Coal and gas outburst risk prediction method based on double-layer Stacking integration algorithm

Abstract

The invention provides a coal and gas outburst risk prediction method based on a double-layer Stacking integration algorithm. The coal and gas outburst risk prediction method comprises the following steps: S1, selecting prediction indexes; s2, dividing the data sample library into two groups, namely security data and highlighted data; s3, preprocessing the highlighted data in the data sample library; S4, dividing the highlighted data into a training set and a verification set according to a certain proportion; s5, dividing the training set into a basic training set and a test set according to a certain proportion, and performing basic model training on the basic training set to obtain a new feature matrix; and S6, the new feature matrixes are respectively transmitted to respective meta-models for learning, and a plurality of different groups of meta-models are selected for voting to determine a final prediction result. The technical problems that the obtained feature information is limited due to the single model, various machine learning algorithms are difficult to fuse and complement, and correlation analysis of various prediction indexes is lacked at present are solved, so that the coal and gas outburst prediction accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of underground mine risk prediction, and in particular relates to a coal and gas outburst risk prediction method based on a double-layer Stacking integrated algorithm. Background Art

[0002] With the rapid development and progress of data mining and machine learning, coal mines have begun to predict the risk of coal and gas outbursts through data mining and machine learning, which can effectively reduce the risk of coal and gas outbursts. However, there are many reasons for coal and gas outbursts, and the mechanism of occurrence is very complex. There are mainly the following influencing factors: coal seam occurrence conditions, coal metamorphism, coal seam permeability, coal rock properties, gas content, gas pressure, temperature, mining methods, mining depth, etc. The degree of influence varies. If coal mine disasters are to be predicted, data from different data sources need to be processed simultaneously to ensure the accuracy of coal mine disaster prediction as much as possible.

[0003] The machine learning algorithms currently used in coal and gas outburst risk prediction all have their own shortcomings, mainly in the following three points:

[0004] (1) Insufficient accuracy: The traditional Stacking ensemble method can combine multiple models to achieve certain generalization and accuracy, but it only uses a set of basic models, and the feature information obtained is limited and fixed, which limits the feature information received by the meta-model. In addition, since there is only one meta-model, different meta-models perform differently on the same feature information, so it cannot be guaranteed that the meta-model will achieve good results when learning new features. In the combination of the base model and the meta-model, due to the influence of factors such as the data type, the result may be poor and the accuracy may be insufficient.

[0005] (2) Single algorithm and poor universality: Various machine learning algorithms are currently used in mine disaster risk prediction. Only one machine learning algorithm can be used for each risk prediction. The algorithms cannot be integrated to complement each other’s strengths, resulting in a narrow scope of application and poor stability of each machine learning algorithm.

[0006] (3) Lack of comprehensive analysis of prediction indicators: The analysis indicators are single, mainly considering the degree of influence of various influencing factors. There is no comprehensive analysis of the integration of various influencing factors. Coal and gas outbursts are caused by the combined effect of various factors. Single data processing fails to establish an interrelated analysis model. When predicting the risk of coal and gas outbursts, it is impossible to dynamically and accurately predict disasters.

[0007] Therefore, it is necessary to provide an improved technical solution to address the above-mentioned deficiencies in the prior art. Summary of the invention

[0008] The purpose of the present invention is to overcome the deficiencies in the prior art, and the present invention provides a.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A coal and gas outburst risk prediction method based on a double-layer Stacking integrated algorithm includes the following steps:

[0011] Step S1, selecting multiple prediction indicators for machine learning based on data acquisition difficulty and representativeness;

[0012] Step S2, selecting multiple groups of data to form a data sample library for machine learning based on the prediction indicators, and dividing the data in the data sample library into two groups: safety data and prominent data;

[0013] Step S3, pre-processing the outstanding data in the data sample library to remove the abnormal data and form an original data set;

[0014] Step S4, dividing the protrusion data into a training set and a validation set according to a certain ratio;

[0015] Step S5, dividing the training set into a basic training set and a test set according to a certain ratio, and performing basic model training on the basic training set to obtain a new feature matrix;

[0016] In step S6, the new feature matrices are respectively sent to respective meta-models for learning, and a plurality of different groups of meta-models are selected to vote and determine the final prediction result.

[0017] Preferably, in step S1, the prediction indexes include gas content, gas pressure, initial velocity of gas release, solidity coefficient of coal, and burial depth;

[0018] In step S2, a data sample library is formed by crawling public data of coal mines.

[0019] Preferably, in step S3, a box plot is used to eliminate abnormal data, and a prediction algorithm for prominent data is established.

[0020] Preferably, in step S4, multiple basic models are trained by a double-layer stacking algorithm, and the original features and each basic model are spliced ​​and then trained again to interact with the original features and the basic models respectively.

[0021] Preferably, the meta-model and the basic model used in the double-layer stacking algorithm are of the same type, and the model hyper-parameters of the same basic model are the same.

[0022] Preferably, after the training set is divided into a basic training set and a test set according to a certain ratio, each group of basic models learns the same original feature information, and the data of the basic training set is used for training during the training process, and the data of the test set is used for k-fold cross validation.

[0023] Preferably, the prediction results of the test set are synthesized through weighted summation to obtain the complete new feature information ultimately used by the meta-model, ensuring that the dimensions of the new feature matrices of the test set and the training set are consistent, so as to retain the prediction information to the maximum extent and obtain a set of new feature matrices for use by the meta-model.

[0024] Preferably, f1-score is used as the evaluation index, and the possible values ​​of each hyperparameter are traversed in turn through a loop algorithm through grid search. The parameters determined by the loop traversal are brought into the algorithm, and the loop traversal of the next parameter is continued to determine each hyperparameter one by one, and the hyperparameter combination with the best prediction result is taken as the final hyperparameter combination.

[0025] Preferably, the output new feature matrix is ​​sent to each meta-model for learning, so that the meta-model can learn more feature information, and multiple different groups of meta-models are selected to vote to determine the final prediction result.

[0026] Preferably, the ratio of training set to validation set is 8:1;

[0027] The ratio of the basic training set to the test set in the training set is 9:1.

[0028] Beneficial effects: It solves the technical problems of limited feature information obtained by the current single model, difficulty in integrating and complementing various machine learning algorithms, and lack of correlation analysis of various prediction indicators, so as to improve the accuracy of coal and gas outburst prediction and reduce the risk of coal mine disasters and accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. Among them:

[0030] Figure 1 This is a schematic diagram of the risk prediction method flow in the specific embodiment provided in this application;

[0031] Figure 2 This is a double-layer stacking integrated algorithm model diagram in the specific embodiment provided in this application;

[0032] Figure 3 This is a working principle diagram of the basic model of the double-layer stacking integration algorithm in the specific embodiment provided in this application;

[0033] Figure 4This is a test result diagram of the double-layer stacking integration algorithm verification set in the specific embodiment provided in this application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0035] In the description of the present invention, the terms "longitudinal", "lateral", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and do not require that the present invention must be constructed and operated in a specific orientation. Therefore, they should not be understood as limitations on the present invention. The terms "connected" and "connected" used in the present invention should be understood in a broad sense. For example, they can be fixed connections or detachable connections; they can be directly connected or indirectly connected through intermediate components. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0036] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present invention can be combined with each other without conflict.

[0037] like Figure 1 As shown, the coal and gas outburst risk prediction method based on the double-layer Stacking integrated algorithm provided by the present application, the so-called Stacking integrated algorithm is an integrated learning technology that integrates multiple classification models through a meta-classifier. The basic model uses the entire training set for training, and the meta-model uses the features of the basic model as features for training. The basic model usually contains different learning algorithms. In addition to using the machine learning and integrated learning basic algorithms of the Scikit-learn machine learning platform, the model of the present application also introduces the deep neural network (DNN) in the PyTorch platform into the model, involving integrated algorithms, basic algorithms, deep neural networks and other learners, namely KNN (K-NearestNeighbor), LGBM (LightGBM), RF (RandomForest), ET (ExtraTrees), CatBoost, XGBoost (eXtreme GradientBoosting), DNN (Deep Neural Networks) 7 different types of machine learning algorithms, each algorithm contains one or more different hyperparameter values, and is optimized and tuned using the same method.

[0038] The specific prediction method includes the following steps:

[0039] Step S1, selecting prediction indicators. The selection of prediction indicators should comprehensively consider the ease of data acquisition and representativeness;

[0040] Step S2, obtain historical coal and gas safety and outburst data to form a data sample library for machine learning. All data sources can be public information published at home and abroad, and the data should contain a complete coal and gas outburst prediction indicator system.

[0041] Step S3, pre-processing the coal and gas outburst data, using a box plot to remove abnormal data, forming an original data set, and only using the original data not affected by the abnormality for basic training.

[0042] Step S4, dividing the coal and gas outburst data into a training set and a validation set in a ratio of 8:1;

[0043] Step S5, dividing the training set into a basic training set and a test set at a ratio of 9:1, and performing basic model training on the 9 basic training sets to obtain a complete new feature matrix for final use;

[0044] In step S6, the new feature matrices are respectively transmitted to respective meta-models for learning, and on the premise of ensuring the prediction performance and learning ability of the Stacking integrated algorithm, multiple different groups of meta-models are selected to vote and determine the final prediction result.

[0045] Furthermore, the output new feature matrix is ​​sent to each meta-model for learning. The meta-model can learn more feature information. Under the premise of ensuring the prediction performance and learning ability of the Stacking integrated algorithm, multiple different groups of meta-models are selected to vote to determine the final prediction result.

[0046] Compared with the traditional Stacking integration method, the double-layer Stacking integration algorithm adopted in this application, each layer of the algorithm is composed of multiple sets of base models and multiple meta-models, which can obtain a large amount of different feature information, so that the accuracy of the feature information provided by the base model is higher than that of the original feature information. The evaluation indicators of each machine learning algorithm in the meta-model are improved, and the prediction performance is significantly improved.

[0047] In an optional embodiment, in step S1, the prediction index includes gas content, gas pressure, gas release initial velocity, coal solidity coefficient, and burial depth; in step S2, a data sample library is formed by grabbing public data of coal mines. In this embodiment, 94 groups of effective coal and gas outburst data are selected to form a large data sample library. The 94 groups of outburst data prediction results are divided into two types, one is safety data and the other is outburst data, which are represented by 0 and 1 respectively. There are 47 data in the group of safety data and 47 data in the group of outburst data, so as to maintain sample balance.

[0048] In an optional embodiment, since some anomalies may occur in actual work, in order to explore the abnormal situation of coal and gas outburst data, in step S3, a box plot is used to analyze the coal and gas outburst data prediction index system, all abnormal data are eliminated, and numerical predictions are performed on them, and a coal and gas outburst data prediction algorithm is established.

[0049] In this embodiment, the data is normalized by adding one to all the data and taking the logarithm to avoid negative values.

[0050] like Figure 2 As shown, in this embodiment, in step S4, multiple basic models are trained by a double-layer stacking algorithm. The integrated algorithm can use all features and defined multiple different types of algorithms to train multiple basic models Model-1, Model-2, and Model-n, and then train again by splicing the original features and each basic model separately to interact with the original features and the basic models respectively.

[0051] In this way, the original features and the model can interact during stacking, which can also be regarded as a skip-connection. This is a design similar to a residual neural network. It is well known that multi-layer Stacking models have serious overfitting phenomena, especially if additional noise may be learned in the training set, which may amplify the risk of overfitting. This application first divides the coal and gas outburst dataset into a training set and a validation set in a ratio of 8:2. The validation set data does not participate in training and is only used as a validation set.

[0052] Among them, in order to improve the generalization ability of the model, the meta-model and basic model used in the two-layer stacking algorithm are of the same type, and different basic models are not used to process different data. The model hyperparameters of the same basic model are the same. Each layer of the algorithm is composed of all base models, and the input value is the original data.

[0053] like Figure 3As shown, in step S5, after the training set is divided into a basic training set and a test set in a ratio of 9:1, each group of base models learns the same original feature information. In order to prevent overfitting, the data is divided into 10 parts, one of which is set aside as a test sample, and the remaining 9 parts of the data are used for training. The test data is scored through k-fold cross validation. Since k-fold cross validation will train 10 models, each piece of data will eventually have the opportunity to become a test sample for other models. This method can make full use of all training samples.

[0054] Furthermore, in case that the 10 prediction results may be different, it is necessary to synthesize the complete new feature information finally used by the meta-model through a simple weighted summation of the prediction results of the test set to ensure that the dimensions of the new feature matrix of the test set and the training set are consistent. This can retain the prediction information to the maximum extent, that is, obtain a set of new feature matrices of the test set used by the meta-model.

[0055] The performance of the integrated algorithm and basic algorithm in the Stacking integrated algorithm metamodel is compared and analyzed with that of the integrated algorithm and basic algorithm in the independent algorithm in the test set.

[0056] The first layer (L1) algorithm consists of 4 basic algorithms, namely KNN, RF, ET, and DNN, and 3 integrated algorithms, namely LGBM, CatBoost, and XGBoost. Table 1 shows that the F1 values ​​of the first layer basic algorithms are generally low. The worst performing model is the integrated model XGBoost, and the best performing model is the integrated model CatBoost and the basic model DNN. The F1 values ​​of the models are all above 0.9, and the prediction effect reaches the prediction effect of the general model. The F1 value of the Stacking algorithm in this layer is the same as the F1 value of CatBoost in the same layer. CatBoost is also given a large weight in the weight calculation, so that the result of the Stacking algorithm is equal to the best result, ensuring that the prediction ability is not reduced.

[0057] The second layer (L2) algorithm consists of three basic algorithms, namely RF, ET, and DNN, and three integrated algorithms, namely LGBM, CatBoost, and XGBoost. As shown in Table 2, due to the acquisition of the additional features generated by the first layer of Stacking, the F1 value of the second layer algorithm is significantly improved. The worst performing model is the integrated model ET, and the best performing models are LGBM, RF, and DNN. The highest F1 value was obtained in this layer of Stacking algorithm, indicating that the prediction ability of the Stacking model is the strongest.

[0058] Table 1 F1 values ​​of various algorithms

[0059]

[0060] In the first layer of the Stacking ensemble algorithm, the CatBoost model with the highest weight is also the best performing model, while in the second layer of the Stacking ensemble algorithm, the LGBM model with the highest weight is the best performing model. For the models with reduced performance in the second layer, the method of discarding them is adopted. When predicting the test set data, the F1 value of the test set is 95.89%, reaching the highest prediction ability in the Stacking ensemble algorithm, which is significantly better than each independent machine learning algorithm (maximum 95.775%, minimum 90.411%).

[0061] like Figure 4 As shown, this application uses a double-layer Stacking model to predict the validation set. The squares represent the actual results, and the diamonds represent the model prediction results. By comparison, it can be concluded that the model prediction results are exactly the same as the actual conditions, indicating that the model is capable of predicting coal and gas outbursts very well, and does not have the overfitting phenomenon and performance degradation of the traditional multi-layer Stacking model.

[0062] This shows that the model can well predict the real results on the validation set, and the feature importance study of the model is of great help in identifying the risk factors of coal and gas outburst.

[0063] In an optional embodiment, in order to reduce the error of the prediction results, the prediction results of the test set are synthesized through weighted summation to obtain the complete new feature information finally used by the meta-model, ensuring that the dimensions of the new feature matrices of the test set and the training set are consistent, so as to retain the prediction information to the maximum extent and obtain a set of new feature matrices for use by the meta-model.

[0064] In this embodiment, in order to avoid the impact of the initial data division on the results, we need to use cross-validation to reduce contingency, that is, grid search and cross-validation are used in combination. Using f1-score as the evaluation index, grid search is used to traverse the possible values ​​of each hyperparameter in turn through a loop algorithm, and the parameters determined by the loop traversal are brought into the algorithm, and the next parameter loop traversal is continued to determine each hyperparameter one by one, and the hyperparameter combination with the best prediction result is taken as the final hyperparameter combination.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are within the scope of protection of the pending claims of the present invention.

Claims

1. A coal and gas outburst risk prediction method based on a double-layer Stacking integrated algorithm, characterized in that: The following steps are involved: Step S1, selecting multiple prediction indicators for machine learning based on data acquisition difficulty and representativeness; Step S2, selecting multiple groups of data to form a data sample library for machine learning based on the prediction indicators, and dividing the data in the data sample library into two groups: safety data and prominent data; Step S3, pre-processing the outstanding data in the data sample library to remove the abnormal data and form an original data set; Step S4, dividing the protrusion data into a training set and a validation set according to a certain ratio; Step S5, dividing the training set into a basic training set and a test set according to a certain ratio, and performing basic model training on the basic training set to obtain a new feature matrix; In step S6, the new feature matrices are respectively sent to respective meta-models for learning, and a plurality of different groups of meta-models are selected to vote and determine the final prediction result.

2. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 1 is characterized in that: In step S1, the prediction indexes include gas content, gas pressure, initial velocity of gas release, solidity coefficient of coal, and burial depth; In step S2, a data sample library is formed by crawling public data of coal mines.

3. The coal and gas outburst risk prediction method based on the double-layer Stacking integrated algorithm according to claim 1 is characterized in that: In step S3, a box plot is used to eliminate abnormal data, and a prediction algorithm for prominent data is established.

4. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 1 is characterized in that: In step S4, multiple basic models are trained by a double-layer stacking algorithm, and the original features and each basic model are spliced ​​and then trained again to interact with the original features and the basic models respectively.

5. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 4 is characterized in that: The meta-model and the basic model used in the double-layer stacking algorithm are of the same type, and the model hyper-parameters of the same basic model are the same.

6. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 5 is characterized in that: After the training set is divided into a basic training set and a test set in a certain ratio, each group of basic models learns the same original feature information. During the training process, the data of the basic training set is used for training, and the data of the test set is used for k-fold cross-validation.

7. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 6 is characterized in that: The prediction results of the test set are weighted summed to synthesize the complete new feature information finally used by the meta-model, ensuring that the dimensions of the new feature matrices of the test set and the training set are consistent, so as to retain the prediction information to the maximum extent and obtain a set of new feature matrices for use by the meta-model.

8. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 6 is characterized in that: Using f1-score as the evaluation indicator, grid search is used to traverse the possible values ​​of each hyperparameter in turn through a loop algorithm. The parameters determined by the loop traversal are brought into the algorithm, and the loop traversal of the next parameter is continued to determine each hyperparameter one by one. The hyperparameter combination with the best prediction result is taken as the final hyperparameter combination.

9. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 1 is characterized in that: The output new feature matrix is ​​sent to each meta-model for learning. The meta-model can learn more feature information and select multiple different groups of meta-models to vote to determine the final prediction result.

10. The method for predicting coal and gas outburst risk based on double-layer Stacking integrated algorithm according to claim 1 is characterized in that: The ratio of training set to validation set is 8:1; The ratio of the basic training set to the test set in the training set is 9:1.

Images