A method and system for dynamically constructing a three-dimensional gas flow field at a working face

By arranging monitoring points in a grid pattern in the three-dimensional space of the coal mine working face, and using multiple quadratic radial basis functions and extreme random tree regression models, the gas flow field is dynamically constructed, solving the problems of insufficient sensor coverage and static interpolation, and realizing high-precision real-time monitoring.

CN123116560APending Publication Date: 2026-11-17CHINA COAL TECH & ENG GRP CHONGQING RES INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611198345.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing coal mine safety monitoring systems, fixed-position sensors cannot cover most areas inside the working face, resulting in prominent monitoring blind spots. Furthermore, existing interpolation methods fail to dynamically reflect the characteristics of gas flow field changes with external conditions, leading to insufficient model generalization performance.

Method used

Monitoring points are arranged in a cross-sectional grid pattern in the three-dimensional space of the working face to collect data on gas concentration, periodic pressure and atmospheric pressure difference. A baseline flow field is generated by multiple quadratic radial basis function interpolation. The optimal feature point set is selected by an extreme random tree regression model to construct a two-level model for dynamic prediction. The model is then updated in real time in conjunction with data from the existing monitoring system.

Benefits of technology

It achieves low-cost, high-precision real-time dynamic construction of three-dimensional gas flow field with a monitoring accuracy of no more than 20%, covering the concentration distribution in the entire space and supporting real-time dynamic updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116560A_ABST
    Figure CN123116560A_ABST
Patent Text Reader

Abstract

This application proposes a method and system for dynamically constructing a three-dimensional gas flow field at a working face. The method includes: collecting measured gas concentration values ​​at monitoring points on the working face, simultaneously collecting periodic pressure data, and preprocessing it to form a time-series sample set; using the measured values ​​of all monitoring points as input, generating a continuous three-dimensional gas concentration distribution field as a reference flow field using multiple quadratic radial basis function interpolation; determining the optimal feature point set using an extreme random tree regression model; constructing and cascading two-level extreme random tree regression models, with the first-level model predicting the feature point concentration using the gas concentration at the upper corner and three types of external variables as input, and the second-level model predicting the concentration distribution across the entire space using the feature point concentration; reading real-time monitoring values ​​and inputting them into the cascaded models, and generating a three-dimensional gas concentration distribution field through interpolation. The technical solution proposed in this application achieves low-cost, high-precision real-time dynamic construction of a three-dimensional gas flow field.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of coal mine safety monitoring technology, and in particular to a method and system for dynamically constructing a three-dimensional gas flow field at a working face. Background Technology

[0002] In underground coal mining, the working face is the area with the highest concentration of gas outbursts. The gas concentration at the working face exhibits a significant non-uniform distribution in three-dimensional space: along the strike of the working face, the gas concentration gradually increases from the intake air side to the return air side, with high concentration areas concentrated in the upper corner and the goaf side; vertically, methane is less dense than air, causing gas to rise and accumulate near the roof, forming obvious concentration stratification; along the dip direction of the working face, the gas concentration on the goaf side is significantly higher than that on the coal face side. Obtaining the gas concentration distribution throughout the entire working face is a fundamental prerequisite for early warning of gas exceedances and for the control of the ventilation system.

[0003] Currently, coal mine safety monitoring systems generally use fixed-location sensors for methane concentration monitoring. These sensors are typically deployed in limited locations such as upper corners and return airways, and the monitoring data only reflects the methane concentration at the sensor's installation point, failing to cover most of the working face. With the increasing number of ultra-long working faces, the problem of monitoring blind spots is becoming more prominent. When methane accumulates in the monitoring blind spot and reaches the explosion limit, fixed-location sensors cannot detect it in time, making it difficult to provide comprehensive spatial situational information for methane disaster prevention and control.

[0004] In the spatial reconstruction of gas concentration, existing studies have used spatial interpolation methods to convert discrete monitoring point data into a continuous distribution field. However, most existing interpolation methods are static estimates, using only the monitoring data at the current moment for spatial interpolation. They fail to incorporate key factors affecting gas emission, such as periodic roof pressure, atmospheric pressure changes, and negative pressure from goaf extraction, into the prediction process. The interpolation results cannot reflect the dynamic evolution of the gas flow field under changing external conditions. Furthermore, existing methods often rely on statistical correlation to select monitoring points for driving full-space concentration prediction. The selected points are often concentrated in a local area, lacking spatial representativeness and failing to comprehensively reflect the concentration gradient changes across the entire flow field. This results in insufficient generalization performance of the model under different working face conditions. Therefore, there is an urgent need to provide a dynamic method for constructing a three-dimensional gas flow field at the working face to address the problems of insufficient monitoring spatial coverage, inability to dynamically update interpolation results, and lack of spatial representativeness in the selection of feature points in existing technologies. Summary of the Invention

[0005] This application provides a method and system for dynamically constructing a three-dimensional gas flow field at a working face, in order to at least solve the above-mentioned technical problems.

[0006] The first aspect of this application proposes a method for dynamically constructing a three-dimensional gas flow field at a working face, the method comprising: Monitoring points are arranged in a cross-sectional grid pattern in the three-dimensional space of the working face. The measured values ​​of gas concentration at each point are collected, and periodic pressure data, atmospheric pressure difference data, and negative pressure data are collected simultaneously. After preprocessing, a time series sample set is formed. Using the measured values ​​of gas concentration at all monitoring points as input, spatial interpolation is performed using the multiple quadratic radial basis function interpolation method to generate a three-dimensional continuous gas concentration distribution field, which serves as the reference flow field. k monitoring points are selected from each monitoring point to form a candidate combination. The concentration of the k monitoring points in the candidate combination is used as the input, and the concentration of the remaining monitoring points in the reference flow field is used as the output. An extreme random tree regression model is used to calculate the prediction error. All candidate combinations and different k values ​​are traversed. The k value with the smallest prediction error and the error tending to level off as the k value increases and the corresponding optimal combination are selected as the optimal feature point set. A two-level extreme random tree regression model was constructed. The first-level model took the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as inputs, and used the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model took the concentration values ​​of each feature point output by the first-level model as inputs, and used the gas concentration at the other monitoring points in the baseline flow field as the prediction target. The two-level models were trained using a time-series sample set, and the two-level models were concatenated after training. The current monitoring values ​​of gas concentration in the upper corner, working resistance of hydraulic support, atmospheric pressure difference, and extraction negative pressure are read and input into the first-level model. After processing by the two-level model, the concentration prediction values ​​of the remaining monitoring points are obtained. These values ​​are then merged with the concentration values ​​of the feature points to form a complete concentration prediction result. The complete concentration prediction result is then used as an interpolation node, and spatial interpolation is performed using the multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

[0007] Preferably, the monitoring points are arranged in the following ways: n monitoring sections are evenly arranged along the direction of the working face. Each section has 3 layers along the vertical direction and 3 columns along the dip direction, forming a 9-point grid, for a total of 9n monitoring points on the entire working face.

[0008] Preferably, the measured values ​​of gas concentration collected at each point include: The gas concentration at each monitoring point was measured three times consecutively, and the arithmetic mean of the three gas concentration measurements was taken as the measured gas concentration value at the monitoring point.

[0009] Preferably, the preprocessing includes: If the gas concentration values ​​of any 9 or more monitoring points are exactly the same in a single measurement, the current measurement data is judged to be abnormal and the entire group is removed. For the data gaps caused by the removal, linear extrapolation or spline interpolation methods are used to fill them.

[0010] Preferably, the step of using the measured gas concentration values ​​of all monitoring points as input, and employing a multiple quadratic radial basis function interpolation method to perform spatial interpolation to generate a three-dimensional continuous gas concentration distribution field as a reference flow field includes: The measured values ​​of gas concentration at each monitoring point are used as input. The interpolation calculation is performed in the three-dimensional spatial coordinate system of the working face using the multiple quadratic radial basis function interpolation method. The concentration value at any interpolation point is obtained by weighted summation of the Euclidean distance from the interpolation point to each monitoring point after mapping with the multiple quadratic radial basis function and the corresponding weight coefficient of each monitoring point. After the interpolation calculation is completed, a three-dimensional continuous distribution field of gas concentration covering the entire working face space is generated, and the three-dimensional continuous distribution field of gas concentration is used as the reference flow field. The weighting coefficients are obtained by solving the radial basis function value matrix between monitoring points and the measured concentration value vector of the monitoring points.

[0011] Preferably, the step of training a two-level model using a time-series sample set includes: The time series sample set is divided into a training set and a test set according to a preset ratio; The training parameters for the first-level and second-level models are set, with 200 regression trees for each and 3 samples for the minimum leaf node. The mean squared error is used as the evaluation metric for node splitting. The first-level model and the second-level model are trained separately using the training set. Training is stopped when the mean squared error on the training set no longer decreases after 10 consecutive iterations, thus obtaining the first-level model and the second-level model that have been initially trained. The prediction accuracy of the first-level model and the second-level model after initial training is verified using a test set. If the mean absolute error and mean relative percentage error on the test set both meet the preset threshold, the current model is taken as the final trained model. If they do not meet the threshold, the model parameters are adjusted and the model is retrained until the preset threshold is met, thus obtaining the trained first-level model and the second-level model.

[0012] Furthermore, the concatenation of the two-level models includes: Connect the output of the first-level model to the input of the second-level model, so that the predicted concentration values ​​of each feature point output by the first-level model can be directly used as the input variables of the second-level model.

[0013] A second aspect of this application proposes a dynamic construction system for a three-dimensional gas flow field at a working face, comprising: The data acquisition module is used to arrange monitoring points in a cross-sectional grid pattern in the three-dimensional space of the working face, collect the measured values ​​of gas concentration at each point, and simultaneously collect periodic pressure data, atmospheric pressure difference data, and negative pressure data. After preprocessing, a time series sample set is formed. The interpolation module is used to take the measured values ​​of gas concentration at all monitoring points as input, and use the multiple quadratic radial basis function interpolation method to perform spatial interpolation to generate a three-dimensional continuous gas concentration distribution field as a reference flow field. The selection module is used to select k monitoring points from each monitoring point to form a candidate combination. The concentration of the k monitoring points in the candidate combination is used as input, and the concentration of the remaining monitoring points in the reference flow field is used as output. The prediction error is calculated using an extreme random tree regression model. The module iterates through all candidate combinations and different k values, and selects the k value with the smallest prediction error and the error tends to level off as the k value increases, and the corresponding optimal combination, as the optimal feature point set. The module is used to construct a two-level extreme random tree regression model. The first-level model takes the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as input, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as input, and uses the gas concentration at the other monitoring points in the baseline flow field as the prediction target. The two-level models are trained using a time-series sample set, and the two-level models are concatenated after training. The generation module is used to read the current monitoring values ​​of gas concentration in the upper corner, working resistance of hydraulic support, atmospheric pressure difference, and extraction negative pressure, input them into the first-level model, and obtain the concentration prediction values ​​of the remaining monitoring points after processing by the two-level model. These values ​​are then merged with the concentration values ​​of the feature points to form a complete concentration prediction result. The complete concentration prediction result is then used as an interpolation node, and spatial interpolation is performed using a multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

[0014] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in the first aspect embodiment.

[0015] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.

[0016] The technical solutions provided by the embodiments of this application bring at least the following beneficial effects: This application proposes a method and system for dynamically constructing a three-dimensional gas flow field at a working face. The method includes: arranging monitoring points in a cross-sectional grid pattern in the three-dimensional space of the working face, collecting measured gas concentration values ​​at each point, and simultaneously collecting periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data; preprocessing these data to form a time-series sample set; using the measured gas concentration values ​​of all monitoring points as input, performing spatial interpolation using a multiple quadratic radial basis function interpolation method to generate a continuous three-dimensional gas concentration distribution field as a reference flow field; selecting k monitoring points from each monitoring point to form a candidate combination, using the concentrations of the k monitoring points in the candidate combination as input, and the concentrations of the remaining monitoring points in the reference flow field as output, and calculating the prediction error using an extreme random tree regression model; traversing all candidate combinations and different k values, selecting the k value with the smallest prediction error and the k value where the error tends to level off as the k value increases, and the corresponding optimal combination, as the optimal feature point set. A two-level extreme random tree regression model is constructed. The first-level model takes the upper corner gas concentration, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as input, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as input, and uses the gas concentration at the remaining monitoring points in the baseline flow field as the prediction target. The two-level models are trained using a time-series sample set, and then concatenated after training. The current monitoring values ​​of the upper corner gas concentration, hydraulic support working resistance, atmospheric pressure difference, and extraction negative pressure are read and input into the first-level model. After processing by the two-level models, the concentration prediction values ​​of the remaining monitoring points are obtained, and merged with the concentration values ​​of the feature points to form a complete concentration prediction result. Then, using the complete concentration prediction result as an interpolation node, spatial interpolation is performed using a multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field. The technical solution proposed in this application realizes low-cost, high-precision real-time dynamic construction of a three-dimensional gas flow field.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0018] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a method for dynamically constructing a three-dimensional gas flow field at a working face according to an embodiment of this application; Figure 2 This is a structural diagram of a three-dimensional gas flow field dynamic construction system for a working face, according to an embodiment of this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0020] This application proposes a method and system for dynamically constructing a three-dimensional gas flow field at a working face. The method includes: arranging monitoring points in a cross-sectional grid pattern in the three-dimensional space of the working face, collecting measured gas concentration values ​​at each point, and simultaneously collecting periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data; preprocessing these data to form a time-series sample set; using the measured gas concentration values ​​of all monitoring points as input, performing spatial interpolation using a multiple quadratic radial basis function interpolation method to generate a continuous three-dimensional gas concentration distribution field as a reference flow field; selecting k monitoring points from each monitoring point to form a candidate combination, using the concentrations of the k monitoring points in the candidate combination as input, and the concentrations of the remaining monitoring points in the reference flow field as output, and calculating the prediction error using an extreme random tree regression model; traversing all candidate combinations and different k values, selecting the k value with the smallest prediction error and the k value where the error tends to level off as the k value increases, and the corresponding optimal combination, as the optimal feature point set. A two-level extreme random tree regression model is constructed. The first-level model takes the upper corner gas concentration, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as input, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as input, and uses the gas concentration at the remaining monitoring points in the baseline flow field as the prediction target. The two-level models are trained using a time-series sample set, and then concatenated after training. The current monitoring values ​​of the upper corner gas concentration, hydraulic support working resistance, atmospheric pressure difference, and extraction negative pressure are read and input into the first-level model. After processing by the two-level models, the concentration prediction values ​​of the remaining monitoring points are obtained, and merged with the concentration values ​​of the feature points to form a complete concentration prediction result. Then, using the complete concentration prediction result as an interpolation node, spatial interpolation is performed using a multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field. The technical solution proposed in this application realizes low-cost, high-precision real-time dynamic construction of a three-dimensional gas flow field.

[0021] The following describes a method and system for dynamically constructing a three-dimensional gas flow field at a working face, based on an embodiment of this application, with reference to the accompanying drawings.

[0022] Example 1 Figure 1 The flowchart below shows a method for dynamically constructing a three-dimensional gas flow field at a working face according to an embodiment of this application. Figure 1 As shown, the method includes: Step 1: Arrange monitoring points in the three-dimensional space of the working face according to the cross-sectional grid method, collect the measured values ​​of gas concentration at each point, and simultaneously collect periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data. After preprocessing, form a time series sample set. In this embodiment of the disclosure, the arrangement of the monitoring points includes: n monitoring sections are evenly arranged along the direction of the working face. Each section has 3 layers along the vertical direction and 3 columns along the dip direction, forming a 9-point grid, for a total of 9n monitoring points on the entire working face.

[0023] It should be noted that the unit measurement method is used to arrange monitoring points in the three-dimensional space of the working face. Nine monitoring sections can be evenly arranged along the strike direction of the working face, with adjacent sections spaced 20m apart. The first section is arranged on the return air side of the working face, and the ninth section is arranged on the intake air side of the working face. Within each monitoring section, three layers are set vertically, at distances of 0.95m, 1.90m, and 2.85m from the floor, respectively; and three columns are set along the dip direction of the working face, at distances of 1.35m, 2.70m, and 4.05m from the coal wall, respectively. The three layers and three columns within each section intersect to form nine monitoring points, for a total of 81 monitoring points across the entire working face, covering the entire mining area.

[0024] In this embodiment of the disclosure, the measured values ​​of gas concentration collected at each point include: The gas concentration at each monitoring point was measured three times consecutively, and the arithmetic mean of the three gas concentration measurements was taken as the measured gas concentration value at the monitoring point.

[0025] It should be noted that each monitoring point uses a wireless gas concentration sensor for data acquisition. Gas concentration is measured three times consecutively at each monitoring point, and the arithmetic mean of the three measurements is taken as the measured gas concentration value for that monitoring point. Data acquisition covers production and maintenance shifts, with three complete sets of data measured per shift. Data is collected continuously for at least 10 working days, accumulating at least 60 complete samples, each containing gas concentration values ​​from 81 monitoring points.

[0026] Three types of external environmental variable data were collected simultaneously: periodic pressure data, which was read from the working face hydraulic support pressure monitoring system to read the working resistance value of the support, in bar; negative pressure data of goaf extraction, which was read from the negative pressure sensor installed in the gas extraction pipeline, in MPa; and atmospheric pressure data, which was obtained from the underground safety monitoring system and the ground meteorological station respectively, and the difference between underground and ground atmospheric pressure was calculated, in kPa.

[0027] In this embodiment of the disclosure, the preprocessing includes: If the gas concentration values ​​of any 9 or more monitoring points are exactly the same in a single measurement, the current measurement data is judged to be abnormal and the entire group is removed. For the data gaps caused by the removal, linear extrapolation or spline interpolation methods are used to fill them.

[0028] It should be noted that the raw data contains outliers due to sensor malfunctions or recording errors, as well as data gaps at certain times. For outliers, the following criteria were established: if the gas concentration values ​​at any nine or more monitoring points are identical in a single measurement of 81 points, the measurement data is considered outlier and the entire set is discarded. Data gaps resulting from outlier removal are filled using linear extrapolation or spline interpolation. After cleaning and completion, the gas concentration data at each time point are aligned with the data of the three types of external environmental variables according to the acquisition time, ensuring that the multi-source data at the same time point have a consistent time reference, forming a regularized time-series sample set for subsequent model training.

[0029] Step 2: Using the measured values ​​of gas concentration at all monitoring points as input, spatial interpolation is performed using the multiple quadratic radial basis function interpolation method to generate a three-dimensional continuous gas concentration distribution field as the reference flow field; In this embodiment of the disclosure, step 2 specifically includes: The measured values ​​of gas concentration at each monitoring point are used as input. The interpolation calculation is performed in the three-dimensional spatial coordinate system of the working face using the multiple quadratic radial basis function interpolation method. The concentration value at any interpolation point is obtained by weighted summation of the Euclidean distance from the interpolation point to each monitoring point after mapping with the multiple quadratic radial basis function and the corresponding weight coefficient of each monitoring point. After the interpolation calculation is completed, a three-dimensional continuous distribution field of gas concentration covering the entire working face space is generated, and the three-dimensional continuous distribution field of gas concentration is used as the reference flow field. The weighting coefficients are obtained by solving the radial basis function value matrix between monitoring points and the measured concentration value vector of the monitoring points.

[0030] It should be noted that the measured values ​​of gas concentration from all 81 monitoring points collected in step 1 are used as input. The multiple quadratic radial basis function interpolation method is used to perform spatial interpolation in the three-dimensional spatial coordinate system of the working face to generate a three-dimensional continuous distribution field of gas concentration covering the entire working face space, which serves as the reference flow field for subsequent feature point selection and model training.

[0031] Using the spatial coordinates of all 81 monitoring points and the corresponding measured values ​​of gas concentration ( ( ) are used as known data points. Several interpolation points are set in the three-dimensional coordinate system of the working surface. For any one of the interpolation points... First, calculate the Euclidean distance from the interpolation point to all 81 monitoring points. Substitute each distance into the quadratic radial basis function. Mapping is performed in , where Here, r is the shape parameter, and r is the Euclidean distance from the interpolation point to a monitoring point. The output value of the radial basis function represents the function value after mapping between the interpolation point and the known monitoring point. It is used for subsequent weighted summation calculation of the gas concentration at the interpolation point, and is optimized within the range [0.1, 2.0] through grid search. Each monitoring point corresponds to a weight coefficient. The concentration value at the interpolation point is obtained by a weighted sum of 81 radial basis function values ​​and their corresponding weighting coefficients, calculated using the following formula: ; Weighting coefficient The solution is as follows: construct an 81×81×81 matrix from the radial basis function values ​​of all 81 monitoring points. ,in Based on the measured concentration values ​​of all 81 monitoring points Constructing vectors By solving the linear equation system Φ W = C This will give you the weight coefficient vector. .

[0032] The above calculation process is repeated point by point for each interpolation point until all interpolation nodes in the entire working face space are covered. After the calculation is completed, a three-dimensional continuous gas concentration distribution field covering the entire working face space is generated. This distribution field is used as the reference flow field for feature point selection in step 3 and model training in step 4.

[0033] Step 3: Select k monitoring points from each monitoring point to form a candidate combination. Take the concentration of the k monitoring points in the candidate combination as the input and the concentration of the remaining monitoring points in the reference flow field as the output. Use an extreme random tree regression model to calculate the prediction error. Iterate through all candidate combinations and different k values, and select the k value with the smallest prediction error and the error tending to level off as the k value increases, and the corresponding optimal combination as the optimal feature point set. Specifically, 1. Candidate combination construction All 81 monitoring points on the working face are used as a candidate pool. k monitoring points are selected from the candidate pool to form a candidate combination. The value of k ranges from 3 to 7. Since it is difficult to cover the main concentration gradient range when the number of feature points is less than 3, and the monitoring cost increases significantly while the accuracy improvement slows down when the number of feature points is more than 7, the value of k ranges from 3 to 7. For each value of k, the total number of candidate combinations is the number of combinations C(81, k).

[0034] 2. Calculation of prediction error For any candidate combination, the gas concentration values ​​of k monitoring points in the combination are used as input features, and the concentration values ​​of the remaining (81-k) monitoring points in the baseline flow field (excluding the k points) are used as prediction targets. An extreme random tree regression model is employed to predict the concentration values ​​of the remaining (81-k) points using the concentration values ​​of the k points in the current candidate combination. The mean absolute error (MAE) and mean relative percentage error (MAPE) between the predicted values ​​and the actual values ​​at the corresponding points in the baseline flow field are calculated as quantitative evaluation indicators of the predictive ability of the candidate combination. Specifically, the data from each time step in the time series sample set are used as training samples, divided into training and testing sets in an 8:2 ratio. The training set is used to train the extreme random tree regression model, and the testing set is used to calculate the prediction error. Five-fold cross-validation is used to take the average value as the final error.

[0035] Extremely random tree regression models randomly select features and splitting thresholds when nodes split. By constructing multiple regression trees and taking the average as the prediction result, they can effectively reduce variance and enhance generalization ability. In this embodiment, the model parameters are set as follows: 200 regression trees, 3 minimum leaf node samples, and mean squared error as the evaluation index for node splitting.

[0036] 3. Determination of the optimal feature point set For each k value, iterate through C(81, k) candidate combinations and select the combination with the smallest mean absolute error and the mean relative percentage error not exceeding a threshold (preferably 20%) as the optimal combination for that k value. Compare the prediction accuracy of the optimal combinations for each k value from k=3 to 7, and select the k value corresponding to the inflection point where the prediction accuracy tends to level off as the k value increases, along with the optimal combination for that k value, as the final feature point set.

[0037] The specific method for determining the "inflection point that tends to level off" is as follows: calculate the decrease in MAE for each optimal combination as k increases from 3 to 7. When the decrease in MAE is less than 5% for the first time after increasing the value of k, the previous k value is the inflection point. The basis for this criterion is that the purpose of increasing the number of feature points is to improve prediction accuracy. However, when the improvement in accuracy slows down, continuing to increase the number of feature points will only increase monitoring costs without obtaining significant benefits. At this point, the previous k value is the optimal balance point between accuracy and cost.

[0038] The combination with the smallest MAE under this k value is selected as the final feature point set. This feature point set is the optimal combination of monitoring points automatically identified by the algorithm that has the best characterization ability for the entire flow field distribution. Experimental verification shows that when k=3, the decrease in MAE has slowed down, and the three feature points can cover the main concentration gradient range from the air inlet side to the air outlet side, meeting the requirements of engineering applications.

[0039] Step 4: Construct a two-level extreme random tree regression model. The first-level model takes the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as input, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as input, and uses the gas concentration at the remaining monitoring points in the baseline flow field as the prediction target. Train the two-level models using a time-series sample set, and then concatenate the two-level models after training. In this embodiment of the disclosure, training a two-level model using a time-series sample set includes: The time series sample set is divided into a training set and a test set according to a preset ratio; The training parameters for the first-level and second-level models are set, with 200 regression trees for each and 3 samples for the minimum leaf node. The mean squared error is used as the evaluation metric for node splitting. The first-level model and the second-level model are trained separately using the training set. Training is stopped when the mean squared error on the training set no longer decreases after 10 consecutive iterations, thus obtaining the first-level model and the second-level model that have been initially trained. The prediction accuracy of the first-level model and the second-level model after initial training is verified using a test set. If the mean absolute error and mean relative percentage error on the test set both meet the preset threshold, the current model is taken as the final trained model. If they do not meet the threshold, the model parameters are adjusted and the model is retrained until the preset threshold is met, thus obtaining the trained first-level model and the second-level model.

[0040] In this embodiment of the disclosure, the concatenation of the two-level models includes: Connect the output of the first-level model to the input of the second-level model, so that the predicted concentration values ​​of each feature point output by the first-level model can be directly used as the input variables of the second-level model.

[0041] Specifically, based on the feature points determined in step 3, a two-level extreme random tree regression model is constructed. Both levels of the model employ the extreme random tree regression algorithm, achieving a step-by-step mapping from external environmental variables to the concentration distribution across the entire space through a cascading approach.

[0042] The first-level model is used to establish the mapping relationship between the upper corner gas concentration and the concentrations of three types of external environmental variables and feature points. The model input consists of four types of data: upper corner gas concentration (real-time monitoring by gas concentration sensors at the intersection of the coal face and return airway), periodic pressure data (obtained from monitoring the working resistance of hydraulic supports), atmospheric pressure difference data (obtained by calculating the difference between underground and surface air pressure), and extraction negative pressure data (obtained from monitoring the negative pressure of the gas extraction pipeline in the goaf). The model's prediction target is the gas concentration value at each feature point determined in step three. During training, the upper corner gas concentration, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data at each moment in the time-series sample set are used as input features, and the measured gas concentration values ​​at each feature point at the corresponding moment are used as output labels to learn the mapping relationship between external variables and feature point concentrations.

[0043] The second-level model is used to establish the mapping relationship between the concentration of feature points and the concentration of other monitoring points in the entire working space. The model input is the concentration values ​​of each feature point output by the first-level model, and the prediction target is the gas concentration values ​​of the other monitoring points in the reference flow field excluding the feature points. During training, the concentration values ​​of feature points at each time step in the time series sample set are used as input features, and the concentration values ​​of the other monitoring points in the reference flow field at the corresponding time step are used as output labels to learn the mapping relationship between the concentration of feature points and the concentration distribution in the entire space.

[0044] The training process for the two-level model is as follows: The time-series sample set was randomly divided into a training set and a test set in an 8:2 ratio. The training set was used for model learning, and the test set was used to test the model's generalization ability. Samples from the test set were not used in the training process.

[0045] Both models were set with the same training parameters: 200 regression trees, 3 minimum leaf node samples, and mean squared error as the evaluation metric for node splitting.

[0046] The first-level model and the second-level model are trained separately using the training set. During training, the model learns the mapping relationship between input and output by minimizing the mean squared error. Training stops when the mean squared error on the training set no longer decreases after 10 consecutive iterations, at which point the two-level model is preliminarily trained.

[0047] After training, the prediction accuracy of the two-level models is verified using a test set, and the mean absolute error and mean relative percentage error of each model on the test set are calculated. If the mean absolute error and mean relative percentage error on the test set both meet the preset thresholds (overall prediction error no greater than 20%, and prediction error of feature points no greater than 11%), then the current model is taken as the final trained model; if not, the parameters such as the number of regression trees and the minimum number of leaf node samples are adjusted and retrained, and the above training and verification steps are repeated until the preset thresholds are met.

[0048] After both levels of models have been trained, the output of the first-level model is connected to the input of the second-level model. This allows the predicted concentration values ​​of each feature point output by the first-level model to be directly used as input variables for the second-level model, forming a unidirectional transmission link from external variable input to overall concentration output. In practical applications, the current monitoring values ​​of the gas concentration in the upper corner and the three types of external environmental variables are input into the first-level model. After processing by the two-level models, the predicted concentration values ​​of the remaining monitoring points on the entire working face are obtained. These values ​​are then combined with the concentration values ​​of each feature point to obtain the complete concentration prediction result for the entire working face.

[0049] Step 5: Read the current monitoring values ​​of gas concentration in the upper corner, working resistance of the hydraulic support, atmospheric pressure difference, and extraction negative pressure, input them into the first-level model, and obtain the concentration prediction values ​​of the remaining monitoring points after processing by the two-level model. Combine them with the concentration values ​​of the feature points to form a complete concentration prediction result. Then, using the complete concentration prediction result as the interpolation node, perform spatial interpolation using the multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

[0050] Specifically, after the two-level models trained in step 4 are concatenated, they are deployed on the host computer of the coal mine safety monitoring system and connected to the real-time data stream of the existing mine safety monitoring system for real-time dynamic prediction and visualization reconstruction of the three-dimensional gas flow field at the working face. The specific implementation process is as follows: The working face safety monitoring system continuously collects four types of data at a fixed frequency and stores them in a real-time database: gas concentration in the upper corner (monitored in real-time by a gas concentration sensor installed at the intersection of the coal face and the return airway), hydraulic support working resistance (collected in real-time by a hydraulic support pressure sensor at the working face, used to characterize the periodic pressure intensity of the roof), underground and surface atmospheric pressure difference (obtained by subtracting the data collected synchronously by an underground pressure sensor and a surface meteorological station), and goaf drainage negative pressure (monitored in real-time by a negative pressure sensor in the goaf gas drainage pipeline). During prediction, the latest monitoring values ​​at the current moment are read from the host computer and used as input data for the two-level models.

[0051] The four current monitoring values ​​are input into the first-level model, which outputs the predicted gas concentration values ​​for each feature point. Then, the predicted values ​​are input into the second-level model, which outputs the predicted gas concentration values ​​for the remaining monitoring points. The predicted values ​​for each feature point are then combined with the predicted values ​​for the remaining monitoring points to obtain the complete concentration prediction results for all 81 monitoring points.

[0052] Using the complete concentration prediction results of the aforementioned 81 monitoring points as interpolation nodes, spatial interpolation was performed using the multiple quadratic radial basis function interpolation method to generate a high-resolution three-dimensional gas concentration distribution field. Using the spatial coordinates of all 81 monitoring points and their corresponding concentration prediction values ​​as known data points, interpolation points were set in the three-dimensional spatial coordinate system of the working face. The Euclidean distance from each interpolation point to the 81 known points was calculated, and these distances were substituted into the multiple quadratic radial basis function interpolation method. In the process, the concentration value at the interpolation point is obtained by weighting and summing the weight coefficients corresponding to each known point. The weight coefficients of each known point are directly substituted into the weight coefficients already solved in step two, without the need for recalculation.

[0053] After point-by-point calculations are completed, a three-dimensional gas concentration distribution map covering the entire working face space is generated. This distribution map is displayed on the monitoring terminal in the form of a three-dimensional cloud map or isosurface map. The gas concentration at any location on the working face can be directly obtained by reading the concentration value of the corresponding interpolation point. When the real-time database of the safety monitoring system is updated, this step is repeated to achieve real-time dynamic updates of the three-dimensional gas flow field.

[0054] In summary, the proposed method for dynamically constructing a three-dimensional gas flow field at the working face in this embodiment obtains gas concentration distribution data at the working face through full-space gridded field measurements. A high-precision reference flow field is generated using multiple quadratic radial basis function interpolation. Based on this, an extreme random tree regression model is used to adaptively select the optimal feature point set from all monitoring points, overcoming the spatial clustering defects of purely statistical point selection. A two-level extreme random tree regression model is constructed and cascaded. The first-level model predicts the feature point concentration using three types of external variables: gas concentration at the upper corner, periodic pressure, atmospheric pressure difference, and extraction negative pressure. The second-level model predicts the full-space concentration distribution based on the feature point concentration. Field measurements verify that the overall prediction error is no greater than 20%, achieving real-time dynamic response to changes in external conditions. All input data reuses existing coal mine monitoring systems, requiring no additional hardware, and features low cost, high precision, and ease of promotion.

[0055] Example 2 Figure 2 This is a structural diagram of a three-dimensional gas flow field dynamic construction system for a working face according to an embodiment of this application, such as... Figure 2 As shown, the system includes: The acquisition module 100 is used to arrange monitoring points in a cross-sectional grid pattern in the three-dimensional space of the working face, collect the measured values ​​of gas concentration at each point, and simultaneously collect periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data. After preprocessing, a time series sample set is formed. Interpolation module 200 is used to take the measured values ​​of gas concentration at all monitoring points as input, and use the multiple quadratic radial basis function interpolation method to perform spatial interpolation to generate a three-dimensional continuous gas concentration distribution field as a reference flow field. Module 300 is used to select k monitoring points from each monitoring point to form a candidate combination. The concentration of the k monitoring points in the candidate combination is used as input, and the concentration of the remaining monitoring points in the reference flow field is used as output. An extreme random tree regression model is used to calculate the prediction error. All candidate combinations and different k values ​​are traversed, and the k value with the smallest prediction error and the error tending to level off as the k value increases and the corresponding optimal combination are selected as the optimal feature point set. Module 400 is used to construct a two-level extreme random tree regression model. The first-level model takes the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as inputs, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as inputs, and uses the gas concentration at the other monitoring points in the baseline flow field as the prediction target. The two-level models are trained using a time-series sample set, and the two-level models are concatenated after training. The generation module 500 is used to read the current monitoring values ​​of gas concentration in the upper corner, working resistance of hydraulic support, atmospheric pressure difference, and extraction negative pressure, input them into the first-level model, and obtain the concentration prediction values ​​of the remaining monitoring points after processing by the two-level model. These values ​​are then merged with the concentration values ​​of the feature points to form a complete concentration prediction result. The complete concentration prediction result is then used as an interpolation node, and spatial interpolation is performed using a multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

[0056] In this embodiment of the disclosure, the acquisition module 100 is further configured to: n monitoring sections are evenly arranged along the direction of the working face. Each section has 3 layers along the vertical direction and 3 columns along the dip direction, forming a 9-point grid, for a total of 9n monitoring points on the entire working face.

[0057] In this embodiment of the disclosure, the acquisition module 100 is further configured to: The gas concentration at each monitoring point was measured three times consecutively, and the arithmetic mean of the three gas concentration measurements was taken as the measured gas concentration value at the monitoring point.

[0058] In this embodiment of the disclosure, the acquisition module 100 is further configured to: If the gas concentration values ​​of any 9 or more monitoring points are exactly the same in a single measurement, the current measurement data is judged to be abnormal and the entire group is removed. For the data gaps caused by the removal, linear extrapolation or spline interpolation methods are used to fill them.

[0059] In this embodiment of the disclosure, the interpolation module 200 is further configured to: The measured values ​​of gas concentration at each monitoring point are used as input. The interpolation calculation is performed in the three-dimensional spatial coordinate system of the working face using the multiple quadratic radial basis function interpolation method. The concentration value at any interpolation point is obtained by weighted summation of the Euclidean distance from the interpolation point to each monitoring point after mapping with the multiple quadratic radial basis function and the corresponding weight coefficient of each monitoring point. After the interpolation calculation is completed, a three-dimensional continuous distribution field of gas concentration covering the entire working face space is generated, and the three-dimensional continuous distribution field of gas concentration is used as the reference flow field. The weighting coefficients are obtained by solving the radial basis function value matrix between monitoring points and the measured concentration value vector of the monitoring points.

[0060] In this embodiment of the disclosure, the construction module 400 is further configured to: The time series sample set is divided into a training set and a test set according to a preset ratio; The training parameters for the first-level and second-level models are set, with 200 regression trees for each and 3 samples for the minimum leaf node. The mean squared error is used as the evaluation metric for node splitting. The first-level model and the second-level model are trained separately using the training set. Training is stopped when the mean squared error on the training set no longer decreases after 10 consecutive iterations, thus obtaining the first-level model and the second-level model that have been initially trained. The prediction accuracy of the first-level model and the second-level model after initial training is verified using a test set. If the mean absolute error and mean relative percentage error on the test set both meet the preset threshold, the current model is taken as the final trained model. If they do not meet the threshold, the model parameters are adjusted and the model is retrained until the preset threshold is met, thus obtaining the trained first-level model and the second-level model.

[0061] In this embodiment of the disclosure, the construction module 400 is further configured to: Connect the output of the first-level model to the input of the second-level model, so that the predicted concentration values ​​of each feature point output by the first-level model can be directly used as the input variables of the second-level model.

[0062] In summary, the three-dimensional gas flow field dynamic construction system proposed in this embodiment achieves low-cost and high-precision real-time dynamic construction of the three-dimensional gas flow field.

[0063] Example 3 To implement the above embodiments, this disclosure also proposes an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in Embodiment 1.

[0064] Example 4 To implement the above embodiments, this disclosure also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0065] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0066] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0067] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for dynamically constructing a three-dimensional gas flow field of a working face, characterized in that, The method includes: Monitoring points are arranged in a cross-sectional grid pattern in the three-dimensional space of the working face. The measured values ​​of gas concentration at each point are collected, and periodic pressure data, atmospheric pressure difference data, and negative pressure data are collected simultaneously. After preprocessing, a time series sample set is formed. Using the measured values ​​of gas concentration at all monitoring points as input, spatial interpolation is performed using the multiple quadratic radial basis function interpolation method to generate a three-dimensional continuous gas concentration distribution field, which serves as the reference flow field. k monitoring points are selected from each monitoring point to form a candidate combination. The concentration of the k monitoring points in the candidate combination is used as the input, and the concentration of the remaining monitoring points in the reference flow field is used as the output. An extreme random tree regression model is used to calculate the prediction error. All candidate combinations and different k values ​​are traversed. The k value with the smallest prediction error and the error tending to level off as the k value increases and the corresponding optimal combination are selected as the optimal feature point set. A two-level extreme random tree regression model was constructed. The first-level model took the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as inputs, and used the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model took the concentration values ​​of each feature point output by the first-level model as inputs, and used the gas concentration at the other monitoring points in the baseline flow field as the prediction target. The two-level models were trained using a time-series sample set, and the two-level models were concatenated after training. The current monitoring values ​​of gas concentration in the upper corner, working resistance of hydraulic support, atmospheric pressure difference, and extraction negative pressure are read and input into the first-level model. After processing by the two-level model, the concentration prediction values ​​of the remaining monitoring points are obtained. These values ​​are then merged with the concentration values ​​of the feature points to form a complete concentration prediction result. The complete concentration prediction result is then used as an interpolation node, and spatial interpolation is performed using the multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

2. The method of claim 1, wherein, The monitoring points are arranged in the following ways: n monitoring sections are evenly arranged along the direction of the working face. Each section has 3 layers along the vertical direction and 3 columns along the dip direction, forming a 9-point grid, for a total of 9n monitoring points on the entire working face.

3. The method of claim 1, wherein, The measured values ​​of gas concentration collected at each point include: The gas concentration at each monitoring point was measured three times consecutively, and the arithmetic mean of the three gas concentration measurements was taken as the measured gas concentration value at the monitoring point.

4. The method of claim 1, wherein, The preprocessing includes: If the gas concentration values ​​of any 9 or more monitoring points are exactly the same in a single measurement, the current measurement data is judged to be abnormal and the entire group is removed. For the data gaps caused by the removal, linear extrapolation or spline interpolation methods are used to fill them.

5. The method of claim 1, wherein, The process involves using the measured gas concentration values ​​from all monitoring points as input, employing a multiple quadratic radial basis function interpolation method for spatial interpolation to generate a three-dimensional continuous gas concentration distribution field as the reference flow field, including: The measured values ​​of gas concentration at each monitoring point are used as input. The interpolation calculation is performed in the three-dimensional spatial coordinate system of the working face using the multiple quadratic radial basis function interpolation method. The concentration value at any interpolation point is obtained by weighted summation of the Euclidean distance from the interpolation point to each monitoring point after mapping with the multiple quadratic radial basis function and the corresponding weight coefficient of each monitoring point. After the interpolation calculation is completed, a three-dimensional continuous distribution field of gas concentration covering the entire working face space is generated, and the three-dimensional continuous distribution field of gas concentration is used as the reference flow field. The weighting coefficients are obtained by solving the radial basis function value matrix between monitoring points and the measured concentration value vector of the monitoring points.

6. The method of claim 1, wherein, The method of training a two-level model using a time-series sample set includes: The time series sample set is divided into a training set and a test set according to a preset ratio; The training parameters for the first-level and second-level models are set, with 200 regression trees for each and 3 samples for the minimum leaf node. The mean squared error is used as the evaluation metric for node splitting. The first-level model and the second-level model are trained separately using the training set. Training is stopped when the mean squared error on the training set no longer decreases after 10 consecutive iterations, thus obtaining the first-level model and the second-level model that have been initially trained. The prediction accuracy of the first-level model and the second-level model after initial training is verified using a test set. If the mean absolute error and mean relative percentage error on the test set both meet the preset threshold, the current model is taken as the final trained model. If they do not meet the threshold, the model parameters are adjusted and the model is retrained until the preset threshold is met, thus obtaining the trained first-level model and the second-level model.

7. The method of claim 6, wherein, The concatenation of the two-level models includes: Connect the output of the first-level model to the input of the second-level model, so that the predicted concentration values ​​of each feature point output by the first-level model can be directly used as the input variables of the second-level model.

8. A working face three-dimensional gas flow field dynamic construction system, characterized in that, The system includes: The data acquisition module is used to arrange monitoring points in a cross-sectional grid pattern in the three-dimensional space of the working face, collect the measured values ​​of gas concentration at each point, and simultaneously collect periodic pressure data, atmospheric pressure difference data, and negative pressure data. After preprocessing, a time series sample set is formed. The interpolation module is used to take the measured values ​​of gas concentration at all monitoring points as input, and use the multiple quadratic radial basis function interpolation method to perform spatial interpolation to generate a three-dimensional continuous gas concentration distribution field as a reference flow field. The selection module is used to select k monitoring points from each monitoring point to form a candidate combination. The concentration of the k monitoring points in the candidate combination is used as input, and the concentration of the remaining monitoring points in the reference flow field is used as output. The prediction error is calculated using an extreme random tree regression model. The module iterates through all candidate combinations and different k values, and selects the k value with the smallest prediction error and the error tends to level off as the k value increases, and the corresponding optimal combination, as the optimal feature point set. The module is used to construct a two-level extreme random tree regression model. The first-level model takes the gas concentration at the upper corner, periodic pressure data, atmospheric pressure difference data, and extraction negative pressure data as input, and uses the gas concentration at each feature point in the optimal feature point set as the prediction target. The second-level model takes the concentration values ​​of each feature point output by the first-level model as input, and uses the gas concentration at the other monitoring points in the baseline flow field as the prediction target. The two-level models are trained using a time-series sample set, and the two-level models are concatenated after training. The generation module is used to read the current monitoring values ​​of gas concentration in the upper corner, working resistance of hydraulic support, atmospheric pressure difference, and extraction negative pressure, input them into the first-level model, and obtain the concentration prediction values ​​of the remaining monitoring points after processing by the two-level model. These values ​​are then merged with the concentration values ​​of the feature points to form a complete concentration prediction result. The complete concentration prediction result is then used as an interpolation node, and spatial interpolation is performed using a multiple quadratic radial basis function interpolation method to generate a three-dimensional gas concentration distribution field.

9. An electronic device, comprising: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.