Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 8, No. 3, 2024 

 

1 

Recognition	and	Prediction	of	Precursory	Feature	
Signals	of	Coal	Mine	Rock	Burst	Based	on	Random	
Forest	and	MK	Trend	Test	
Tianhang Yang * 

Sun Yueqi Honors College, China University of Mining and Technology, Xuzhou, Jiangsu, China 

* Corresponding author: Tianhang Yang 

 

Abstract: This paper aims to identify and predict the precursory characteristic signals of coal mine rock burst by employing 
the Random Forest and MK (Mann-Kendall) trend test methods. Initially, the study conducts data preprocessing and analysis on 
indicators such as electromagnetic radiation intensity, acoustic wave intensity, and type. It uses clustering methods to distinguish 
various data types and analyzes the data after visual representation. The visualization of data illustrates the distribution and trends, 
which aids in understanding the characteristics of the data, such as the upward trend observed in precursory feature data. During 
the data preprocessing phase, diagrams of various classes of data for Acoustic Emission (AE) and Electromagnetic Radiation 
(EMR) intensity are presented, providing an intuitive reference for subsequent analysis. The research utilizes the Random Forest 
algorithm and MK trend test to recognize precursor signals and predict their occurrence intervals. The Random Forest model is 
chosen for its efficiency and accuracy in handling classification and regression issues, while the MK trend test provides a 
statistical basis for identifying precursory signals by analyzing monotonic trends within the dataset. This study not only enhances 
the accuracy of precursory signal identification for coal mine rock bursts but also offers scientific early warning and control 
measures for coal mine safety production, which is of significant practical value. 

Keywords: Coal Mine Rock Burst; Precursory Characteristic Signals; Random Forest; MK Trend Test; Perceptron. 
 

1. Introduction 
Coal, as a vital component of global energy, plays a crucial 

role in economic and social development. However, as coal 
mining deepens, safety issues in coal mines become 
increasingly prominent, especially rock burst accidents that 
not only threaten the lives of miners but also lead to 
significant economic losses [1]. Therefore, effective 
prediction and early warning of coal mine rock bursts have 
become a key issue in the field of mine safety [2-3]. 

Rock bursts, as a complex mining dynamic phenomenon, 
are often accompanied by a variety of precursory feature 
signals. These signals include, but are not limited to, 
electromagnetic radiation, acoustic wave activity, and other 
physical phenomena. Identifying and interpreting these 
precursory signals are of great significance for predicting rock 
bursts. However, due to the complexity and variability of 
these signals, traditional prediction methods often struggle to 
accurately identify and predict rock burst events [4-5]. 

To improve the accuracy of predictions, this paper proposes 
an integrated method based on the Random Forest and MK 
(Mann-Kendall) trend test. The Random Forest, as an 
ensemble learning technique, enhances model performance 
and generalization ability by constructing multiple decision 
trees and voting or averaging [6]. The MK trend test is a non-
parametric statistical method that can effectively detect trend 
changes in a dataset [7]. By combining these two methods, 
this paper aims to build a robust recognition and prediction 
model to improve the identification rate and prediction 
accuracy of precursor signals of coal mine rock bursts. 

2. Data Preprocessing 

2.1. Data Cleaning 

 
Fig 1. AE intensity map of Class A and Class B data 

 
In this paper, the data are preprocessed and analyzed first. 

We preprocessed electromagnetic radiation intensity, acoustic 
wave intensity and type as indicators, distinguished various 
data by clustering method, and analyzed the data after data 
visualization. In the visualization of data, we show the 
distribution and trend of data by drawing to help understand 



 

2 

the characteristics of data, such as the upward trend of 
precursor feature data. Fig. 1 are AE intensity graphs of data 
A and B, and Fig. 2 are EMR intensity graphs of two types of 
data. 

2.2. Precursor Feature Signals 
Within about 7 days before the earth burst, the 

electromagnetic radiation and acoustic emission signals have 
a tendency to increase with time cycle, which is called 
precursor characteristic signals. Rock burst may occur within 
about 7 days after the occurrence of precursor characteristic 
signals, so in general, after the occurrence of precursor 
characteristic signals, certain measures will be taken to 
prevent rock burst as much as possible. 

 

 
 

Fig 2. EMR intensity map of Class A and Class B data 
 

This paper first analyzed the data, visualized the normal 
signal and the precursor signal, and found that the precursor 
characteristic signal strength has a slight upward trend, which 

is the same as the definition of the precursor characteristic 
signal. Fig. 3 and Fig. 4 respectively show the comparison 
results between the normal acoustic emission signal and the 
precursor characteristic signal strength and the normal 
electromagnetic radiation signal and the precursor 
characteristic signal strength. 

 

 
Fig 3. Comparison of the intensity of acoustic emission normal 

signal and precursor characteristic signal 
 

 
Fig 4. Comparison of the intensity of normal and precursory 

characteristic signals of electromagnetic radiation 

3. Model Construction 
In the analysis of precursor data, this paper uses KS test, 

MK trend test and ADF stationary series test to test all the data 
in Annex 2, distinguish precursor data from normal data, 
calculate and compare, and get their characteristics. Then, the 
features are substituted into the random forest model for 
machine learning and interval calculation. 

3.1. KS Test 
 

Table 1. Comparison of normal and precursory signals of electromagnetic radiation 
 Normal EMR EMR Precursor Normal AE AE Precursor

Quantity 73418 12532 15587 2212
Mean Value 49.81542148 71.05480777 37.4329889 41.66865778

Standard Deviation 18.27932445 91.89984446 3.72736829 8.303728496
Minimum Value 9.61 11.67 29 29

25% Quantile 42.451 30.14875 35.29 35.41
50% Quantile 46.4895 41 36.661 39.105
75% Quantile 52 71.674 38.2655 45.16375

Maximum Value 270 491 80 81

The KS test is a test to see whether a distribution f(x) is 
consistent with the theoretical distribution g(x), or whether 
two observed distributions are significantly different. It can 
be seen from the title that the precursory characteristic signal 
is the signal of electromagnetic radiation and acoustic 
emission that has a tendency to increase with time cycle, and 
this upward trend signal can be screened by KS test. Therefore, 

in this paper, KS test is carried out on all the data, the 
precursor feature data is selected, and a series of feature 
values are obtained by comparing it with the normal data. The 
data comparison results of normal electromagnetic radiation 
signals and precursor signals are shown in the table 1. Among 
them, the more obvious is that the mean value and standard 
difference are large. 

 



 

3 

 
Fig 5. Autocorrelation function diagram of acoustic emission 

normal signal and precursor characteristic signal 
 

 
 

Fig 6. Autocorrelation function diagram of normal signal and 
precursory characteristic signal of electromagnetic radiation 

 
Then, we visualized the autocorrelation function of the 

normal signal and precursor signal distinguished after KS test, 
and the results were shown in Fig. 5 and Fig. 6. We find that 
the precursory feature signal is like the normal signal, which 
is not conducive to the further prediction as a feature. 
Therefore, this paper further tests the data. 

3.2. MK Trend Test and ADF Stationary Series 
Test 

The machine learning of random forest needs more 
significant features to predict the interval of the next 
precursor signal more accurately, so this paper selects MK 
trend test and ADF stationary sequence test to obtain features. 
The MK test is a nonparametric trend test used to analyze 
monotonic trends in a data set. It is based on symbolic 
comparison of data pairs. 

Set a time series , for any (i,j) pair where 
i<j, defines as follows. 

 

                 (1) 

 

The test statistic S is the sum of all  pairs. 
 

                       (2) 

 

In the absence of equivalent values, the standard normal 
variable Z is used to test. 

 

                (3) 

 

The calculation formula of Var(S) is as follows. 
 

            (4) 

 

The ADF test is used to detect the unit root in the time 
series to determine whether the series is non-stationary. The 
basic model for testing is an autoregressive model (AR 
model). 

 

                             (5) 

 

Where   is the time series and Δ  =   is 

the first difference of the series.  is the constant term and 

 is the time trend term.  tests the coefficient of the 

unit root, if =0 then there is a unit root, indicating that 

the sequence is non-stationary.   is the 

coefficient of the difference term, and  is the error term. 
The article creates a function and defines a constant time 

interval of 2 minutes for the time series of our data to perform 
ADF (Augmented Dickey-Fuller) and MK (Mann-Kendall) 

tests. After the tests, corresponding P-values are obtained, 
which are used to measure the probability of observing the 
data or more extreme data under the null hypothesis being true. 
The P-values are recorded, and their dynamic changes are 
analyzed. It is found that when precursor features of the data 
appear, the P-values exhibit a specific trend. This trend is 
similar to the described trend in the question, which increases 
cyclically over time. Therefore, the article also defines this 
trend in P-values as a characteristic. 



 

4 

3.3. Random Forest 
Through KS test, MK trend test and ADF stationary series 

test, this paper obtained many characteristics of precursor 
signals, including the eigenvalues in the table obtained in the 
above KS test, including large mean and standard deviation, 
as well as the P-value trend characteristics of ADF test and 
MK test. Then, by means of random forest, this paper trains 
the model with the above features to predict the occurrence 
time of precursor signals. Finally, the precursory 
characteristic signal time interval of electromagnetic 
radiation and the precursory characteristic signal interval of 
acoustic emission are obtained by the model. 

 

 
 

Fig 7. Random forest specific flow chart 
 

Random forest is the combination of multiple decision trees, 
each data set is randomly selected, and some features are 
randomly selected as input. The specific flow chart of random 
forest is shown in Fig. 7. In this paper, the data is divided into 
training set and test set by means of random forest, and the 
feature scaling and feature engineering processing are carried 
out to train the model and get the prediction result. The 
electromagnetic radiation interference signal time interval 
and acoustic emission interference signal interval obtained by 
this model are shown in Table 2 and Table 3. 

 

Table 2. Time interval of electromagnetic radiation interference 
signal 

Number Start of Time Interval End of Time Interval
1 2022/05/01 00:00:00 2022/05/14 15:10:30
2 2022/05/14 15:16:15 2022/05/14 15:58:22
3 2022/05/14 16:18:48 2022/05/14 17:39:34
4 2022/05/14 17:41:23 2022/05/14 18:15:13
5 2022/05/14 18:18:14 2022/05/14 19:10:02

 
Table 3. Time interval of acoustic emission interference signal 

Number Start of Time Interval End of Time Interval
1 2022/10/10 00:00:00 2022/10/16 01:32:21
2 2022/10/16 01:32:21 2022/10/16 01:32:21
3 2022/10/25 14:30:56 2022/10/31 03:22:59
4 2022/10/31 03:22:59 2022/11/09 09:58:22
5 2022/10/31 04:53:48 2022/10/31 05:07:21

4. Probabilistic Prediction 
We use the model to predict the probability of precursor 

feature data after some data. Since we do not analyze all the 
data, but the data before the precursor signal occurs, we no 
longer use the method of setting threshold and KS test, but 
directly use the MK trend test and ADF stationary series test 
to obtain the characteristics of their data. Then, the feature is 
inserted into the perceptron model of sigmoid activation 

function, and the possibility of precursor data after output a 
certain data. 

The Sigmoid activation function is an activation function 
commonly used in neural networks to map the input of a 
neuron to the output between 0 and 1. This property makes 
the sigmoid function particularly suitable for interpreting the 
output as probability, especially in binary classification 
problems. Therefore, for problem 3, this paper selects a 
perceptron model using sigmoid activation function, and 
calculates the occurrence probability of precursor feature 
signals at the last moment of a certain time period by 
analyzing the characteristics of the data. 

Specifically, the output of a perceptron model using the 
sigmoid activation function can be expressed as follows. 

 

                    (6) 

 

Where x is the input, w is the weight, and b is the bias term. 
The output value of this function is between 0 and 1, which 
can be used to represent the probability of an event, that is, 
the probability of the occurrence of the precursor 
characteristic signal requested by the problem. 

With the MK trend test and ADF stationary series test in 
3.2 above, the data is first tested to obtain its corresponding 
P-value, and its dynamic situation is analyzed to obtain the 
unique trend of P-value. After replacing it into the perceptron 
model using sigmoid activation function, the corresponding 
probability can be obtained through calculation, as shown in 
Table 4. 

 
Table 4. The probability of precursor features occurring at the time 

of data collection 

Time of 
electromagnetic 
radiation data 

Probability 

The time of 
the acoustic 

emission 
data 

Probability 

2023-1-24 23:58:36 0.001195% 
2023-1-24 
23:58:36 

0.07968% 

2023-2-11 23:59:20 79.90% 
2023-2-11 
23:59:20 

0.6814% 

2023-2-26 23:59:27 99.42% 
2023-2-26 
23:59:27 

40.21% 

2023-3-10 23:58:14 9.950% 
2023-3-10 
23:58:14 

2.205% 

2023-3-30 23:58:13 99.85% 
2023-3-30 
23:58:13 

81.99% 

5. Conclusion 
This paper effectively identifies and predicts the precursory 

feature signals of coal mine rock bursts by constructing a 
model based on the Random Forest and MK trend test. 
Through in-depth analysis of electromagnetic radiation and 
acoustic wave intensity data, we have revealed specific 
patterns of these signals prior to the occurrence of rock bursts. 
By using clustering methods for data preprocessing and 
visualization, we successfully distinguished between normal 
and precursory signals, thereby establishing a predictive 
model. 

The Random Forest model, known for its robust 
classification capabilities, was employed in this study, while 
the MK trend test provided a solid statistical foundation for 
identifying trend changes in the signals. The combined 
application of these two methods not only improved the 
accuracy of precursory signal identification but also predicted 



 

5 

the time intervals in which these signals are likely to occur, 
providing valuable early warning information for the safety 
of coal mine production. 

Furthermore, the model proposed in this study 
demonstrated high accuracy and good generalization ability 
in empirical analysis, verifying the effectiveness and 
practicality of the model. However, the complexity of the coal 
mine environment suggests that there is still room for 
improvement. Future research can further explore the 
applicability of the model under different mining conditions 
and consider incorporating more types of sensor data to 
enhance the accuracy and reliability of predictions. 

References 
[1] Fu-xing J, Xian-feng S H I, Cun-wen W, et al. Mechanical 

mechanism of rock burst accidents in slice mining face under 
high pressure[J]. Chinese Journal of Geotechnical Engineering, 
2015, 37(6): 1123-1131. 

[2] Peixoto A, Ribeiro e Sousa L, Leal e Sousa R, et al. Prediction 
of rockburst based on an accident database[C]//ISRM Congress. 
ISRM, 2011: ISRM-12CONGRESS-2011-224. 

[3] Chen Y, Zhang Z, Cao C, et al. Research on the causal 
mechanism of a rock burst accident in a longwall roadway and 
its prevention measures[J]. Scientific reports, 2023, 13(1): 
22312. 

[4] Varotsos C A, Tzanis C, Cracknell A P. Precursory signals of 
the major El Niño Southern Oscillation events[J]. Theoretical 
and Applied Climatology, 2016, 124: 903-912. 

[5] Sánchez‐Pastor P, Obermann A, Schimmel M. Detecting and 
locating precursory signals during the 2011 El Hierro, Canary 
Islands, submarine eruption[J]. Geophysical Research Letters, 
2018, 45(19): 10,288-10,297. 

[6] Rigatti S J. Random forest[J]. Journal of Insurance Medicine, 
2017, 47(1): 31-39. 

[7] Sang Y F, Wang Z, Liu C. Comparison of the MK test and 
EMD method for trend identification in hydrological time 
series[J]. Journal of Hydrology, 2014, 510: 293-298.