Architecture and Application of Mine Ventilation System Safety Knowledge Graph Based on Neo4j

Abstract

To improve the safety management and accident prevention capabilities of mine ventilation systems, the application of knowledge graph technology is proposed. By employing methodologies such as data analysis, entity relationship definition, and entity relationship extraction, and entity extraction using BERT + BiLSTM + CRF model, a safety knowledge graph for the mine ventilation system is constructed. This facilitates the structured processing of historical accident-related textual data and enables the visual analysis and application of accidents based on the knowledge graph. The research results demonstrate that knowledge graph technology can effectively integrate unstructured data and present it in visual graphs or tables. By utilizing Cypher query statements, multi-dimensional accident statistics and the frequency analysis of specific information can be generated, contributing to a comprehensive understanding of accident occurrence patterns. Leveraging the node-to-node characteristics of the knowledge graph, a correlation analysis between entities is conducted, deeply exploring relationships among different types of data, thereby providing new insights to prevent accidents in mine ventilation systems. Moreover, the analysis of mine ventilation accidents and system failure characteristics offers valuable guidance for the safety management of mine ventilation systems.

1. Introduction

Mine ventilation systems are essential for maintaining a safe underground environment, boosting productivity, and safeguarding worker health, making them a critical component of mine operations [1]. Inefficient ventilation systems can contribute to hazardous situations, such as gas accumulation, oxygen deficiency, and poor air quality, all of which are major contributors to mining accidents [2]. For instance, the 2002 Shanxi coal mine explosion in China, which claimed the lives of over 200 miners, was caused by inadequate ventilation, leading to the accumulation of explosive gases. In contrast, the 2010 Copiapó mine disaster in Chile underscored the critical role of the ventilation system. When 33 miners were trapped for 69 days, the mine’s ventilation system ensured a continuous supply of breathable air, preventing oxygen deficiency and sustaining the miners’ survival underground, ultimately contributing to their successful rescue. Consequently, the efficient management and performance enhancement of mine ventilation systems have consistently been a focal point of interest for both industry and academia. Studies include reviews of ventilation monitoring and control technologies [3], the development of quantitative assessment methods for airflow stability [4], and dynamic ventilation modeling that incorporates thermal and gas pollutant simulations [5]. Additionally, the following were proposed: ventilation management procedures to ensure regulatory compliance [6], the use of ventilation-on-demand technologies [7], and the optimization of oxygen supply through computational fluid dynamics [8]. Furthermore, research by [9] explored the integration of real-time data monitoring with predictive analytics to improve system reliability, all aimed at enhancing the safety and economic efficiency of ventilation systems in mines. However, most of these efforts have primarily concentrated on the optimization of ventilation parameters and system configurations, with relatively little attention paid to the comprehensive management and analysis of accident-related data within ventilation systems.

Traditionally, the safety management of mine ventilation systems relies on manual inspection records, spreadsheet-based documentation, and relational database systems. These conventional methods typically focus on monitoring key parameters, such as airflow volume, gas concentration, and equipment status [10,11]. However, they face significant limitations when handling heterogeneous, complex, and dynamically evolving safety data. Specifically, rigid data structures, limited scalability, and reliance on manual analysis make it challenging to integrate multi-source information and uncover implicit relationships among personnel, equipment, and environmental factors. These limitations can result in inefficiencies in accident analysis, knowledge discovery, and decision-making processes.

As a novel database architecture, knowledge graphs organize complex information in a structured manner and comprehensively uncover intricate relationships within data texts [12,13,14]. This demonstrates widespread applicability across various industries, including, but not limited to, smart factories, smart mines, and the digital transformation of healthcare and finance [15]. Liu et al. [16] explored the application of knowledge graphs, specifically within risk pre-control management systems, demonstrating how knowledge graphs can improve risk identification and decision-making in high-risk industries like underground coal mining. Liu et al. [17] also investigated the use of knowledge graphs for enhancing situation awareness and decision-making in industrial control systems, highlighting the benefits of this technology in improving system security and operational efficiency. Zhang et al. [18] used knowledge graph technology to establish a knowledge graph system for coal mine equipment maintenance, addressing the need for a question-and-answer system in this domain. Jiang et al. [19] constructed a medical question-answering system based on knowledge graphs, which improves the accuracy and efficiency of responses to medical questions through entity linking, intent classification, and question–answer matching. Experimental validation demonstrated its effectiveness and practicality. Hu et al. [20] studied the storage and analysis of building fire cases using knowledge graph technology and concluded that this approach could significantly improve the display and utilization of fire information compared with traditional text-based storage, thus enhancing the effectiveness of fire accident analysis. Liu et al. [21] proposed a Chinese mineral question-and-answer system based on a mineral knowledge graph to retrieve mineral entities and relationships. The system achieved an accuracy rate of 91.2% for 2000 test questions. Pei et al. [22] constructed a deposit knowledge graph based on raw text data from different gold deposits in a gold belt, and they proposed methods for the visual construction of the deposit knowledge graph and for calculating the similarity between deposits based on their metallogenic geological characteristics.

The Neo4j graph database, a sophisticated NoSQL graph database system, enables the construction of highly structured graph databases by defining nodes, relationships, and properties [23]. It is widely used for building and storing knowledge graphs. Saad et al. [24] developed a semantic graph database for Life Cycle Inventory using Neo4j (version 5.2.0), which provides a versatile, queryable, scalable, intuitive, and interchangeable data format. Tuck [25] summarized research by integrating data on non-small cell lung cancer and constructing a graph database to study genomic variations and their relationships with clinical outcomes. Compared with traditional safety management methods based on manual data analysis, spreadsheets, or relational databases, knowledge graph technology offers several key advantages [26,27]. First, it allows for the semantic representation of unstructured and multi-source data, overcoming the rigidity and schema limitations of relational databases [28,29]. Second, the graph-based structure of Neo4j enables the explicit modeling of complex relationships among personnel, equipment, environmental factors, and procedures, which are often implicit or overlooked in conventional systems [30]. Third, knowledge graphs support real-time updates and dynamic expansion, making them ideal for managing evolving safety information [31]. These characteristics make Neo4j-based knowledge graphs a robust solution for addressing the limitations of traditional approaches in mine ventilation systems. Furthermore, Neo4j was selected over other available graph database solutions (such as JanusGraph, OrientDB, or GraphDB) due to its balanced combination of querying efficiency, ease of use, stable performance, and robust visualization capabilities [18,32]. Its intuitive Cypher query language facilitates flexible and readable complex queries, supporting both deep correlation analysis and traceability analysis. These features align well with the specific scale, complexity, and operational requirements of mine ventilation system safety data, ensuring practical applicability and system scalability.

To enhance the integration of textual data related to mine ventilation systems, we applied knowledge graph technology and the Neo4j graph database to the safety management of these systems. By collecting and organizing relevant accident records, we constructed a knowledge graph for mine ventilation system safety, transforming unstructured data into structured formats. This provided essential structured data for the safe management of mine ventilation systems.

2. Construction of the Knowledge Graph for Mine Ventilation System Safety

2.1. Construction of the Knowledge Graph Framework

A knowledge graph, functioning as a structured semantic network, efficiently represents, organizes, and utilizes entities, concepts, and their interrelationships in the real world. The fundamental unit of a knowledge graph is the triple, which is typically structured as <Entity1, Relationship, Entity2>, and is used to represent the relationships between entities [33]. Entities are fundamental units within a knowledge graph, representing “objects” or “concepts” in the real world. Each entity typically has a unique identifier and a set of attributes. Relationships connect two or more entities, define the type of association or interaction between them, and are generally characterized by a specific directionality. Attributes are data that describe the characteristics of entities or relationships, providing additional information.

The basic process of constructing a knowledge graph is as follows: First, an appropriate method is selected to build the knowledge graph framework based on the specific application scenario. Subsequently, relevant data are collected and processed, with knowledge elements such as entities and relationships extracted using knowledge extraction techniques, and then are represented as triples. Finally, the knowledge graph is constructed using triple-formatted data. A variety of applications, such as semantic searches, question–answering systems, and recommendation systems, can be developed based on knowledge graphs. Additionally, the knowledge graph can also be continuously optimized and updated for practical applications [34]. The specific process is shown in Figure 1.

The construction of a knowledge graph primarily involves two main approaches: top-down and bottom-up [35]. The top–down approach begins with the highest-level concepts and incrementally refines them to establish a well-structured taxonomic hierarchy. Entities are then added to the knowledge base according to these predefined patterns. This method is often suitable for building knowledge graphs in specific or vertical domains, as it effectively represents the hierarchical relationships between concepts. However, designing and maintaining a comprehensive ontology upfront requires extensive domain expertise, which can be time-consuming and inflexible, especially when dealing with dynamic or heterogeneous data sources.

In contrast, the bottom–up approach starts with the data themself, gradually refining and expanding the graph to ensure accuracy and flexibility. Given that the raw data collected in this study consist of unstructured accident reports, characterized by diverse formats, incomplete information, and the lack of predefined data models, the top–down approach presents significant challenges in modeling. The bottom–up approach, however, offers the following several advantages in this context: (1) it accommodates the heterogeneity and variability of the accident text data without requiring a rigid predefined schema; (2) it enables the automatic discovery and validation of relationships and patterns inherent in the data, which are crucial for accurately capturing the complex interactions among personnel, equipment, environment, and procedures in mine ventilation systems; and (3) it reduces the reliance on extensive manual input from domain experts, thus lowering modeling and maintenance costs. Therefore, to ensure flexibility, scalability, and the ability to handle the inherent complexity of the data, the bottom–up approach is adopted for constructing the mine ventilation system safety knowledge graph.

2.2. Data Sources

The construction of the mine ventilation system safety knowledge graph is primarily grounded in historical accident data related to mine ventilation systems, specifically those occurring during mining operations as a result of unsafe conditions of the mine ventilation system equipment or unsafe behaviors of employees. Based on the causes of accidents, incidents related to mine ventilation systems can be categorized into three main types:

• Accidents involving gas, coal dust, or asphyxiation that occurred due to the failure to activate or the improper configuration of ventilation facilities, such as air ducts and auxiliary fans, which prevented the ventilation system from functioning properly.
• Accidents occurring during maintenance work on ventilation equipment, where unsafe behavior by workers led to incidents such as being struck by objects or suffering mechanical injuries.
• Accidents resulting from mechanical failure or aging of ventilation equipment, leading to mechanical injuries.

The accident data were primarily obtained from the Safety Management Network (https://www.safehoo.com/), a platform dedicated to collecting safety-related information, including accident cases from various industries, commonly used safety education materials, and the latest safety policies. The raw accident data were crawled using Octoparse 8 software. A total of 404 accident reports were retrieved as raw data by performing a full-text search using the keyword ‘ventilation’. After manually reviewing the reports, only those explicitly involving mine ventilation system failures or contributing factors were retained, based on the defined classification criteria, resulting in 202 complete accident reports. It is important to note that the dataset may exhibit potential biases due to its reliance on publicly available online sources. Specifically, limitations in temporal coverage, geographic distribution, and data completeness may exist, as the dataset reflects records published on the website during the collection period. Despite these constraints, the dataset provides a representative basis for analyzing ventilation system safety issues, and future work will focus on expanding data sources to enhance coverage and reduce potential biases.

Most accident reports are comprehensive investigation reports characterized by a clear structure, including six main sections: an overview of the situation, basic information about the accident unit, a description of the accident’s occurrence and rescue process, an analysis of the accident’s causes and nature, recommendations for handling the personnel and units responsible for the accident, and suggestions for accident prevention and corrective measures.

2.3. Definition of Entity Labels and Relationship Labels

Based on the basic structure of the raw accident text data and the characteristics of accident analysis, the entity labels were divided into the following three parts, with a total of twenty entity label categories defined (Figure 2):

• Basic Information: This section provides an overview of the situation and basic information about the accident unit from the raw accident text, clearly reflecting the general circumstances of the accident. The defined entity labels include the accident name, type, time, location, mine involved, casualties, and economic losses.
• Specific Causes: This section focuses on the causes and nature of the accident, as described in the original text, emphasizing unsafe behaviors of individuals and unsafe conditions of objects. The defined entity labels include direct cause, indirect cause, spatial location, personnel, work environment, equipment, ventilation equipment, process phenomena, and work tasks.
• Recommendations and Preventive Measures: This section pertains to the recommendations for handling responsible personnel and units, as well as accident prevention and rectification suggestions in the original text. The defined entity labels include penalized personnel, penalties, preventive measures, and regulations and standards.

Following the definition of entity labels, the relationship labels were also defined and classified into the following two types, resulting in a total of sixteen relationship label types (Figure 2):

• Relationships Centered on the Accident Name (Part1): This type focuses on the accident name as the central entity, establishing relationships with other entity labels related to the basic information of the accident. For example, in the triple <Accident Name, Relatype, Accident Type>, “Accident Name” and “Accident Type” represent entity labels, while “Relatype” denotes the relationship type connecting these two entities. This structure follows the standard (Entity1, Relation, Entity2) format used in knowledge graph representation.
• Relationships Based on Specific Causes (Part2): This category involves relationships typically associated with multiple factors, such as personnel, equipment, environment, and specific locations. The following four types of relationship labels are defined: “cause”, “relate to”, “send out”, and “use”.
  • Cause: This label denotes a direct or indirect causal relationship between two entities. For example, the entities Direct Reason, Indirect Reason, Environment, and Exact Place may act as causal factors leading to a particular phenomenon. It captures how specific environmental conditions, work procedures, or locations contribute to the occurrence of incidents.
  • Relate to: This label captures associative relationships without implying direct causality. In the diagram, entities such as Person1, Rule, and Phenomenon are connected through the “relate to” relationship, indicating that, while they influence each other, they may not be the direct cause of an incident. It helps illustrate the broader operational context.
  • Send out: This label describes cases where an entity actively emits, triggers, or generates a phenomenon. For instance, Person1 and Equipment may “send out” a specific phenomenon, such as a hazardous condition or observable event, emphasizing their role in initiating certain outcomes.
  • Use: This label represents functional usage relationships. Specifically, Person1 may “use” entities such as Ventilation or Equipment, highlighting how human interaction with tools or facilities impacts safety and operations.

2.4. Data Preprocessing and Labeling

First, the 202 retained accident text records were pre-processed. Using Python, unnecessary characters, spaces, and line breaks within the texts were removed. Additionally, irrelevant content in the accident reports—such as section or paragraph headings, as well as descriptions of emergency response procedures—was deleted. The focus was placed on preserving three key sections: the detailed accident description, the accident occurrence process, and consequence management.

Subsequently, the pre-processed text data were annotated for entities and relationships using the “Label Studio Assistant” software(version 2.0.4), following the predefined 20 entity label categories and 16 relationship label categories. Upon completion of the annotation process, the labeled data were exported directly in ann format. The annotated data were then converted into the BIO labeling format using Python.

2.5. Knowledge Extraction

Knowledge extraction is the process of automatically extracting structured information from unstructured or semi-structured data sources. It involves identifying entities and relationships from raw data, often using techniques such as natural language processing, machine learning, and data mining [36]. In entity extraction tasks, commonly used models include BERT, CRF, and BiLSTM.

2.5.1. Entity Extraction

To efficiently perform entity extraction, this study compared BERT, BERT + CRF, and BERT + BiLSTM + CRF models. The performance of these models was primarily evaluated using Precision, Recall, and F1-score metrics [37].

The proposed model was implemented using the TensorFlow framework, within an experimental environment configured with an Intel Core i5-8250U CPU (1.60 GHz, 4 cores), Python 3.7, and TensorFlow 1.14.0. The dataset was randomly split into training, validation, and test sets at ratios of 7:1.5:1.5. The parameter settings for the BERT + BiLSTM + CRF model developed in this study are presented in

Table 1. BERT + BiLSTM + CRF model parameter settings.

Model Parameters	Model Parameter Values
Epoch	20
Max_len	70
Batch_size	16
Lstm_units	128
Drop_rate	0.1
Learning_rate	1 × 10−5

.

The results of the entity extraction task are presented in

Table 2. Named entity recognition experiment results.

Model	Precision	Recall	F1
BERT	0.59304	0.63148	0.61166
BERT + CRF	0.67409	0.65241	0.66210
BERT + BiLSTM + CRF	0.73486	0.69911	0.71654

, demonstrating that the BERT + BiLSTM + CRF model achieved the best performance across all evaluation metrics. Despite some limitations, such as high computational cost and long training times, the BERT + BiLSTM + CRF model remained the best-performing model, combining the contextual representation power of BERT, the bidirectional sequence modeling capabilities of BiLSTM, and the sequence label optimization ability of CRF, which collectively enhance performance in the named entity recognition tasks. In addition, models such as RoBERTa and ALBERT offer improvements in performance and efficiency, while Transformer-XL and GPT-2 may be better suited for tasks involving long-range dependencies or sequence generation.

2.5.2. Relation Extraction

Compared to entity extraction, relation extraction tasks tend to yield better results. Relation extraction was performed using the BERT model. The parameters of the BERT model for relation extraction are shown in

Table 3. BERT model parameter settings.

Model Parameters	Model Parameter Values
Epoch	10
Max_len	70
Batch_size	4
Learning_rate	1 × 10−5

.

The overall relation extraction task metrics were calculated based on the labels and predictions on the validation set, as shown in

Table 4. Relation extraction experiment results.

Model	Precision	Recall	F1
BERT	0.9210	0.9373	0.9260

.

2.6. Knowledge Storage

The extracted entities and relationships were saved in the form of triples and stored in a CSV file.

Data were imported into the Neo4j database using Cypher query statements, and a safety knowledge graph of the mine ventilation system was constructed [38,39]. In Neo4j, data are represented as a graph, which not only enhances the flexibility and efficiency of data representation but also effectively handles dynamic changes, multi-layered relationships, and highly interconnected data scenarios [40]. Furthermore, Neo4j utilizes Cypher, a query language specifically designed for graph data. Cypher is robust and highly capable of efficiently executing complex graph traversals, path queries, and aggregation operations, thereby uncovering potential connections and revealing hidden patterns within the data [41]. Its concise syntax reduces the complexity of querying graph data, allowing users to express queries related to graph structures intuitively and conveniently.

3. Analysis and Application of Knowledge Graph in Mine Ventilation Safety

3.1. Visualization of Mine Ventilation Accident Information

After completing the tasks of entity label extraction and relationship label extraction, the resulting triple dataset was imported into Neo4j, creating a safety knowledge graph with a total of 3318 entity nodes and 4291 relationship nodes. This abundant data can depict the knowledge network in the field of mine ventilation system safety relatively comprehensively and in detail, providing data support for the subsequent application of the knowledge graph. Figure 3 shows a partial view of the safety knowledge graph of the mine ventilation system. In this view, the knowledge graph adopts a typical graph-based architecture, where key entities are represented as nodes and interconnected through various types of relationships. The central node corresponds to the accident event, from which multiple branches extend to capture associated information such as personnel, equipment, environmental factors, and accident causes. This multi-level, multi-dimensional construction of the knowledge network not only allows managers to quickly grasp the fundamental details of accidents but also facilitates the in-depth exploration of complex underlying factors through the analysis of related nodes. Moreover, this approach can be integrated with existing mine safety management systems, enhancing decision-making by providing data-driven insights. By incorporating historical accident data and predictive analytics into the existing framework, this knowledge network can offer proactive recommendations for risk mitigation and safety improvements, seamlessly aligning with the operational workflows of safety managers.

Focusing on a specific node, Neo4j’s visualization functionality was employed to query relevant node information and generate a visual representation. Figure 4 demonstrates the query results using the node “Gas explosion in Pinghu coal mine in Fengcheng” as an example. Through the query, a visual representation of the accident’s location, type, direct cause, casualties, and economic losses can be obtained. For instance, the accident location is identified as “Jiangxi Province”, the accident type as “gas explosion”, and the direct cause as “Gas accumulation”. Additionally, data on casualties and economic losses can be traced. This information not only enables managers to quickly grasp the basic details of the accident but also provides critical insights for accident analysis and prevention. Based on research or analytical needs, the query can be further extended to indirectly related nodes, such as “penalty details”, to retrieve information on the penalties imposed on responsible individuals. This extended query enriches the comprehensiveness and depth of accident information, offering support for tracing accountability and improving management practices. Building upon the query of the “Gas explosion in Pinghu coal mine in Fengcheng”, the analysis can be expanded to include nodes related to environmental factors, such as “Environment” and “Exact place ”, to examine whether these factors contributed to the accident. Through such correlation analysis, a more comprehensive understanding of the accident’s context and environment can be achieved, providing valuable references for enhancing safety management in similar mines.

3.2. Query and Analysis of Mine Ventilation Accident Information

3.2.1. Multi-Dimensional Statistics of Accident Information

Multi-dimensional statistics of accident information are an essential tool in the safety management of mine ventilation systems. Not only do they provide an initial overview of accidents through traditional single-dimensional analysis, but they also enable a detailed examination of accidents from multiple perspectives, revealing connections that might be overlooked in single-dimensional analyses [42]. This approach offers more precise support for management decision-making and risk warning. By constructing a knowledge graph for the mine ventilation system and utilizing the Cypher query language for fuzzy searches, it is possible to quantify the most frequently occurring entity names under different entity labels (e.g., accident causes, accident types, accident locations, etc.). This process allows managers to perform comprehensive statistical analyses of accident occurrences across various dimensions (such as time, location, and type), thereby uncovering deeper patterns and potential hazards behind accidents.

Statistical analysis of accident data reveals that coal mines are areas with a high incidence of accidents in the mine ventilation system (see

Table 5. Multi-dimensional accident statistics.

Entity Label	Entity Fuzzy Search	Number of Accidents
Direct reason	Gas accumulation	70
Indirect reason	Poor safety management	12
Accident time	Year 2008	19
Accident place	Hunan	30
Accident type	Gas explosion	70
Accident mine	Coal mine	105
Exact place	Working face	26

). This phenomenon is closely related to the geological conditions of coal mines, the mining environment, and the complexity of the ventilation system. Gas accumulation in coal mines is one of the main causes of frequent accidents, with gas explosions being the most common direct cause. Gas accumulation is not only closely related to the ventilation conditions in the mine but also to factors such as mining depth, mining technology, and the use and maintenance of ventilation equipment. When the ventilation system is poorly designed or equipment is aged, gas concentrations may reach dangerous levels, triggering an explosion. Therefore, the early detection and management of gas accumulation are critical to accident prevention.

In addition, the statistical data also reveal the geographical concentration of accidents. Shanxi Province has the highest frequency of accidents in mine ventilation systems, a trend closely associated with the large number of coal mines in the province, its extensive mining history, and the large scale of its mining operations. The in-depth analysis of accidents in Shanxi Province reveals that these frequent incidents are often related to factors such as complex geological conditions, inadequate mine ventilation systems, and insufficient management measures. The regional concentration of accidents emphasizes the need for tailored safety management plans and preventive measures specific to the characteristics of mines in different regions. The year 2008, which saw a peak in mine ventilation system accidents, may be closely linked to the domestic economic situation, policy environment, and the production and operational status of mining enterprises that year. Increased economic activity and heightened demand for mining contributed to the emergence of hidden risks that ultimately led to accidents.

Through the multi-dimensional statistics of accidents, especially the analysis of cases with high death tolls and significant economic losses, the severity of accidents in the mine ventilation system is further highlighted. These major accidents, often with numerous casualties and high economic losses, typically reflect failures in the mine ventilation system or inadequate management at the time of the incident, leading to widespread impacts. The severity of accidents is measured not only by the number of casualties but also by their profound impacts on production, business operations, social reputation, legal liabilities, and other factors. Consequently, a detailed analysis of the causes of these major accidents can provide critical insights and improvement measures for mine safety management.

3.2.2. Frequency Statistics of Specific Information

By conducting frequency statistics on key factors such as equipment failure nodes, personnel violation nodes, and ventilation facility faults, high-risk areas and critical risk factors within the mine ventilation system can be accurately identified, providing a solid data foundation for timely intervention and accident prevention. As a core element of the mine ventilation system, the operating status of ventilation equipment directly impacts the safety of the entire mine. Malfunctions or instability within the ventilation system can lead to hazardous conditions, such as gas accumulation and insufficient oxygen supply, triggering a series of safety accidents. Therefore, frequent monitoring and analysis of faults and unsafe states in ventilation equipment are essential for enhancing the mine’s safety.

Taking “ventilation equipment” as an example, the statistical analysis of its unsafe states using the Cypher query language reveals that local fans and air ducts are the components most prone to safety hazards within the ventilation system (see

Table 6. Statistical analysis of the number of unsafe condition nodes involving ventilation equipment.

Ventilation Equipment	Number of Unsafe Condition Nodes	Examples
Auxiliary fans	28	Not operating properly, stopped functioning, not turned on, poor quality, etc.
Air ducts	16	Poor installation, disconnection, not properly connected, multiple tears, etc.
Main fans	3	Not turned on
Ventilation facilities	3	Incomplete, non-compliant

). The specific Cypher query used is as follows: MATCH (t: Ventilation) − [r] − > (p: Phenomenon), RETURN t.name AS Ventilation, count (p) AS Phenomenon_Count, ORDER BY Phenomenon_Count DESC.

Local fans show 28 unsafe state nodes, with issues such as abnormal operation, shutdown, failure to start, and poor quality. These problems directly affect the ventilation efficiency, potentially leading to poor air circulation, thereby increasing the risk of accidents such as gas explosions. Failures in local fans are often related to factors such as design flaws, quality issues, service life, and improper operation and maintenance. Air ducts, an essential part of the ventilation system, also present 16 unsafe state nodes, with common problems including improper installation, fractures, insecure connections, and ruptures. Failures in air ducts can result in poor airflow, reduced ventilation effectiveness, and safety risks, such as equipment damage or fire.

Given the critical nature of air ducts, their maintenance and inspection are crucial. Any improper installation or structural defects can trigger a chain reaction. In contrast, the number of unsafe state nodes for main fans and ventilation facilities is relatively small, with only three nodes each. This suggests that the failure frequency of these components is low. However, considering their vital role in the mine ventilation system, even a minor failure could pose significant safety risks. For instance, the failure of a main fan may cause the entire ventilation system to collapse, directly affecting air quality and increasing the risk of harmful gas accumulation. Therefore, despite the lower failure frequency, regular inspections, maintenance, and early warning systems for main fans and other ventilation equipment are critical to ensure the stable, long-term operation of the system.

Furthermore, frequency statistics of equipment faults and unsafe states not only assist in identifying potential high-risk areas but also inform equipment management and maintenance decision-making. By categorizing these issues, managers can prioritize which equipment and components require urgent attention, allowing for more targeted maintenance and overhaul plans. Additionally, the analysis of historical data can help enterprises proactively identify risks and take preventive measures. For example, upon identifying that local fans and air ducts frequently exhibit unsafe states, timely inspections and repairs can be conducted, or new equipment can be purchased to enhance the stability and safety of the mine ventilation system.

3.2.3. Specific Content Relevance Analysis

Content correlation analysis plays a crucial role in uncovering the root causes of deep-seated issues, enabling managers to gain a comprehensive understanding of the complex environment in which various factors in the ventilation system interact with each other. Through the in-depth analysis of these correlation relationships, it is possible to identify potential risks that may be overlooked through single-dimensional analysis, as well as provide new perspectives and ideas for accident prevention in the mine ventilation system.

In Neo4j, deep correlation analysis typically starts with a specific type of entity node, using the query pattern “entity node-any relationship-any node-any relationship-entity node”. This type of query not only reveals direct connections between entity nodes but also uncovers indirect and potential associations through other intermediate nodes.

For example, Figure 5 shows the node associations between the entity types “Equipment” and “Phenomenon”, indicating that there are not only direct connections between the two, such as equipment failures directly affecting changes in process phenomena, but also indirect connections through other nodes. For instance, equipment failures may affect factors like airflow rate and ventilation effects, which then lead to changes in process phenomena, such as gas accumulation and abnormal gas concentrations. The revelation of these indirect relationships provides richer information for safety management of the mine ventilation system, enabling managers to gain a more comprehensive understanding of the background and environment in which accidents occur.

In addition to direct and indirect node associations, correlation analysis can also perform traceability analysis. By selecting a key node such as t as the central element and utilizing the entity relationships constructed in the knowledge graph, upstream and downstream nodes related to this node can be identified, and the interactions between these nodes, as well as their impacts on the overall system, can be analyzed in depth. Traceability analysis helps uncover the potential causes of accidents and deep-rooted risk factors. For example, by analyzing the relationships between equipment failures and other related factors, such as involved personnel, production environment, and process phenomena, it is possible to trace back to the specific causes of failures, such as design defects, quality issues, or improper operations.

In practical applications, correlation analysis serves two key purposes. First, it helps identify high-risk areas within the mine ventilation system. Second, it assists in predicting key factors that may lead to accidents. For example, the failure of a particular ventilation device may not be solely due to equipment aging, but could also be closely related to factors like environmental conditions inside the mine, climate changes, and the maintenance habits of operators. By analyzing these multi-dimensional relationships, potential safety hazards within the mine ventilation system can be identified in advance, allowing managers to take targeted preventive measures.

Additionally, correlation analysis can help uncover systemic problems within the mine ventilation system. Various components, including devices, personnel, environmental factors, and operational processes, are interconnected, and any change in one factor can have a profound impact on overall safety. By conducting a correlation analysis across these factors, managers can identify the key risk factors. For instance, the failure of a specific device is often the result of multiple factors, such as equipment aging, operational errors, and inadequate maintenance. This type of analysis can help managers optimize the design and management of the ventilation system from a global perspective, improving its operational efficiency and safety.

Traditional data management methods, such as relational databases or static spreadsheets, often fail to capture the dynamic and multi-dimensional relationships inherent in mine ventilation systems. In contrast, the Neo4j-based knowledge graph provides substantial benefits by uncovering hidden correlations and supporting complex decision-making processes. Overall, the Neo4j knowledge graph enables more efficient identification of potential risks and root causes through multi-dimensional correlation analysis, offering a marked improvement in safety decision-making compared to traditional data management approaches.

3.3. Assisted Safety Management for Mine Ventilation Accidents

3.3.1. Fault Prediction

In the safety management of the mine ventilation system, fault prediction and safety training are essential auxiliary tools that help identify potential risks in advance and reduce the probability of accidents. Fault prediction analyzes the historical data of equipment in the mine ventilation system using the Cypher query language, enabling the early identification of performance issues and potential failures based on past accidents.

Taking the “local fan” as an example, a detailed analysis of its unsafe states has revealed multiple factors that could lead to equipment failure. For instance, poor-quality local fans may exhibit unstable performance due to issues with material, manufacturing processes, or design defects, which in turn affects the equipment’s safety. Another issue is that the local fan might fail to start, which could be caused by equipment malfunctions, improper operation by workers, or power outages underground. Additionally, improper installation of the local fan can lead to abnormal operation. Deviations in the design and installation process may negatively impact the overall efficiency of the fan. Finally, the “intermittent operation” of local fans also reflects the weak safety awareness of miners. Random operation can cause instability in the equipment’s performance.

Through an in-depth analysis of these potential fault states, data support can be provided for the maintenance and operational specifications of local fans. Targeted improvement measures can then be formulated to enhance the safety and stability of the equipment, ultimately reducing the frequency of faults. This proactive approach not only ensures the reliable operation of the ventilation system but also contributes to the overall safety management of the mine by identifying potential risks early and facilitating timely interventions to prevent accidents.

3.3.2. Safety Training

Safety management is a critical indirect factor contributing to accidents in mine ventilation systems. Historical accident data indicate that the occurrence of many mine accidents is closely related to deficiencies in safety management by enterprises. For example, imperfect operating procedures, insufficient safety awareness among employees, and poor equipment management act as catalysts for accidents. Deficiencies in the enterprise’s safety management system, personnel training, and safety supervision expose the weaknesses in existing management practices.

By using multi-dimensional statistics on accident data, managers can identify these weak links and implement targeted improvements to enhance the overall safety management level. This process not only helps discover and eliminate potential management loopholes but also improves the safety of mine ventilation systems by optimizing management processes.

Using Cypher queries, managers can identify common unsafe behaviors in historical accidents within the mine ventilation system and design targeted safety training content based on these behaviors. For high-risk personnel, violation records in the knowledge graph can be used to mark, track, and manage individuals. For instance, common unsafe behaviors include illegal blasting, live-working, illegal disassembly of miner’s lamps, absence from duty, missed inspections, and illegal live-maintenance. These behaviors not only reflect safety awareness issues among employees during daily operations but may also directly lead to major accidents.

Therefore, in safety training, these common unsafe behaviors should be emphasized and analyzed as case studies to enhance employees’ safety awareness. At the same time, for high-risk personnel, managers can improve their safety operation skills and risk prevention awareness through regular training and assessments, effectively reducing accidents caused by human factors. Through systematic and regular safety training, the overall safety literacy of miners can be improved, the accident rate can be reduced, and the stable operation of the mine ventilation system can be ensured.

4. Conclusions

By collecting and processing accident case data from the mine ventilation system, a Neo4j-based safety knowledge graph of the mine ventilation system is constructed. The chaotic data is organized into a database with a specific structure and applied to the safety management of the mine ventilation system. The main research conclusions are as follows:

• Visual storage and efficient management of historical accident data were achieved through the Neo4j-based knowledge graph. Accident records and related information are intuitively represented in the form of interconnected nodes and relationships. This structured representation not only enhances the operability and practicality of safety information management but also significantly improves data accessibility and interpretation efficiency compared to traditional manual or relational database approaches.
• Accident information retrieval and multi-dimensional correlation analysis were conducted based on the constructed knowledge graph. By leveraging the explicit interrelationships among various entities, such as personnel, equipment, environmental factors, and operational processes, the system facilitates an in-depth analysis of accident causation pathways. It enables safety managers to identify indirect risk factors and systemic safety hazards that are often overlooked in conventional systems. Additionally, the flexible querying capabilities of Neo4j considerably reduce the time required for accident cause tracing and pattern discovery, providing new insights and supporting informed decision-making.
• The detailed text content within the safety knowledge graph of the mine ventilation system is fully utilized to collect and analyze the unsafe states of objects and unsafe behaviors of individuals. The system aids in fault prediction for equipment management and provides targeted guidance for safety training. By delivering actionable insights for preventive interventions, it supports proactive risk mitigation, enhances regulatory compliance, and improves employee safety.

Future research could further explore the potential application of the mine ventilation system safety knowledge graph in real-world mine safety management through field deployment and testing, thereby validating its effectiveness and application potential in practical environments. This would provide a valuable experience for the system’s practical implementation.