A coal mine gas risk identification method based on multi-source monitoring daily report sample reconstruction

Abstract

The application discloses a coal mine gas risk identification method based on multi-source monitoring daily report sample reconstruction, comprising: acquiring coal mine multi-source monitoring daily report data, standardizing the multi-source monitoring daily report data, merging and mapping original monitoring types in different daily reports into standard monitoring categories, and constructing date-measuring point samples indexed by report dates and measuring point numbers; constructing statistical characteristics based on the date-measuring point samples; generating weakly supervised abnormal labels based on the coal mine multi-source monitoring daily report data, training an abnormal identification model based on the statistical characteristics and the weakly supervised abnormal labels, and obtaining abnormal probabilities of each date-measuring point sample; calculating a rule risk score according to the statistical characteristics, and fusing the rule risk score with the abnormal probabilities to obtain a final risk score value; and outputting an abnormal identification result and a grading early warning result according to the final risk score value. The application improves the accuracy, interpretability and engineering application value of coal mine gas risk analysis.

Description

Technical Field

[0001] This invention belongs to the field of coal mine safety monitoring and early warning technology, and particularly relates to a method for identifying coal mine gas risk based on the reconstruction of daily monitoring data from multiple sources. It is applicable to the processing, analysis, and early warning output of daily monitoring data such as analog statistical daily reports, gas monitoring daily reports, and ventilation monitoring daily reports, where parameters such as oxygen, wind speed, pressure, and temperature are used to further characterize the gas risk status. Background Technology

[0002] Coal mine safety monitoring systems continuously generate numerous daily monitoring reports during production. These reports typically contain monitoring information on various parameters such as methane, carbon monoxide, oxygen, wind speed, pressure, dust, and temperature, as well as statistical results including maximum values, average values, alarm counts, power outage counts, fault counts, and abnormal durations. This data reflects the daily changes in mine gas and ventilation conditions and has high analytical value. However, current technologies largely rely on manual table lookups, single-item statistics, and post-outbreak analysis, lacking a systematic method for anomaly identification and tiered early warning for multi-source daily monitoring data. The main problems include: multiple data sources, inconsistent field formats, and complex monitoring types, making it difficult to directly form standardized analysis samples; existing analyses often rely on single-point maximum values, average values, or whether limits are exceeded as the basis for judgment, failing to comprehensively reflect gas anomaly risks; a lack of unified integration and quantitative evaluation methods among multi-source monitoring information such as methane, carbon monoxide, oxygen, wind speed, and pressure; and existing daily monitoring results mostly remain at the statistical display level, making it difficult to further develop anomaly identification results and tiered early warning conclusions.

[0003] Therefore, it is necessary to propose a technical method that can perform data standardization processing, anomaly identification, and risk classification and early warning for daily monitoring data from multiple sources in coal mines, so as to improve the accuracy and practicality of coal mine gas anomaly analysis. Summary of the Invention

[0004] Purpose of the invention: In view of the problems that existing daily coal mine monitoring data sources are numerous, field structures are inconsistent, monitoring types are complex, and it is difficult to directly form modelable samples, and that existing daily analysis is mostly limited to manual table lookup, single-item statistics and post-exceedance analysis, making it difficult to form anomaly identification and graded early warning results, this invention proposes a coal mine gas risk identification and graded early warning method based on multi-source monitoring daily data.

[0005] Technical solution: The present invention provides a method for identifying coal mine gas risks based on reconstruction of daily monitoring data from multiple sources, comprising the following steps:

[0006] Step 1: Obtain daily multi-source monitoring data of coal mines, standardize the daily multi-source monitoring data, merge and map the original monitoring types in different daily reports into standard monitoring categories, and construct a date-monitoring point sample indexed by report date and monitoring point number;

[0007] Step 2: Construct statistical features based on the date-measurement point samples;

[0008] Step 3: Generate weakly supervised anomaly labels based on daily multi-source monitoring data of coal mines, train anomaly identification models based on statistical features and weakly supervised anomaly labels, and obtain the anomaly probability of each date-measuring point sample;

[0009] Step 4: Calculate the rule risk score based on statistical characteristics, and then integrate the rule risk score with the anomaly probability to obtain the final risk score;

[0010] Step 5: Output the anomaly identification result and graded early warning result based on the final risk score.

[0011] Furthermore, in step 1, the daily multi-source monitoring data for the coal mine includes monitoring point number, monitoring point name, monitoring type, unit, maximum value, average value, maximum value time, upper limit alarm threshold, upper limit power outage threshold, lower limit alarm threshold, alarm count, over-limit count, power outage count, fault count, power supply anomaly count, and corresponding anomaly duration information; the standardization processing includes: uniformly mapping the field names in the original daily report; converting numerical fields into a computable format; uniformly converting alarm duration, over-limit duration, power outage duration, fault duration, and power supply anomaly duration into seconds; and generating system type and regional grouping attributes in combination with the monitoring point name.

[0012] Furthermore, in step 2, the statistical features include threshold proximity features, peak-to-average difference features, historical deviation features, statistical features of abnormal events, and features of abnormal persistence.

[0013] Furthermore, the threshold proximity feature and peak-to-average difference feature include: alarm threshold proximity. Power outage threshold proximity and peak-to-average difference characteristics The calculation methods are as follows:

[0014] ;

[0015] ;

[0016] ;

[0017] in, This is the maximum value for the day. This is the average value for the day. This is the alarm threshold. ε is the power-off threshold, and ε is a constant to prevent the denominator from being zero.

[0018] Furthermore, the historical deviation features include the previous day's lag value, first-order difference value, past 7-day rolling mean, past 7-day rolling standard deviation, past 30-day rolling mean, and deviation features based on a 7-day window, all constructed on a per-unit measurement point basis; among these, the maximum historical deviation feature... The calculation method is as follows:

[0019] ;

[0020] in, and These are the rolling mean and rolling standard deviation of the same measuring point over the past 7 days, respectively. ε is the maximum value for the day, and ε is a constant to prevent the denominator from being zero.

[0021] Furthermore, in step 3, the weak supervision anomaly label is generated in the following way: when the date-measuring point sample meets any of the following conditions: alarm count greater than zero, over-limit count greater than zero, power outage count greater than zero, fault count greater than zero, power supply anomaly count greater than zero, anomaly duration greater than zero, maximum value reaching or exceeding alarm threshold, maximum value reaching or exceeding power outage threshold, or existence of a lower limit alarm gap, the sample is marked as an anomaly candidate sample.

[0022] Furthermore, in step 3, the anomaly identification model is trained by dividing the training set and test set in chronological order, and outputs the anomaly probability of each date-measurement point sample based on the threshold proximity feature, peak-to-average difference feature, historical deviation feature, anomaly event statistical feature, anomaly persistence feature, system type, and regional grouping attribute. .

[0023] Furthermore, in step 4, the rule-based risk score is determined by the proximity to the alarm threshold. Power outage threshold proximity Degree of deviation from historical maximum value Historical deviation of average value Peak-to-mean-difference characteristics Lower limit risk factor and abnormal event factors The weighted summation yields the expression:

[0024] ;

[0025] in, This represents the abnormal event factor, which consists of the number of alarms, the number of power outages, the number of faults, and the duration of abnormalities. ~ These are the corresponding weighting coefficients.

[0026] Furthermore, in step 5, the final risk score is obtained by a weighted fusion of the rule-based risk score and the anomaly probability, and its expression is:

[0027] ;

[0028] in and To merge weights, and + =1, Assess the risk of rules. The probability of anomalies for each date-measurement point sample;

[0029] The graded early warning results include at least four levels: normal, attention, warning, and severe warning. The warning level is determined based on the correspondence between the final risk score and the preset graded threshold. When a power outage event occurs in the date-measuring point sample, it is directly classified as a severe warning.

[0030] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0032] This invention makes full use of the daily summary data accumulated by coal mining enterprises over a long period of time, such as analog quantity statistical daily reports, gas monitoring daily reports, and ventilation monitoring daily reports. Without relying on the original monitoring time series data at the second or minute level, it realizes the standardized processing of multi-source daily reports, extraction of abnormal features, risk quantification and evaluation, abnormal identification modeling, and output of early warning levels, thereby improving the accuracy, interpretability, and engineering application value of coal mine gas risk analysis.

[0033] This invention addresses daily monitoring data from multiple sources in coal mines. It enables unified extraction, standardized processing, and structured organization of various monitoring information from analog statistical daily reports, thereby improving the utilization efficiency of daily data. It not only considers the maximum or average value of a single measuring point but also comprehensively incorporates features such as threshold proximity, peak-to-average difference, historical deviation, frequency of abnormal events, and duration of abnormalities, thus providing a more comprehensive reflection of coal mine gas anomaly risks.

[0034] This invention can integrate and analyze multi-source monitoring information such as methane, carbon monoxide, oxygen, wind speed, pressure, dust, and temperature to form a unified risk assessment result, enhancing the accuracy and comprehensiveness of anomaly identification. By utilizing existing statistical information from daily reports, such as alarms, power outages, faults, and abnormal power supply, weak tags are constructed, allowing for the establishment of anomaly identification models without extensive manual annotation, thus demonstrating strong feasibility.

[0035] This invention integrates rule-based risk scoring with the output of an anomaly detection model, simultaneously outputting risk scores, anomaly detection results, and warning levels. This facilitates high-risk point screening and key area management for coal mining enterprises. This invention is applicable to the daily report data scenarios accumulated by coal mining enterprises over a long period, without relying on second- or minute-level raw time-series data, and has good engineering adaptability and application value. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall process for identifying and warning of abnormal coal mine gas emissions according to the present invention.

[0037] Figure 2 A flowchart for standardized processing of daily monitoring data from multiple sources;

[0038] Figure 3 Flowchart for feature construction and rule scoring;

[0039] Figure 4 This is a diagram showing the results of the case risk analysis. Detailed Implementation

[0040] The present invention will be further illustrated below with reference to actual daily monitoring data from coal mines. This embodiment is only used to illustrate the technical solution of the present invention and is not intended to limit the scope of protection of the present invention.

[0041] like Figure 1 As shown, the present invention provides a method for identifying coal mine gas risks based on the reconstruction of daily monitoring data from multiple sources, comprising the following steps:

[0042] Step 1: Obtain daily multi-source monitoring data of coal mines, standardize the daily multi-source monitoring data, merge and map the original monitoring types in different daily reports into standard monitoring categories, and construct a date-monitoring point sample indexed by report date and monitoring point number;

[0043] Step 2: Construct statistical features based on the date-measurement point samples;

[0044] Step 3: Generate weakly supervised anomaly labels based on daily multi-source monitoring data of coal mines, train anomaly identification models based on statistical features and weakly supervised anomaly labels, and obtain the anomaly probability of each date-measuring point sample;

[0045] Step 4: Calculate the rule risk score based on statistical characteristics, and then integrate the rule risk score with the anomaly probability to obtain the final risk score;

[0046] Step 5: Output the anomaly identification result and graded early warning result based on the final risk score.

[0047] Among them, such as Figure 2As shown, step 1, data standardization, is used to uniformly convert heterogeneous fields, different measurement formats, and abnormal event records in the multi-source monitoring daily reports into daily report-level date-measuring point samples; for example... Figure 3 As shown, steps 2 to 4, involving feature construction, rule scoring, and risk fusion, are used to generate interpretable risk scores; Figure 4 As shown, the output of step 5 can be further used to generate daily risk trends, ranking of key monitoring points, distribution of category warnings, and risk evolution results of representative monitoring points.

[0048] (1) Data sources and sample construction

[0049] This embodiment selects the daily continuous monitoring reports of a coal mining enterprise from July 1, 2021 to January 28, 2022 as the original data source, covering a total of 211 statistical days, forming 37,501 valid samples involving 215 monitoring points. The samples are constructed using "report date + monitoring point number" as the basic unit, and each sample corresponds to the summary monitoring results of one monitoring point within one statistical day.

[0050] The main fields extracted from the original daily report include: measurement point number, measurement point name, monitoring type, unit of measurement, maximum value, average value, maximum value time, upper limit alarm threshold, upper limit power outage threshold, lower limit alarm threshold, number of alarms, alarm duration, number of times exceeding the limit, duration of exceeding the limit, number of times power outage, duration of power outage, number of faults, duration of faults, number of times power supply is abnormal, duration of power supply is abnormal, and remarks.

[0051] According to statistics, 3,121 weakly supervised anomalies were found in the entire sample, accounting for 8.3224% of the total sample. The main monitoring categories included in the modeling included methane, carbon monoxide, oxygen, wind speed, dust, temperature, carbon dioxide, and pressure.

[0052] (2) Data preprocessing and standardization

[0053] like Figure 2 As shown, the standardization processing of multi-source monitoring daily data takes original reports such as analog quantity statistical daily reports, gas monitoring daily reports, and ventilation monitoring daily reports as input, and sequentially completes report reading, field mapping, numerical conversion, duration conversion, invalid record removal, and monitoring category merging. Through this processing, information such as the maximum value, average value, alarm threshold, power outage threshold, number of anomalies, and duration of anomalies at the same measuring point on different dates is organized into a date-measuring point sample with a unified structure, providing basic data for subsequent feature construction and model training.

[0054] The original daily report files are uniformly cleaned and standardized, including: reading the mine name, report date, and main content of the report; standardizing the mapping of the original header; converting numerical fields into a computable format; uniformly converting fields such as alarm duration, power outage duration, fault duration, and power supply anomaly duration into seconds; and deleting invalid records with missing report dates, missing measurement point numbers, and abnormal key fields.

[0055] Furthermore, the original monitoring types are merged and mapped, converting similar monitoring items from different daily reports into standard monitoring categories. Based on the monitoring point names and location descriptions, system type and regional grouping attributes are automatically generated. The system type is used to characterize gas monitoring, environmental monitoring, or ventilation monitoring, while the regional grouping is used to characterize mining faces, return air systems, ventilation systems, extraction systems, safety facilities, transportation and transshipment areas, or other areas.

[0056] (3) Feature construction and anomaly label generation

[0057] like Figure 3 As shown, the feature construction and rule scoring process takes standardized samples as input. First, it extracts basic features such as peak-to-average difference, threshold proximity, lower limit alarm gap, total number of events, and duration of abnormality. Then, it calculates the lag value, difference value, rolling mean, rolling standard deviation, and historical deviation according to the time series of the measurement points. Finally, it weights and fuses alarm threshold proximity, power outage threshold proximity, historical deviation, peak-to-average difference, lower limit risk, and abnormal event factors to obtain the rule risk score.

[0058] Daily features are constructed based on four aspects: threshold proximity, historical deviation, intensity of abnormal events, and persistence of abnormalities.

[0059] Based on the maximum value, average value, alarm threshold, and power outage threshold, features such as peak-to-average difference, alarm threshold proximity, power outage threshold proximity, and lower limit alarm gap are constructed. Based on alarm, over-limit, power outage, fault, and power supply anomaly information, features such as the total number of abnormal events per day, the total duration of abnormal events per day, the proportion of abnormal duration, and event density are constructed. Taking a single measuring point as a unit, the previous day's lag value, first-order difference value, the rolling average of the past 7 days, the rolling standard deviation of the past 7 days, the rolling average of the past 30 days, and deviation features based on a 7-day window are further constructed.

[0060] A sample is marked as an anomalous candidate sample when it meets any of the following conditions: an alarm or over-limit event exists, there is an abnormal duration, the maximum value reaches or exceeds the alarm threshold, the maximum value reaches or exceeds the power outage threshold, or there is a lower limit alarm gap. This forms a weakly supervised anomaly label for subsequent model training.

[0061] (4) Risk scoring, model training and early warning classification

[0062] First, a rule risk score is calculated based on factors such as the proximity of alarm thresholds and power outage thresholds, the degree of historical deviation, the number of abnormal events, and the duration of abnormality. This score is then adjusted by combining the weights of monitoring categories to obtain the rule risk value.

[0063] Subsequently, modeling was performed by dividing the training and test sets chronologically, with the first 168 days serving as the training set and the last 43 days as the test set, corresponding to 30,006 training samples and 7,495 test samples, respectively. The model input consisted of two parts: numerical features and categorical features. The numerical features were imputed using the median, and the categorical features were imputed using the mode and then encoded.

[0064] This embodiment uses a random forest classification model for training, with 400 decision trees, a minimum number of samples per leaf node, and class balancing to address the issue of a low proportion of outlier samples. After the model is trained, it outputs the outlier probability for each sample.

[0065] The rule-based risk score and the model's anomaly probability are weighted and fused to obtain the final risk score. An optimal judgment threshold is automatically searched based on the training set, and a final anomaly judgment threshold of 0.38 is obtained. If the final risk score is greater than or equal to 0.38, or if the sample exhibits power outage-related anomalies, the sample is determined to be an anomaly.

[0066] Based on this, the final risk score is divided into four warning levels: Level 0 Normal, Level 1 Attention, Level 2 Warning, and Level 3 Severe Warning. When a power outage occurs in the sample or the final risk score reaches the severe risk range, it is directly classified as Level 3 Severe Warning.

[0067] (5) Implementation results

[0068] like Figure 4 As shown, the case risk analysis results include four types of outputs: daily average risk score and changes in the number of abnormal monitoring points, ranking of persistently high-risk monitoring points, distribution of warning levels for different monitoring categories, and risk trends of representative high-risk monitoring points. These results can interpret the identification results from four perspectives: the overall mine, the monitoring point objects, the monitoring categories, and the temporal evolution, thus assisting in on-site verification and tiered handling.

[0069] The evaluation results of this embodiment on the test set are as follows: the ROC-AUC of the fused risk score is 0.9817, the PR-AUC is 0.9032, the Precision is 0.9928, the Recall is 0.7740, and the F1 score is 0.8698. In the corresponding confusion matrix, there are 6961 true negative samples, 3 false positive samples, 120 false negative samples, and 411 true positive samples. The results show that this invention has high recognition accuracy and a low false positive rate in daily report-level anomaly identification scenarios.

[0070] Based on the results of feature importance, the rule risk score, alarm threshold proximity feature, peak-to-average difference variation feature, power outage threshold proximity feature, and historical deviation feature contribute significantly to anomaly identification, indicating that the rule features and historical features constructed in this invention have good effectiveness and interpretability.

[0071] In the analysis of persistent high-risk monitoring points, representative monitoring points with high average risk values ​​include: 008A01 (15305 low-level extraction roadway preparation team T1), 013A01 (15314 return airway excavation team 2 T1), 018A16 (15314 intake airway excavation team 1 T2), 013A01 (15314 cutting roadway excavation team 2 T1), and 150A10 (15# refuge chamber east transition door O2), etc., which shows that the present invention can screen out key objects with persistent high risk from the full daily report.

[0072] The warning level distribution of the full sample is as follows: Level 0 normal samples: 34,705 samples; Level 1 attention samples: 458 samples; Level 2 warning samples: 232 samples; Level 3 severe warning samples: 2,106 samples. As shown in the attached figures, this invention can not only output the anomaly identification results of individual monitoring points, but also generate daily risk change trends at the mine level, risk distribution characteristics at the monitoring category level, and risk evolution trajectories at the key monitoring point level.

[0073] (6) Project Implementation Method

[0074] This invention can be deployed in the existing daily safety monitoring management system of coal mining enterprises, or it can run as an independent batch processing program on the dispatch center server or office terminal. After the daily monitoring report is generated, the system automatically reads the analog quantity statistical daily report, gas monitoring daily report, and ventilation monitoring daily report, and extracts fields such as report date, measuring point number, measuring point name, monitoring type, unit of measurement, maximum value, average value, maximum value time, alarm threshold, power outage threshold, lower limit threshold, number of alarms, number of over-limits, number of power outages, number of faults, number of abnormal power supply, and corresponding duration. It also completes unified field mapping, missing record removal, numerical format conversion, and duration unit conversion to form a standardized sample table indexed by "report date-measuring point number".

[0075] Based on a standardized sample table, the system calculates features such as threshold proximity, peak-to-average difference, historical deviation, total number of abnormal events, and total duration of abnormalities according to the measurement point and date order. It then generates weakly supervised labels by combining alarm, over-limit, power outage, fault, and power supply anomaly information. Subsequently, it calculates rule-based risk scores for the samples and fuses them with the anomaly probabilities output by the random forest model to obtain the final risk score. Samples with power outage-related anomalies are directly classified as severe warnings; other samples are categorized into four levels—normal, watch out, warning, and severe warning—according to a preset score range.

[0076] The system outputs include a monitoring point-level risk result table, a daily risk summary table, and a list of key high-risk monitoring points. The monitoring point-level risk result table must include at least the report date, monitoring point number, monitoring category, rule-based risk score, model probability, final risk score, anomaly determination result, and warning level. The daily risk summary table is used to count the number of warnings at each level and the main anomaly types. The list of key high-risk monitoring points assists managers in reviewing and handling these situations. When adding new monitoring points or new daily report types, only the corresponding field mapping relationships and monitoring point attribute information need to be added; the identification and warning output can then be completed following the same process.

[0077] The above embodiments are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and equivalent substitutions without departing from the principle of the present invention. All such improvements and equivalent substitutions to the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A method for identifying coal mine gas risk based on reconstruction of daily monitoring reports from multiple sources, characterized in that, Includes the following steps: Step 1: Obtain daily multi-source monitoring data of coal mines, standardize the daily multi-source monitoring data, merge and map the original monitoring types in different daily reports into standard monitoring categories, and construct a date-monitoring point sample indexed by report date and monitoring point number; Step 2: Construct statistical features based on the date-measurement point samples; Step 3: Generate weakly supervised anomaly labels based on daily multi-source monitoring data of coal mines, train anomaly identification models based on statistical features and weakly supervised anomaly labels, and obtain the anomaly probability of each date-measuring point sample; Step 4: Calculate the rule risk score based on statistical characteristics, and then integrate the rule risk score with the anomaly probability to obtain the final risk score; Step 5: Output the anomaly identification result and graded early warning result based on the final risk score.

2. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction as described in claim 1, characterized in that, In step 1, the daily multi-source monitoring data of the coal mine includes monitoring point number, monitoring point name, monitoring type, unit, maximum value, average value, maximum value time, upper limit alarm threshold, upper limit power outage threshold, lower limit alarm threshold, number of alarms, number of over-limits, number of power outages, number of faults, number of power supply anomalies, and corresponding anomaly duration information. The standardization process includes: uniformly mapping the field names in the original daily report; converting the numerical fields into a computable format; uniformly converting the alarm duration, over-limit duration, power outage duration, fault duration, and power supply anomaly duration into seconds; and generating system type and regional grouping attributes in combination with the monitoring point name.

3. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 1, characterized in that, In step 2, the statistical features include threshold proximity features, peak-to-average difference features, historical deviation features, statistical features of abnormal events, and features of abnormal persistence.

4. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 3, characterized in that, The threshold proximity feature and peak-to-average difference feature include: alarm threshold proximity. Power outage threshold proximity and peak-to-average difference characteristics The calculation methods are as follows: ； ； ； in, This is the maximum value for the day. This is the average value for the day. This is the alarm threshold. ε is the power-off threshold, and ε is a constant to prevent the denominator from being zero.

5. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 3, characterized in that, The historical deviation features include the previous day's lag value, first-order difference value, past 7-day rolling mean, past 7-day rolling standard deviation, past 30-day rolling mean, and deviation features based on a 7-day window, all constructed on a per-point basis; among these, the maximum historical deviation feature... The calculation method is as follows: ； in, and These are the rolling mean and rolling standard deviation of the same measuring point over the past 7 days, respectively. ε is the maximum value for the day, and ε is a constant to prevent the denominator from being zero.

6. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 1, characterized in that, In step 3, the weak supervision anomaly label is generated in the following way: when the date-measuring point sample meets any of the following conditions, the number of alarms is greater than zero, the number of over-limits is greater than zero, the number of power outages is greater than zero, the number of faults is greater than zero, the number of power supply anomalies is greater than zero, the duration of anomalies is greater than zero, the maximum value reaches or exceeds the alarm threshold, the maximum value reaches or exceeds the power outage threshold, or there is a lower limit alarm gap, the sample is marked as an anomaly candidate sample.

7. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 1, characterized in that, In step 3, the anomaly identification model is trained by dividing the training set and test set in chronological order, and outputs the anomaly probability of each date-measurement point sample based on the threshold proximity feature, peak-to-average difference feature, historical deviation feature, anomaly event statistical feature, anomaly persistence feature, system type, and regional grouping attribute. .

8. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 1, characterized in that, In step 4, the rule-based risk score is determined by the proximity to the alarm threshold. Power outage threshold proximity Degree of deviation from historical maximum value Historical deviation of average value Peak-to-mean-difference characteristics Lower limit risk factor and abnormal event factors The weighted summation yields the expression: ； in, This represents the abnormal event factor, which consists of the number of alarms, the number of power outages, the number of faults, and the duration of abnormalities. ~ These are the corresponding weighting coefficients.

9. The method for identifying coal mine gas risk based on multi-source monitoring daily report sample reconstruction according to claim 1, characterized in that, In step 5, the final risk score is obtained by weighted fusion of rule-based risk scoring and anomaly probability, and its expression is: ； in and To merge weights, and + =1, Assess the risk of rules. The probability of anomalies for each date-measurement point sample; The tiered early warning results include at least four levels: normal, attention, warning, and severe warning; the warning level is determined based on the correspondence between the final risk score and the preset tiered threshold. When a power outage event occurs in the date-measurement point sample, it is directly classified as a severe warning.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images