An automatic matching method for coal mine accident emergency plan

Abstract

The application discloses a kind of automatic matching methods of coal mine accident emergency plan, belong to coal mine safety production and emergency management technical field, this method includes: to the pre-processing of multi-source monitoring data, construct numerical characteristics and text features and carry out fusion;Based on fusion feature, use classification model to calculate disaster occurrence probability, determine target disaster type;Conduct confidence evaluation, when lower than first threshold, trigger manual intervention, otherwise generate accident situation information;Rule screening is carried out to plan library to obtain candidate set, and with situation information is semantically matched, determine the highest matching degree of target plan;According to second threshold value, judge matching result, meet the condition when output target plan, otherwise output general plan or trigger manual intervention.The present application realizes comprehensive perception by multi-source data fusion, adopts double-threshold grading decision to ensure matching reliability, and can be based on feedback dynamic optimization threshold, significantly improve the intelligent level and disposal efficiency of coal mine emergency response.

Description

Technical Field

[0001] This application relates to the field of coal mine safety production and emergency management technology, and in particular to an automatic matching method for coal mine accident emergency plans. Background Technology

[0002] As coal mining progresses deeper, the underground working environment becomes increasingly complex, easily leading to various accidents such as fires, floods, roof collapses, gas explosions, and dust explosions. Traditional coal mine safety monitoring mainly relies on setting fixed thresholds for single or a few sensor parameters to trigger alarms. Once a parameter (such as gas concentration) exceeds the set value, an alarm is triggered. However, this approach has the following problems:

[0003] Unable to distinguish disaster types: Different disaster types may exhibit similar parameter anomalies (e.g., elevated CO may be caused by fire or explosion), making it impossible to determine the specific nature of the disaster.

[0004] False alarms and false alarms: A single threshold is prone to false alarms (parameter fluctuations caused by non-catastrophic reasons) or false alarms (weak changes in multiple parameters together indicate a disaster).

[0005] Delayed or inaccurate emergency response: Due to the inability to accurately determine the type of disaster, only general emergency plans (such as evacuating the entire mine) can be activated, which is inefficient and may cause the best time for response to be missed.

[0006] Relying on manual judgment: Complex abnormal situations still require safety management personnel to make judgments based on their experience, which is time-consuming and carries the risk of human error.

[0007] Existing technologies are insufficient for accurately predicting potential coal mine disasters and providing targeted emergency responses. Therefore, there is an urgent need for an automatic emergency plan matching method that can fuse and model multi-source information, is interpretable and adaptively optimize, and automatically match the optimal emergency plan for different disaster types to improve the speed and accuracy of emergency response. Summary of the Invention

[0008] This invention addresses the technical problems existing in the background art by proposing an automatic matching method for emergency response plans for coal mine accidents. It utilizes the fusion of multi-source sensor data and text data to identify coal mine accident disaster types, and automatically matches and pushes emergency response plans based on a rule engine and semantic similarity retrieval. It also involves dynamic threshold adaptive optimization to improve the accuracy, timeliness, and pertinence of emergency response.

[0009] To solve the technical problem, the technical solution of the present invention is as follows:

[0010] An automatic matching method for coal mine accident emergency response plans, comprising:

[0011] Multi-source monitoring data are preprocessed and feature constructed to obtain numerical features and text features. The numerical features and text features are then fused to form fused features for accident identification.

[0012] Based on the fusion features, the probability of occurrence of each disaster type is calculated using a disaster classification model to determine the target disaster type;

[0013] Confidence assessment is performed based on the probability of occurrence of each disaster type. When the classification confidence is lower than the first threshold, manual intervention is triggered or a general contingency plan is output. When the classification confidence meets the preset conditions, accident scenario information corresponding to the target disaster type is generated.

[0014] Based on the accident scenario information, the contingency plan database is filtered according to rules to obtain a set of candidate contingency plans;

[0015] The accident scenario information and candidate contingency plans are semantically matched to determine the target contingency plan with the highest matching degree.

[0016] The matching result is compared with the second threshold. When the matching degree meets the preset conditions, the target plan is output; when it does not meet the conditions, a general plan is output or manual intervention is triggered.

[0017] Furthermore, before preprocessing and feature construction of the multi-source monitoring data, the method also includes: acquiring multi-source monitoring data from underground coal mines, wherein the multi-source monitoring data includes sensor numerical data and textual information data; the sensor numerical data includes gas concentration, carbon monoxide concentration, temperature, humidity, wind speed, delamination parameters, stress parameters, and microseismic parameters; the textual information data includes inspection records, work logs, manual reports, and dispatch instruction summaries.

[0018] Furthermore, after the contingency plan is implemented, the second threshold is dynamically adjusted based on the feedback results to achieve adaptive optimization of the contingency plan matching strategy; wherein, the dynamic adjustment includes updating the matching threshold online based on the manual confirmation results or post-evaluation results after the contingency plan is implemented, and limiting the updated threshold within a preset range.

[0019] Furthermore, the preprocessing of multi-source monitoring data includes: aligning sensor data according to timestamps, and performing missing value compensation, outlier handling, and smoothing within a unified time window.

[0020] Furthermore, the numerical features include: statistical features, trend features, threshold crossing features, and multi-sensor linkage features; the construction of the text features includes: word segmentation, stop word removal, synonym normalization, key entity extraction and weight calculation, and generating vector representation.

[0021] Furthermore, the numerical features and text features are normalized before fusion, and fused features are formed by splicing or weighting, wherein the fusion weight is a preset or adjustable parameter.

[0022] Furthermore, the disaster classification model is a probabilistic classification model trained based on historical samples, and the probabilistic classification model can output the occurrence probability of multiple disaster types.

[0023] Furthermore, the confidence assessment is made by comparing the probability difference between the disaster type with the highest probability and the second highest probability disaster type. When the difference is lower than the preset confidence threshold, at least one of the following measures is performed: manual review, parallel recommendation of multiple contingency plans, or activation of a general contingency plan.

[0024] Furthermore, the accident scenario information includes the predicted disaster type, key abnormal features, accident location, and abnormal duration, and the accident scenario information is converted into a query representation for contingency plan matching.

[0025] Furthermore, the rule-based filtering is performed based on at least one of the applicable disaster type, applicable area, and triggering conditions of the contingency plan to form a set of candidate contingency plans.

[0026] This application has the following advantages:

[0027] First, by fusing sensor numerical features with text semantic features, a comprehensive perception of underground coal mine accident scenarios is achieved, overcoming the limitation of easy misjudgment from a single data source and significantly improving the accuracy and robustness of disaster identification.

[0028] Second, a dual-threshold hierarchical decision-making mechanism was constructed: the first threshold (classification confidence threshold) assesses the credibility of disaster identification results, triggering timely manual intervention when confidence is low to avoid erroneous decisions; the second threshold (matching similarity threshold) controls the quality of contingency plan matching, ensuring that the pushed contingency plans are highly consistent with the current situation. The dual-threshold series control forms a robust safety defense, effectively balancing the risks of false alarms and missed alarms.

[0029] Third, it has dynamic adaptive optimization capabilities, and can update the matching threshold online based on the feedback results after the execution of the plan, so that the system can adapt to changes in different mine environments and mining conditions and continuously improve the matching accuracy.

[0030] In summary, this invention achieves full-process automation from accident perception and intelligent decision-making to contingency plan matching, significantly improving the timeliness and scientific nature of coal mine emergency response, and has significant safety benefits and application value. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This application provides a technical roadmap for an automatic matching method for emergency response plans in coal mine accidents. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Example 1:

[0035] like Figure 1 As shown, this embodiment aims to provide an automatic matching method for coal mine accident emergency plans. It achieves real-time determination of disaster type and provides posterior probability based on multi-source sensors and text input; automatically constructs an accident scenario summary based on the disaster type and current context; and retrieves the most matching plan from the plan database. This is achieved through dual threshold control (classification confidence threshold). + Matching similarity threshold This reduces false triggering and missed matching; through feedback-driven dynamic threshold optimization, the system can continuously adapt to different mining environments. The method includes the following steps:

[0036] S1. Multi-source data acquisition;

[0037] It collects data from various sensors, including gas, CO, temperature, humidity, wind speed, delamination, stress, and micro-vibration; and also collects text data such as inspection records, work logs, and scheduling information.

[0038] S2: Data preprocessing and feature construction;

[0039] S2.1 Sensor data preprocessing;

[0040] Align each sensor sequence according to its timestamp to form a unified time window. The missing and outlier conditions are handled (e.g., interpolation, amplitude limiting, outlier removal), and smoothing / filtering is performed.

[0041] Extract features within each time window, including at least (one or more of the following combinations):

[0042] Mean, variance, maximum / minimum, slope, abrupt change, quantiles;

[0043] Threshold outage duration and rate of increase;

[0044] Microseismic counting / energy statistics;

[0045] Multi-sensor linkage features (such as CO rising synchronously with temperature).

[0046] Obtain numerical eigenvectors .

[0047] S2.2 Text Data Preprocessing and TF-IDF Vectorization;

[0048] The text is segmented, stop words are removed, synonyms are grouped, and key entities (such as location, equipment, and job type) are extracted. The corpus is used for... Each document in With terms definition:

[0049] (1)

[0050] in, For terms, For document, For terms In the document Number of times it appears in For document The total number of occurrences of all terms within the text.

[0051] (2)

[0052] in, For corpus; The number of documents in the corpus; For included terms The number of documents.

[0053] (3)

[0054] in, For terms In the document The TF-IDF weights on the vector are used to obtain the text vector. .

[0055] S2.3 Multi-source feature fusion;

[0056] Normalize S and T and then concatenate or weighted merge (when there is no text). ):

[0057] (4)

[0058] in, For min-max normalization, For weight fusion.

[0059] Step S3: Disaster Classification;

[0060] Let the disaster category set be defined. The fusion feature is The classification output is as follows:

[0061] (5)

[0062] in, It is a priori probability, which can be corrected by historical sample statistics or expert experience; This represents the conditional probability (likelihood).

[0063] S3.1 Discretization and Laplace Smoothing;

[0064] If some sensor features are discretized into level / binary values ,but:

[0065] (6)

[0066] in, For category Lower features Value The count, For category The number of samples, This represents the number of possible values ​​for this feature.

[0067] S3.2 Continuous feature Gaussian likelihood;

[0068] Gaussian Naive Bayes can be used for continuous sensor features:

[0069] (7)

[0070] in, , Estimated from historical samples.

[0071] S4: Classification confidence assessment;

[0072] make and These are the first and second largest categories in terms of posterior probability, defined as follows:

[0073] (8)

[0074] Among them, when If the classification is deemed insufficiently reliable, the following strategy will be triggered:

[0075] Enter manual review;

[0076] Output Top-N disaster types in parallel recommendation;

[0077] The general contingency plan was activated and the on-duty expert was notified to intervene.

[0078] Step S5: Constructing the context summary (defining the retrieval query vector q);

[0079] To ensure complete input for matching contingency plans, this invention constructs an accident scenario summary q, which includes at least:

[0080] Predicting disaster types ;

[0081] Key anomalies (such as CO increase, temperature gradient, sudden drop in wind speed, gas fluctuation, sudden increase in delamination, surge in microseismic counts, etc.);

[0082] Location / tunnel / working face number / equipment number;

[0083] Time window and duration of anomalies;

[0084] The above structured fields are concatenated into normalized short text (or the structured key-value pairs are re-textualized), and then quantized to obtain the query vector q.

[0085] Step S6: Rule engine filtering;

[0086] For each contingency plan in the contingency plan database (Including metadata such as applicable disaster type, location range, triggering conditions, response steps, linkage strategies, priority, etc.) Execution rule filtering:

[0087] Hard requirements: For example, fire emergency plans must meet at least one of the following abnormalities: CO / temperature / smoke / airflow; water hazard emergency plans must meet at least one of the following abnormalities: water inflow / water level / humidity / water pressure, etc.

[0088] Priority: Similar contingency plans are sorted by the mine's policy priority or version number;

[0089] Obtain a set of candidate solutions .

[0090] The role of the rule engine is to reduce full database traversal and improve real-time performance and accuracy.

[0091] S7: Semantic matching of contingency plans;

[0092] Candidate proposal text vectors (Offline pre-computation and caching are possible) Calculate the similarity with the query vector q, using cosine similarity:

[0093] (9)

[0094] Take the maximum similarity And obtain the optimal contingency plan:

[0095] (10)

[0096] Decision-making rules:

[0097] like Manual review / parallel recommendation / general contingency plan as a backup;

[0098] Otherwise if Push notifications ;

[0099] Otherwise: Push the general contingency plan and prompt expert intervention, while recording the low similarity samples that were not hit for optimization.

[0100] S8: Execute push and linkage control;

[0101] Will (Or general contingency plan) Key instructions are pushed to: underground personnel terminals, dispatch and command center, ventilation / power supply / drainage control room, etc.;

[0102] It can be linked to control functions such as: fan start / stop / damper / window adjustment, public evacuation announcements, lighting indicators, dust suppression spraying, and drainage systems. During execution, it continuously monitors and can be dynamically updated by triggering S2–S7 again.

[0103] S9: Dynamic threshold optimization (τ updated based on feedback);

[0104] Define feedback variables :

[0105] y=1: This contingency plan has been confirmed as valid (either through manual confirmation or post-event review and evaluation).

[0106] y=0: Invalid (mismatch / inapplicable / missed match).

[0107] Adopt an online update strategy:

[0108] (11)

[0109] Where: η>0 is the step size; γ∈(0,1) is the target failure rate / risk preference parameter (the smaller the value, the more conservative the approach, and the more likely it is to trigger and push notifications); Limit the threshold to Within this context, drift is avoided. The intuitive meaning of this update form is: when y=0 (failure), the threshold is tended to be increased to reduce future false matches; when y=1 (success), the threshold is tended to be decreased to increase the hit rate. γ,η can also be adjusted according to a system that prioritizes false positives over false negatives.

[0110] Optimization of δ: δ can be slowly updated based on the historical misjudgment rate or fixed to an empirical value.

[0111] In step S7, the design of the contingency plan library includes:

[0112] Contingency plan structure fields (structured or semi-structured):

[0113] Contingency plan number, version number, effective date, applicable disaster type, applicable area / location, triggering conditions, response steps, material and personnel dispatch, joint control strategy, precautions, and priority.

[0114] Offline pre-calculation of contingency plan text vectors: When updating the contingency plan, generate according to equations (1)–(3). It is cached; online, only q and similarity are calculated to meet real-time requirements.

[0115] Versioning and Approval: New or revised plans require expert approval and historical versions must be retained for traceability and evaluation.

[0116] Example 2:

[0117] This embodiment uses abnormal gas outbursts in coal mines (hereinafter referred to as abnormal gas surges) as a specific application scenario to elaborate on the implementation process of the method in a real production environment. Abnormal gas surges are a typical dynamic disaster in coal mines, with obvious precursor characteristics. However, due to the coupling influence of multiple factors, a single sensor is difficult to accurately determine the cause. It is necessary to integrate information from multiple sources to achieve rapid identification and accurate emergency response plan delivery.

[0118] S1: Multi-source data acquisition;

[0119] The following sensors are deployed at the mining face (such as the 12301 longwall mining face):

[0120] Gas concentration sensor: Five monitoring points are set up at locations such as the return airway, upper corner, and tunneling head, with a sampling frequency of 1Hz.

[0121] Wind speed sensor: Installed in the intake and return air passages to monitor changes in wind speed.

[0122] Temperature and humidity sensors: to help detect abnormal environmental conditions.

[0123] Coal stress sensor: Embedded in the coal wall to monitor sudden stress changes.

[0124] Microseismic monitoring system: picks up microseismic events caused by coal and rock fracturing, and records the time, energy, and frequency of the events.

[0125] Other: CO sensor (used to identify associated phenomena such as spontaneous combustion of coal).

[0126] At the same time, the following text data was collected:

[0127] Gas inspector’s Gas Inspection Record for each shift (including location, time, concentration value and remarks).

[0128] Drilling record of the tunneling face (including the location and depth of the borehole, whether there are blowouts, and descriptions of abnormal gas outbursts).

[0129] The dispatch room's "Shift Handover Log" records information such as abnormal gas conditions and equipment start-up and shutdown during the shift.

[0130] Alarm records from the safety monitoring system.

[0131] S2: Data preprocessing and feature construction;

[0132] S2.1 Sensor data preprocessing;

[0133] At the current moment (e.g., 14:00 on March 10, 2025) is used as the benchmark to select a time window. (i.e., 10 minutes of data). Timestamps were aligned for each sensor sequence (missing points were filled using linear interpolation), and outliers were removed using median filtering. The following numerical features were extracted (forming a feature vector). ):

[0134] Gas concentration: mean, variance, maximum, minimum, mean of the last 5 seconds (reflecting recent trends), slope (slope of the least squares fitted line), abrupt change (maximum first-order difference), duration of exceeding limits (cumulative seconds exceeding 0.8%).

[0135] Wind speed: mean, minimum, rate of change (synchronous with gas concentration).

[0136] Temperature: mean, rate of change.

[0137] Microseismic events: total number of events, total energy, maximum energy, and energy-weighted frequency.

[0138] Linkage characteristics: correlation coefficient between gas and wind speed (a stronger negative correlation may indicate a prominent trend), and index of synchronous rise in gas and temperature.

[0139] S2.2 Text Data Preprocessing and TF-IDF Vectorization;

[0140] The text records are segmented, stop words are removed, and synonyms are unified based on a coal mining dictionary (e.g., "blowhole" and "drill hole ejection" are considered the same concept). All text records generated within the last 30 minutes are merged into a single document. The TF-IDF weights are calculated using the following formula to obtain the text vector. (corpus) (This represents all historical text records; dimension 100 is determined by feature selection).

[0141]

[0142]

[0143] S2.3 Multi-source feature fusion;

[0144] right and Perform min-max normalization separately, and then fuse according to equation (4):

[0145]

[0146] in The normalization function is the fused feature vector. .

[0147] S3: Disaster Classification;

[0148] Let the disaster category set be defined. A Naive Bayes classifier is used, with prior probability... Obtained from historical accident statistics ( The feature conditional probability is calculated as follows:

[0149] Continuous features (such as mean gas concentration, slope, microseismic energy, etc.) are represented using Gaussian Naive Bayes, with the mean... and variance Estimated from historical samples, calculated according to equation (7) :

[0150]

[0151] Discrete features (such as the "nozzle appearance" marker extracted from the text) are Laplace smoothed and estimated according to Equation (6):

[0152]

[0153] Substitute the current fusion features Calculate the posterior probability:

[0154]

[0155] Assume that:

[0156]

[0157] S4: Classification confidence assessment;

[0158] Take the class with the highest posterior probability (Gas outburst) and the second category (Gas explosion), calculated according to formula (8):

[0159]

[0160] Set classification confidence threshold (Experience points). Because If the classification is deemed reliable, it can be directly adopted. As a type of disaster. If If so, manual review or Top-N recommendations will be triggered, but this is not necessary in this embodiment.

[0161] S5: Context Summary Construction (Defining Retrieval Query Vectors) );

[0162] Construct an accident scenario summary, including the following structured fields, and concatenate them into normalized text:

[0163] Disaster type: Gas outburst;

[0164] Anomaly characteristics: maximum gas concentration 1.8% (exceeding limit), rise rate 0.15% / min, total microseismic energy J. A nozzle record appears (text prompt);

[0165] Location: Return airway of working face 12301;

[0166] Time window: 2025-03-10 13:50~14:00;

[0167] Abnormal duration: Gas concentration >0.8% lasts for 320 seconds;

[0168] Input the text into the same TF-IDF model as S2.2 (using the same corpus and feature set) to obtain the query vector. .

[0169] S6: Rule engine filtering;

[0170] Contingency Plan Database It contains over 200 contingency plans of various types, each with metadata: applicable disaster type, applicable area, triggering conditions, priority, etc. The rule engine executes:

[0171] Hard criteria: 50 contingency plans were selected based on the applicable disaster type of "gas outburst".

[0172] Location matching: Filter the applicable areas to include "12301 working face" or "applicable to the whole mine", 12 options remain.

[0173] Priority ranking: Candidate plans are ranked according to the mine's regulations (specific plans > comprehensive plans > general plans) to obtain the set of candidate plans. (In order of priority from high to low).

[0174] S7: Semantic matching of contingency plans;

[0175] The text vector of each contingency plan in the contingency plan library Pre-computed and cached offline (using the same method) (The same TF-IDF model). For each contingency plan, calculate the cosine similarity according to formula (9):

[0176]

[0177] The obtained similarity sequence is: 0.94, 0.91, 0.88, 0.85, 0.82, ... . Maximum similarity. Corresponding contingency plan For the "12301 Working Face Gas Outburst Special Emergency Plan (2024 Edition)":

[0178]

[0179] Set the initial matching similarity threshold .because Decision push .like If so, a general gas emergency plan will be pushed out and experts will be notified to intervene.

[0180] S8: Execute push and linkage control;

[0181] The system will Key instructions are simultaneously pushed through underground broadcasts, personnel positioning terminals, and dispatch screens:

[0182] Personnel instructions: All personnel at working face 12301 shall immediately put on self-rescue devices and evacuate to the intake ventilation roadway along the disaster avoidance route.

[0183] Linkage control:

[0184] Automatically cuts off non-intrinsically safe power supplies to the working face.

[0185] Turn on the local ventilation fan to its maximum power and adjust the damper to increase the airflow at the working face.

[0186] Activate the audible and visual alarm.

[0187] A dispatch order was issued to the mine rescue team. During the operation, the monitoring system continuously collected data. If the gas concentration continued to rise or signs of an explosion appeared, S2 to S7 would be repeated, and the response plan would be dynamically updated.

[0188] S9: Dynamic Threshold Optimization (Based on Feedback Update) );

[0189] Afterwards, the mine's chief engineer organized experts to evaluate the effectiveness of the contingency plan and its implementation. The evaluation results showed that the contingency plan was accurately implemented, the measures were appropriate, and casualties and the escalation of the accident were successfully avoided. Therefore, the contingency plan was confirmed to be effective, i.e., a feedback variable. The matching similarity threshold is updated online using equation (11). :

[0190]

[0191] in (Step length) (Target failure rate 10%) , , .calculate:

[0192]

[0193] Restricted by the clip function Within, a new threshold is obtained. This adjustment slightly lowers the threshold, making it easier to trigger contingency plans in similar scenarios in the future, while retaining a certain safety margin. If this assessment is deemed invalid ( If the threshold is raised, the risk of false matches will be reduced. Classification confidence threshold We will temporarily fix the value at 0.5 based on experience, and then slowly update it after accumulating enough data on misjudgments.

[0194] Through the above steps, this embodiment realizes real-time disaster identification, high-confidence classification, context summary construction, rule and semantic dual matching, automatic linkage control, and feedback-based adaptive optimization based on multi-source sensors and text input in the case of abnormal gas surge, effectively improving the intelligence level and accuracy of coal mine gas disaster emergency response.

[0195] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0196] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An automatic matching method for emergency response plans for coal mine accidents, characterized in that, include: Multi-source monitoring data are preprocessed and feature constructed to obtain numerical features and text features. The numerical features and text features are then fused to form fused features for accident identification. Based on the fusion features, the probability of occurrence of each disaster type is calculated using a disaster classification model to determine the target disaster type; Confidence assessment is performed based on the probability of occurrence of each disaster type. When the classification confidence is lower than the first threshold, manual intervention is triggered or a general contingency plan is output. When the classification confidence meets the preset conditions, accident scenario information corresponding to the target disaster type is generated. Based on the accident scenario information, the contingency plan database is filtered according to rules to obtain a set of candidate contingency plans; The accident scenario information and candidate contingency plans are semantically matched to determine the target contingency plan with the highest matching degree. The matching result is compared with the second threshold. When the matching degree meets the preset conditions, the target plan is output; when it does not meet the conditions, a general plan is output or manual intervention is triggered.

2. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, Before preprocessing and feature construction of the multi-source monitoring data, the method further includes: acquiring multi-source monitoring data from underground coal mines, wherein the multi-source monitoring data includes sensor numerical data and textual information data; the sensor numerical data includes gas concentration, carbon monoxide concentration, temperature, humidity, wind speed, delamination parameters, stress parameters, and microseismic parameters; the textual information data includes inspection records, work logs, manual reports, and dispatch instruction summaries.

3. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, After the contingency plan is implemented, the second threshold is dynamically adjusted based on the feedback results to achieve adaptive optimization of the contingency plan matching strategy. The dynamic adjustment includes updating the matching threshold online based on the manual confirmation results or post-evaluation results after the contingency plan is implemented, and limiting the updated threshold to a preset range.

4. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The preprocessing of multi-source monitoring data includes: aligning sensor data according to timestamps, and performing missing value compensation, outlier handling, and smoothing within a unified time window.

5. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The numerical features include: statistical features, trend features, threshold crossing features, and multi-sensor linkage features; the construction of the text features includes: word segmentation, stop word removal, synonym normalization, key entity extraction and weight calculation, and generating vector representation.

6. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The numerical features and text features are normalized before fusion, and then formed into fused features by splicing or weighting. The fusion weights are preset or adjustable parameters.

7. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The disaster classification model is a probabilistic classification model trained based on historical samples, and the probabilistic classification model can output the occurrence probability of multiple disaster types.

8. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The confidence assessment is determined by comparing the probability difference between the most probable disaster type and the second most probable disaster type. When the difference is lower than a preset confidence threshold, at least one of the following measures is implemented: manual review, parallel recommendation of multiple contingency plans, or activation of a general contingency plan.

9. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The accident scenario information includes the predicted disaster type, key abnormal features, accident location and abnormal duration, and the accident scenario information is converted into a query representation for contingency plan matching.

10. The automatic matching method for coal mine accident emergency plans according to claim 1, characterized in that, The rule-based filtering is based on at least one of the applicable disaster type, applicable area, and triggering conditions of the contingency plan to form a set of candidate contingency plans.

Images