Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 11, No. 3, 2025 

 

63 

Construction	and	Implementation	of	Coal	Mine	Safety	
Knowledge	Graph	Based	on	Historical	Accidents	
Minmin Li *, Yanyu Diao, Xin Wang 

School of Big Data and Artificial Intelligence, Anhui Xinhua University, Hefei Anhui, 230088, China 

* Corresponding author: Minmin Li (Email: 464130929@qq.com) 

 

Abstract: With the rapid development of intelligent coal mine, the field of coal mine also ushered in a large number of 
heterogeneous data information. Aiming at the data information existing in the field of coal mine, a coal mine safety knowledge 
atlas centered on coal mine accident cases is constructed. The construction of knowledge graph mainly uses crawler technology 
to obtain text data of coal mine, Albert-BilSTM-CRF model to realize entity recognition, Albert-BilSTM-ATT to complete 
relation extraction, and obtain triplet form in the field of coal mine safety. Finally, the obtained triplet form is imported into 
Neo4J graph database. Form a structured knowledge graph. 

Keywords: Coal Mine Accident; Entity Recognition; Relation Extraction. 
 

1. Introduction 
Coal mining industry, as an important economic pillar 

industry in China, has always been vigorously developed by 
the state, and while the coal mining industry is rapidly 
transforming, its safety problems have also been attached 
great importance. Although in recent years in the country's 
continuous transformation of the coal mining industry, coal 
mine safety accidents have been reduced year by year, but in 
order to better prevent the occurrence of coal mine accidents, 
this paper through the coal mine safety network on the coal 
mine accident case text data, to build the knowledge graph in 
the field of coal mine safety, through the knowledge graph 
will be associated between the accident, so as to analyze it, to 
find out the similarities between the accidents, and to avoid 
similar accidents from reoccurring. 

In this paper, the accident cases in the coal mine safety 
production network are used as the data source for the 
construction of coal mine safety knowledge graph, and the 
text mainly analyzes the common seven kinds of coal mine 
accident cases, including: gas accidents, roof accidents, 
transportation accidents, electromechanical accidents, water 
penetration accidents, fire accidents, and coal dust explosion 
accidents. Data crawling is realized by using java language, 
combined with deep learning algorithm ALBERT-BiLSTM-
CRF model, ALBERT-BiLSTM-ATT model realizes entity 
recognition as well as relationship extraction of coal mine text 
data, and the ternary data obtained by relationship extraction 
is saved in CSV format, then imported into Neo4j database, 
and through Neo4j graph database to visualize the coal mine 
safety knowledge graph. 

Knowledge graph is a knowledge base in the form of a 
graph that represents the relationships between entities in the 
real world [1]. Through data mining technology and then 
analyze, organize, and discover the association between 
entities and entities, which constitutes a semantic network 
graph. With the gradual maturity of the development of 
knowledge graph technology, knowledge graph has been 
gradually applied to various fields. Peng Yunong[2] and 
others applied knowledge graph in the field of corn 
agriculture planting, using knowledge graph technology to 
build the completed corn production process, knowledge 

graph can enhance the efficiency of users to obtain knowledge 
of corn planting, and improve the management level of corn 
planting. Zeng Ziling[3] and others stored massive medical 
knowledge in the form of mapping, which reduces the loss of 
information, preserves the original meaning of knowledge, 
and contributes to the in-depth application of knowledge in 
the Chinese medicine industry. Pu Tianjiao[4] and others fully 
mining the data information in the power Internet of Things, 
to achieve the construction of knowledge map in the field of 
power, the concepts, entities and events that exist in the power 
system through the map of the relationship between the 
structured form of display, the proposed method makes the 
power industry chain has a more effective cross-media big 
data organization that is management. 

As the coal mining industry data information is messy and 
heterogeneous, the content is not only extensive but also 
complex. Therefore, the application of knowledge graph to 
the coal mining industry is less and incomplete. Li Zhe[5] and 
others mainly focus on the class of coal mine mechanical and 
electrical equipment accidents, through the analysis of coal 
mine mechanical and electrical historical accident cases, 
analyze the historical data of coal mine mechanical and 
electrical equipment accident cases, and correlate the 
accidents obtained, the equipment involved in the accidents, 
the causes, and the treatment measures in the face of the 
accidents, to achieve the diagnosis of coal mine mechanical 
and electrical equipment accidents and risk management. Liu 
Peng[6] and others made a preliminary exploration of 
intelligent query for coal mine safety knowledge graph 
construction by introducing knowledge graph technology into 
the field of coal mine safety, categorizing and modeling 
relevant knowledge concepts, and proposing a language-
based query method based on the constructed knowledge 
graph. Based on the above research, this paper proposes to 
construct a knowledge graph in the field of coal mine safety 
based on seven types of historical coal mine accident cases, 
to correlate the same accidents and analyze the implied 
information between the accidents, so as to better prevent the 
occurrence of coal mine accidents. 



 

64 

2. Coal Mine Safety Knowledge Graph 
Construction Process and 
Techniques 

Knowledge graph construction generally adopts both top-
down and bottom-up approaches [7]. Top-down construction 
of the knowledge graph needs to first define the ontology and 
data schema, and add new entities to the existing knowledge 
base, i.e., top-down construction can continuously add new 
entities to the existing knowledge base, and construct the 
knowledge graph for the existing knowledge base. Bottom-up 
construction first obtains data from various web pages, and 
after processing, screens out entities and relationships that 
meet the requirements and adds them to the knowledge base. 
In this paper, the bottom-up construction method is chosen for 
the construction of knowledge graph in the field of coal mine 
safety. The construction process is shown in Figure 1. 

 

 
Figure 1. Construction process of knowledge map of coal mine 

safety 

2.1. Coal Mine Accident Text Data Acquisition 
and Processing 

Acquisition of coal mine accident text data is the first step 
of constructing knowledge graph, the acquisition of data in 
this paper is mainly through the crawler technology to the coal 
mine accident cases in the coal mine safety network, the use 
of java language to realize the crawler technology, through 
the connection of the webpage URL, the webpage is identified 
cyclically until the termination condition is reached, the 
specific process is shown in Figure 2. 

 
Figure 2. Data acquisition flow chart 

Each web page page is parsed as shown in Figure 3. Get 
every incident link on every page with “<div 
class="imgr">...</div>” and then visit each incident link to 
get the specific text content. 

 

 
Figure 3. Coal mine accident case webpage source code 



 

65 

 
The text data obtained through the web crawler are saved 

in txt text, because the data obtained by the crawler will have 
some repetitive text, as well as the text contains a lot of 
useless punctuation, in order to improve the quality of the 
subsequent entity recognition as well as relationship 
extraction, so it is necessary to clean the text content. 

2.2. Coal Mine Accident Data Entity 
Recognition 

Text entity recognition of coal mine accidents is a key step 
in the construction of coal mine safety knowledge graph. As 
a research hotspot in natural language processing, entity 
recognition can be roughly divided into three categories : 
named entity recognition based on rules and dictionaries, 
named entity recognition based on statistical models and rules, 

and named entity recognition based on deep learning. In 
recent years, with the development of natural language 
technology, deep learning has been widely used in the field of 
named entity recognition, and has achieved good results. 
Therefore, this paper uses the named entity recognition 
method based on deep learning to obtain the entities in the 
field of coal mine safety. According to the characteristics of 
coal mine accident text, eight entities are set up, namely : time, 
place, mine name, number of casualties, fault phenomenon, 
fault location, direct cause and indirect cause. And the BIO 
format is used to label the acquired coal mine data, where B 
represents the beginning of the entity, I represents the middle 
or end of the entity, and O represents the non-entity. The 
ALBERT-BiLSTM-CRF model is used to extract entities 
from data texts in the field of coal mine safety. The model is 
shown in Figure 4. 

 

 
Figure 4. ALBERT-BiLSTM-CRF model 

 
As shown in Figure 4, the whole model is divided into three 

parts: ALBERT layer, BiLSTM layer and CRF layer.  
(1) The ALBERT layer is used to obtain the word vector 

( 1x  , 2x  , 3x  … nx  ) of the sentence as the input of the next 

layer of the model. 
(2) The BiLSTM layer is used as the coding layer of the 

model. By combining the output of the forward LSTM with 
the output of the backward LSTM, the context semantics of 
the input sentence is obtained, and the implicit state sequence 
of the input sentence is obtained, such as formula (1). 

*
1 2( , , ) n m

nh h h R              (1) 

(3) The CRF layer is used to receive the output of the 
BiLSTM layer. The output of the BiLSTM layer is input to 
the CRF layer to obtain the label sequence of the sentence 

1 2( , , )ny y y y  , where the label score of y is : 

1
1 1

( ) ( , , , )
m n

i j i i
j i

s y x f x i y y 
 

       (2) 

In the formula, i  represents the position of the word in 

the sentence, iy  is the label of the current word, 1iy  is the 

label of the previous word, m corresponds to the number of 
features, n represents the length of the input sentence. 

The obtained score is converted into a probability value 
after normalization, and the maximum probability value in the 
output result is used as the final sequence label. 

When training on the ALBERT-BiLSTM-CRF model, the 

standard used in the model is Precision, Recall, 1f  -score, 

and the specific formula is : 
TP

P
TP FP




                 (3) 

TP
R

TP FN



                 (4) 

1

2* *P R
f

P R



                 (5) 

 
Table 1. Experimental training results 

Model P R F1 

ALBERT-BiLSTM-CRF 71.05 70.26 70.65 

BERT-BiLSTM-CRF 62.43 61.73 62.08 

BiLSTM-CRF 58.13 60.14 59.12 

 
In order to prove the validity of the model proposed in this 

paper, two sets of comparative experiments are selected to 



 

66 

compare with the model. Through the training of 
experimental data, the results are shown in Table 1. 

It can be seen from Table 1 that the ALBERT-BiLSTM-
CRF model proposed in this paper has an increase of 8.62 % 
in P value, 8.53 % in R value, and 8.57 % in F1 value 
compared with the BERT-BiLSTM-CRF model. Compared 
with the BiLSTM-CRF model, the ALBERT-BiLSTM-CRF 
model increased the P value by 12.93 %, the R value by 
10.12 %, and the F1 value by 11.53 %. Therefore, ALBERT-
BiLSTM-CRF has a better recognition effect on named entity 
recognition in the field of coal mine safety. 

2.3. Entity Relationship Extraction 
Relation extraction is an important part of knowledge graph 

construction. In this section, ALBERT-BiLSTM-Att model is 
used to extract text data of coal mine accidents, and the 
relationship between entities is stored in the form of triples. 
According to the entity category, eight relationships can be 
defined, including accident time ( the time when the accident 
occurred ), accident location ( the place where the accident 
occurred ), accident coal mine ( the name of the mine where 
the accident occurred ), result ( the number of deaths caused 
by the accident ), specific location ( the working face where 
the accident occurred ), accident phenomenon ( the 
phenomenon of accident occurrence ), primary cause ( the 

direct cause of the accident ), and secondary cause ( the 
indirect cause of the accident ). 

The model algorithm mainly uses the ALBERT layer to 
complete the pre-training task, and obtains a low-dimensional 
dense vector containing word feature information. The 
BiLSTM layer is used to obtain the semantic context features, 
and the input word vector is bidirectionally encoded to obtain 

the output 1 2 3[ , , , ]Nh h h h h  . The Att layer is used as 

the last layer of the model, which is mainly used to assign the 
weight of the words in the sentence. The greater the 
contribution to the entire sentence, the higher the weight 
obtained. The specific calculation formula is : 

i q i qq W h b                   (6) 

i k i kk W h b                   (7) 

i v i vv W h b                   (8) 

In the formula, iq , ik , iv  are query vector, key vector, 

value vector, qW  , kW  , vW   is the corresponding 

parameter matrix, qb  , kb  , vb   is the corresponding 

parameter matrix. The specific calculation process of ATT is 
shown in Figure 5. 

 

 
Figure 5. Self-attention mechanism calculation process 

 

As shown in Figure 5, where iq , ik  are used to calculate 

the correlation between the two vectors, the formula is as 
follows : 

( , ) ( )Ti j i js q k q k                  (9) 

Normalize ( , )i js q k  and calculate ,i ja . 

,

1

exp( ( , ))

exp( ( , ))

i j
i j N

i tt

s q k
a

s q k





            (10) 

Through the formula calculation in the above steps, the 
final output vector is: 

,
1

N
att
i i j j

j

h a y


                  (11) 

The model algorithm flow chart is shown in Figure 6. 
The experimental data are trained by ALBERT-BiLSTM-

Att model, and the evaluation indexes of the experimental 
results are still taken to measure P, R, and F1 in 2.2, and the 

experimental results of 8 relationship categories are obtained, 
as shown in Table 2. 

 
Table 2. Extraction results of relation types 

Relation type P R F1 

Occurrence time 78.63 76.12 77.38 

Accident location 80.00 81.02 80.51 

Accident mine 84.15 82.13 83.13 

Result in 72.16 70.84 71.50 

Specific location 63.42 61.05 62.21 

Accident 
phenomenon 

60.72 58.26 59.46 

Primary cause 54.71 56.87 55.77 

Secondary cause 51.67 53.29 52.47 

 

 

 



 

67 

 

 
Figure 6. Relational extraction algorithm model 

 

As can be seen from Table 2, the P, R, and F1 of the primary 
and secondary causes are lower, mainly due to the fact that 
the primary and secondary cause sentences are longer and 
more complex, so the P, R, and F1 obtained when performing 
the relational extraction are lower. The experimental results 
of the ALBERT-BiLSTM-Att model are shown in Table 3. 

 
Table 3. Results of model experiments 

  P R F1 

BERT-BiLSTM-Att 78.63 76.12 77.38 

ALBERT-BiLSTM-CRF 54.71 56.87 55.77 

ALBERT-BiLSTM-Att 51.67 53.29 52.47 

 
In order to prove the effectiveness of the proposed model, 

three groups of comparison experiments are set up, which can 
be seen in Table 3 that the ALBERT-BiLSTM-Att model 
improves P, R, and F1 by 7.05%, 0.74%, and 3.86%, 
respectively, compared with the ALBERT-BiLSTM-CRF 
model, which shows that the introduction of the self-attention 
mechanism model in the relational extraction and the 
enhancement of the keyword's weight in the sentences can be 
effectively improve the performance of relation extraction. 
Comparing the BERT-BiLSTM-Att model with the ALBERT-
BiLSTM-Att model, the latter has a slightly higher relational 
extraction performance in the field of coal mine safety, and it 
can be seen that the ALBERT model has a better model 
performance in relational extraction compared to the BERT 
model. 

3. Coal Mine Safety Knowledge Graph 
Visualization 

In this paper, coal mine accident text case data is used to 
get coal mine ternary data through two techniques of entity 
recognition and relation extraction, and Neo4j database is 
used to construct coal mine safety knowledge graph, which 
associates all the same accident cases as well as displays them 
in the form of graphs. The constructed graph contains 8 kinds 
of entities and corresponding 8 kinds of relationships, using 

Neo4j to store the ternary, and querying the trend of accidents 
in a certain region through Cypher. 

In the coal mine safety knowledge graph, all the same 
accidents are associated with each other, so that practitioners 
can view all the same accidents at the same time and analyze 
them to get the hidden information in the accidents, so as to 
better prevent accidents from occurring. 

4. Conclusion 
This paper takes seven kinds of coal mine accident text data 

as the research object, processes the unstructured coal mine 
text data through entity recognition technology and 
relationship extraction technology, gets the structured ternary 
data imported into the Neo4j database, constructs the coal 
mine safety knowledge mapping, and visualizes the coal mine 
knowledge in the form of mapping. By visualizing the map, 
we can analyze the correlation between the same accidents 
and prevent the recurrence of accidents. Due to the 
preliminary construction of the coal mine safety knowledge 
map, there are deficiencies in the paper, the coal mine 
knowledge is not comprehensive enough, and we will 
continue to improve and expand the coal mine knowledge 
base in the future to build a more complete coal mine safety 
knowledge map. 

Acknowledgments 
The authors gratefully acknowledge the financial support 

from Anhui Provincial Education Department's Higher 
Education Quality Project 'Artificial Intelligence New 
Professional Quality Improvement Project' (2023xjzlts088) 
and Anhui Xinhua University's school-level scientific research 
project (2024zr016). 

References 
[1] Ma Rui-Xin, Li Ze-Yang, Chen Zhi-Kui, et al. Review of 

reasoning on knowledge graph[J]. Computer Science, 2022, 49 
(S1): 74-85. 



 

68 

[2] Peng Yunong, Liu Pingzeng, Zhang Yan. Construction of 
knowledge mapping of maize production processes based on 
deep learning [J]. Journal of Chinese agricultural 
Mechanization, 2025, 46(2): 245-252. 

[3] Zeng Ziling, Zhang Huamin, Yu Tong, et al. Review on the 
Research and Application of Knowledge Graph and Its Key 
Technologies in the Field of Traditional Chinese Medicine[J]. 
World Science and Technology -Modernization of Traditional 
Chinese Medicine, 2022, 24(02): 780-788. 

[4] Pu Tianjiao, Han Xiao. Key Technologies of Artificial 
Intelligence Application in New Power Systems [J]. Electric 
Power Information and Communication Technology, 2024, 
22(01): 1-13. 

[5] LI Zhe, ZHOU Bin, LI Wenhui, et al. Construction and 
application of mine electromechanical equipment accident 
knowledge graph[J]. Industry and Mine Automation, 2022, 
48(1): 109-112. 

[6] LIU Peng, YE Shuai, SHU Ya, et al. Coalmine safety: 
knowledge graph construction and its QA approach [J].  
Journal of Chinese Information Processing, 2020, 34(11): 49-
59. 

[7] MAOR B, ZHUJ, LI A W, et al. Construction of knowledge 
graph of industry chain based on natural language processing 
[J]. Journal of the China Society for Scientific and Technical 
Information, 2022, 41(3): 287-299.