours, and proximity to prohibited areas. The risk embedding layer 29 converts cumulative exposure and region risk encoding 26 into risk semantic features.

[0034] The Long Short-Term Memory (LSTM) network 31 receives spliced ​​multimodal features and models the state evolution relationship within a continuous time window. For slow-variable states such as fatigue, heat stress, and signs of poisoning, the LSM network 31 can utilize the gradually accumulated physiological, environmental, and action changes within the historical window; for event-type states such as falls and intrusion into restricted areas, the model can combine two-dimensional trajectory features and action impact features for rapid discrimination.

[0035] The regional risk-guided attention layer 32 generates modal attention weights based on the regional risk code. Specifically, the regional risk code R can be input into the fully connected layer to obtain the attention query vector, and then the correlation score is calculated with each modal feature vector. After normalization, the attention weights are obtained. If the worker is located in an area with high ventilation risk, the weights of gas concentration, wind speed, and cumulative exposure features are increased; if the worker is located in an area with high heat load, the weights of skin temperature rise, heart rate change, and thermal environment features are increased; if the worker is located in an area with difficult escape, the weights of long-term stillness, path deviation, and prohibited access features are increased.

[0036] The classifier 33 outputs state categories and confidence levels 34. State categories can include normal operation, fatigue, heat stress, signs of poisoning, and abnormal behavior. Abnormal behavior can further include falls, prolonged stillness, intrusion into restricted areas, abnormal deviation from the trajectory, and entering dangerous areas in the wrong direction. The model output not only provides the state with the highest confidence level but also outputs the probability distribution of each state, so that the graded early warning module 8 can make a comprehensive judgment.

[0037] like Figure 4 As shown, the graded early warning module 8 receives the state category and model confidence level 35, duration 36, area risk level and personnel location 37, and executes threshold and hysteresis judgment 38 and event priority rule 39. The alert level 40 is used for short-term low-confidence anomalies or slight state fluctuations; the medium-level alert 41 is used for states such as fatigue and heat stress that persist for multiple consecutive time windows; and the high-level alert 42 is used for high-risk events such as signs of poisoning, falls, trespassing into restricted areas, and prolonged periods of stillness.

[0038] Warning output 43 can be achieved through wearable terminal vibration, mine lamp flashing, voice broadcasting, lamp strip direction indication, and dispatch platform map display. For high-level warnings, the system can simultaneously send personnel identification, area grid location, status category, confidence level, duration, and recent environmental parameters to on-site personnel, team leaders, and the dispatch center; for medium-level warnings, the system can prompt workers to rest, replenish water, evacuate high-temperature areas, or accept team verification; for alert-level warnings, the system can provide local notifications and continue monitoring.

[0039] During the model training phase, training samples can be generated from on-site annotations, shift inspection records, dispatch and handling records, wearable terminal event tags, and environmental exceedance records. To address the issue of a limited number of abnormal samples, training can be conducted using methods such as category resampling, loss function weighting, hard example mining, or data augmentation. During model deployment, basic feature extraction can be completed at the edge gateway, while complex model inference can be performed on the ground server or the mine dispatch host. When communication quality is limited, a lightweight model can be run at the edge to output temporary early warnings.

[0040] The focus of this invention is on constructing cumulative exposure features and regional risk codes based on personnel-environment-region fusion samples, and utilizing a regional risk-guided attention mechanism for multimodal state identification and hierarchical early warning. This invention does not limit the fusion samples to be generated by a specific data fusion system; it can be implemented as long as the input data meets the requirements of a unified time axis and a unified spatial grid representation.

[0041] Example 2: Specific Application Example

[0042] This embodiment uses the inspection operation of a -520 m level return air connecting roadway in an underground mine as an example to illustrate the complete process from raw data acquisition, spatiotemporal alignment, fusion sample generation, cumulative exposure and regional risk calculation, to state identification and graded early warning output. The following parameters are a set of reproducible configurations for this embodiment, used to illustrate the implementation of the present invention, and do not constitute a limitation on sensor model, sampling frequency, risk threshold, or network size.

[0043] Step 1: Raw Data Input. The worker, designated P017, wears a positioning tag, a wrist-worn physiological terminal, and an inertial measurement unit (IMU) on a safety helmet. The positioning tag outputs local mine coordinates at 1 Hz, the physiological terminal outputs heart rate, blood oxygen saturation, and skin temperature at 1 Hz, and the IMU acquires triaxial acceleration and attitude angles at 50 Hz. Environmental monitoring station E23 outputs carbon monoxide, hydrogen sulfide, methane, oxygen, temperature, relative humidity, and wind speed at 1 Hz. The regional attribute database stores information on roadway type, local ventilation status, frequency of gas anomalies over the past 30 days, temperature baseline, distance to the refuge chamber, no-entry signs, and communication coverage. The system uses a uniform sampling period of 1 second, with a spatial grid size of 5 m × 5 m × 3 m.

[0044] Taking the 60-second data from 14:30:00 to 14:31:00 as an example. For ease of illustration, a representative record is extracted from the original 1-second data every 10 seconds, with the starting times of the six segments being 14:30:00, 14:30:10, 14:30:20, 14:30:30, 14:30:40, and 14:30:50 respectively. The corresponding heart rate sequences were 102, 105, 108, 111, 114, and 117 beats / min; blood oxygen saturation sequences were 96%, 96%, 95%, 95%, 94%, and 94%; skin temperature sequences were 34.8 ℃, 34.9 ℃, 35.0 ℃, 35.2 ℃, 35.3 ℃, and 35.4 ℃; and cadence sequences were 88, 92, 96, 100, 102, and 105 steps / min. The normalized composite acceleration index was based on the gravitational acceleration at rest, with a baseline value of 1. The normalized composite acceleration indices for the six sub-segments were 1.247, 1.303, 1.360, 1.417, 1.487, and 1.500, respectively. The positioning coordinates are (126.4, 56.1, -520.4), (128.1, 56.2, -520.4), (131.6, 56.3, -520.4), (134.8, 56.5, -520.4), (137.2, 56.7, -520.4), and (139.0, 56.8, -520.4), in meters.

[0045] The environmental data for the six sub-segments are as follows: carbon monoxide 15, 17, 19, 21, 23, 24 ppm; hydrogen sulfide 2.5, 3.0, 3.5, 4.0, 4.5, 5.0 ppm; methane 0.35%, 0.38%, 0.41%, 0.43%, 0.46%, 0.48%; oxygen 20.4%, 20.3%, 20.2%, 20.2%, 20.1%, 20.0%; temperature 30.0, 30.2, 30.5, 30.7, 31.0, 31.2 ℃; relative humidity 82%, 83%, 83%, 84%, 85%, 86%; wind speed 0.48, 0.45, 0.42, 0.40, 0.37, 0.35 m / s. The area's attributes are recorded as follows: return air connecting alley, local ventilation status score 0.80, 6 gas anomalies in the past 30 days, baseline temperature 29.4 ℃, distance to the disaster refuge chamber 860 m, disaster refuge route accessible, no-entry sign 0, and communication coverage 92%.

[0046] Step 2: Spatiotemporal alignment and sample fusion output. The edge gateway uses the location tag time as a reference, correcting for the +0.42 s clock deviation of the physiological terminal relative to the reference and the -0.18 s clock deviation of the environmental monitoring station relative to the reference. Single-point missing data of no more than 2 s is interpolated linearly, and channels with consecutive missing data exceeding 3 s are invalidated. All channels within this time window are valid. Spatial mapping is performed according to the grid origin (0, 0, -525) and a grid size of 5 m × 5 m × 3 m. P017 occupies G25-11-01, G26-11-01, and G27-11-01 sequentially. Environmental monitoring station E23 is 12.4 m away from the current person, located in the same ventilation zone, and has an association radius of less than 30 m; therefore, its environmental data is associated with the personnel sample. The sample confidence level was calculated as q = 0.35qpos + 0.30qphy + 0.20qact + 0.15qenv, where the confidence levels for location, physiology, action, and environment channels were 0.96, 0.94, 0.92, and 0.91, respectively, resulting in q = 0.9385.

[0047] Taking 14:30:40 as an example, a 1 Hz personnel-environment-region fusion sample output by the system is as follows: {Personnel=P017, Time=14:30:40, Coordinates=(137.2, 56.7, -520.4), Grid=G27-11-01, Heart Rate=114 beats / min, Blood Oxygen=94%, Skin Temperature=35.3 ℃, Step Frequency=102 steps / min, Equivalent Acceleration=1.487, Carbon Monoxide=23 ppm, Hydrogen Sulfide=4.5 ppm, Methane=0.46%, Oxygen=20.1%, Ambient Temperature=31.0 ℃, Relative Humidity=85%, Wind Speed=0.37 m / s, Region Attributes=(Return Airway, 0.80, 6 times, 29.4 ℃, 860 m, Accessible, Not Prohibited, 92%), Sample Confidence=0.9385}. The remaining 59 1-second fused samples are output using the same field structure, thus completing the transformation from heterogeneous original data to fused samples with a unified time axis and a unified spatial grid.

[0048] Step 3: Sliding Time Window and Multimodal Feature Construction. The window length is set to 60 s and the sliding step size to 30 s. The following data were calculated for the window: mean heart rate 109.5 bpm, heart rate change rate (117-102) / 50 = 0.30 bpm, blood oxygen saturation decrease 96%-94% = 2 percentage points, skin temperature rise rate (35.4-34.8) / 50 = 0.012 ℃ / s, average cadence 97.2 steps / min, number of posture changes 0, maximum impact intensity 1.36 g, duration of prolonged stillness 0 s, and maximum deviation of trajectory from the planned route 6.8 m. The personnel trajectory was projected onto a 10×10 area occupancy matrix. The occupancy ratios of G25-11-01, G26-11-01, and G27-11-01 were all 20 / 60 = 0.333, with the remaining matrix elements being 0.

[0049] Step four, cumulative exposure feature calculation. In this embodiment, the activity intensity is set as A(t) = clip[0.60 + 0.004f]. step (t)+0.60(a r [(t)-1), 0.60, 1.50], where, For step frequency, This is a normalized composite acceleration index with the static gravitational acceleration as the reference value of 1. To characterize the additional dynamic acceleration intensity generated by personnel movement, the A(t) values ​​for the six segments are 1.10, 1.15, 1.20, 1.25, 1.30, and 1.32, respectively. The gas normalization parameter T is set. CO =24 ppm, T H2S =10 ppm, T CH4=1.0%, calculate rCO=min(C) respectively. CO / T CO ,1.5), rH2S=min(C H2S / T H2S ,1.5) and rCH4=min(C CH4 / TC H4 , 1.5), and let Rgas = 0.60rCO + 0.25rH2S + 0.15rCH4. The hypoxia risk term is taken as Ro2 = clip[(20.9-C O2 [) / (20.9-19.5), 0, 1]; Thermal environment risk item takes R heat =0.50clip[(T-26) / 8,0,1]+0.30clip[(R H -60) / 40, 0, 1]+0.20clip[(0.8-v) / 0.8, 0, 1]. Here, clip means restricting the calculation result to between 0 and 1.

[0050] Given β1=0.45, β2=0.20, and β3=0.35, and using Renv=0.45R gas +0.20R o2 +0.35R heat Calculate. (R) of the six sub-segments gas R o2 R heat R env The values ​​are (0.4900, 0.3571, 0.4950, 0.4652), (0.5570, 0.4286, 0.5225, 0.5192), (0.6240, 0.5000, 0.5487, 0.5729), (0.6895, 0.5000, 0.5737, 0.6111), (0.7565, 0.5714, 0.6075, 0.6673), and (0.7970, 0.6429, 0.6325, 0.7086). Substituting Δt = 10 / 60 min into E = Σ[R] env(t) The six exposure increments are 0.0853, 0.0995, 0.1146, 0.1273, 0.1446 and 0.1559 respectively. Therefore, the cumulative exposure characteristic E of this 60 s window is 0.7272 risk equivalent·min. After normalization with a 1 min window length, the cumulative exposure intensity is 0.7272.

[0051] Step 5, Regional Risk Coding. The return air attribute, local ventilation status, frequency of gas anomalies, temperature baseline, escape distance, accessibility of refuge chambers, no-entry signs, and communication coverage are linearly normalized according to the mine's preset range, resulting in ventilation risk (0.81), gas risk (0.70), heat load risk (0.85), escape difficulty (0.75), no-entry attribute (0), and communication loss risk (0.08). The regional risk code is R = (0.81, 0.70, 0.85, 0.75, 0, 0.08). To facilitate early warning display, the regional comprehensive risk value Q = 0.709 is calculated using weights (0.25, 0.20, 0.25, 0.20, 0.05, 0.05). In this embodiment, Q < 0.30 is set as low risk, 0.30 ≤ Q < 0.60 as medium risk, and Q ≥ 0.60 as high risk. Therefore, the current grid is determined to be a high-risk area.

[0052] Step Six: Model Input and Inference Parameters. A 60×13 one-dimensional sequence is input into a one-dimensional convolutional subnetwork. The 13 channels represent heart rate, blood oxygen, skin temperature, cadence, activity intensity, concentrations of three gases, oxygen, ambient temperature, relative humidity, wind speed, and cumulative exposure. The one-dimensional convolutional subnetwork uses 32 convolutional kernels of length 5, 2x max pooling, and 64 convolutional kernels of length 3. A 10×10 region occupancy matrix is ​​input into a two-dimensional convolutional subnetwork, using 16 3×3 convolutional kernels, 2×2 max pooling, and 32 3×3 convolutional kernels. Region risk encoding is embedded through a 16-dimensional fully connected layer, and the features from each branch are concatenated and input into a 64-unit Long Short-Term Memory (LSTM) network. ReLU is used as the activation function, and Softmax is used as the classifier. The modal weights of the five groups of physiological, action, trajectory, environment and cumulative exposure outputs of the risk-guided attention layer are 0.29, 0.12, 0.08, 0.27 and 0.24 respectively, reflecting the enhancement of physiological, environmental and cumulative exposure characteristics in areas with high heat load and high ventilation risk.

[0053] Step 7: Status Category and Graded Warning Output. The classifier outputs probabilities of 0.06, 0.12, 0.74, 0.05, and 0.03 for normal operation, fatigue, heat stress, signs of poisoning, and abnormal behavior, respectively. Therefore, the status category for this window is heat stress, with a confidence level of 0.74. The confidence levels for heat stress in the previous two adjacent windows (14:29:00 to 14:30:00 and 14:29:30 to 14:30:30) were 0.68 and 0.71, respectively. A first threshold of 0.80 and a second threshold of 0.65 are set. Fatigue or heat stress triggers a medium-level warning when the second threshold is reached for three consecutive windows; falls, signs of poisoning, intrusion, or prolonged stillness trigger a high-level warning directly when the first threshold is reached. In this example, there are no event-type high-risk indicators, and heat stress has been above 0.65 for three consecutive windows, therefore a medium-level warning is output. The hysteresis release condition is set to a state confidence level that is below 0.45 for three consecutive windows to avoid frequent changes in the warning level.

[0054] The final output record is as follows: {Personnel = P017, Time = 14:31:00, Grid = G27-11-01, Regional Comprehensive Risk = 0.709, Cumulative Exposure Characteristics = 0.7272, Status = Heat Stress, Confidence Level = 0.74, Duration = 90 s, Warning Level = Medium, Main Basis = (Skin Temperature Rise 0.012 ℃ / s, Blood Oxygen Decrease 2 percentage points, Ambient Temperature 31.2 ℃, Wind Speed ​​0.35 m / s), Handling Recommendation = Stop High-Intensity Operations, Replenish Water and Evacuate to a Well-Ventilated Area}. Simultaneously, the system controls the P017 wearable terminal to vibrate twice continuously and the miner's lamp to flash yellow at 1 Hz for 30 s, and sends the personnel identification, location, risk value, status probability, and handling recommendation to the team leader's terminal and the dispatch platform. This completes a closed-loop application from raw data input, fused sample output to cumulative exposure calculation, regional risk guidance identification, and graded warning output.

[0055] It should be noted that, in order to facilitate an accurate understanding of the common meanings of the English abbreviations in this invention document, the following explanations are provided in this invention document, specifically referring to the English abbreviations and their Chinese meanings in Table 1.

[0056] Table 1: Comparison of English Abbreviations and Chinese Meanings

[0057] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for classifying and issuing early warnings of miners' status by integrating regional risk and cumulative exposure, characterized in that, Includes the following steps: S1. Obtain personnel-environment-region fusion samples of workers under a unified time axis and a unified spatial grid. The fusion samples include workers' location information, physiological information, action information, regional environmental information, and regional attribute information. S2. Construct a sliding time window based on the fused samples, and extract physiological time sequence features, action behavior features, and personnel trajectory features within each sliding time window; S3. Calculate the cumulative exposure characteristics of workers within the sliding time window based on regional environmental information, worker activity intensity, and time spent in the area; S4. Construct regional risk codes based on regional attribute information. Regional risk codes include ventilation risk codes, gas risk codes, heat load risk codes, escape difficulty codes, and restricted access attribute codes. S5. Input physiological time sequence characteristics, action and behavior characteristics, personnel trajectory characteristics, cumulative exposure characteristics, and regional risk codes into the status recognition model; S6. The state recognition model includes a one-dimensional convolutional sub-network, a two-dimensional convolutional sub-network, a risk embedding layer, a feature concatenation and normalization module, a long short-term memory network, a region risk-guided attention layer, and a classifier; the outputs of the one-dimensional convolutional network, the two-dimensional convolutional network, and the risk embedding layer are concatenated by the feature concatenation and normalization module and then input into the long short-term memory network. S7. The regional risk-guided attention layer uses regional risk coding to generate attention weights and performs weighted fusion of features from each modality to obtain multimodal fusion features. S8. The classifier outputs the worker's status category and confidence level based on the multimodal fusion features. The status categories include normal work, fatigue, heat stress, signs of poisoning, and abnormal behavior. S9. Generate graded early warning results based on status category, confidence level, duration, regional risk code, and event priority rules, and send the graded early warning results to wearable terminals, mining lamps, voice broadcasting systems, light strip indicator systems, or dispatch platforms.

2. The method for classifying and warning miners' status based on integrated regional risk and cumulative exposure according to claim 1, characterized in that, In step S2, the physiological time-series characteristics include mean heart rate, rate of change of heart rate, magnitude of decrease in blood oxygen, and rate of increase in skin temperature; Motor behavior characteristics include cadence, number of posture changes, fall impact intensity, duration of prolonged stillness, and path deviation. Personnel trajectory characteristics include area stay time, number of area switching, trajectory heatmap, and area occupancy matrix.

3. The miner status classification and early warning method based on the integration of regional risk and cumulative exposure as described in claim 1, characterized in that, In step S3, the cumulative exposure features are calculated according to E=Σ[R env(t) Calculate [×A(t)×Δt]; Where E represents the cumulative exposure characteristic, R env(t) Let A(t) be the environmental risk intensity at time t, A(t) be the activity intensity of the workers at time t, and Δt be the sampling time interval.

4. The miner status classification and early warning method based on the integration of regional risk and cumulative exposure as described in claim 3, characterized in that, Environmental risk intensity R env(t) It is obtained by weighting the risk items of harmful gases, oxygen deficiency, and thermal environment. The hazardous gas risk item is determined by the normalized degree of exceedance of the concentration of at least one of the gases, namely carbon monoxide, hydrogen sulfide, methane, and nitrogen dioxide, relative to the corresponding safety threshold; the thermal environment risk item is determined by temperature, humidity, wind speed, and the intensity of personnel activity.

5. The method for classifying and warning miners' status based on integrated regional risk and cumulative exposure according to claim 1, characterized in that, In step S4, the regional risk code is generated based on the ventilation conditions, historical gas anomaly frequency, temperature baseline, escape distance, accessibility of disaster refuge chambers, no-entry signs, and communication coverage level of the work area, and is input into the state recognition model in any form of discrete level code, continuous normalized code, or embedded vector.

6. The method for classifying and warning miners' status based on integrated regional risk and cumulative exposure according to claim 1, characterized in that, In step S6, the one-dimensional convolutional subnetwork includes at least one one-dimensional convolutional layer and pooling layer, used to extract local change features of physiological and environmental sequences, including heart rate, blood oxygen, temperature, gas concentration and cumulative exposure sequence. The two-dimensional convolutional subnetwork includes at least one two-dimensional convolutional layer and a pooling layer, used to extract the spatial distribution features of personnel trajectory heatmaps or region occupancy matrices. The spatial distribution features include spatial clustering, path deviation, and proximity to restricted areas. The risk embedding layer is used to convert cumulative exposure features and regional risk codes into risk semantic features; Long Short-Term Memory (LSTM) networks are used to extract state evolution features within continuous time windows.

7. The method for classifying and early warning of miners' status by integrating regional risk and cumulative exposure according to claim 1, characterized in that, In step S7, the regional risk-guided attention layer generates modal attention weights based on the regional risk code, and increases the weights of environmental features in high ventilation risk areas, physiological temperature rise features in high heat load areas, and abnormal behavior features in areas with difficult escape.

8. The method for classifying and warning miners' status based on integrated regional risk and cumulative exposure according to claim 1, characterized in that, In step S9, the graded early warning results include alert-level early warning, medium-level early warning, and high-level early warning; A high-level warning is triggered when the status category is a fall, signs of poisoning, trespassing into a restricted area, or prolonged stillness and the confidence level is higher than the first threshold. A medium-level warning is triggered when the state category is fatigue or heat stress and occurs within multiple consecutive time windows with a confidence level higher than the second threshold. When the status category is transient anomaly and the confidence level is below the second threshold, a warning or continued observation is triggered.

9. The method for classifying and warning miners' status based on integrated regional risk and cumulative exposure according to claim 1, characterized in that, The training samples for the state recognition model are generated based on on-site annotations, team inspection records, dispatch and handling records, wearable terminal event tags, and environmental over-limit records. Class resampling, loss function weighting, or hard example mining are used to reduce the class bias caused by the scarcity of abnormal state samples.

Images