The International Journal of Coal Science & Technology is a peer-reviewed open access journal. It focuses on key topics of coal scientific research and mining development, serving as a forum for scientists to present research findings and discuss challenging issues.
Coverage includes original research articles, new developments, case studies and critical reviews in all aspects of scientific and engineering research on coal, coal utilizations and coal mining. Among the broad topics receiving attention are coal geology, geochemistry, geophysics, mineralogy, and petrology; coal mining theory, technology and engineering; coal processing, utilization and conversion; coal mining environment and reclamation and related aspects.
The International Journal of Coal Science & Technology is published with China Coal Society, who also cover the publication costs so authors do not need to pay an article-processing charge.
The journal operates a single-blind peer-review system, where the reviewers are aware of the names and affiliations of the authors, but the reviewer reports provided to authors are anonymous.
A forum for new research findings, case studies and discussion of important challenges in coal science and mining development
Offers an international perspective on coal geology, coal mining, technology and engineering, coal processing, utilization and conversion, coal mining environment and reclamation and more
Published with the China Coal Society
Research Article
Open Access
Published: 29 December 2024
0 Accesses
International Journal of Coal Science & Technology Volume 11, article number 87, (2024)
1.
Beijing Key Laboratory for Precise Mining of Intergrown Energy and Resources, China University of Mining and Technology (Beijing), Beijing, China
2.
School of Energy and Mining Engineering, China University of Mining and Technology (Beijing), Beijing, China
3.
Ministry of Emergency Management Key Laboratory for Prevention and Control of Coal Mine Disasters, Beijing, China
Coal bursting liability refers to the mechanical property of the degree and possibility of coal burst. The bursting liability is important to evaluate coal burst in mining. In this paper, the needle penetration test was carried out to determinate the coal bursting liability, and the empirical criterion of coal bursting liability was proposed. Moreover, the machine learning method was applied to coal bursting liability determination. Through analyzing the elastic strain energy release and failure time, the residual elastic strain energy release rate index KRE was proposed to evaluate the coal bursting liability. According to the relationship between needle penetration index (NPI), KRE and the critical value of KRE, the Needle Penetration Test-based Empirical Classification Criterion (NPT-ECC) was obtained. In addition, four machine learning classification models were constructed. After training and testing of the models, Needle Penetration Test-based Machine Learning Classification Model (NPT-MLCM) was proposed. The research results show that the accuracy of NPT-ECC is 6.66% higher than that of China National Standard Comprehensive Evaluation (CNSCE) according to verification of the coal fragment ejection ratio F. Gridsearch cross validation-extreme gradient boosting (GSCV-XGBoost) has the best prediction performance among all the models, and accuracy, Macro-Precision, Macro-Recall and Macro-F1-score of which were 86.67%, 88.97%, 87.50% and 87.37%. Based on this, the Needle Penetration Test-based GSCV-XGBoost (NPT-GSCV-XGBoost) was proposed. After comparative analysis and discussion, NPT-GSCV-XGBoost is superior to NPT-ECC and CNSCE in the comprehensive prediction ability.
Coal burst is a kind of dynamic disasters caused by underground coal mining. Bursting liability is a very important parameter to evaluate the prone of coal burst. Previous researches prove that bursting liability is an inherent mechanical property of coal, which relates to the energy accumulation and release ability during the loading process (Li et al. 2023; Wang et al. 2017; Wei et al. 2022; Zhu et al. 2023).
Until now, numerous criteria on calculating the bursting liability of coal have been proposed. The criteria can be divided into single index criteria and multi-index criteria. In the aspect of single index criteria, researchers put forward the determination of bursting liability from various perspective. For examples, according to the elastic energy characteristics, Lu et al. (2019) raised Effective Elastic Strain Energy Release Rate (EESERR) by analyzing the energy evolution and failure time of coal samples during loading, which can comprehensively and universally evaluate bursting liability. Zhang et al. (2020) proposed modified bursting energy index KED considering rock rebound and damage through uniaxial compression test and PFC2D numerical simulation of combined samples with different paste rock ratios. Gong et al. (2021) calculated the residual elastic energy index CEF based on the linear energy storage law, and measured the actual coal busting liability with the far field ejection mass ratio. According to the deformation characteristics, Homand et al. (1990) adopted the decreasing modulus index DMI to predict the probability and degree of dynamic failure, and verified it by experiments. Gao et al. (2020) used the ratio of elastic modulus Ee before peak strength to declining modulus Ed after peak strength to evaluate the bursting liability of coal-rock assembly. In addition to these single index criteria, many multi-index criteria of coal bursting liability comprehensive evaluation have also been researched. The China National Standard Comprehensive Evaluation (CNSCE) (GB/T 25217.2–2010 2010) utilized indexes including uniaxial compressive strength RC, bursting energy index KE, elastic strain energy index WET and dynamic failure time DT to comprehensive evaluation of coal busting liability. Guo et al. (2012) assessed the bursting liability of 11 − 2 coal seam in Zhuji Coal Mine by combining the indexes of CNSCE and the roof bending energy index. Tan et al. (2018) conducted experiments on coal and rock assembly samples, and proposed coal-rock impact energy speed index (CRIES) comprehensively evaluating bursting liability. Wang et al. (2021) proposed a prediction method for the location and risk degree of coal burst in underground mine considering nine typical evaluation indexes such as mining height, uniaxial compressive strength, elastic modulus. Obviously, coal bursting liability index has become the focus of the research in the prediction of coal burst.
Normally, coal bursting liability index is calculated based on uniaxial compression tests of the coal samples in the laboratory, which is a time-consuming process. Moreover, the preparation of coal samples is difficult sometimes. A testing method to obtain burst liability easily with portable instrument will be very useful. Many researches on mechanical property tests of rocks were reported by using portable method, such as point load test, needle penetration test, Schmidt hammer test (Rabat et al. 2020; Bilgin et al. 2016). Among these indirect test methods, needle penetration test is a very convenient and portable method for uniaxial compressive strength of rock (Ulusay et al. 2014; Ulusay and Erguler 2012), and has been extended to test various physical and mechanical parameters of rock (Aydan 2012; Aydan et al. 2014). Depth of needle penetration test is not more than 10 mm, the damage to rock would be generally negligible (Aydan et al. 2014). Hence, needle penetration test can provide fast, non-destructive and high-resolution testing in the field according to the prediction relation and the characteristics of the equipment. Zhao et al. (2023) reported the relationship between needle penetration index (NPI) and coal bursting liability index based on needle penetration test. However, the data discreteness of needle penetration test is relatively large due to the heterogeneity of coal samples. Therefore, how to effectively analyze these data and establish an accurate criterion of coal bursting liability is a challenge to be faced.
Thus, machine learning has been put forward for coal and rock bursting liability prediction, because of the advantages in processing a large number of nonlinear and discrete data. Researchers have established effective bursting liability prediction models by collecting data from experiment and cases. For examples, Wang et al. (2019) introduced Mahalanobis distance discrimination analysis (DDA) to establish model based on the database of 95 groups of coal bursting liability indexes. This model can distinguish the bursting liability of coal samples better than the fuzzy evaluation method. Guo et al. (2021) collected 344 sets of bursting liability indexes, and combined the multivariable adaptive regression spline (MARS) with deep forest algorithm to classify the rockburst intensity. Meanwhile, Shapley additive interpretation method was used for analyzing the relative importance of features. Guo et al. (2022) established database of 420 sets with tangential stress, brittleness index and elastic energy index, and built models based on multi-layer feedforward neural networks and support vector sets to predict coal burst. Zhou et al. (2022) proposed an improved connected cloud model (ICCM) to evaluate the coal bursting liability, which effectively eliminated the subjective influence and designed a graphical user interface. Some researchers also compared with the prediction effect of different machine learning model. Zhou et al. (2016) used ten kinds of supervised machine learning algorithms to build models respectively, and the advantages and disadvantages of different models were analyzed. Pu et al. (2019) summarized the effect of typical machine learning models for the prediction of coal and rock bursting liability, and thought that deep learning has advantages over machine learning under the condition of large amount of data. Xue et al. (2020) established the particle swarm optimization (PSO)-extreme learning machine (PSO-ELM) using the database of 344 cases. Moreover, the prediction effect of PSO-ELM and the support vector machine algorithm model optimized by PSO, genetic algorithm were compared. Li et al. (2022) proposed a classification model based on deep forest algorithm and Bayesian optimization, which aimed at the shortcomings of ensemble learning algorithm. In addition, compared with models based on random forest and complete random forest algorithm, the prediction performance of Bayesian optimized-deep forest model was more excellent.
Overall, needle penetration test and machine learning can achieve the fast acquisition in situ and the accurate prediction of coal bursting liability respectively. However, there are few researches to coal bursting liability determination combining needle penetration test and machine learning. There is desirable to establish this method, which will give full play to the advantages of the needle penetration test and machine learning.
The purpose of this paper is to achieve quick and accurate coal bursting liability prediction in situ. Through the needle penetration test and uniaxial compression experiment, the NPI and bursting liability indexes of coal samples were obtained. The experimental data were fitted and analyzed, and the Needle Penetration Test-based Empirical Classification Criterion (NPT-ECC) was obtained. Four machine learning classification models were established with SVM, RF, BP and XGBoost as the basic algorithms and GridSearchCV as the hyperparameter optimization strategy. A coal bursting liability database including training set (240 groups of data) and test set (30 groups of data) was established by literature collection and mechanical tests. The prediction performance and robustness of four machine learning classification models were analyzed. The Needle Penetration Test-based Machine Learning Classification Model (NPT-MLCM) can be built by the relationship between NPI and coal bursting liability indexes and the optimal model. The advantages and disadvantages of NPT-MLCM and empirical criteria (NPT-ECC and CNSCE) were discussed.
Coal samples from Talahao Coal Mine (TLH), Wanglou Coal Mine (WL), Sijiazhuang Coal Mine (SJZ), Zhaogu No.2 Coal Mine (ZG), Daling Coal Mine (DL) and Fugu-Yian Coal Mine (FG) in the north of China (Table 1). The samples were cylindrical shape with a diameter of 50 mm, a height of 100 mm. The non-parallelism of the samples end face were less than 0.1 mm and their axial deviation were less than 0.25 ° to ensure the experimental effect. In addition, the mineral composition obtained from XRD testing of coal samples is shown in Table 2.
Name | Geological condition | Type of coal | Mine |
|---|---|---|---|
TLH | 3− 1 coal seam of Jurassic Yan’ an Formation, Dongsheng coalfield, Erdos, Inner Mongolia | Long flame coal | Talahao Coal Mine |
SJZ | 15# coal seam of Taiyuan Formation, Qinshui Coalfield, Shanxi | Blind coal | Sijiazhuang Coal Mine |
WL | 3Upper coal seam of Permian Shanxi Formation, Jining coalfield, Shandong Province | Gas-Fat coal | Wanglou Coal Mine |
ZG | 21 coal seam of Permian Shanxi Formation, Jiaozuo coalfield, Henan province | Anthracite | Zhaogu No.2 Coal Mine |
DL | 22 coal seam of Jurassic Yan’ an Formation, Rujigou coal field, Ordos, Inner Mongolia | Anthracite | Daling Coal Mine |
FG | 5− 1 coal seam of Jurassic Yan’ an Formation in Permian Carboniferous coalfield, northern Shaanxi Province | Long flame coal | Fugu-Yian Coal Mine |
Name | Quartz (%) | Calcite (%) | Dolomite (%) | Aragonite (%) | Siderite (%) | Pyrite (%) | Non-crystalline (%) | Clay (%) |
|---|---|---|---|---|---|---|---|---|
TLH | 27.4 | / | / | / | / | / | 70.0 | 2.6 |
SJZ | / | 24.0 | / | / | 2.1 | 2.8 | 40.0 | 31.1 |
WL | 1.3 | 35.5 | 12.5 | / | / | 4.4 | 10.0 | 36.3 |
ZG | / | / | / | / | / | / | / | 100.0 |
DL | 1.1 | 68.9 | / | 6.6 | / | 1.1 | 20.0 | 2.3 |
FG | 13.6 | / | 8.4 | / | 22.5 | / | 30.0 | 25.5 |
The experimental procedure involves the two parts, as shown in Fig. 1. Firstly, needle penetration test was performed. Five test points were marked on the flat surface of coal samples with a spacing of at least 10 mm apart (Ulusay et al. 2014). When crack growth occurred during the needle penetration process leading to smaller value, or the test value had a large degree of dispersion, the test value should be discarded. NPI is calculated by Eq. (1).
where P is the penetration force, L is the penetration depth.
Secondly, uniaxial compression test was carried out after needle penetration test. The loading mode of uniaxial compression test was set under displacement control loading mode with the loading rate of 0.06 mm/min. F (coal fragment ejection ratio) is the ratio of the coal sample mass before loading to the coal sample ejection mass after loading, and indicates the efficiency of converting the elastic energy into the energy of ejection after coal sample failure in physical sense. According to the damage image after loading, actual coal bursting liability of F was classified (Gong et al. 2021; Zhao et al. 2023).
Mechanical property test of coal samples. a Needle penetrometer utilized in this study b Test points marked c Needle penetration test d Uniaxial compression experiment
Through uniaxial compression test, the coal bursting liability indexes including RC, KE, WET and DT according to CNSCE are acquired. Moreover, the elastic modulus E of the linear elastic stage is used to approximately replace the pre-peak unloading modulus to calculate the peak elastic strain energy index WE. The reason is the elastic strain energy accumulation stage concentrated in the pre-peak elastic stage (Lu et al. 2019; Wang 2001). Figure 2 shows the loading energy calculation diagram and the Eqs. (2)-(4) shows calculation formula of indexes, as follows:
Diagram of loading energy calculation in mechanical property test
where W1 is the total energy of pre-peak stage, and W2 is the total energy of post-peak stage.
Besides, the residual elastic strain energy release rate index KRE is proposed as shown in Eq. (5) for the calculation on the basis of the elastic strain energy released per unit time during coal sample loading (Zhao et al. 2023). Through mathematical analysis, KRE can comprehensively represent the coal bursting liability indexes stipulated by the CNSCE. It shows that KRE can effectively solve the complicated and conflicting problems in the analysis process, as shown in Eq. (6). Therefore, the NPT-ECC is deduced by the relationship between NPI and KRE.
Machine learning algorithms normally have four categories, which are unsupervised, semi-supervised, supervised, and reinforced algorithms. According to the literatures (Guo et al. 2022; Qiu and Zhou 2023), support vector machine, back propagation networks, random forest and extreme gradient boosting in supervised learning algorithm were selected to build the classification model respectively. Moreover, these algorithms can be divided into ensemble tree structure algorithms (random forest and extreme gradient boosting) and non-tree structure algorithms (support vector machine and back propagation networks). Figure 3 shows the algorithm diagram. The algorithm principle as following:
(1) Support vector machine (SVM) was originally developed as an algorithm for binary classification problems, which has been extended to regression and multi-classification research (Li et al. 2011). The basic model of SVM is a linear classifier based on data feature space, whose goal is to find a hyperplane. Transforming the problem into a quadratic programming solution problem and maximizing the distance between different classes of data points and the hyperplane, the solution will be obtained. In addition, SVM has the ability of nonlinear classification by introducing the concept of kernel function to map data to high dimensions. Objective function of SVM (Yang et al. 2013):
where x and y are the input data and label values respectively, w and b are the normal vector and intercept of the hyperplane, and λ is the regularization term constant.
(2) Random forest (RF) is an ensemble learning algorithm based on decision trees, which extends the application of classification and regression tree algorithm (Ziegler and Konig 2014). The basic principle of RF is multiple subsets constructed from the data set by sampling with put back, and a decision tree is constructed for each subset. Meanwhile, according to the Gini, the optimal feature is selected to split on the subset until the condition is met. Objective function of RF (Strobl et al. 2007):
where D is the input data set.
(3) Back propagation network (BP) is a multi-layer feedforward neural network (MLP) trained by error back propagation algorithm. The learning method of neural network is proposed according to simulating biological nervous system (Chen et al. 1998). BP can learn and store a large number of input and output mapping relationships, including input layer, hidden layer and output layer network structure. The basic principle of BP is to minimize the sum of squares error of the network for adjustment of the weight and threshold of the network by using the gradient descent method and back propagation. Objective function of BP (Liu et al. 2007):
where Er is the cumulative error of the data set, tk is the kth expected output, and yk is the kth actual output.
Diagram of machine learning algorithms
(4) Extreme Gradient Boosting (XGBoost) is an ensemble learning algorithm based on gradient boosting trees (Chen and Guestrin 2016). The principle of XGBoost is to construct a group of weak evaluators on data and analyze the prediction of the weak evaluators (Qiu and Zhou 2023). Compared with the other models, the operation efficiency of XGBoost is higher due to its parallel computation and efficient processing of sparse data. Objective function of XGBoost (Pan et al. 2022):
where the former term is the loss function, and the latter term is the regularization term, \({\hat {y}_i}\) is the sample prediction result after ith iteration.
It is necessary to adjust and optimize hyperparameters in machine learning algorithm, because building prediction model by direct selection of machine learning algorithm tends to overfitting or underfitting. Gridsearch cross validation (GridSearchCV) is a common hyperparameter tuning method in machine learning. The principle of gridsearch is to iterate through gridded hyperparameters by step size in value range is selected, then trying each hyperparameter combination to find the optimal hyperparameter combination (Zou et al. 2022). The principle of cross validation is to divide the data set into k parts in equal proportion, and a part of which is used as the validation set and the rest as the training set. After repeating experiments for k times, k hyperparameter combinations of models were obtained. Finally, the average value of evaluation score of each hyperparameter combination was calculated. The highest score hyperparameter combination was the final optimization result (Ahmad et al. 2022), as shown in Fig. 4. Therefore, GridSearchCV can effectively improve the prediction performance and generalization ability of machine learning classification model.
GridSearchCV can traverse all hyperparameter combinations to greatly guarantee the optimal hyperparameter combinations, even though this will result in slower training of prediction model. In order to ensure the efficiency of obtaining the optimal hyperparameter combinations, the hyperparameters with greater influence in the structure of machine learning algorithm and their value ranges are selected before training.
Diagram of cross validation method
A dataset of 240 groups was established as a training set by utilizing the experimental data in this study and collecting the literature data of coal bursting liability (Gong et al. 2021; Lu et al. 2019; Ju et al. 2021; Zhou et al. 2021; Feng et al. 2022; Yang et al. 2021; Zhang et al. 2019; Mo et al. 2020). In order to ensure the integrity and unity of the data, the characteristics of training set of data were composed of RC, KE, WET and DT as specified in the CNSCE, and the label is the coal bursting liability.
In term of the relationship between NPI and coal bursting liability indexes obtained by experiment and the actual coal bursting liability of samples, a test set was established. The test set can simulate engineering examples to verify the prediction performance of the machine learning classification models in this study.
The machine learning classification model training and testing in this study was compiled by Jupyter Notebook software in Anaconda3 configuration environment via python version 3.7.6. The hardware information of utilized computer was as follows: CPU-i5 11400 F 2.6 GHz, RAM-16GB, GPU-RTX 3060 12GB. Figure 5 shows the implementation procedure as following:
Step 1: Importing coal bursting liability database training set and identifying machine learning algorithms and optimize hyperparameter method.
Step 2: Data preprocessing. The order of the row index of the training set is scrambled, and the value of the training set is normalized. The formula is as follows:
The iterative algorithm of SVM and BP is gradient descent. In order to improve the prediction accuracy and convergence speed, it is necessary to normalize the training set. However, the training sets of the models based on ensemble tree structure algorithms were not normalized data processing in this study. The reason is the optimization of ensemble tree structure algorithm is completed by finding the optimal split point of features, which is insensitive to the size of feature values. On the contrary, the accuracy of model based on the ensemble tree structure algorithms can be probably reduced because of missing abnormal and extreme eigenvalues.
Step 3: Training machine learning classification models. The machine learning algorithms of scikit-learn library were used to build the models. GridserchCV is used to optimize and adjust the hyperparameters with great influence. The 5 folds cross validation was adopted in GridserchCV with accuracy as evaluation index. The hyperparameters and value range of the adjustment were determined according to the user manual and experience. The hyperparameters and value ranges of each machine learning algorithm are shown in Table 3.
Diagram of machine learning classification model training and testing process
Step 4: Testing machine learning classification models. The test set was imported into the trained Gridsearch cross validation-support vector machine (GSCV-SVM), Gridsearch cross validation-random forest (GSCV-RF), Gridsearch cross validation-back propagation networks (GSCV-BP) and Gridsearch cross validation-extreme gradient boosting (GSCV-XGBoost) respectively. The prediction performance of the above models was analyzed by evaluation indexes.
Algorithm | Hyperparameter | Value range |
|---|---|---|
SVM | Kernel | Linear, poly, rbf, sigmoid |
C | 1, 10, 100, 1000 | |
Gamma | 0.001, 0.01, 0.1,1 | |
Degree | 1, 2, 3, 4 | |
RF | Max_features | 1 ~ 10 |
Min_samples_leaf | 1 ~ 10 | |
BP | Solver | Lbfgs, adam, sgd |
Activation | Identity, logistic, tanh, relu | |
Alpha | 0.0001, 0.001, 0.01, 0.1 | |
Hidden_layer_sizes | 10, 50, 100, 200 | |
Learning_rate_int | 0.001, 0.01, 0.1, 0.3 | |
Momentum | 0.3 ~ 1 | |
XGBoost | Max_depth | 3 ~ 10 |
Learning_rate | 0.01 ~ 0.10 | |
Min_child_weight | 1 ~ 10 |
According to experiment data and calculation formula Eqs. (1)-(5), NPI and the coal bursting liability indexes of coal samples were obtained, as shown in Table 4. The KRE numerical calculation of TLH-3 and SJZ-1 are negative. The reason is that the Eq. (4) regards the coal sample as an ideal elastic body, leading to the smaller value of the elastic strain energy of the coal sample with weak elastic energy storage capacity (Zhao et al. 2023). Therefore, the KRE of TLH-3 and SJZ-1 should be greater than and close to 0 numerically, which should be judged as being at no bursting liability.
Number | NPI (N/mm) | RC (MPa) | KE | WET | DT (ms) | KRE (kJ/m3 s) | F | NPT-ECC | CNSCE | Actual coal bursting liability |
|---|---|---|---|---|---|---|---|---|---|---|
TLH-1 | 5.40 | 0.98 | 4.06 | 1.31 | 6229 | 0.42 | 0.02 | N | N | N |
TLH-2 | 9.00 | 3.85 | 2.00 | 2.69 | 4978 | 1.45 | 0.02 | N | N | N |
TLH-3 | 9.20 | 1.24 | 0.84 | 5.09 | 5356 | -0.43 | 0.01 | N | N | N |
TLH-4 | 10.20 | 2.07 | 1.68 | 2.37 | 4150 | 0.35 | 0.03 | N | N | N |
TLH-5 | 11.20 | 1.85 | 5.30 | 1.37 | 4850 | 1.03 | 0.01 | N | N | N |
SJZ-1 | 28.16 | 2.28 | 0.75 | 3.22 | 900 | -5.01 | 0.02 | W | N | N |
SJZ-2 | 41.34 | 2.93 | 31.19 | 0.99 | 517 | 16.52 | 0.01 | W | N | N |
SJZ-3 | 63.33 | 8.65 | 2.12 | 2.19 | 268 | 26.33 | 0.04 | W | W | N |
WL-1 | 65.79 | 12.15 | 3.94 | 4.22 | 164 | 113.94 | 0.09 | W | W | W |
WL-2 | 66.67 | 12.35 | 11.12 | 4.78 | 399 | 68.97 | 0.12 | W | W | W |
WL-3 | 67.57 | 9.70 | 1.22 | 7.66 | 400 | 4.36 | 0.09 | W | W | W |
WL-4 | 61.73 | 8.67 | 2.14 | 1.08 | 927 | 1.82 | 0.10 | W | N | W |
ZG-1 | 51.55 | 10.51 | 1.44 | 4.88 | 641 | 8.37 | 0.14 | W | W | W |
ZG-2 | 84.75 | 14.97 | 3.81 | 6.47 | 199 | 155.73 | 0.11 | S | W | W |
ZG-3 | 69.44 | 11.28 | 2.24 | 1.05 | 1562 | 2.92 | 0.17 | W | N | W |
ZG-4 | 53.19 | 11.31 | 4.40 | 5.16 | 200 | 127.34 | 0.10 | W | W | W |
DL-1 | 81.30 | 18.44 | 20.63 | 2.89 | 53 | 1118.85 | 0.16 | W | S | W |
DL-2 | 84.75 | 21.81 | 4.71 | 8.54 | 73 | 779.14 | 0.13 | S | S | W |
DL-3 | 71.94 | 18.67 | 28.45 | 6.48 | 200 | 320.33 | 0.13 | W | S | W |
DL-4 | 73.53 | 14.06 | 20.75 | 2.99 | 150 | 231.80 | 0.06 | W | S | W |
FG-1 | 101.01 | 45.40 | 1441 | 26.28 | 40 | 11097.39 | 1.00 | S | S | S |
FG-2 | 94.34 | 25.58 | 1289.62 | 8.93 | 24 | 6276.61 | 0.32 | S | S | S |
FG-3 | 94.34 | 30.37 | 300.37 | 12.41 | 45 | 4862.53 | 0.76 | S | S | S |
FG-4 | 96.15 | 36.21 | 5.65 | 15.56 | 33 | 8354.85 | 0.98 | S | S | S |
FG-5 | 97.09 | 40.91 | 105.77 | 17.09 | 53 | 7373.02 | 0.98 | S | S | S |
FG-6 | 86.21 | 28.10 | 100.00 | 3.44 | 59 | 2631.53 | 0.47 | S | S | S |
FG-7 | 90.91 | 30.03 | 232.63 | 9.90 | 16 | 14457.50 | 0.99 | S | S | S |
FG-8 | 101.01 | 50.67 | 1962.53 | 22.34 | 44 | 14507.50 | 1.00 | S | S | S |
FG-9 | 104.17 | 52.82 | 2098.60 | 26.82 | 58 | 12202.76 | 1.00 | S | S | S |
FG-10 | 94.34 | 38.18 | 208.46 | 4.85 | 28 | 12150.36 | 0.99 | S | S | S |
Through data fitting and regression analysis, the relationship between NPI and coal bursting liability indexes were acquired, as shown in Fig. 6. It can be found that NPI increases with RC, KE, WET, KRE by power function, and decreases with DT by exponential function. The critical value of KRE classification criterion is calculated by the critical value of RC, KE, WET, DT, WE and Eq. (6). Besides, NPT-ECC is obtained by combining the NPI-KRE relationship with KRE classification criterion (Zhao et al. 2023), as shown in Table 5. In order to measure the actual coal bursting liability of the sample, the coal bursting liability classification of F is obtained according to the failure characteristic and F distribution. When F < 0.05, coal sample can be classified as no bursting liability; When 0.05 ≤ F < 0.30, coal sample can be classified as weak bursting liability; When F ≥ 0.3, coal sample can be classified as strong bursting liability. Figure 7 shows the coal bursting liability classification of F.
The results of NPT-ECC and CNSCE were compared and analyzed by the coal bursting liability classification of F. The accuracy of NPT-ECC (83.33%) is higher than that of CNSCE (76.67%). During the evaluation, the weights of RC, KE, WET and DT are 0.3, 0.2, 0.2 and 0.3 in CNSCE respectively, which was difficult to evaluate for some samples (such as WL-4). Therefore, the coal bursting liability prediction of NPT-ECC is more convenient and accurate than CNSCE.
Relationship of NPI and coal bursting liability indexes
Index | No bursting liability | Weak bursting liability | Strong bursting liability | Reference source |
|---|---|---|---|---|
RC (MPa) | < 7 | 7 ≤ RC<14 | ≥ 14 | GB/T 25217.2–2010 2010 |
KE | < 1.5 | 1.5 ≤ KE<5 | ≥ 5 | |
WET | < 2 | 2 ≤ WET<5 | ≥ 5 | |
DT (ms) | > 500 | 50 < DT≤500 | ≤ 50 | |
WE (kJ/m3) | < 50 | 50 ≤ WE<200 | ≥ 200 | Wang (2001) |
KRE (kJ/m3 s) | < 1 | 1 ≤ KRE<3000 | ≥ 3000 | Zhao et al. (2023) |
NPI (N/mm) | < 27 | 27 ≤ NPI < 83 | ≥ 83 |
F distribution, classification and failure characteristic of the coal samples
On 5 folds cross validation random equiproportional data sets during model training process, the average validation accuracy of GSCV-SVM, GSCV-RF, GSCV-BP and GSCV-XGBoost is 83.75%, 87.91%, 85% and 87.5% respectively. The average validation accuracy of the four machine learning classification models is above 80%, which reflects that all the models have good training effect and generalization ability. In addition, the average validation accuracy of the ensemble tree structure models (GSCV-RF and GSCV-XGBoost) are higher than that of the non-tree structure models (GSCV-SVM and GSCV-BP). It is evident that the ensemble tree structure models have better coal bursting liability prediction potential than non-tree structure models.
In order to test the prediction performance of four machine learning classification models, a test set is established according to the relationship between NPI and coal bursting liability index. Meanwhile, the coal bursting liability of the test set should be consistent with the actual coal bursting liability in Table 4, as shown in Table 6. The evaluation of coal bursting liability is calibrated digitally, and 0, 1 and 2 represent no bursting liability, weak bursting liability and strong bursting liability respectively.
Number | NPI (N/mm) | RC (MPa) | KE | WET | DT (ms) | Coal bursting liability |
|---|---|---|---|---|---|---|
TLH-1 | 5.40 | 1.39 × 10− 3 | 6.17 × 10− 20 | 6.5 × 10− 9 | 6217.21 | 0 |
TLH-2 | 9.00 | 8.49 × 10− 3 | 4.86 × 10− 16 | 2.97 × 10− 7 | 5038.89 | 0 |
TLH-3 | 9.20 | 9.18 × 10− 3 | 7.15 × 10− 16 | 3.51 × 10− 7 | 4980.41 | 0 |
TLH-4 | 10.20 | 1.32 × 10− 2 | 4.38 × 10− 15 | 7.59 × 10− 7 | 4698.02 | 0 |
TLH-5 | 11.20 | 1.84 × 10− 2 | 2.26 × 10− 14 | 1.53 × 10− 6 | 4431.64 | 0 |
SJZ-1 | 28.16 | 0.48 | 2.44 × 10− 7 | 1.52 × 10− 3 | 1646.74 | 0 |
SJZ-2 | 41.34 | 1.88 | 2.07 × 10− 4 | 2.68 × 10− 2 | 762.98 | 0 |
SJZ-3 | 63.33 | 8.54 | 0.37 | 0.65 | 211.33 | 0 |
WL-1 | 65.79 | 9.77 | 0.72 | 0.87 | 183.11 | 1 |
WL-2 | 66.67 | 10.24 | 0.91 | 0.96 | 173.97 | 1 |
WL-3 | 67.57 | 10.74 | 1.15 | 1.06 | 165.06 | 1 |
WL-4 | 61.73 | 7.79 | 0.24 | 0.54 | 232.09 | 1 |
ZG-1 | 51.55 | 4.12 | 0.01 | 0.14 | 420.51 | 1 |
ZG-2 | 84.75 | 23.96 | 61.68 | 5.78 | 60.56 | 1 |
ZG-3 | 69.44 | 11.83 | 1.87 | 1.30 | 147.93 | 1 |
ZG-4 | 53.19 | 4.60 | 0.02 | 0.18 | 382.01 | 1 |
DL-1 | 81.30 | 20.68 | 29.76 | 4.24 | 74.05 | 1 |
DL-2 | 84.75 | 23.96 | 61.68 | 5.78 | 60.56 | 1 |
DL-3 | 71.94 | 13.41 | 3.47 | 1.70 | 127.86 | 1 |
DL-4 | 73.53 | 14.49 | 5.10 | 2.00 | 116.55 | 1 |
FG-1 | 101.01 | 44.62 | 1346.61 | 21.52 | 23.43 | 2 |
FG-2 | 94.34 | 35.03 | 405.66 | 12.90 | 34.59 | 2 |
FG-3 | 94.34 | 35.03 | 405.66 | 12.90 | 34.59 | 2 |
FG-4 | 96.15 | 37.48 | 566.82 | 14.88 | 31.12 | 2 |
FG-5 | 97.09 | 38.78 | 671.64 | 16.00 | 29.46 | 2 |
FG-6 | 86.21 | 25.45 | 83.28 | 6.57 | 55.61 | 2 |
FG-7 | 90.91 | 30.72 | 211.66 | 9.78 | 42.26 | 2 |
FG-8 | 101.01 | 44.62 | 1346.61 | 21.52 | 23.43 | 2 |
FG-9 | 104.17 | 49.76 | 2311.76 | 27.09 | 19.49 | 2 |
FG-10 | 94.34 | 35.03 | 405.66 | 12.90 | 34.59 | 2 |
The test set is preprocessed (consistent with the training set) and imported into the machine learning classification models. In term of the results, the predicted data samples are divided into true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The corresponding distribution is shown in Table 7, where T indicates the category predicted correctly by the machine learning classification model, and F indicates the other categories predicted. The confusion matrixes of the four machine learning classification models were acquired, as shown in Fig. 8. The confusion matrix of GSCV-SVM was the same as GSCV-BP, indicating that their coal bursting liability prediction ability are similar.
Accuracy, precision, recall and F1-score are used as evaluation indexes for the test set of this study. Accuracy is global evaluation index, while precision, recall and F1-score are within-class evaluation indexes (Yin et al. 2021). Accuracy reflects the degree of correct prediction of the machine learning classification models for all samples. Precision represents the probability of actually being positive samples in all predicted positive samples. Recall demonstrates the probability of being predicted as positive samples in actual positive samples. Therefore, precision and recall reflect the prediction accuracy and recognition ability of samples respectively. F1-score is an evaluation index that fully considers precision and recall of samples. The calculation formula for the above evaluation indexes are as follows (Guo et al. 2022):
Actual | Predict | |
|---|---|---|
T | F | |
T | TP | FN |
F | FP | TN |
Confusion matrix of the machine learning classification models
In light of the confusion matrix of machine learning classification model, the evaluation indexes of coal samples with different coal bursting liabilities in test set was calculated. Macro-averaging is a means of averaging within-class evaluation indexes, which was used to evaluate the global performance of each machine learning classification model in this study, as shown in Table 8.
Classification model | Evaluation index | No bursting liability (%) | Weak bursting liability (%) | Strong bursting liability (%) | Macro-averaging (%) |
|---|---|---|---|---|---|
GSCV-SVM | Accuracy | 80.00 | |||
Precision | 77.78 | 87.50 | 76.92 | 80.73 | |
Recall | 87.50 | 58.33 | 100.00 | 81.94 | |
F1-score | 82.35 | 70.00 | 86.95 | 79.77 | |
GSCV-RF | Accuracy | 83.33 | |||
Precision | 87.50 | 88.89 | 76.92 | 84.44 | |
Recall | 87.50 | 66.67 | 100.00 | 84.72 | |
F1-score | 87.50 | 76.19 | 86.95 | 83.55 | |
GSCV-BP | Accuracy | 80.00 | |||
Precision | 77.78 | 87.50 | 76.92 | 80.73 | |
Recall | 87.50 | 58.33 | 100.00 | 81.94 | |
F1-score | 82.35 | 70.00 | 86.95 | 79.77 | |
GSCV-XGBoost | Accuracy | 86.67 | |||
Precision | 100.00 | 90.00 | 76.92 | 88.97 | |
Recall | 87.50 | 75.00 | 100.00 | 87.50 | |
F1-score | 93.33 | 81.82 | 86.95 | 87.37 | |
Compared with the overall prediction results of different machine learning classification models, obviously the evaluation indexes of the ensemble tree structure models are higher than those of the non-tree structure models, as shown in Fig. 9. Macro-Precision, Macro-Recall and Macro-F1-score are the Macro-averaging of precision, recall and F1-score, respectively. The evaluation indexes of GSCV-XGBoost are comprehensively higher than those of other models. The accuracy and Macro-F1-score of GSCV-XGBoost are 3.34% and 3.82% higher than that of GSCV-RF respectively. The above result indicates that GSCV-XGBoost has the best comprehensive prediction and classification ability of coal bursting liability among all the models. Compared with the results of test set and cross validation, the accuracy of all the machine learning classification models shows a decreasing trend. The accuracy of GSCV-XGBoost decreased by 0.83% after migrating from cross validation to test set, showing the good generalization performance of GSCV-XGBoost.
Analysis of overall prediction effect of the machine learning classification models
The prediction results of the machine learning classification models for coal samples with different coal bursting liability are shown in Fig. 10. All the machine learning classification models have good effect on prediction of coal samples with no bursting liability and strong bursting liability, and the F1-score of those are greater than 80%. However, in the prediction of weak bursting liability, the recall of GSCV-BP, GSCV-SVM is 58.33%, and the recall of GSCV-RF is 66.67%. The F1-score of GSCV-SVM, GSCV-BP, GSCV-RF are less than 80%. Markedly GSCV-SVM, GSCV-BP, GSCV-RF are weaker in learning the characteristics of weak bursting liability. In contrast, the F1-score of GSCV-XGBoost for the three types of coal bursting liability are above 80%, which has the best prediction effect among all the models.
Prediction effect analysis of coal samples with different bursting liability based on the machine learning classification models
In summary, four machine learning classification models have excellent prediction ability on the whole after testing. Nevertheless, GSCV-SVM, GSCV-BP and GSCV-RF are slightly inadequate in the prediction and classification of coal samples with weak bursting liability. In terms of testing results, the recall of all the models for the strong bursting liability reaches 100%, showing a strong ability to identify coal bursting risk. It is apparent that all the models meet the requirements of identifying coal mass with strong bursting liability in coal burst prevention. Furthermore, the evaluation indexes of GSCV-XGBoost are the highest, reflecting it has the strongest prediction ability of samples with different bursting liability and the best model migration generalization ability among all the models. Therefore, combining GSCV-XGBoost with needle penetration test (the relationship between NPI and RC, KE, WET, DT), it is proposed that NPT-GSCV-XGBoost is a coal bursting liability prediction method fully utilizes the advantages of both.
The data processing method in Sect. 4.1 is used to analyze the prediction results of coal bursting liability based on NPT-ECC (Zhao et al. 2023) and CNSCE. The confusion matrix is shown in Fig. 11. Moreover, the prediction evaluation indexes of NPT-GSCV-XGBoost are summarized in Table 9, which will be used to discuss the advantages and disadvantages of the prediction methods.
Confusion matrix of empirical criteria
Method | Evaluation index | No bursting liability (%) | Weak bursting liability (%) | Strong bursting liability (%) | Macro-averaging (%) |
|---|---|---|---|---|---|
NPT-ECC | Accuracy | 83.33 | |||
Precision | 62.50 | 100.00 | 83.33 | 81.94 | |
Recall | 100.00 | 66.67 | 100.00 | 88.89 | |
F1-score | 76.92 | 80.00 | 90.91 | 82.61 | |
CNSCE | Accuracy | 76.67 | |||
Precision | 77.78 | 85.71 | 71.43 | 78.31 | |
Recall | 87.50 | 50.00 | 100.00 | 79.17 | |
F1-score | 82.35 | 63.16 | 83.33 | 76.28 | |
NPT-GSCV-XGBoost | Accuracy | 86.67 | |||
Precision | 100.00 | 90.00 | 76.92 | 88.97 | |
Recall | 87.50 | 75.00 | 100.00 | 87.50 | |
F1-score | 93.33 | 81.82 | 86.95 | 87.37 | |
The overall prediction results of empirical criteria (NPT-ECC and CNSCE) and NPT-GSCV-XGBoost were analyzed. As shown in Fig. 12, NPT-ECC is slightly stronger than NPT-GSCV-XGBoost for coal bursting liability recognition ability, and the difference in Macro-Recall was 1.39%. However, NPT-GSCV-XGBoost is superior to empirical criteria in accuracy, Marco-Precision and Macro-F1-score (Accuracy, Marco-Precision and Macro-F1-score of NPT-GSCV-XGBoost lead NPT-ECC by 3.34%, 7.03%, 4.76% and CNSCE by 10.00%, 10.66%, 11.09% respectively). This result reveals that NPT-GSCV-XGBoost has better comprehensive prediction performance of coal bursting liability than empirical criteria.
Analysis of overall prediction effect of NPT-ECC, CNSCE, NPT-GSCV-XGBoost
As shown in Fig. 13, in the prediction of no bursting liability, F1-score of NPT-GSCV-XGBoost is 16.41% and 10.98% higher than NPT-ECC and CNSCE, which shows obvious prediction advantages for no bursting liability. In the prediction of weak bursting liability, NPT-GSCV-XGBoost is slightly better than NPT-ECC (F1-score difference between the two is 1.82%), and significantly higher than CNSCE (F1-score difference between the two is 18.66%). The recall of the two empirical criteria is lower than 70%, resulting in poorer identification ability of weak bursting liability than NPT-GSCV-XGBoost. In the prediction of strong bursting liability, the recall of NPT-GSCV-XGBoost and the empirical criteria is 100%, indicating that the three methods can identify all the coal samples with strong bursting liability in this experiment. Moreover, the precision of NPT-ECC is higher than that of NPT-GSCV-XGBoost (the precision difference between the two is 6.41%) and CNSCE (the precision difference between the two is 11.90%). Therefore, the comprehensive prediction ability of NPT-ECC is obviously stronger than that of NPT-GSCV-XGBoost and CNSCE (F1-score difference is 3.96% and 7.58% respectively) for strong bursting liability. The results show that the precision and reliability of NPT-ECC are slightly higher than NPT-GSCV-XGBoost for predicting the burst risk of coal mass.
To sum up, compared with NPT-ECC, NPT-GSCV-XGBoost has a slight disadvantage in the prediction of strong bursting liability, the comprehensive prediction performance of which still is significantly ahead of the empirical criteria. Meanwhile, NPT-GSCV-XGBoost can improve the prediction effect by increasing the database and optimizing the algorithm hyperparameters, which shows greater development potential. Obviously, NPT-GSCV-XGBoost is a more objective, effective prediction method for coal bursting liability than empirical criteria. Therefore, compared with traditional data statistics and fitting methods, machine learning has greater advantages and development potential in processing data with multiple indexes. Furthermore, machine learning is an important means of analysis for future disaster prediction.
Prediction effect analysis of coal samples with different coal bursting liability based on NPT-ECC, CNSCE, NPT-GSCV-XGBoost
In this study, prediction methods of empirical criterion and machine learning by needle penetration test were proposed for coal bursting liability. The NPI and bursting liability indexes of coal samples were obtained after experiments. According to statistical analysis of experimental data, the prediction method of combining NPT and empirical criterion is established. Meanwhile, four machine learning algorithms were selected to build machine learning classification models by collecting literature and summarizing experimental data. The prediction performance of the machine learning classification models was compared, and prediction method of combining needle penetration test and machine learning classification model was proposed. In addition, the above prediction methods were discussed. The main conclusions are as follows:
(1) A comprehensive coal bursting liability index KRE was proposed by analyzing the releasing characteristic of elastic strain energy and the mathematical expression of RC, KE, WET, DT, WE. Through the relation between NPI-KRE and the critical value of coal bursting liability indexes (RC, KE, WET, DT, WE), the NPT-ECC was determined. When NPI < 27 N/mm, coal can be classified as no bursting liability; When 27 ≤ NPI < 83 N/mm, coal can be classified as weak bursting liability; When NPI ≥ 83 N/mm, coal can be classified as strong bursting liability. The accuracy of NPT-ECC (83.33%) is higher than CNSCE’s (76.67%), which verified by coal fragment ejection ratio F of coal samples.
(2) After the migration of the machine learning classification models to the test set, the accuracy of GSCV-XGBoost reached 86.67%. Compared with the average validation accuracy of GSCV-XGBoost, its accuracy decreased by 0.83%. The Macro-Precision, Macro-Recall and Macro-F1-score of GSCV-XGBoost reached 88.97%, 87.50% and 87.37% respectively, which shows the best prediction performance and robustness among all the models. Therefore, NPT-GSCV-XGBoost is proposed for coal bursting liability prediction.
(3) Compared with the NPT-ECC, the NPT-GSCV-XGBoost is slightly weaker for the prediction effect of strong bursting liability. However, NPT-GSCV-XGBoost is better than the empirical criteria in the comprehensive prediction performance (the accuracy and Macro-F1-score is 86.67% and 87.37%). Meanwhile, machine learning classification model can further improve the prediction performance by expanding the database and improving the algorithm. Therefore, NPT-MLCM is an important direction for the prediction of coal bursting liability in the future.
| [1] | Ahmad GN, Fatima H, Ullah S, Saidi AS, Imdadullah (2022) Efficient medical diagnosis of human heart diseases using machine learning techniques with and without GridSearchCV. IEEE ACCESS 10:80151–80173 |
| [2] | Aydan Ö (2012) The inference of physico-mechanical properties of soft rocks and the evaluation of the effect of water content and weathering on their mechanical properties from needle penetration tests. J Agricultural Chem Soc Japan 29(5):707–711 |
| [3] | Aydan Ö, Sato A, Yagi M (2014) The inference of geo-mechanical properties of soft rocks and their degradation from needle penetration tests. Rock Mech Rock Eng 47(5):1867–1890 |
| [4] | Bilgin N, Copur H, Balci C (2016) Use of Schmidt Hammer with special reference to strength reduction factor related to cleat presence in a coal mine. Int J Rock Mech Min Sci 84:25–33 |
| [5] | Chen TQ, Guestrin C (2016) XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 13–17 |
| [6] | Chen JY, Chen ZX, Lu YY, Xu SC (1998) Circle BP algorithm for MLP neural network. Semicond Photonics Technol 4(3):46–49 |
| [7] | Feng ZJ, Su FQ, Su CD, Guo JQ (2022) Experimental study on impact proneness and acoustic emission characteristics of coal samples from Zhaogu 21 mine. J Saf Sci Technol 18(3):131–136 |
| [8] | Gao MT, Song ZQ, Duan HQ, Xin HQ, Tang JQ (2020) Mechanical properties and control rockburst mechanism of coal and rock mass with bursting liability in deep mining. Shock and Vibration 8833863 |
| [9] | GB/T 25217.2–2010 (2010) Methods for test, monitoring and prevention of rock burst-part 2: classification and laboratory test method on bursting liability of coal. Standards Press of China, Beijing |
| [10] | Gong FQ, Wang YL, Wang ZG, Pan JF, Luo S (2021) A new criterion of coal burst proneness based on the residual elastic energy index. Int J Min Sci Technol 31(4):553–563 |
| [11] | Guo DM, Xue HJ, Li LJ, Xue JL, Li GH (2012) Research on bursting liability and its preventive measures of -906m deep coal and rock in Zhuji coal mine. In: Proceedings of 2nd International Conference on Civil Engineering, Architecture and Building Materials 170–173:428–433 |
| [12] | Guo DP, Chen HM, Tang LB, Chen ZX, Samui P (2021) Assessment of rockburst risk using multivariate adaptive regression splines and deep forest model. Acta Geotech 17(4):1183–1205 |
| [13] | Guo J, Guo JW, Zhang QL, Huang MJ (2022) Research on rockburst classification prediction based on BP-SVM model. IEEE ACCESS 10:50427–50447 |
| [14] | Homand F, Piguet JP, Revalor R (1990) Dynamic phenomena in mines and characteristics of rocks. In: Proceedings of the 2nd Rockbursts and seismicity in mines 139–142 |
| [15] | Ju WJ, Lu ZG, Gao FQ, Zhao YX, Li WZ, Sun ZY, Hao XJ (2021) Research progress and comprehensive quantitative evaluation index of coal rock bursting liability. Chin J Rock Mechan Eng 40(9):1839–1856 |
| [16] | Li GQ, Wen CY, Huang GB, Chen Y (2011) Error tolerance based support vector machine for regression. Neurocomputing 74:771–782 |
| [17] | Li YF, Wang C, Liu Y (2023) Classification of coal bursting liability based on support vector machine and imbalanced sample set. Minerals 13(1):15 |
| [18] | Li DY, Liu ZD, Armaghani DJ, Xiao P, Zhou J (2022) Novel ensemble tree solution for rockburst prediction using deep forest. Mathematics 10(5):787 |
| [19] | Liu LH, Chen J, Xu LX (2007) Realization and application research of BP neural network based on MATLAB. In: Proceedings of International Symposium on Distributed Computing and Applications to Business, Engineering and Science 626–628 |
| [20] | Lu ZG, Ju WJ, Gao FQ, Feng YL, Sun ZY, Wang H, Yi K (2019) A new bursting liability evaluation index for coal the effective elastic strain energy release rate. Energies 12(19):3734 |
| [21] | Mo YL, Li HY, Deng ZG, Qi QX, Li HT, Zhang GH (2020) Initial damage effect of energy response for coal with different bump proneness. J Saf Sci Technol 37(6):1205–1212 |
| [22] | Pan SW, Wang ZY, Zhang Y, Cai WB (2022) Lithology identification based on LSTM neural networks completing log and hybrid optimized XGBoost. J China Univ Petroleum (Edition Nat Science) 46(3):62–71 |
| [23] | Pu YY, Apel DB, Liu V, Mitri H (2019) Machine learning methods for rockburst prediction-state-of-the-art review. Int J Min Sci Technol 29(4):565–570 |
| [24] | Qiu YG, Zhou J (2023) Short-term rockburst prediction in underground project: insights from an explainable and interpretable ensemble learning model. Acta Geotechnica 18(12):6655–6685 |
| [25] | Rabat Á, Cano M, Tomás R, Tamayo AE, Alejano LR (2020) Evaluation of strength and deformability of soft sedimentary rocks in dry and saturated conditions through needle penetration and point load tests: a comparative study. Rock Mech Rock Eng 53(6):2707–2726 |
| [26] | Strobl C, Boulesteix AL, Augustin T (2007) Unbiased split selection for classification trees based on the Gini Index. Comput Stat Data Anal 52(1):483–501 |
| [27] | Tan YL, Liu XS, Ning JG (2018) New approaches to testing and evaluating the rockburst risk of coal seam with hard roof and/or floor in coal mines. Geomech Eng 14(4):367–376 |
| [28] | Ulusay R, Erguler ZA (2012) Needle penetration test: evaluation of its performance and possible uses in predicting strength of weak and soft rocks. Eng Geol 149:47–56 |
| [29] | Ulusay R, Aydan Ö, Erguler ZA (2014) ISRM suggested method for the needle penetration test. Rock Mech Rock Eng 47(3):1073–1085 |
| [30] | Wang JA, Park HD (2001) Comprehensive prediction of rockburst based on analysis of strain energy in rocks. Tunn Undergr Space Technol 16(1):49–57 |
| [31] | Wang HW, Jiang YD, Xue S, Pang XF, Lin ZN, Deng DX (2017) Investigation of intrinsic and external factors contributing to the occurrence of coal bumps in the mining area of western Beijing, China. Rock Mech Rock Eng 50(4):1033–1047 |
| [32] | Wang C, Song DZ, Zhang CL, Liu L, Zhou ZH, Huang XC (2019) Research on the classification model of coal’s bursting liability based on database with large samples. Arab J Geosci 12(13):411 |
| [33] | Wang J, Apel DB, Dyczko A, Walentek A, Prusek S, Xu HW, Wei C (2021) Investigation of the rockburst mechanism of driving roadways in close-distance coal seam mining using numerical modeling method. Min Metall Explor 38(5):1899–1921 |
| [34] | Wei C, Zhang C, Canbulat I, Song Z, Dai L (2022) A review of investigations on ground support requirements in coal burst-prone mines. Int J Coal Sci Technol 9(1):13. https://doi.org/10.1007/s40789-022-00485-1 |
| [35] | Xue YG, Bai CH, Qiu DH, Kong FM, Li ZQ (2020) Predicting rockburst with database using particle swarm optimization and extreme learning machine. Tunn Undergr Space Technol 98:103287 |
| [36] | Yang XW, Yu QZ, He LF, Guo TJ (2013) The one-against-all partition based binary tree support vector machine algorithms for multi-class classification. Neurocomputing 113:1–7 |
| [37] | Yang L, Wang XQ, Li JZ (2021) Energy evolution and damage characteristics of coal with different bursting liability under uniaxial compression. Coal Sci Technol 49(6):111–118 |
| [38] | Yin X, Liu QS, Huang X, Pan YC (2021) Real-time prediction of rockburst intensity using an integrated CNN-Adam-BO algorithm based on microseismic data and its engineering application. Tunn Undergr Space Technol 117:104133 |
| [39] | Zhang TJ, Bao RY, Wang Q, Yu Y (2019) A research on the index system and its experimental verification of the coal seam impact tendency. J Hunan Univ Sci Technol (Natural Sci Edition) 34(3):1–10 |
| [40] | Zhang DX, Guo WY, Zang CW, Gong XF, Li ZH, Qiu Y, Chen WG (2020) A new burst evaluation index of coal-rock combination specimen considering rebound and damage effects of rock. Geomatics Nat Hazards Risk 11(1):984–999 |
| [41] | Zhao YX, Xie RH, Gao YR (2023) Coal bursting liability evaluation by needle penetration test. J China Coal Soc 48(5):1932–1942 |
| [42] | Zhou J, Li XB, Mitri HS (2016) Classification of rockburst in underground projects: comparison of ten supervised learning methods. J Comput Civil Eng 30(5):4016003 |
| [43] | Zhou J, Chen C, Wang MZ, Khandelwal M (2021) Proposing a novel comprehensive evaluation model for the coal burst liability in underground coal mines considering uncertainty factors. Int J Min Sci Technol 31(5):799–812 |
| [44] | Zhou J, Chen C, Wei C, Du K (2022) An improved connection cloud model of an updated database: a multicriteria uncertainty model for coal burst liability evaluation. Nat Resour Res 31(3):1687–1704 |
| [45] | Zhu ZJ, Guan SS, Tang GS, Chen K, Lv F (2023) Bursting liability criteria of coal mass based on energy storage and release characteristics. ACS Omega 8(32):29725–29734 |
| [46] | Ziegler A, Konig IR (2014) Mining data with random forests: current options for real-world applications. Wiley Interdisciplinary Reviews-Data Min Knowl Discovery 4(1):55–63 |
| [47] | Zou M, Jiang WG, Qin QH, Liu YC, Li ML (2022) Optimized XGBoost Model with small dataset for predicting relative density of Ti-6Al-4V parts manufactured by selective laser melting. Materials 15(15):5298 |
08 November 2023
06 May 2024
12 October 2024
November -0001
https://doi.org/10.1007/s40789-024-00738-1
Springer
