Coal mine high-energy mine earthquake prediction and main control factor judgment method based on machine learning

Abstract

The invention discloses a coal mine high-energy mine earthquake prediction and main control factor judgment method based on machine learning. The method comprises the following steps: acquiring micro-earthquake data, geological data and mining data during mining of a working face, and determining indexes corresponding to various types of data; constructing a plurality of machine learning models and carrying out model training; performing comprehensive evaluation on each model through a set prediction index, and obtaining an optimal model; and explaining a machine learning model prediction result by adopting an SHAP method, and sorting various prediction indexes by taking the size of an SHAP value as a main control factor judgment standard to obtain a main control factor. The method can comprehensively analyze the factors causing the mine earthquake, improves the accuracy of mine earthquake prediction, and guarantees the safe mining of a coal mine and the precise prevention and control of the mine earthquake.

Description

Technical Field

[0001] The invention relates to a method for predicting high-energy mine earthquakes in coal mines and determining main control factors based on machine learning, and belongs to the technical field of mine safety monitoring. Background Art

[0002] Mine tremors are the changes in the originally stable stratum structure during the mining process. The original rock stress field in the coal rock mass around the mining area is affected by mining, and the stress is abnormally concentrated and unstable. When the stress concentration in the local area reaches the bearing limit of the coal rock mass, the energy is suddenly released, causing violent rock stratum vibration. Among them, high-energy strong mine tremors are prone to induce secondary dynamic disasters such as rock burst.

[0003] Mine earthquake prediction is a complex process. Existing prediction methods often ignore the influence of nonlinear factors. Most mine earthquake predictions rely on experience, which easily ignores the characteristics contained in massive data and is greatly affected by subjective factors. With the rapid development of big data and machine learning technology, machine learning algorithms are becoming an alternative method to solve such nonlinear fuzzy problems in the field of traditional rock mechanics engineering, providing a development direction for mine earthquake prediction. However, current machine learning algorithms often only consider microseismic data and cannot predict mine earthquakes more accurately, and cannot fully meet the requirements of safe mining of coal mines and precise prevention and control of mine earthquakes. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting high-energy mine earthquakes in coal mines and determining the main controlling factors based on machine learning. The method can comprehensively analyze the factors causing mine earthquakes, improve the accuracy of mine earthquake prediction, and ensure the safe mining of coal mines and the precise prevention and control of mine earthquakes.

[0005] In order to achieve the above object, the present invention provides a method for predicting high-energy mine earthquakes in coal mines and determining main controlling factors based on machine learning, comprising the following steps:

[0006] S1. Obtain microseismic data, geological data and mining data during the mining period of the working face and determine the indicators corresponding to each type of data;

[0007] S2. Build multiple machine learning models and perform model training;

[0008] S3, comprehensively evaluate the models in S2 through the set prediction indicators and obtain the optimal model;

[0009] S4. Use the SHAP method to explain the prediction results of the machine learning model. Use the SHAP value as the criterion for judging the main controlling factors, sort the prediction indicators in S3, and obtain the main controlling factors.

[0010] Furthermore, the indicators corresponding to the microseismic data in S1 include b value, total fault area, spatiotemporal diffusion, lack of earthquakes, A(b) value, Z value, equivalent energy level, activity, time information entropy, and algorithm complexity; the indicators corresponding to the geological data include the thickness of the main mining coal seam, the buried depth of the coal seam, the thickness of the main thick hard roof above the coal seam and the distance from the coal seam, the characteristic parameters of the roof rock thickness, folds and fault structures; the indicator corresponding to the mining data is the working face advancement speed.

[0011] Furthermore, it is characterized in that the construction and training process of the S2 machine learning model is:

[0012] S2.1, performing normalization preprocessing on various data obtained in S1, the method is: first taking lg E of the energy data, and then using minimum-maximum normalization to make the data fall into the region [0, 1]; the energy data is data related to the energy of microseismic events, such as average energy, maximum energy, etc.;

[0013] S2.2, divide the data set processed in S2.1 into training set, validation set and test set in a ratio of 7:2:1;

[0014] S2.3. Construct decision tree model DT, random forest model RF, lightweight gradient boosting machine algorithm model LightGBM, extreme gradient boosting model XGBoost and classification boosting model CatBoost according to machine learning algorithm data package and Python tools;

[0015] S2.4, substitute the data in the training set into each model in S2.3, and obtain a preliminary model by learning the patterns and features in the data;

[0016] S2.5. Based on the trained machine learning models, the hyperparameters of each machine learning model are optimized through the validation set by means of ten-fold cross validation. The key hyperparameters include the number of decision trees n_estimators, the maximum depth of the decision tree max_depth, the minimum number of separation samples min_samples_split, the minimum number of leaf node samples min_samples_leaf, the maximum number of separation features max_features, and whether to perform random sampling bootstrap;

[0017] The model effect of ten-fold cross validation was evaluated using three error indicators: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square (RMSE). The value that minimized the three error indicators was obtained through traversal search, which was the optimal hyperparameter value, and a trained machine learning model that met the prediction effect requirements was obtained.

[0018] Furthermore, the indicators used by S3 to comprehensively evaluate each model are accuracy, precision, recall and F1 score; the evaluation process is as follows:

[0019] S3.1. Use the test set to calculate the accuracy, precision, recall and F1 score indicators of each machine learning model. The calculation formulas for each indicator are:

[0020]

[0021] Among them, TP is the number of times that high-energy mine tremors are accurately predicted to occur, TN is the number of times that high-energy mine tremors are accurately predicted not to occur, FP is the number of times that high-energy mine tremors are incorrectly predicted to occur, and FN is the number of times that high-energy mine tremors are incorrectly predicted not to occur;

[0022] S3.2. Sort the index values ​​calculated in S3.1 according to the F1 score, precision, accuracy, and recall rate, and select the optimal model.

[0023] Furthermore, the specific process of S4 is as follows:

[0024] S4.1. Use the optimal model obtained in S3 as the prediction model to predict high-energy mine earthquakes at the working face;

[0025] S4.2. Calculate the SHAP value of each set prediction index, including maximum energy, total energy, average energy, energy deviation, cumulative energy, total frequency, resource concentration, daily frequency, daily footage, total footage of the previous three days, coal seam thickness, burial depth, fold and date;

[0026] The calculation formula is as follows:

[0027]

[0028] Where c is the number of prediction indicators, A={x1,x2,…,x c}, if there is a set S that satisfies and Assume there is a sample x k =[x1=a1,x2=a2,…,x c =a c ], then it is called For x i =a i The SHAP value is a union of feature x i The weighted sum of contributions across all possible coalitions;

[0029] S4.3. Sort the SHAP values ​​of the prediction indicators obtained in S4.2 from large to small to obtain their main controlling factors.

[0030] Furthermore, the indicators corresponding to the microseismic data include b value, total fault area, spatiotemporal diffusion, missing earthquakes, A(b) value, Z value, equivalent energy level, activity, time information entropy, and algorithm complexity. The calculation formulas are:

[0031] logN(M)=a-bM; where N(M) is the cumulative number of microseismic events with magnitude greater than M, b is the b value, and a and b are both constants;

[0032] Where k0 is the lower limit of the statistical microseismic energy level, k is the microseismic energy level of each microseismic event, and N(k) is the event count of microseismic energy level k;

[0033] in, The average distance between consecutive microseismic events, is the average time between microseismic events;

[0034] Among them, M mean is the average number, M min is the minimum amplitude of a given microseismic event;

[0035] Where, b is the b value, M i is the magnitude of the microseismic event, and the A(b) value is the conversion amplitude of a microseismic event set;

[0036] in, is the arithmetic mean size of all microseismic events during the entire monitoring period; is the arithmetic mean size of the microseismic events in the time window; N and n are the number of microseismic events in the entire detection period and the number of microseismic events in the time window, respectively; ΣM and Σm are and The corresponding standard deviation; when Z = 0, the occurrence rate of microseismic events is the same as the background conditions, and when Z < 0, it indicates that the probability of high-energy microseismic events will increase; when the Z value is less than 0, it can be defined as an indication of microseismic anomaly, and the larger the absolute value, the higher the risk level of rock burst;

[0037] Among them, E i is the energy of the i-th microseismic event, M E is the amplitude of the microseismic event after conversion, M Eavg M E It can be inferred that there will be an abnormal high value of EEM before the rock burst occurs;

[0038] Where N is the number of microseismic events in the entire detection period, M maxis the maximum amplitude of the microseismic event; it can be concluded that when S D When the value increases or remains high, rock bursts will be extremely likely to occur;

[0039] in, Its value is 0 to 1, t i is the time of occurrence of the i-th microseismic event; this indicator mainly shows the clustering characteristics of the time distribution of microseismic events, or the heterogeneity or order characteristics in the time evolution process. The more systems there are, the lower the information entropy. Therefore, if Q t Reduced, rock bursts will be extremely likely;

[0040] Where n is the number of microseismic events in the time window; M AC =M MAX -M MIN +1. This parameter does not require the sequence to be a synchronous one, which avoids artifacts and noise, so the AC value will decrease before the rock burst occurs.

[0041] The present invention obtains microseismic data, geological data and mining data during the mining of the working face and determines the indicators corresponding to each type of data, then constructs multiple machine learning models and performs model training, comprehensively evaluates each model through the set prediction indicators, and obtains the optimal model; the SHAP method is used to interpret the prediction results of the machine learning model, and the SHAP value is used as the standard for judging the main control factors, and the various prediction indicators are sorted to obtain the main control factors. The present invention organically combines microseismic data, geological data, mining data and machine learning, which makes up for the problem of poor accuracy in predicting mine earthquakes using only microseismic data, and at the same time, it also determines the main control factors of mine earthquakes, improves the accuracy of mine earthquake predictions, and ensures the safe mining of coal mines and the precise prevention and control of mine earthquakes. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a work flow chart of the method of the present invention;

[0043] Figure 2 is a time series variation curve of the b-value warning indicator of the embodiment of the present invention;

[0044] Figure 3 is a time series variation curve of the total fault area early warning indicator in an embodiment of the present invention;

[0045] Figure 4 is a time series variation curve of the spatiotemporal diffusion warning indicator of the embodiment of the present invention;

[0046] Figure 5 is a time series variation curve of the earthquake absence warning indicator in the embodiment of the present invention;

[0047] Figure 6 is a time series variation curve of the A(b) value warning indicator of the embodiment of the present invention;

[0048] Figure 7 is a time series variation curve of the Z value warning indicator of the embodiment of the present invention;

[0049] Figure 8 is a time series variation curve of the equivalent energy level warning indicator of the embodiment of the present invention;

[0050] Fig. 9 is a time series variation curve of the activity warning indicator according to an embodiment of the present invention;

[0051] Fig.10 is a time series variation curve of the time information entropy warning indicator of the embodiment of the present invention;

[0052] Fig.11 is a time series variation curve of the algorithm complexity warning indicator of the embodiment of the present invention;

[0053] Fig.12 SHAP value result diagram of each influencing factor in the embodiment of the present invention;

[0054] Fig.13 is a graph showing the effect of coal seam thickness on maximum energy according to an embodiment of the present invention;

[0055] Fig.14 is a diagram showing the influence of the burial depth on the maximum energy according to an embodiment of the present invention;

[0056] Fig.15 Graph showing the effect of folding on maximum energy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings.

[0058] like Figure 1 As shown, a method for predicting high-energy mine earthquakes in coal mines and determining main controlling factors based on machine learning includes the following steps:

[0059] S1. Obtain microseismic data, geological data and mining data during the mining period of the working face and determine the indicators corresponding to each type of data;

[0060] S2. Build multiple machine learning models and perform model training;

[0061] S3, comprehensively evaluate the models in S2 through the set prediction indicators and obtain the optimal model;

[0062] S4. Use the SHAP method to explain the prediction results of the machine learning model. Use the SHAP value as the criterion for judging the main controlling factors, sort the prediction indicators in S3, and obtain the main controlling factors.

[0063] Example: Take the 204 working face of a coal mine in Shaanxi as an example. The strike length of the working face is 1455 meters, and the dip length is 200 meters wide. Coal seam No. 4 is a single-layer coal seam with a thickness of 5.9 to 15.0 meters and a maximum depth of more than 1000 meters. The coal seam has an altitude of +72m to +147m, and there are two synclines and one anticline. The two synclines are almost perpendicular and parallel to the strike of the working face, respectively, and the anticline is 45° to the strike of the working face. There are two secondary key layers and one main key layer above the coal seam. The thickness of the key layers are 21.1m, 54.0m and 101.3m, respectively, and the distances from the coal seam are 13.4m, 139.1m and 193.1m, respectively. The original microseismic monitoring data and related information of the working face during the mining period (May 1, 2019 to October 31, 2020) were collected: a total of 239 large-energy mining earthquake events (microseismic energy greater than 1E+5J) occurred during the mining process of the working face, and early warning was issued for them. The specific steps are as follows:

[0064] (1) Obtain all microseismic data, geological data, and mining data during the working face mining period and determine the corresponding indicators

[0065] (1.1) Collect real-time monitoring data from the on-site microseismic monitoring system and pre-process the raw data using a 15-day time window and a 1-day slip step. Figure 2 to Figure 11 Shown is the time series variation curve of each rock burst precursor warning indicator;

[0066] (1.2) Collect geological data, including the thickness of the main coal seam, the depth of the coal seam, the thickness of the main thick hard roof above the coal seam and its distance from the coal seam, the characteristic parameters of the thickness of the roof rock layer (within 100m above the coal seam), folds and fault structures;

[0067] (1.3) Collect the working face advancement speed;

[0068] (2) Build five tree-based machine learning models and perform model training

[0069] (2.1) Since the indicators have different dimensions, they should be normalized before entering the model: first, the energy data is taken as lg E, and then min-max normalization is used to make the data fall into a specific range [0, 1];

[0070] (2.2) Divide the processed data set into training set, validation set and test set in a ratio of 7:2:1;

[0071] (2.3) Construct five tree-based machine learning models: decision tree model (DT), random forest model (RF), lightweight gradient boosting machine algorithm model (LightGBM), extreme gradient boosting model (XGBoost), and classification boosting model (CatBoost) based on machine learning algorithm data packages and Python tools;

[0072] (2.4) Substitute the data in the training set into all models and obtain a preliminary model by learning the patterns and features in the data;

[0073] (2.5) Based on the five machine learning models after training, the hyperparameters of the machine learning models are optimized by a ten-fold cross-validation method. The key hyperparameters include the number of decision trees n_estimators, the maximum depth of the decision tree max_depth, the minimum number of separation samples min_samples_split, the minimum number of leaf node samples min_samples_leaf, the maximum number of separation features max_features, and whether to perform random sampling bootstrap.

[0074] The model effect of the ten-fold cross-validation was evaluated using the three error indicators of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square (RMSE). The value that minimized the three error indicators was searched traversally, which was the optimal hyperparameter value, thus obtaining a trained machine learning model that met the prediction effect requirements.

[0075] (3) Comprehensively evaluate the model through indicators such as accuracy, precision, recall and F1 score to obtain the optimal model

[0076] (3.1) Use the test set to evaluate the model performance and calculate the accuracy, precision, recall and F1 score of the five machine learning models; compare the predicted data with the real data, calculate the evaluation indicators of the model and sort them in order of F1 score as shown in Table 1:

[0077] Table 1 Evaluation indicators of various machine learning models

[0078] Model Accuracy Accuracy Recall F1 Score LightGBM 0.8831 0.7621 0.8273 0.7921 XGBoost 0.8757 0.7711 0.7636 0.7650 DT 0.8733 0.7675 0.7545 0.7553 CatBoost 0.8636 0.7364 0.7727 0.7498 RF 0.8514 0.7220 0.7364 0.7248

[0079] (3.2) Comprehensively evaluate the model according to the indicators, sort it according to the F1 score, precision, accuracy, and recall rate, and select the optimal model; As shown in Table 1, the lightweight gradient boosting machine algorithm model (LightGBM) is selected as the optimal model;

[0080] (4) The SHAP method is used to interpret the prediction results of the machine learning model. The SHAP value is used as the criterion for judging the main controlling factors, and the prediction indicators are ranked and their main controlling factors are determined.

[0081] (4.1) Use the lightweight gradient boosting machine algorithm model (LightGBM) to predict high-energy mine earthquakes at the working face;

[0082] (4.2) Calculate the SHAP value of each prediction index: Select maximum energy, total energy, average energy, energy deviation, cumulative energy, total frequency, resource concentration, daily frequency, daily footage, total footage of the previous three days, coal seam thickness, burial depth, fold, and date as prediction indicators, and calculate the SHAP value of each prediction index by the following formula:

[0083]

[0084] Where c is the number of prediction indicators, A={x1,x2,…,x c}, if there is a set S that satisfies and Assume there is a sample x k =[x1=a1,x2=a2,…,x c =a c ], then it is called For x i =a i The SHAP value is a union of feature x i The weighted sum of contributions across all possible coalitions;

[0085] (4.3) Sort the SHAP values ​​of each prediction indicator and obtain its main controlling factor: Sort the SHAP values ​​of each prediction indicator and plot them as follows Fig.12 As shown, it can be seen that the maximum energy is its main controlling factor;

[0086] By separately plotting the effects of coal seam thickness, burial depth and folding on maximum energy, such as Figure 13 to Figure 15 As shown, we can see that: the greater the thickness of the coal seam, the greater the maximum energy that appears, which proves that the thickness of the coal seam has a positive effect on the maximum energy; the greater the burial depth, the greater the maximum energy that appears, which proves that the burial depth has a positive effect on the maximum energy; the smaller the fold, the greater the maximum energy that appears, which proves that the fold has a negative effect on the maximum energy.

Claims

1. A method for predicting high-energy mine earthquakes in coal mines and determining main controlling factors based on machine learning, characterized in that: The steps include: S1. Obtain microseismic data, geological data and mining data during the mining period of the working face and determine the indicators corresponding to each type of data; S2. Build multiple machine learning models and perform model training; S3, comprehensively evaluate each model in S2 through the set prediction indicators and obtain the optimal model; S4. Use the SHAP method to explain the prediction results of the machine learning model. Use the SHAP value as the criterion for judging the main controlling factors, sort the prediction indicators in S3, and obtain the main controlling factors.

2. The method for predicting high-energy mine earthquakes and determining main controlling factors in coal mines based on machine learning according to claim 1 is characterized in that: The indicators corresponding to the microseismic data in S1 include b value, total fault area, spatiotemporal diffusion, lack of earthquakes, A(b) value, Z value, equivalent energy level, activity, time information entropy, and algorithm complexity; the indicators corresponding to the geological data include the thickness of the main mining coal seam, the buried depth of the coal seam, the thickness of the main thick hard roof above the coal seam and the distance from the coal seam, the characteristic parameters of the roof rock thickness, folds and fault structures; the indicator corresponding to the mining data is the working face advancement speed.

3. The method for predicting high-energy mine earthquakes and determining main controlling factors in coal mines based on machine learning according to claim 1 or 2, characterized in that: The construction and training process of the S2 machine learning model is as follows: S2.1, perform normalization preprocessing on the various data obtained in S1, the method is: first take lg E for the energy data, then use minimum-maximum normalization to make the data fall into the region [0, 1]; S2.2, divide the data set processed in S2.1 into training set, validation set and test set in a ratio of 7:2:1; S2.

3. Construct decision tree model DT, random forest model RF, lightweight gradient boosting machine algorithm model LightGBM, extreme gradient boosting model XGBoost and classification boosting model CatBoost according to machine learning algorithm data package and Python tools; S2.4, substitute the data in the training set into each model in S2.3, and obtain a preliminary model by learning the patterns and features in the data; S2.

5. Based on the trained machine learning models, the hyperparameters of each machine learning model are optimized through the validation set by means of ten-fold cross validation. The key hyperparameters include the number of decision trees n_estimators, the maximum depth of the decision tree max_depth, the minimum number of separation samples min_samples_split, the minimum number of leaf node samples min_samples_leaf, the maximum number of separation features max_features, and whether to perform random sampling bootstrap; The model effect of ten-fold cross validation was evaluated using three error indicators: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square (RMSE). The value that minimized the three error indicators was obtained through traversal search, which was the optimal hyperparameter value, and a trained machine learning model that met the prediction effect requirements was obtained.

4. The method for predicting high-energy mine earthquakes and determining main controlling factors in coal mines based on machine learning according to claim 3 is characterized in that: The indicators used by S3 to comprehensively evaluate each model are accuracy, precision, recall and F1 score; the evaluation process is as follows: S3.

1. Use the test set to calculate the accuracy, precision, recall and F1 score indicators of each machine learning model. The calculation formulas for each indicator are: Among them, TP is the number of times that high-energy mine tremors are accurately predicted to occur, TN is the number of times that high-energy mine tremors are accurately predicted not to occur, FP is the number of times that high-energy mine tremors are incorrectly predicted to occur, and FN is the number of times that high-energy mine tremors are incorrectly predicted not to occur; S3.

2. Sort the index values ​​calculated in S3.1 according to the F1 score, precision, accuracy, and recall rate, and select the optimal model.

5. The method for predicting high-energy mine earthquakes and determining main controlling factors in coal mines based on machine learning according to claim 4 is characterized in that: The specific process of S4 is as follows: S4.

1. Use the optimal model obtained in S3 as the prediction model to predict high-energy mine earthquakes at the working face; S4.

2. Calculate the SHAP value of each set prediction index, including maximum energy, total energy, average energy, energy deviation, cumulative energy, total frequency, resource concentration, daily frequency, daily footage, total footage of the previous three days, coal seam thickness, burial depth, fold and date; The calculation formula is as follows: Where c is the number of prediction indicators, A={x1,x2,…,x c }, if there is a set S that satisfies and Assume there is a sample x k =[x1=a1,x2=a2,…,x c =a c ], then it is called For x i =a i The SHAP value is a union of feature x i The weighted sum of contributions across all possible coalitions; S4.

3. Sort the SHAP values ​​of the prediction indicators obtained in S4.2 from large to small to obtain their main controlling factors.

6. The method for predicting high-energy mine earthquakes in coal mines and determining main controlling factors based on machine learning according to claim 2 is characterized in that: The indicators corresponding to the microseismic data include b value, total fault area, spatiotemporal diffusion, missing earthquakes, A(b) value, Z value, equivalent energy level, activity, time information entropy, and the calculation formulas of algorithm complexity are: logN((M) = a-bM; where N(M) is the cumulative number of microseismic events greater than M, b is the b value, and a and b are both constants; Where k0 is the lower limit of the statistical microseismic energy level, k is the microseismic energy level of each microseismic event, and N(k) is the event count of microseismic energy level k; in, The average distance between consecutive microseismic events, is the average time between microseismic events; Among them, M mean is the average number, M min is the minimum amplitude of a given microseismic event; Where, b is the b value, M i is the magnitude of the microseismic event, and the A(b) value is the conversion amplitude of a microseismic event set; in, is the arithmetic mean size of all microseismic events during the entire monitoring period; is the arithmetic mean size of the microseismic events in the time window; N and n are the number of microseismic events in the entire detection period and the number of microseismic events in the time window, respectively; ΣM and Σm are and The corresponding standard deviation; when Z = 0, the occurrence rate of microseismic events is the same as the background condition, and when Z < 0, it means that the probability of high-energy microseismic events will increase; Among them, E i is the energy of the i-th microseismic event, M E is the amplitude of the microseismic event after conversion, M Eavg M E The average value of Where N is the number of microseismic events in the entire detection period, M max is the maximum amplitude of a microseismic event; in, Its value is 0 to 1, t i is the occurrence time of the i-th microseismic event; Where n is the number of microseismic events in the time window; M AC =M MAX -M MIN +1.

Images