Early Warning of Coal Mine Production Environment Safety Risks Based on Multi-Source Information Feature Fusion

Abstract

As intelligent mining develops, utilizing coal mine production monitoring data for early warnings has become a crucial means of ensuring safety in mining operations. Assisting decision-makers in making scientific choices through multi-source and massive data is a complex yet vital task. Based on multi-source information fusion, a model for the coal mine production environment is proposed in this paper. It is designed to provide early warnings regarding the safety status of coal production environments in order to assist management and control personnel in making scientific decisions. Firstly, data integration of multi-source heterogeneous datasets was conducted. Multi-source heterogeneous data collected by various types of monitoring sensors in coal mines were analyzed, including temperature, dust, wind speed, vibration energy, and gas. Based on this, the factors influencing coal mine production safety were identified. These factors were then screened through factor analysis to determine the index. An early warning index system for coal mine production environment safety was established. The index weight was established by the principal component analysis method, and the index system for coal mine production environment safety and early warning systems was established. Secondly, based on BP neural networks, a multi-input single-output feature-level fusion model and a multi-input multi-output feature-level fusion model were constructed. Based on the above model, the safety warning for coal mine production environments was implemented. The accuracy of model was 89.29%. Based on multi-source information fusion, the early warning system for coal mine production environments was constructed. The system exhibited good feasibility. It could assist management and control personnel in making scientific decisions.

1. Introduction

The stable supply of coal is the cornerstone of ensuring China’s energy security, and coal mine production safety is the basic prerequisite for achieving a stable coal supply [1]. With the development of intelligent and information technologies, early warnings based on coal mine production monitoring data have become an important means of ensuring safe production.

On 28 April 2024, the State Administration of Mine Safety and six other departments issued the “Guiding Opinions on Deepening the Intelligent Development of Mines and Promoting Mine Safety”, which emphasizes advancing the intelligent construction of mines to promote mine safety. The main safety issues involve the “human–machine–environment–management” system. The unsafe states of “environment and machine” can be easily monitored and measured, and unsafe human behaviors can be reflected in unsafe material states. Analysis of monitoring results can provide effective early warnings.

As a result, coal enterprises are equipped with safety production monitoring systems capable of real-time monitoring of underground environmental parameters such as temperature, carbon monoxide, methane, and dust. However, the vast amount of data obtained from monitoring has not yet been effectively mined and analyzed, remaining in the early stages of transmission, reading, and storage.

Using information fusion technology to analyze mining production monitoring data effectively and applying it to the field of coal mine production safety and early warning systems is of great significance to improving mine safety management and reducing accident rates. Feature-level fusion involves extracting representative features from the raw data provided by each sensor, reducing the amount of data processing, improving system response speed, and solving real-time issues.

In the field of coal mine safety risk assessment and early warning, scholars [2] have proposed a safety risk evaluation method based on the GRA-RBF neural network model, which can provide robust technical support and decision-making foundations for coal mine safety risk pre-control management. However, due to the complexity of pre-control management factors in coal mine safety risks, the practical implementation of this method remains suboptimal. Additionally, other researchers [3] have introduced a variable weight-cloud model to construct an evaluation index system and assessment model for evaluating coal mine safety conditions. Nevertheless, the early warning mechanisms in this approach are insufficiently developed and require further refinement. These studies highlight the necessity for enhanced integration of multi-dimensional risk factors and improved dynamic adaptability in future research on coal mine safety risk management systems.

Currently, many scholars have applied information fusion theory to the coal mine safety field, achieving substantial results. Scholars have used neural networks to fuse multi-sensor data, overcoming the accuracy issues of traditional models [4,5,6]. The data sources for fusion mainly include Zhong Ming [7], Liang Yueqiang [8,9], Shen Zhuo [10], Wu [11], and others, who applied data fusion technology to coal mine gas early warning systems, improving the accuracy of gas early warnings and overall mine safety; Zhang Jing [12] established an information fusion-based mine ventilation and gas disaster early warning model, with simulation results showing improved early warning accuracy for mine ventilation and gas disasters; Gu Baoze [13] used multi-source information fusion for evaluating old water hazard risks in mines.

These studies demonstrate the advantages and feasibility of feature-level information fusion technology in coal mine early warning systems from different perspectives. However, most of the related research focuses on early warning analysis for a single disaster based on production monitoring data, lacking comprehensive early warning systems for the overall safety status of coal mines.

This study first analyzes the production characteristics of the research object, identifies key parameters affecting coal mine production safety (such as CH₄, CO, coal dust, temperature, and wind speed etc.), and uses information fusion technology to process multi-source heterogeneous data to realize early warnings for the underground production environment. The goal is to accurately understand the overall safety status of the underground production environment and provide scientific support for safety production management personnel, ensuring the safe production of coal enterprises.

2. Research Methods

Information fusion is an emerging technology formed at the intersection of traditional disciplines, including monitoring technology, artificial intelligence, neural networks, pattern recognition, and decision theory [14]. It can be logically divided into three levels: data-level fusion, feature-level fusion, and decision-level fusion.

Feature-level fusion is considered the middle layer. It involves the extraction, comprehensive analysis, and processing of feature information from the raw data of various sensors. In multi-level information fusion, the results of feature-level fusion provide the foundational data for decision-level information fusion.

Therefore, feature-level information fusion typically involves three processes: classification, summarization, and integration of information. The fusion process is shown in Figure 1.

Feature-level fusion involves extracting feature information from data sources, analyzing and processing it, and retaining important information to support later decision-making and analysis. The advantages of feature-level fusion include extracting raw data features, reducing the amount of data to be processed, and improving real-time performance. The most common approach is using neural networks to optimally fuse different features, enhancing warning performance.

The underground production environment includes sensor data for gas, temperature, wind speed, dust, and other parameters. Due to its complexity and insignificant changes, it is difficult to express with explicit mathematical models. Therefore, neural networks, which excel in handling nonlinear problems, are fully applied. Neural networks can leverage computer technology to process vast amounts of data, and their powerful nonlinear mapping ability can scientifically and accurately reflect the complexity of the underground production environment.

Compared to other neural networks, such as CNN and RNN, BP neural networks are more flexible and can handle various types of data. They automatically extract features from input data, support parallel computing to speed up training and prediction, and possess strong self-learning, self-organizing, and adaptive capabilities [15].

For these reasons, BP neural networks were chosen as the feature-level information fusion algorithm in this study.

The learning process of a BP neural network involves automatic learning from a given sample dataset, continuously adjusting the connection weights and thresholds between neurons to minimize output error. This allows the neural network to gradually reach the mapping relationship contained in the sample dataset, which cannot be directly expressed by linear programming models. It not only has a complete nonlinear mapping function but also possesses stronger “robustness” and more accurate predictive abilities. Due to these characteristics, BP neural networks have been widely applied and developed in disaster prediction and control [16].

Neural networks have strong self-learning capabilities, especially with the forward network based on the Sigmoid (S-shaped) activation function, which gives them a unique advantage in solving nonlinear problems with arbitrary complexity and precision.

The BP neural network consists of an input layer, hidden layers, and an output layer. The input and output layers are visible, with corresponding input and output values. Figure 2 shows a typical BP neural network structure with two hidden layers.

After the structure and initial parameters of the BP neural network are defined, the sample dataset needs to be divided into a training set and a testing set. The training set is used to continually train and optimize the BP neural network until the output results meet the desired requirements. The testing set is then used to test the optimized network to verify whether it satisfies practical needs. When solving real-world problems, the core steps of BP neural networks are as follows:

(1)

Define Input and Output Nodes and Layer Structure to Initialize the Network:

Set the weights between the input layer, hidden layers, and output layer to random numbers and normalize them within the range [−1, 1].

(2)

Forward Propagation:

As sample data pass forward layer by layer from the input layer to the output layer, calculate the input and output of each neuron step by step. The activation function is applied according to the following Equations (1) and (2) [17]:

$$z_j^l=\sum _{i=1}{w}_{ij}^l{a}_j^{l-1}+b_j^l$$

(1)

$$a_j^l=f\left(z_j^l\right)$$

(2)

(3)

Use the Error BP Mechanism:

Gradually update the weights and thresholds between neurons layer by layer based on the error BP mechanism. The specific mathematical calculation is shown in Equations (3) and (4) [17]:

$$w_{ij}^l=w_{ij}^l-\mu {\displaystyle \frac{\mathsf{\partial }c}{\mathsf{\partial }{w}_{ij}^l}}$$

(3)

$$b_j^l=b_j^l-\mu {\displaystyle \frac{\mathsf{\partial }c}{\mathsf{\partial }{b}_j^l}}$$

(4)

Finally, calculate the error between the network’s output $T$ and the desired output $a^L$. Determine whether the error satisfies the predefined target value. If the predefined accuracy is met, the model training will terminate; otherwise, the process will return to step 2 and continue the iterative training. The error calculation is shown in Equation (5) [17]:

$$C={\displaystyle \frac{1}{2n}}\sum _{n=1}{(a^L-T)}^2$$

(5)

By incorporating nonlinear units into the network during training, it overcomes the issue of the input sample data and the desired output being merely a single linear combination. This enhances the capability of the BP neural network to handle nonlinear problems associated with the research subject.

3. Case Study

3.1. Basic Information of the Research Object

A coal mine is located on the southern edge of the Junggar Basin. The mine’s terrain is characterized by a high southern side and a low northern side. The mining method used is horizontal segmented integrated mechanized top coal caving, with full caving applied for roof management. The main coal seams currently being mined are the B₁₊₂ seam with an average thickness of 28 m and the B₃₊₆ seam with an average thickness of 40 m. The entire coal seam is nearly vertical, with an average dip angle of 87°. Between the two coal seams is a rock pillar with a thickness of over 10 m, which gradually narrows from west to east, ranging from 50 m to 110 m in thickness.

Due to the coal seam strike and mining method, the tunnel axes of the working faces in the B₁₊₂ and B₃₊₆ seams are approximately perpendicular to the direction of maximum horizontal stress. The working face is located in the fissure-pore weak aquifer of the Middle Jurassic Xishanyao Formation. The specific yield of the aquifer is only 0.0001 to 0.0089 L/s·m, with a normal water inflow of 0.37 m³/h and a maximum water inflow of 0.83 m³/h. The aquifer exhibits weak water abundance and poor permeability, with generally limited recharge conditions. Under normal circumstances, the primary source of water inflow is fissure water from the coal and rock strata, which is mainly concentrated at the coal–rock interface and areas with well-developed fissures. The mining method of the mine is upper layer mining and lower layer tunneling, with the sequence being to mine the B₃₊₆ seam first, followed by the B₁₊₂ seam. The mining above the +450 m level has been completed, and the current coal mining face is located at the +425 m level of the B₃₊₆ seam, with the tunneling face at the +400 m level, as shown in Figure 3, which displays a layered mining diagram of the mine.

The mining sequence is to first extract the B₃₊₆ coal seam, followed by the B₁₊₂ coal seam. The southern area above the +450 m level has been fully mined, and the current mining face is located at the +425 m level. The B₃₊₆ coal seam mining face has a strike length of 2493 m, a dip length of 41 m, and a vertical height of 25 m. The B₃₊₆ coal seam belongs to the same coal group as the B₁₊₂ coal seam, with a distance of 98 m between them. The B₃₊₆ coal seam is a gas-bearing seam. According to the gas grade evaluation results in 2017, the maximum absolute gas emission is 0.55 m³/min, with a maximum relative gas emission of 0.13 m³/t. The maximum absolute gas emission for the B6 coal seam is 0.21 m³/min, and for the B₃ coal seam, it is 0.19 m³/min. Tests have shown that the gas content in this coal seam is low, and there is no danger of coal and gas outbursts. It is a spontaneous combustion coal seam with a tendency for natural ignition. Additionally, this coal mine experiences severe rock burst phenomena, with several significant rock burst incidents having occurred. Statistics show that all rock burst incidents during the mining process occurred in the previously mined B₃₊₆ working face.

3.2. Coal Mine Production Environment Safety and Warning Indicator Systems

(1)

Establishment of the Indicator System

Constructing a scientific and reasonable indicator system is the foundation for evaluation and early warning. When selecting indicators, excessive attention to details may lead to an overly large number of indicators, thereby increasing the complexity of the indicator system and complicating subsequent evaluation or early warning processes. This could also potentially obscure some critical influencing factors. On the other hand, selecting too few indicators, while making the evaluation and early warning processes more manageable, introduces new issues, such as the inability of the evaluation and early warning results to comprehensively and objectively reflect the true state of the research subject. For early warning systems in the context of coal mine production environments, the ultimate goal is to propose control measures through early warnings to reduce the occurrence of various accidents and achieve safe production. When selecting indicators, it is crucial to consider their impact on the overall early warning results. Typically, it is advisable to use as few key indicators as possible. By reviewing the literature and analyzing the actual conditions of the research subject, a preliminary evaluation indicator system can be established in conjunction with relevant industry standards. Subsequently, factor analysis and correlation analysis methods can be employed to analyze and screen the indicators. The process of establishing the indicator system is illustrated in Figure 4.

Many scholars have established indicator systems and studied coal mine early warning models from the “human–machine–environment–management” perspective. These studies cover critical aspects of coal mine safety production. However, the indicator systems built from this perspective are often too large to meet the requirements for real-time and accurate warnings.

The “human–machine–environment–management” indicator system includes many indicators that cannot be directly measured and require manual inspection, such as personnel information and enterprise management levels. These data lack real-time monitoring capabilities and show little variation over time, making them more suitable for static warning and evaluation within fixed timeframes.

Environmental parameters, however, can be monitored in real-time using sensors. When anomalies occur in the “human–machine–management” dimensions, such as operator errors, fan failures, or poor personnel management, these issues are reflected in environmental parameter changes.

This study considers “environment” as a tangible representation of the “human–machine–management” dimensions. Monitoring environmental parameter changes can directly reveal safety-related anomalies across all four dimensions. This approach reduces the complexity of the warning indicators while effectively achieving early warning systems for the underground production environment safety.

Many scholars have conducted early warning research based on the “human–machine–environment–management” or “human–machine–environment–management–information” dimensions. Corresponding indicator systems were established, achieving coal mine safety production early warning.

The ultimate goal of early warning systems for coal mine production environment is to propose control measures to reduce accidents and ensure safety. Currently, most researchers focus on gases, temperature, wind speed, and dust to establish early warning indicators for coal mine production environment safety, as shown in

Table 1. Establishment of coal mine environmental indicators.

Author	Index	Remarks
Ding Zhen [18]	CH4, CO, Coal dust, wind speed	Auxiliary coal face environment warning and governance
Wang Lei [19]	CO, C2H4, C2H2	Regarding a certain mine
Zhou Li [20]	CO, O2, temperature, CH4
Li Zhenhua [21]	CH4, Coal dust, wind speed	Coal mine safety forecasting based on multi-sensor data fusion
Liu Kai [22]	CH4, temperaturet, wind speed	Hierarchical multi-sensor data fusion model for coal mine safety monitoring
Xiong Bojie [23]	CH4, CO, temperature, Coal dust, wind speed	Assessment of coal mine safety status based on multi-sensor data fusion
Sun Yanfei [24]	Multi-sensor data	Bayes estimation; fuzzy set theory
Xie liming [25]	CH4, CO, temperature, wind speed	Early warning of gas data fusion based on concept lattice
Wang Chuanyin [26]	temperature, wind speed	Fuzzy data fusion algorithms in coal mine safety systems

.

Based on the definition of coal mine safety monitoring systems in the “General Technical Requirements for Coal Mine Safety Monitoring Systems” and considering the actual situation of the research object, five monitoring areas were selected: the +425 m level B₃₊₆ working face, the +400 m level B₃₊₆ tunneling face, the +400 m level B₁₊₂ tunneling face, the B₁₊₂ goaf, and the B₃₊₆ goaf, as shown in

Table 2. Production environment monitoring indicators.

Monitoring Object	Monitoring Area	Sensor	Acquisition Frequency
Temperature	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+2 heading face, Mined-out areaB1+2, B3+6	Infrared temperature sensor	30 min
Dust	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+2 heading face	Dust concentration sensor	1 min
Wind speed	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+2 heading face	Mining air duct wind speed sensor	1 min
O2	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+2 heading face	Infrared spectrum sensor	1 s
CH4	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+2 heading face, Mined-out areaB1+2, B3+6	Optical interference methane analyzer	1 s
CO2	+425 level B3+6 coal face, +400 level B3+6 heading face, +400 level B1+ 2heading face, Mined-out areaB1+2, B3+6	Infrared spectrum sensor	1 s
CO	+425 level B3+6 cola face, +400 level B3+6 heading face, +400 level B1+2 heading face, Mined-out areaB1+2, B3+6	Infrared spectrum sensor	1 s
Vibrational energy	+425 level B3+6 coal face	Mining vibration sensor	1 s

.

Different monitoring areas have different characteristics, and thus the focus of monitoring varies. For the serious issue of rock burst in the working face, vibration energy was added as a monitoring indicator. For the goaf, the main concern is fire prevention and extinguishment, with the focus on gas and temperature monitoring. For both the working face and tunneling face, the safety and comfort of underground personnel are considered, with gas and temperature monitoring, along with wind speed and dust as additional monitoring indicators.

Based on previous research on coal mine production environment safety and early warning indicators, relevant gases related to spontaneous combustion, such as C₂H₂ and C₂H₄, were selected to establish an early warning indicator system with 37 secondary indicators, including oxygen, methane, carbon monoxide, temperature, wind speed, and dust.

Based on factor analysis and correlation analysis, highly correlated indicators were removed to reduce information redundancy.

Taking working face indicator selection as an example, KMO and Bartlett tests were conducted before factor analysis. After meeting the basic requirements, factor loadings were calculated, as shown in

Table 3. Factor load coefficient.

Name	Factor 1	Factor 2	Factor 3	Factor 4	Commonality (Common Factor Variance)
Temperature	0.085	0.131	0.029	0.913	0.859
Dust	0.169	−0.068	0.699	0.242	0.580
Wind speed	0.073	0.787	0.255	0.056	0.692
O2	0.884	−0.033	0.037	−0.050	0.787
CH4	−0.102	0.882	−0.240	0.083	0.853
CO2	0.389	0.091	−0.015	−0.444	0.457
CO	−0.212	0.045	0.770	−0.185	0.675
N2	−0.920	−0.002	−0.040	0.160	0.874

.

Then, through correlation analysis, the value of the correlation coefficient ranges between −1 and 1. A value of 1 indicates a perfect positive correlation, −1 indicates a perfect negative correlation, and 0 indicates no linear relationship. As shown in

Table 4. Index correlation coefficient.

	Temperature	Dust	Wind Speed	O2	CH4	CO2	CO	N2	Vibration Energy
Temperature	1	0.113	0.016	−0.069	0.073	−0.202	−0.039	0.118	−0.005
Dust	0.113	1	0.131	0.046	−0.102	0.109	0.130	−0.089	0.030
Wind speed	0.016	0.131	1	0.117	0.321	0.141	0.145	−0.154	0.178
O2	−0.069	0.046	0.117	1	0.041	0.194	0.012	−0.950	0.327
CH4	0.073	−0.102	0.321	0.041	1	0.046	−0.060	−0.055	−0.062
CO2	−0.202	0.109	0.141	0.194	0.046	1	−0.110	−0.457	0.158
CO	−0.039	0.130	0.145	0.012	−0.060	−0.110	1	0.015	−0.108
N2	0.118	−0.089	−0.154	−0.950	−0.055	−0.457	0.015	1	−0.369
Vibration energy	−0.005	0.030	0.178	0.327	−0.062	0.158	−0.108	−0.369	1

, the correlation coefficient between N₂ and O₂ is −0.95, the correlation coefficient between N₂ and CO₂ is −0.457, and the correlation coefficient between N₂ and vibration ability is −0.369, indicating that N₂ has correlations with the other three influencing factors. Therefore, the nitrogen gas (N₂) indicator is deleted.

(2)

Establishment of Warning Levels

The purpose of establishing warning levels is to help coal mine management personnel assess the safety status of the entire environment system based on different warning levels. This enables appropriate control measures to be implemented to address and eliminate potential hazards.

The warning levels required in this study were determined by integrating the Coal Mine Safety Regulations, General Technical Requirements for Coal Mine Safety Monitoring Systems, and previous research findings. Four warning levels were defined in

Table 5. Early warning thresholds for indicators.

Serial Number	Monitoring Area	Index	Level 1	Level 2	Level 3	Level 4
GZ1	+425 level B3+6 coal face	Coal face temperature (°C)	≥30	26~30	22~26	16~22
GZ2		Wind speed in the intake airway (m/s)	≤1	1~1.5	1.5~2	2~6
GZ3		Coal face Dust (mg/m3)	≥4	3~4	2~3	0~2
GZ4		O2 (%)	≤17	17~18	18~20	20~21
GZ5		CH4 (%)	≥1.5	1~1.5	0.5~1	<0.5
GZ6		CO2 (%)	≥1.5	1~1.5	0.5~1	<0.5
GZ7		CO (ppm)	≥24	18~24	8~18	<8
GZ8		Vibrational energy (105 J)	≥1	0.5~1	0.1~0.5	<0.1
JA1	+400 level B3+6 Heading face	Heading face temperature (°C)	≥30	26~30	22~26	16~22
JA2		Wind speed in the intake airway (m/s)	≤1	1~1.5	1.5~2	2~6
JA3		Heading face dust (mg/m3)	≥4	3~4	2~3	0~2
JA4		O2 (%)	≤17	17~18	18~20	20~21
JA5		CH4 (%)	≥1.5	1~1.5	0.5~1	<0.5
JA6		CO2 (%)	≥1.5	1~1.5	0.5~1	<0.5
JA7		CO (ppm)	≥24	18~24	8~18	<8
CB1	B3+6 Mined-out Area	Fiber optic temperature measurement (°C)	≥150	90~150	30~90	<30
CB2		C2H2 (ppm)	≥1	0~1	-	0
CB3		C2H4 (ppm)	≥1	0~1	-	0
CB4		CH4 (%)	≥1	0.5~1	0.3~0.5	<0.3
CB5		CO2 (%)	≥10	6~10	4~6	<4
CB6		CO (ppm)	≥24	18~24	8~18	<8

: Level 4 corresponds to no danger, Level 3 to minor danger, Level 2 to moderate danger, and Level 1 to severe danger.

3.3. Feature-Level Fusion Model Based on BP Neural Network

According to relevant research, underground production environment monitoring data form a continuous time series. These data exhibit characteristics such as nonlinearity and non-stationarity. Within a certain timeframe, the data change slowly, with no significant large-scale fluctuations.

The coal mine production environment system is a complex system composed of multiple subsystems. Each subsystem contains several influencing factors. Strong nonlinear relationships exist both between subsystems and among influencing factors, directly increasing system complexity. For such a complex system, achieving efficient, real-time, and accurate early warnings is challenging. Traditional methods, such as mathematical models, struggle to provide effective warnings. Conventional evaluation and warning methods demonstrate limitations.

In this study, BP neural networks were employed as a feature-level information fusion algorithm. By learning from multi-source heterogeneous sensor data samples, the network model was continuously trained until the desired accuracy was achieved. Actual production monitoring data were used as the data source. Single-indicator warning levels corresponding to production monitoring data were calculated. The maximum value principle for single-indicator warning levels was used to determine the safety status of the coal mine production environment. Warning level values for the mine were set as expected values for subsequent neural network model training. The sample dataset and expected values are shown in

Table 6. Initial set of sample data.

Serial Number	GZ1	GZ2	GZ3	GZ4	…	JA2	JA5	Expected Value
1	19.8	1.4	1.6	20.66	…	1.7	1.18	4
2	19.5	1.33	5.75	20.64	…	1.85	0.74	3
3	19.8	1.35	2	20.66	…	1.8	1.31	4
…	…	…	…	…	…	…	…	…
1510	20.1	1.25	4.33	20.61	…	2.8	1.07	3
1511	19.5	1.28	1.67	20.68	…	2.3	0.14	4
1512	20.7	1.25	8.33	20.66	…	2.2	0.31	4

.

(1)

BP Neural Network Regression Model

Twenty-eight sample data sets were selected to verify the model’s accuracy. The comparison between prediction results and actual values is shown in Figure 5. A BP neural network model was constructed by importing the sample dataset of the research object. Based on this, the neural network model was trained. The accuracy of each training result was compared to determine the most accurate neural network model.

Based on different sample output types, parameters were adjusted, and continuous training and testing were conducted. The prediction accuracy was 89.29% based on sensor monitoring data from the mine. The feasibility of applying BP neural networks to early warning systems in coal mine production environment safety was demonstrated.

(2)

BP Neural Network-Based Classification Model

A feature-level fusion model based on the BP neural network was constructed for the complex system of coal mine production environment safety. After testing and training, high prediction accuracy was achieved for both multi-input single-output regression models and multi-input multi-output classification models. The model was found to be well-suited for predicting nonlinear problems in such environments. A multi-input multi-output BP neural network early warning model with thirty-four inputs and four output groups was built to verify its applicability. Single-output values were transformed into forms such as (0,0,0,1) or (0,0,1,0). The expected outputs for the four-level warnings were defined as follows: (0,0,0,1) for level 4, (0,0,1,0) for level 3, (0,1,0,0) for level 2, and (1,0,0,0) for level 1. As shown in Figure 6, the prediction accuracy of the multi-input multi-output BP neural network reached 85.71%.

The prediction results of four output groups in the multi-input multi-output model, as shown in Figure 7, indicate the advantages of the BP neural network in addressing nonlinear problems. It is demonstrated that the network can be effectively applied to early warning systems in coal mine production environments. Using the feature-level fusion model based on the BP neural network constructed in this study, the overall safety status of the underground production environment can be obtained. Scientific theoretical references can be provided to coal mine safety management personnel for daily safety assurance.

4. Conclusions

This study focuses on the complex system of coal mine production environment safety. Based on a multi-source information feature fusion method, a case mine was selected as the research object. Based on the analysis of its production characteristics, key parameters affecting coal mine production safety are identified. Utilizing the monitoring data of coal mine production, a targeted multi-index fusion early warning indicator system has been established. A feature-level fusion model was constructed using a BP neural network. After testing and training, high prediction accuracy was achieved by both multi-input single-output regression models and multi-input multi-output classification models. The models were shown to be well-suited for predicting nonlinear problems in coal mine production environment safety. The results indicate that the BP neural network, with its advantages in addressing nonlinear problems, can be effectively applied to early warning systems in such environments.