Characteristic Analysis of Unsafe Behavior by Coal Miners: Multi-Dimensional Description of the Pan-Scene Data

As a high-risk occupation, coal mining has many accidents, primarily due to the unsafe behavior of coal miners. Based on the research of analysis of unsafe behavior and pan-scenario data of miners, a theoretical framework for the analysis of unsafe behavior characteristics was proposed in this paper. The collected data were divided into realistic scenes and abstract scenes according to different manifestations; the pan-scene data were described from the eight dimensions of time, behavioral trace, location, behavioral property, behavioral individual, degree, unsafe action, and specialty using a quantitative method for the structure conversion; and the rules were discovered through cluster analysis and association analysis. A total of 225 coal mine gas explosion accidents were used for analysis, and the pan-scene data description and structure conversion of unsafe behavior that caused these accidents were realized. In a certain cluster, the distribution rules of dimensions and the interaction between different dimensions of unsafe behavior were explored after analysis. The results show that the proposed eight dimensions can fully explain the basic characteristics and attributes of the unsafe behavior of coal miners. The structure conversion can reduce the workload of managers and effectively improve the safety data processing capabilities, and the result of data analysis can provide data support and a management basis for safety management. A new method and thought for the data analysis of miners’ unsafe behavior is provided.


Introduction
In recent years, safety and health of work have gradually become the focus of attention in the development of industry [1]. Especially the coal industry, which has the most serious accidents in industrial production. Approximately one-half of Chinese major disasters occur in coal mines. The death toll from coal mine accidents exceeds that of all other accidents combined in China [2]. Coal mining is also considered one of the most dangerous jobs in the world [3]. Compared with other industries, coal mine workers face a relatively more dangerous working environment [4]. The frequent occurrence of coal mine accidents seriously affects the security situation of China's coal industry and the sustainable development of the whole society's economy [5].
Over the years, the Chinese government has continuously increased the intensity of safety management of coal enterprises, to a large extent, has improved the performance of coal mine safety production [6]. Many experts and scholars have also made a lot of important contributions in coal mine safety management and accident prevention. Mahdevari et al. put forward an evaluation and action, taking the unsafe behavior of the human in the scene as the core, the internal information of the scene data is extensively explored, and the multi-dimensional description of the scene is realized to form richer pan-scene data. Wang et al. put forward the concept of emergency scene in the framework of structural error behavior analysis method in the research of subway accident investigation traffic dispatcher error behavior and put forward the structural transformation of subway accident data. It is easy for safety managers to understand the common error behavior more clearly and provide a more detailed method to collect and store the wrong behavior data in the subway traffic scheduling system [27]. Kumar et al. put forward a framework of data mining for road traffic accidents in the analysis and research of road traffic accidents. The inherent laws and hidden characteristics of road traffic accidents are analyzed by cluster analysis and other data processing methods. Based on the research of structured data, this study provided a new idea for data feature analysis [28].
The multiple accidents of a coal miner, the predominance of accidents caused by unsafe worker behavior, the complexity of the environment, and the explosion of data all propose new requirements for the traditional methods of statistical data on behavior-based safety. Improving the processing ability and analyzing the information resources to form effective management strategies and improve the level of safety management has become a topic for many scholars. Meanwhile, the theory of pan-scene data remains imperfect at home and abroad for the security sciences' generalized data mining. There is a lack of research on the characteristics of unsafe behavior of miners based on structured data. Therefore, in this study, the unsafe behaviors of coal mine workers are taken as the research object to construct the theoretical framework for analysis of unsafe behavior characteristics based on pan-scene data. Finally, with China's 225 coal mine gas explosion accident reports as an example, the description of the scene data and structural transformation are realized, and the regularities and the characteristics of the unsafe behavior that caused these accidents are obtained through cluster analysis and association analysis.

Theoretical Framework
The theoretical framework for analysis of unsafe behavior characteristics based on pan-scene data shown in Figure 1 primarily includes four parts: pan-scene data source; pan-scene data description; pan-scene data structure conversion; and pan-scene data analysis. The first part is for collecting the realistic scene and abstract scene data from a coal mine enterprise or a certain type of accident and screening the useful data resources for subsequent data processing. The second part is the description of the pan-scene data, which is done using the eight dimensions. The third part is the structural transformation of the pan-scene data. The level of concept and granularity of each dimension are clarified, and the value of each attribute is encoded and quantified. The fourth part is the pan-scene data analysis. Cluster and association analyses are used to mine data rules to realize the visualization of unsafe behavior data. action, taking the unsafe behavior of the human in the scene as the core, the internal information of the scene data is extensively explored, and the multi-dimensional description of the scene is realized to form richer pan-scene data. Wang et al. put forward the concept of emergency scene in the framework of structural error behavior analysis method in the research of subway accident investigation traffic dispatcher error behavior and put forward the structural transformation of subway accident data. It is easy for safety managers to understand the common error behavior more clearly and provide a more detailed method to collect and store the wrong behavior data in the subway traffic scheduling system [27]. Kumar et al. put forward a framework of data mining for road traffic accidents in the analysis and research of road traffic accidents. The inherent laws and hidden characteristics of road traffic accidents are analyzed by cluster analysis and other data processing methods. Based on the research of structured data, this study provided a new idea for data feature analysis [28].
The multiple accidents of a coal miner, the predominance of accidents caused by unsafe worker behavior, the complexity of the environment, and the explosion of data all propose new requirements for the traditional methods of statistical data on behavior-based safety. Improving the processing ability and analyzing the information resources to form effective management strategies and improve the level of safety management has become a topic for many scholars. Meanwhile, the theory of panscene data remains imperfect at home and abroad for the security sciences' generalized data mining. There is a lack of research on the characteristics of unsafe behavior of miners based on structured data. Therefore, in this study, the unsafe behaviors of coal mine workers are taken as the research object to construct the theoretical framework for analysis of unsafe behavior characteristics based on pan-scene data. Finally, with China's 225 coal mine gas explosion accident reports as an example, the description of the scene data and structural transformation are realized, and the regularities and the characteristics of the unsafe behavior that caused these accidents are obtained through cluster analysis and association analysis.

Theoretical Framework
The theoretical framework for analysis of unsafe behavior characteristics based on pan-scene data shown in Figure 1 primarily includes four parts: pan-scene data source; pan-scene data description; pan-scene data structure conversion; and pan-scene data analysis. The first part is for collecting the realistic scene and abstract scene data from a coal mine enterprise or a certain type of accident and screening the useful data resources for subsequent data processing. The second part is the description of the pan-scene data, which is done using the eight dimensions. The third part is the structural transformation of the pan-scene data. The level of concept and granularity of each dimension are clarified, and the value of each attribute is encoded and quantified. The fourth part is the pan-scene data analysis. Cluster and association analyses are used to mine data rules to realize the visualization of unsafe behavior data.

Pan-Scene Data Source
Scene is the purpose of statistical data application. The occurrence of unsafe coal miner behavior can be seen as a scene. According to different manifestations, pan-scene data sources can be divided into two categories: Behavior monitoring and observation can be used as tools for internal inspection and feedback [29]. Combined with the safety management of coal mines in China, the safety managers of coal mine enterprises will take photos to obtain evidence when they conduct safety supervisions and then fill in the inspection certificate according to the process of BBS. At the same time, coal mining enterprises in China install several camera devices for video surveillance as part of the construction of intelligent mines. Therefore, a considerable amount of unsafe behavior video data is preserved. The research value of inspection photos was gradually discovered [30]. Field inspection photos and video surveillance data can fully reflect some factors of behavior, such as place and time, which is first-hand information on the unsafe behavior pan-scene data.

Abstract Scene
Safety standards and operating regulations are important for determining unsafe behavior [31]. Studies show an internal relationship between "rules and regulations" and behavior, and this is a finding that implies "a restriction of rules and regulations on the behavior" exists [32]. China regulates coal miner behavior by revising and improving the legal status of "coal mine safety regulations." Coal enterprises and relevant departments have formulated "the coal mine working standard" and "the coal miners' unsafe behavior management manual" to manage the behavior of coal miners; scholars and experts have also obtained statistics and conducted analyses on the unsafe behavior of coal miners, and they have proposed relevant management countermeasures and targeted preventive measures [33]. Furthermore, the process of the accident, the cause of the accident, the rectification measures, and the result of accident handling have been recorded in the accident investigation reports [34], which are an important source of unsafe behavior pan-scene data that includes the unsafe behavior of the human and the unsafe conditions of the material.

Description Model
Through the analysis of data sources of unsafe behavior, it is found that the difference between the recording mode of the realistic scene and the abstract scene and the difference between the realistic and abstract scenes in the form of unstructured or semi-structured data have made it difficult to describe unsafe behavior. Therefore, it is important to explore a method to express the coal mine unsafe behavioral pan-scene data and to standardize and quantify the data from different sources.
The unsafe coal miner behavior pan-scene data structured description should reflect the behavior itself as comprehensively as possible. Based on the Lasswell "5W analysis method," the unsafe coal miner behavior pan-scene data description model was proposed to describe the characteristics of unsafe behavior through practice and communication with safety management workers at the coal mine site and the summary of the related literature about unsafe behavior in this paper; moreover, this method highlighted the following eight dimensions: • Specialty: describe the stage where the unsafety behavior occurs and the specialty of the coal mine work operation.

Structure Conversion
The description of unsafe behavior in coal mines is more systematic and standardized through the model. However, it does not achieve a structured and coded quantitative description. Therefore, it is necessary to analyze the level of concept and the value of attributes for each dimension of the unsafe behavior guided by the method of quantification.

Time (T)
The continuity of time is a common characteristic in the development of anything. Some scholars have extensively studied the rules of time for human behavior, and these rules are used in various fields [35,36]. It is also important for safety management to define the rules for the occurrence time of unsafe behavior through pan-scene data mining. The time dimension has a different division for granularity and levels, such as quarter, month, day, and shift. The smaller the granularity is, the lower the level. Predictions from low levels to high levels can be achieved to improve the accuracy of information based on statistical data. Based on the characteristics of the information expressed by the source data, the structural conversion of the time dimension is shown in Table 1 combined with the reality of the coal mine. Table 1. Structured conversion of the time dimension of unsafe coal miner behavior pan-scene data.

Coding
Time Value Note: "three and eight" working form: three shifts including two working shifts and one maintenance shift, which work eight hours per shift every day; "four and six" working form: four shifts including three working shifts and one maintenance shift, which work six hours per shift every day.

Behavioral Trace (BT)
Behavioral trace indicates the traceability after the occurrence of the behavior. Based on the behavioral trace, the behavior can be divided into traced unsafe behavior and non-traced unsafe behavior. The characteristic of traced unsafe behavior is that the certain behavioral traces will be left after the occurrence of the behavior within a certain period of time; in contrast, the characteristic of non-traced unsafe behavior is that the unsafe behavior can only be found as it occurs and will not be traceable. Safety managers can use appropriate methods to infer the cause of unsafe behavior according to the characteristics of different behavioral traces. For the traced unsafe behaviors, the focus is on the identification of responsibility and the corresponding timely punishment. However, for the non-traced unsafe behaviors, the focus is to strengthen the supervision and inspection of the field. Therefore, it is particularly necessary to take behavioral traces as a dimension of unsafe behavioral pan-scene data. The structural conversion of the behavior trace dimension is relatively simple: Bt1 represents traced unsafe behavior, and Bt2 represents non-traced unsafe behavior.

Location (L)
This dimension is used to identify the location of the region where the unsafe behavior occurs. Relevant studies have shown that work environments have a significant effect on the behavioral choices of workers [37,38]. Therefore, the study on the dimensions of work sites is of great significance. Based on the characteristics of accidents in the past and the location where unsafe behavior in coal mines occurs frequently, similar to the time dimension, the regional locations are stratified according to different levels of granularity to clarify the differences in the occurrence of unsafe behavior and to improve safety management efficiency. Combined with the reality of the coal mine, the structural conversion of the location dimension is shown in Table 2. The names of the accident sites in this research are standard, for they are derived from Terms relating to coal mining (GB/T 15663-2008), National Standard of the People's Republic of China.

Behavioral Property (BP)
Behavioral property reflects the category of unsafe coal miner behavior combined with relevant research [39,40], including violation of commands, violation of operation, violation of action and non-violation unsafe behavior. Among them, violation of commands refers to the behavior of management personnel who order other people to conduct illegal operations; violation of operation refers to the behavior of the workers who violate the operating procedures to operate the specific object, such as "the sledgehammer of dangerous rock"; violation of action refers to the unsafe actions of the workers who do not involve the object, equipment or facility in the work, such as "the distance from the blasting does not meet the requirement"; and non-violation unsafe behavior refers to behavior that does not violate laws and regulations (excluding the company's internal regulations), but the action itself is unsafe and provides impetus to the occurrence of accidents, such as "without careful inspection after blasting." The structural conversion of the behavioral property is as follows: Bp1 represents violation of command, BP2 represents violation of operation, BP3 represents violation of action, and Bp4 represents non-violation unsafe behavior.

Behavioral Individual (BI)
The behavioral individual dimension refers to the different individual attributes of the worker's unsafe behavior. Factors such as age [41,42], working age, and physical status at work all affect unsafe behavior [22,43]. This paper only analyzes the three attributes of age, working years, and job type in the individual dimension. According to the division of the property of unsafe behavior and the function, behavioral individuals can be divided into three categories: managers, field commanders, and grassroots workers. The structural conversion of the behavioral individual is shown in Table 3. Age and working years can be queried in the coal company employee management system. However, the classification of coal mines for job type remains unclear. Through the investigation of key coal mines in major coal-producing regions in the provinces of Shanxi, Anhui, Inner Mongolia Autonomous Region, and Henan in China, the classification of job types is shown in Table 4 combined with the "Management Regulations for the Examination and Evaluation of Safety Technologies for Special Operators" promulgated by the "State Administration of Work Safety" on 26 April 2010 in China. Thus, the grassroots workers are coded according to job type in Table 4. The degree dimension is used to reflect the severity of unsafe behavior. Based on the more mature division method of accident classification and hidden risk classification, and the analysis and assessment of the risk [44], the definition of degree dimension is the level of risk combining the practice of coal mine safety management, which is based on the direct or indirect potential severity of the consequences of unsafe behavior. Thus, the unsafe behavior is divided into five levels: serious-risk, major-risk, medium-risk, general-risk, and low-risk levels. The structural conversion is shown in Table 5.

Unsafe action (UA)
The unsafe action dimension describes the specific unsafe behavior that may lead to accidents, casualties and environmental disruption. According to the purpose of the action, the unsafe actions are divided into four categories: safety, operation, management, and general type. The structural conversion of unsafe action is shown in Table 6.  The specialty dimension is used to represent the work stage and the specialized category of the coal miner when unsafe behavior occurs. According to the "Coal Mine Safety Risk Pre-control Management System" promulgated by the "State Administration of Work Safety" in 2011, the coal mine production system is divided into 14 management elements. Combining the characteristics of the unsafe coal miner behavior, causes of accidents, and previous research results [45], the specialty dimensions are divided into heading, mining, electric, transportation, "one ventilation and three preventions," waterproofing, blasting, and others. The structural conversion of the specialty is shown in Table 7. Table 7. Structured conversion of the specialty dimension of unsafe coal miner behavior pan-scene data.

Coding
Specialty Explanation

S1 Heading
The unsafe behavior occurs during operations such as heading, roof supporting, and refuse transportation at the tunneling working site

S2 Mining
The unsafe behavior occurs during operations such as coal mining, supporting, and transportation primarily at the coal face

S3 Electric
The unsafe behavior occurs during the operation, maintenance, and inspection of machinery and equipment

S4 Transportation
The unsafe behavior occurs during the transportation of coal or refuse at the main transportation roadway

S5
One ventilation and three preventions The unsafe behavior occurs during the ventilation, dust prevention, fire prevention, and gas prevention

S6 Waterproof
The unsafe behavior occurs during operations such as water inspection, water exploration, water discharge, water blocking, and water interception

Analysis and Visualization
The pan-scene data analysis focuses on the study of the rules of distribution for unsafe coal miner behavior and the interaction between the dimensions to realize the explicitness of the interaction relationship, which can not only deeply explore the miners' unsafe behavior in the specific coal mine enterprise but also lay a foundation for the discovery of the inherent nature of the unsafe behavior.
Before mining accident data, it is necessary to eliminate the heterogeneous nature of the data [46]. Cluster analysis is used to preliminarily explore the distribution of accidents and to prepare for multi-dimensional correlation analysis. The use of multi-dimensional interaction analysis in different clusters can deeply explore the interaction relationship between different dimensions, which is of great significance for discovering the potential characteristics of unsafe behavior.
According to the eight dimensions of the unsafe coal miner behavior pan-scene data, theoretically, any two or more dimensions can be analyzed. The interaction between different dimensions has different practical significance, which can explore the deep rules of unsafe behavior and improve safety management efficiency. Some representative dimensions for association analysis and the meaning of safety management are shown in Table 8. Table 8. An example of unsafe coal miner behavior pan-scene data association analysis. Identify when and where different individuals are prone to unsafe behavior.

Data and Structure Conversion
The accident investigation reports were used for analysis. This paper selects 225 cases of gas explosion accidents in China [47]. To ensure the accuracy of the data, this paper follows the principles of accident completeness and case authoritativeness: The completeness refers to the complete content of the accident investigation report and to facts that are expressed clearly to analyze the human factors according to the report content. Furthermore, authoritativeness refers to the accident report that must be issued by the state-accredited accident investigation or safety regulatory institutions.
Because the description of unsafe behavior in the original accident investigation report is in unstructured text records, it must be adjusted to the coal mine unsafe behavioral pan-scene data description model to achieve the structural conversion. Three coal mine gas explosion accidents were selected as an example, and the results of the description and structural conversion processes are shown in Table 9. At this point, the coal miners' unsafe behavior pan-scene data (MUBD) based on the gas accident investigation report was constructed to include a total of 871 data points as shown in Formula 1. The pan-scene data can enable coal mining companies to record the unsafe behavior of workers more accurately and provide a method for evaluating the record for unstructured or semi-structured data on unsafe behavior and a way to analyze and visualize it. At this point, the coal miners' unsafe behavior pan-scene data (MUBD) based on the gas accident investigation report was constructed to include a total of 871 data points as shown in Formula 1. The pan-scene data can enable coal mining companies to record the unsafe behavior of workers more accurately and provide a method for evaluating the record for unstructured or semi-structured data on unsafe behavior and a way to analyze and visualize it.

Clustering Method
Data mining is a fast and effective way to find unknown, implicit and potentially useful information from large-scale data to guide decision making. It is necessary to select a suitable tool to mine useful information and knowledge with strong application value from a large amount of unsafe behavior data of miners. As an unsupervised learning method, clustering plays an important role in the data natural grouping structure and has been widely used. In the efficiency of the algorithm and clustering effect, the existing cluster methods have achieved excellent performance. The k-means algorithm has been widely used for many years because of its high efficiency in data processing and good clustering effect in numerical data. However, the k-means algorithm can only evaluate a data set with continuous attributes, and the structured data presented here represent a discrete type dataset that contains many classification attributes. Thus, the k-modes algorithm is selected, which realizes the high-efficiency of the k-means algorithm while realizing the clustering of the discretetype data [48].
In the k-modes algorithm, the difference factor is used to replace the distance in the k-means algorithm, and the smaller the difference factor is, the closer the distance will be. The difference factor between a sample and a cluster center is the number of different attributes: different values are represented by 1, identical values are represented by 0, and finally, the sum of the different values is calculated. The resulting value is the difference factor between a sample and the corresponding cluster center.
Given a data set Z, each data point is described by n classification variables, and the difference factor between X and Y is calculated as follows:

Clustering Method
Data mining is a fast and effective way to find unknown, implicit and potentially useful information from large-scale data to guide decision making. It is necessary to select a suitable tool to mine useful information and knowledge with strong application value from a large amount of unsafe behavior data of miners. As an unsupervised learning method, clustering plays an important role in the data natural grouping structure and has been widely used. In the efficiency of the algorithm and clustering effect, the existing cluster methods have achieved excellent performance. The k-means algorithm has been widely used for many years because of its high efficiency in data processing and good clustering effect in numerical data. However, the k-means algorithm can only evaluate a data set with continuous attributes, and the structured data presented here represent a discrete type dataset that contains many classification attributes. Thus, the k-modes algorithm is selected, which realizes the high-efficiency of the k-means algorithm while realizing the clustering of the discrete-type data [48].
In the k-modes algorithm, the difference factor is used to replace the distance in the k-means algorithm, and the smaller the difference factor is, the closer the distance will be. The difference factor between a sample and a cluster center is the number of different attributes: different values are represented by 1, identical values are represented by 0, and finally, the sum of the different values is calculated. The resulting value is the difference factor between a sample and the corresponding cluster center.
Given a data set Z, each data point is described by n classification variables, and the difference factor between X and Y is calculated as follows: where X i and Y i are the values of attribute i in X and Y, respectively. The k-modes process is as follows: Step 1. Select k objects as the cluster center.
Step 2. Use Equation (1) to calculate the difference value of each sample and cluster center.
Step 3. Distribute the samples to the cluster with the lowest difference factor between the cluster center and the cluster center. After all samples are divided, the cluster center is redetermined.
Step 4. Repeat steps 2 and 3 until the sum of the distance between each cluster sample and each cluster center is no longer reduced and retain the final cluster results.

Association Rules Method
Association rules are the techniques applied to large-scale database mining. The core of the association rules is to show the rules of association and the correlations between different items, whicht is an implication of the form: A⇒B. A⇒B represents the association between A and B: A is the left-hand side (LHS); B is the right-hand side (RHS). Support, confidence, and lift are three important parameters in the association rules. I = {I 1 , I 2 , . . . , I m } is a set including m different items, and I is the itemset. Support is the probability that the elements included in the LHS and RHS are simultaneously present, that is, the ratio of the number of occurrences of itemsets A and B in the database (A∩B) to the total itemset, which can be expressed as: The support reflects the strength of the association rules. The minimum support is the minimum support threshold of the itemset, which is denoted SUPmin and represents the minimum importance of the association rules.
Confidence is the probability that B is extrapolated by the association rule A⇒B under the condition of containing A, which can be expressed as: The association rules with high support and confidence can more clearly illustrate the problem in general. The minimum confidence is denoted CONFmin. When the association rule A⇒B satisfies the support ≥SUPmin and the confidence ≥CONFmin, the A⇒B is called the strong association rule. The purpose of association rule mining is to find strong association rules and provide decision making assistance.
The lift reflects the size of the RHS influenced by the LHS. When the lift exceeds 1, it means that the LHS has a great influence on the RHS, and this association rule has obvious practical significance; in contrast, when the lift is less than 1, it means that the probability of RHS under the condition of LHS influence is smaller than the prior probability, and this association rule has no meaning in reality. Moreover, when the lift equals 1, it means that the leader and the successor are independent, and there is no relationship between them. The lift is the ratio of the confidence of A⇒B to the support of B, which can be expressed as: In this paper, we use the Apriori algorithm proposed by Agrawal to mine mining association rules from the clustering data set for the unsafe behavior of miners [49].

Cluster Results
The primary problem of clustering analysis is determining the number of clusters formed by clustering algorithms. In this paper, Akaike information criteria (AIC), Bayesian information criteria (BIC) and consistent AIC (CAIC) are used to judge [28]. After trying to divide several models, the data clustering number was determined to be 4. After determining the number of clusters, the k-modes algorithm was used to classify the data sets using R software. Then, each of the separated clusters was characterized. Brief descriptions of the clusters and their sizes are shown in Table 10. Such unsafe behaviors are primarily in the tunneling working site and coal face, including 78% of the high-risk behavior; additionally, approximately 53% are traced unsafe behaviors, and most of the time is distributed in March, April, November, and December, and the operation class accounted for 77%.

Cluster 2
Such unsafe behaviors in the airway and main haulage roadway account for 69%, and safety inspection work accounts for 82%. Additionally, 66% of behaviors occur in November and December, and most of the work types are gas inspectors and mine ventilators.

Cluster 3
Approximately 70% of these unsafe behaviors occurred in the skip loading pocket and underground locomotive repair room, and they involved electric specialty unsafe behavior, accounting for 56%. Additionally, the working age in 0-5 years accounted for 68%.

Cluster 4
Among these unsafe behaviors, managers and field commanders accounted for 71%, low-risk unsafe behavior accounted for 60%, general-risk unsafe behavior accounted for 58%, and the distributions of region and time were relatively wide.
The k-modes clustering method was used to analyze the unsafe behavior data of 871 miners. After detailed analysis of all the classification results, it was found that the eight dimensions describing the unsafe behavior of miners could be effectively used as the main variables of the data gathering class, and there was a significant correlation among the dimensions of the unsafe behavior of the miners.

Association Results
Selecting the data set of Cluster 1 for association analysis and taking the time and location dimensional association analysis as an example, the statistics for the number of unsafe behaviors at different locations in different months are shown in Figure 2. It is seen from the figure that the number of unsafe behaviors in the tunneling working site and coal face was highest in November. Additionally, the number of unsafe behaviors in the tunneling working site in March, April, and December was significantly larger than the number that occurred in other locations. Moreover, for the room and airway, the number of unsafe behaviors per month was relatively small compared to the other locations. Therefore, when safety managers conduct behavior interventions, such as behavioral observations, they can determine from this figure the intensity and proportion of the number of unsafe behaviors in different locations at different times. IBM SPSS Modeler 18.0 software (International Business Machines Corporation, Armonk, NY, USA) was selected as an association rule mining tool. Through the Apriori algorithm, the unsafe action dimension is used as the LHS, and the other dimensions are the RHS. The most prevalent unsafe behaviors in different dimensions are explored. SUPmin and CONFmin were 8% and 20%, respectively, and this information was used to obtain effective strong association rules. The results are shown in Table 11. Combining the meaning of the attributes in each dimension, the results of the association IBM SPSS Modeler 18.0 software (International Business Machines Corporation, Armonk, NY, USA) was selected as an association rule mining tool. Through the Apriori algorithm, the unsafe action dimension is used as the LHS, and the other dimensions are the RHS. The most prevalent unsafe behaviors in different dimensions are explored. SUPmin and CONFmin were 8% and 20%, respectively, and this information was used to obtain effective strong association rules. The results are shown in Table 11.
Combining the meaning of the attributes in each dimension, the results of the association analysis of the data are analyzed. This rule indicates that unsafe behavior occurring in November usually includes casually turning the auxiliary ventilating fan on and off. The confidence rate of this rule is 23.92%. According to this rule, the safety management efficiency can be improved 1.87-fold by behavioral intervention on workers who operate auxiliary ventilating fans casually.

Rule 2
This rule indicates that traced unsafe behavior of gas explosion accidents usually include not evacuating in the case of reduced or stopped air. The confidence rate of this rule is 35.67%. According to this rule, the efficiency of inspection of this unsafe behavior can be improved 1.76-fold.

Rule 3
This rule indicates that unsafe behavior at the coal face usually involves not examining the gas concentration at shift time. The confidence rate of this rule is 43.71%. According to this rule, the efficiency of controlling unsafe behavior can be improved 2.70-fold.

Rule 4
This rule indicates that the most vulnerable unsafe behavior of gas inspectors is blasting without examining the gas concentration. The confidence rate of the rule is 52.34%. According to this rule, the efficiency of targeted safety training for gas inspectors can be improved 2.15-fold.

Rule 5
This rule indicates that when there is a major risk of unsafe behavior, it usually involves blasting without examining the gas concentration. The confidence rate of this rule is 26.86%. According to this rule, the efficiency of controlling unsafe behavior that is of major-risk can be improved 1.89-fold.

Rule 6
This rule indicates that when unsafe behavior occurs during the heading process, it usually involves unqualified hole-sealing. The confidence rate of this rule is 44.73%. According to this rule, the efficiency of the behavior observation of the tunneling working site can be improved 1.56-fold.

Discussion
In this paper, the investigation reports of coal mine gas explosion accidents are taken as the data source. Through the theoretical framework of unsafe behavior characteristic analysis based on the pan-scene data, the eight dimensions of time, behavioral trace, location, behavioral property, behavioral individual, degree, unsafe action and specialty are used to describe unsafe behavior, and the data structure conversion is realized. Through cluster analysis and association analysis, the unsafe behavior characteristics of coal miners are obtained and the guiding significance of the results for coal mine safety management are analyzed. Coal mining remains a high-risk occupation [50], and coal mine safety data have complex, dynamic and heterogeneous, fuzzy and random characteristics [51]. Meanwhile, digital and intelligent mine construction propose new requirements for safety management work. Based on this situation, the findings in this paper have significant implications for safety research.
The process of describing miner unsafe behavior and constructing structured pan-scene data are conducive to promoting coal enterprises to collect unsafe behavior data and manage safety information resources. In addition, this information can be used to explore the value of existing resources of behavior-based safety.
The application of new technologies and the increased emphasis on safety by the government has resulted in a decrease in the fatality rate of coal mines in China [52,53]. The construction of digital and intelligent mines proposes new requirements for coal mine safety management. The intelligent safety management must pay more attention to the application and mining of data resources. The massive data sources of unsafe behavior contain a wealth of hidden knowledge of rules and laws. However, due to the lack of extraction of coal mine unsafe behavior data, these knowledge treasures cannot be used to guide safety management. However, the analysis and mining of data primarily focuses on the construction of a data mining model and algorithm and the application of computer technology [54,55]. For the large safety production dataset, these methods are relatively weak, which restricts the extraction of safety data and the improvement of safety management. In response to this type of problem, the description of pan-scene data and the process of data structure conversion in the framework greatly improve the processing capabilities for large datasets and is of great significance to the effective use of information resources and the intelligent management of a coal mine.
For coal mine safety managers, the structured processing of behavior safety data eases the burden on managers in terms of collating photos and records of unsafe behavior. In addition, the statistical analysis and multi-dimensional association analysis results help safety managers to clarify the characteristics and rules of unsafe behavior of workers, which improves the efficiency of safety management and optimizes the allocation of management resources.
The occurrence of unsafe coal miner behavior has intrinsic complexity and is greatly affected by the natural environment, coal mining technology, personality traits, and management level [22]. Using clustering analysis, the potential rules among the dimensions of unsafe miner behavior can be preliminarily excavated, and the heterogeneity of much unsafe miner behavior data can be effectively eliminated at the same time. It is of great significance to determine the potential features of unsafe behaviors using association analysis to explore the relationship between the different dimensions in a particular single cluster. The interaction between different dimensions has different meanings. Safety managers can select the different dimensions according to their needs and judge whether there is regularity between them through the interaction results to improve the efficiency of safety management and stimulate the advancement of empirical safety management to intelligent safety management.
For the workers of coal mining enterprises, the visual presentation of the data analysis results has improved their level of information cognition. The results of cluster analysis and association analysis of eight dimensions are mostly displayed in visualization, which has characteristics that can be summarized as visibility, interactivity, and multidimensionality. Based on the perspective of safety cognition and the different individual attributes, the results will be provided to the workers to improve their information absorption and analysis ability and their safety behavior selection ability.

Conclusions
The theoretical framework for analysis of unsafe behavior characteristics based on pan-scene data in this study has very important application value. The requirement of the pan-scene data theory to the accident data source can maximize the comprehensive description of human unsafe behavior factors. The accuracy and completeness of unsafe behavior data are guaranteed. Thus, the unsafe behavior information associated with the accident can be extracted more thoroughly. The data structure conversion process can be used not only to follow-up analysis but also to improve the management and storage of accident data in coal mining enterprises.
Through cluster analysis and correlation analysis, the safety managers of coal mining enterprises can directly understand the probability distribution of all kinds of accidents and the interrelations and interaction rules between the various dimensions of the unsafe behavior of the workers. In practical work, the analysis results of unsafe behavior characteristics can provide a scientific basis for the rational allocation of safety management and accident prevention resources and the adjustment of safety training programs in coal mining enterprises. These targeted guidance functions are conducive to greatly improving the safety performance of coal mining enterprises.
Based on the study of unsafe behavior of miners [19,20], this paper realizes the structure conversion of data based on the theory of pan-scene data [24][25][26] and introduces the framework thought of data characteristic analysis into the field of coal mine [27,28]. A framework for analyzing unsafe behavior characteristics of workers in coal mines is formed, and the feasibility and effectiveness of the framework are verified by accident cases. In this paper, there is a solid theoretical foundation and a sufficient empirical analysis, the description of data sources is more comprehensive, and the design of structure conversion and characteristic analysis process are more suitable for coal mines. A new method and thought for the data analysis of workers' unsafe behavior is provided in coal mines.
This paper hopes to contribute to the characteristics analysis of unsafe behaviors of coal mining enterprises by proposing a data processing and analysis method. However, there are limitations in this study. The data sample selected in this paper is only the coal mine gas explosion accident from 2001 to 2016 in China. There are various types of coal mine accidents with different characteristics. Therefore, the results of this study have certain limitations, which will become the focus of our subsequent research.