Use of Knowledge Graphs for Construction Safety Management: A Systematic Literature Review

Abstract

Effective safety management is crucial in the construction industry. The growing interest in employing Knowledge Graphs (KGs) for safety management in construction is driven by the need for efficient computing-aided safety practices. This paper systematically reviews the literature related to automating safety management processes through knowledge base systems, focusing on the creation and utilization of KGs for construction safety. It captures current methodologies for developing and using KGs in construction safety management, outlining the techniques for each phase of KG development, including scope identification, integration of external data, ontological modeling, data extraction, and KG completion. This provides structured guidance on building a KG for safety management. Moreover, this paper discusses the challenges and limitations that hinder the wider adoption of KGs in construction safety management, leading to the identification of goals and considerations for future research.

1. Introduction

The construction industry is recognized globally as one of the most hazardous sectors. Persistent safety challenges are highlighted by stark fatality statistics. In 2020, fatalities included 1008 in the USA [1], 27 in the UK [2], and 794 in China [3]. The occurrence of injuries and occupational diseases is similarly alarming. The US construction sector reported 174,000 nonfatal injuries and illnesses in 2020 [2]. In the UK, the construction industry averaged 59,000 injuries annually between 2019 and 2021, with a notable increase following the coronavirus pandemic [4].

Both academia and industry have advanced a range of safety management theories and methods, covering performance measurement, event documentation, policy formulation, and risk evaluation [5]. With particular interests in technological advancements for safety, researchers like Fang et al. [6] and Ye et al. [7] have introduced computer vision and intelligent platforms for hazard identification and response. These automation-oriented approaches to safety management, grounded in digital technology, necessitate enhanced gathering, processing, and analyzing of safety data from diverse sources, including sensors [8], accident reports [9], and surveillance systems [10].

The rise of digitalization and data science has broadened the sources of safety information. Technologies like Virtual Reality (VR), Geographic Information System (GIS), 3D & 4D Computer-Aided Design (CAD), Building Information Modeling (BIM), and the Internet of Things (IoT) have significantly diversified the data supporting safety management in construction [11]. While these varied techniques offer vast opportunities for collecting various types of safety-related information, the diversity and volume of data presented new challenges: (1) incompatible file formats—VR, CAD, or BIM IFC file formats vary across software, complicating data integration and analysis; (2) knowledge entity ambiguity—the different uses of construction terms, such as “Toolbox” in “Toolbox Meeting” or “Toolbox Talks”, which denote brief safety meetings, may cause confusion in processing such textual data; (3) information redundancy—safety regulations, frequently updated by authorities or project management teams, can render some data obsolete or redundant; (4) interoperability issues—employing multiple tools for creating and processing safety-related data complicates gaining comprehensive insights, especially due to interoperability issues among software programs and platforms [12].

Knowledge Graph (KG) technology offers a method for representing and integrating such heterogeneous safety-related data. A KG is structured as a semantic graph comprising nodes (or vertices) and edges, which encapsulate and denote data from various sources. Nodes can symbolize concepts or entities, such as workers, tools, or environmental factors, while edges represent the relationships among them, such as operations, commands, or causation [13], in the context of construction safety. For instance, as Figure 1 depicts, a regulation about working on a roof could be represented in a KG as follows. Rule 1 requires fall protection (including harnesses, guardrails, lanyards, and safety nets) and PPE (helmet and non-slip footwear). The relationship “requires” links Rule 1 to these safety measures. A worker, linked by the “equipped” relationship, must use this equipment. The “supervise” relationship connects the safety manager to the worker, ensuring compliance. Finally, the worker is linked by the “work” relationship to the task “Work on roof”. This KG structure clarifies the dependencies and requirements for safe roof work. By organizing data into graphs, a KG also enables data retrieval based on semantic relationships between specific data. Also, this method aids in uncovering insights from datasets that are not apparent in the original data [14].

The appeal of Knowledge Graphs (KGs) has garnered widespread interest. Commercial IT giants like Google, IBM, Amazon, and LinkedIn have each developed their own proprietary KGs [15,16]. Moreover, there are various open-source KGs available, such as DBpedia, Google’s Knowledge Graph, and YAGO, drawing on general knowledge and official databases [17]. Several researchers proposed the use of KGs in health and safety management in multiple domains. Wu et al. [18] integrated Bayesian networks with KGs for fault prognosis in high-speed railway traction devices. In aviation, KGs have been utilized to analyze risk reports and assist in flight [19,20]. Wang et al. [21] created a KG for analyzing railway electrical accidents. Furthermore, KGs have supported safety management in electrical grids, including incident analysis [22], fault handling [23], and safety planning [24]. Notably, the development of KGs in the healthcare domain accelerated due to the COVID-19 pandemic, aiding in the integration of information from various organizations and identifying transmission patterns [25].

In the domain of construction safety management in particular, several researchers also proposed developing Knowledge Graphs (KGs) for organizing different types of data. Jiang et al. [26] conducted a case study on developing a KG for safety standards in the construction industry. Chen et al. [27] proposed to integrate linguistic and visual information about on-site occupational hazards, extracted through AI models, into a KG. Preliminary experiments have also explored converting Industry Foundation Classes (IFC) from BIM models to develop safety ontologies and map them onto graphs [28]. As such, research on KGs within construction safety management is emerging, but there is a lack of research that systematically analyzes the application, development, trends, and challenges of using KGs in the construction safety management domain.

Against this backdrop, this paper aims to thoroughly review the application of Knowledge Graphs (KGs) in the construction safety management domain and to propose a detailed, step-by-step process for developing KGs specifically for this domain. More specifically, this review sets out with the following objectives to fulfill the study’s aim:

(1)

To identify global trends and research focal points within scientific publications on KGs in construction safety management;

(2)

To categorize current research efforts in applying KGs in the domain.

(3)

To model the methods employed at each stage of constructing KGs for construction safety management.

The rest of the paper is organized as follows. The next section details the methodology for the literature retrieval process. The third section presents a scientometric analysis of 139 publications selected for this review paper. The fourth section critically examines the categorization, construction procedures, and applications of KGs. The fifth section summarizes the main methods used in KG development and highlights five major issues impeding the use of KGs in the construction safety management domain. The conclusion is presented in the sixth section.

2. Selection of Papers for Review

Web of Science (WoS) and Scopus are two renowned and extensively utilized databases in the engineering field. Previous studies have indicated that, while Web of Science offers rich metadata for analysis, Scopus boasts a wider array of literature resources [29]. To ensure a comprehensive database for preliminary analysis, the authors opted for these two databases. The search timeframe was set from 2000 onwards to encompass the relevant literature and developments in the field [30], as the application of KGs in safety management domain was scarce before then.

The search strategy encompassed three themes: KG, safety, and construction. For the safety theme, terms like “accident”, “incident”, “job hazard”, “risk OR risk management”, “safety OR safety management”, and “occupational health” were identified. The construction theme included keywords such as “construction”, “civil engineering”, and “construction industry”, while “knowledge graph” was deemed sufficient to represent its own theme. As such, the final search formula executed in Scopus was ALL (“knowledge graph”) AND TITLE-ABS-KEY (“accident” OR “incident” OR “risk” OR “safety” OR “occupational health”) AND TITLE-ABS-KEY (“construction” OR “civil engineering” OR “construction industry”). Adjustments were made to the search syntax for application in WoS such that the papers relevant to all three themes are searched in the database.

As a result of the initial search, a total of 368 publications were retrieved. Subsequent filtering was conducted to eliminate duplicates across the databases. Following the PRISMA guideline [31], the contents were meticulously reviewed to discard papers unrelated to the topic, resulting in 139 papers suitable for review, as presented in Figure 2 (the research project was registered in Open Science Framework). These papers span various publication years and categories. The bulk of the publications were in journals (67.7%), while conference proceedings accounted for the remaining 32.3%. An upward trend in publications began in 2016, with a significant increase noted in 2022, and the momentum remained high into 2023 compared to earlier years.

Table 1. Publication statistics (2016–2023).

Source		Total	2019	2020	2021	2022	2023
Top-8 Most Productive Journals	Automation in Construction	14	2	3	1	5	3
	Advanced Engineering Informatics	7	1	2	0	3	1
	Buildings	5	0	0	0	3	2
	Journal of Management in Engineering	2	0	0	0	0	2
	Computer in Industry	2	0	0	0	1	1
	Engineering Applications of Artificial Intelligence	2	0	0	0	0	2
	Expert System with Applications	2	0	0	0	1	1
	Mathematical Problems in Engineering	1	0	0	0	1	0
Total Journals	-	93	4	11	7	35	36
Proceedings	-	43	9	3	9	16	6
Total Publications	-	136	13	14	16	51	42

summarizes the distribution over the years and the categories of the publications selected for review.

Furthermore, we identified journals that have published more than two papers on this topic in recent years. The top eight journals accounted for 24.5% of the total papers, with “Automation in Construction”, “Advanced Engineering Informatics”, and “Buildings” emerging as the most prolific in terms of publications on KGs in construction safety management.

3. Scientometric Analysis

The scientometric approach is a statistical method that elucidates the knowledge structure of a specific domain by analyzing extensive literature sources. Constructing a comprehensive knowledge structure from large-scale bibliographic data is both time-consuming and challenging for researchers. In review studies, mapping and clustering techniques prove useful for revealing relationships and characteristics among articles, keywords, authors, and journals [32]. To support scientometric analysis, various tools have been developed, including Bibexcel, VOSviewer, and CiteSpace [33]. This study employed CiteSpace 6.2R4, a robust Java-based software that is capable of processing data from multiple sources to generate networks of heterogeneous nodes [34].

Utilizing CiteSpace, the study extracted the co-occurrence of keywords directly from the bibliographic data. The criteria for filtering papers were defined by a modified g-index with a scale factor value of 50. Figure 3 showcases the co-occurrence links among various keywords, with the size of the label font and node corresponding to the frequency of publications containing the keyword. The six most-frequent keywords identified were “knowledge graph”, “deep learning”, “system”, “ontology”, “management”, and “risk assessment”. The centrality of the knowledge network’s nodes was calculated to assess their role within the network. The analysis revealed that the five keywords with the highest connectivity were “deep learning”, “data mining”, “BIM”, “construction method”, and “design”, indicating their significant influence on other nodes within the network. The keywords could be categorized into two groups: the first group pertains to techniques, including “deep learning”, “data mining”, “BIM”, “named entity recognition”, and “ontology”, among others; the second group relates to the field’s application, featuring keywords such as “risk assessment”, “construction safety”, “knowledge representation”, and “design”.

Cluster analysis is frequently employed to extract themes and research insights from extensive bibliographic data [32]. In this study, the modularity Q value of the keyword network is 0.7887, suggesting that the network can be effectively segmented into clusters for more detailed analysis (Q value quantifies the strength of how a network can be divided into distinct communities [35]). Figure 4 illustrates eight clusters identified. Furthermore, this paper applied silhouette score (range from −1 to 1) to evaluate the quality of clustering; a closer value to 1 indicates a clear clustering [36]. The silhouette score for these clusters exceeds 0.827, affirming their internal homogeneity and suitability for further investigation. The cluster analysis reveals that construction site-related terms are a focal point in the papers, while deep learning and complex system analysis are frequently linked with the topic as methodical approaches. Moreover, the representation of information through Knowledge Graphs appears to be a primary goal across the publications.

Overall, the analysis of keyword co-occurrence and cluster formation within the selected papers highlights two principal research dimensions. The first concerns artificial intelligence (AI) techniques, featuring keywords like “deep learning”, “name entity recognition”, and “data mining” and clusters like “deep neural network”. The second concerns safety management themes, with keywords such as “BIM” and “risk assessment” and clusters like “accident scenario” and “healthy operation index”. KGs provides a structured representation of information possibly from multiple sources by connecting data points through nodes and edges. Structured information makes AI algorithms such as natural language processing and machine learning easier to understand the relationships of information and the context of data. On the other hand, KGs enable semantic search by both the meanings and the relationships of words, not just word matching, which enriches the information retrieval capabilities of AI systems. The trend towards leveraging AI for knowledge acquisition and utilizing KGs in construction safety management, which is consistent with the direction of AI development.

4. Categories of KGs for Construction Safety Management

This section is structured into three subsections. The first subsection delineates the categories of KGs in safety management, shedding light on the dominant research perspectives in academia. The second subsection provides a comprehensive overview of the processes and methods required to develop a KG in the context of safety management. The third and final subsection outlines the application scenarios of KGs in construction safety management research, demonstrating the current trends in their use.

4.1. The Categories of Knowledge Graphs in Safety Management

KGs in safety management can be classified from several angles. A notable approach is the incorporation of time series elements into KGs, leading to their classification into static and dynamic categories. Furthermore, the data source, whether singular or multiple, impacts data-processing methods, graph scope, and information diversity, resulting in single-source and fusion-source KG categories. From a usage standpoint, KGs are differentiated into domain-specific KGs and general KGs, depending on whether their entities and edges are confined to a specific domain or encompass a broader range. In the field of safety management, the focus is often on developing and analyzing domain-specific KGs rather than general ones. Consequently, the first two classifications are explored in further detail below.

Time consideration plays a vital role in differentiating between static and dynamic KGs. In conventional KGs, relationships among nodes are usually depicted without regard for temporal changes. However, in the safety management domain, time is an essential factor due to the dynamic nature of accident occurrence scenarios. Recent research has delved into two primary methods for incorporating time elements into KGs. The first method involves incorporating time information directly into the definition of entities and relations in the KG. Leblay et al. [37] introduced a KG where time intervals are used to represent relationships between entities, along with utilizing event timelines to forecast future connections, thus providing enhanced understanding of event relationships evolving over time. Moreover, several studies have embedded timestamps into nodes, enabling these KGs to capture temporal changes effectively and facilitate dynamic analysis [38]. The second method updates the graph contents at regular intervals. Duan et al. [39] developed a crowd congestion KG to depict real-time evacuation scenarios, extracting data from sensors and videos and employing a dynamic algorithm to refresh node attributes at each time interval.

The differentiation between single-source KGs and fusion-source KGs lies in their data integration process. Single-source KGs rely on homogeneous data types, such as standardized accident reports, official statistical reports, and field logs. These data are consistent and standardized, facilitating straightforward retrieval and processing from the respective sources. KGs based on a single source are, therefore, typically more concise and clearer due to this uniformity. In contrast, fusion-source KGs have gained more attention as access to a variety of data types has increased, including videos, event news, meteorological data, and safety regulations. For purely textual data, researchers employ natural language processing techniques to extract and convert related data into a uniform format [38,40,41]. Researchers have explored the use of computer vision alongside natural language processing to manage different data types [27,42]. However, the integration of varied data sources through KGs in the safety management domain is still in its early stages.

4.2. Application Scenarios of KG

Most use cases of KGs in the construction safety management domain are primarily for data representation and storage, which are considered initial phases of KG utilization [43,44,45,46]. However, several pioneering researchers have recently explored a more advanced use of KGs in this domain:

(1)

Creation of Accident Evolution Networks for Key Factor Identification: KGs have been used to depict risk factors as entities and connect different risk factors with directed or undirected edges, forming a network that models accident evolution. The underlying idea is that identifying critical risk factors along potential accident paths before an incident can facilitate targeted preventive measures to avert accident escalation and recurrence [40];

(2)

KG-Based Construction Safety Recommendation System: Adherence to safety regulations is essential for the competence of construction practitioners. However, gaining knowledge of these regulations often requires significant time and educational investment. Some studies have focused on extracting regulatory requirements into KGs for application in safety scenarios, aiding in comprehension and compliance [47];

(3)

Entity-Based Risk Assessment: Utilizing the structural characteristics of KGs, algorithms that compute the shortest path and node connections can facilitate risk assessments [48]. For example, in their study, hazard-related entities such as incident cause, consequence, time, location, and speed were extracted as nodes, with causation serving as links to construct KGs. Consequently, nodes closer to the incident’s nodes indicate a higher risk and occurrence associated with these nodes. Network path techniques such as closeness centrality have been employed to quantify the impact of entity occurrences, with thresholds set to evaluate entities’ risk levels. Direct causation and indirect causation elements were assessed as active and passive causal closeness to calculate the probability of accident occurrence.

These applications highlight ongoing efforts to leverage KGs for advanced analysis and problem-solving in construction safety management, pointing toward a future where KGs play a central role in enhancing safety measures and reducing risk.

4.3. Trends in KG Research in the Construction Safety Management Domain

From a technological perspective, KGs are increasingly being integrated with natural language processing, machine learning, and recommender [49]. Specifically, research in construction safety management by leveraging KGs is advancing in three main directions:

(1)

Transformation of data into knowledge: By transforming construction data into nodes and edges, the semantic connections between construction management elements become clearer. For example, risk analysts decompose project entities into nodes and their interactions into links to illustrate complex relationships [6,27]. Most publications employ KGs as a method to bolster their findings, allowing for concise expression that facilitates easy understanding and modification by others;

(2)

KG as a database for complex AI models: KGs are constructed as a base layer to support generative AI systems, such as large language models (LLM) and multi-modal models, such that AI algorithms can access, retrieve, and process real-world facts and data in the most efficient way. KGs, with their inherent machine-readable structure, are suited as input databases. Consequently, developing safety management KGs by grounding a generative AI system to assist in construction safety management tasks (e.g., [50,51]) has recently emerged as a research direction;

(3)

New dimensional knowledge discovery: KGs are a part of graph science, making graph attributes and theories directly applicable. Node metrics like degree, centrality, and clustering indicate the significance of nodes and their relation to other nodes. Thus, the development of KGs not only aids in data structuring but also paves the way for gaining new insights from the knowledge base through graph science principles [30,52].

These three research directions highlight the technological trends in advancing KGs in the construction safety management domain, underscoring their role in clarifying complex relationships among project data, enhancing AI model development, and discovering insights from knowledge bases.

5. Analysis of KG Development Process

This section provides a thorough examination of the KG development process as found in the construction safety management literature, essentially discussing the state-of-the-art practices for constructing KGs within this domain. It begins by outlining the most common process for developing KGs. Then, it delves into a comprehensive description of the methods employed in KG development, as discovered through the literature review.

5.1. Overview of KG Development Process

There is no standard procedure for building Knowledge Graphs (KGs) in either computer science or the construction safety domain. The literature analysis shows that, at a high level, the development of a KG for construction safety management typically follows the process depicted in Figure 5.

The first step in developing a Knowledge Graph for construction safety management involves integrating both internal and external project data. In addition to utilizing safety files created on-site and used by stakeholders, some researchers suggest that incorporating external knowledge not directly related to the project is essential for broadening the knowledge base. This external information may include safety data from official publications, similar companies, academic research, and industry management bureaus [53]. Accessing a wider range of information resources aids in refining technical terminology, enhancing semantic context, and enriching domain-specific knowledge.

Ontological templates function as foundational blueprints for constructing a KG, with their design primarily aimed at knowledge sharing and enabling semantic interoperability. This ensures that various parties and computer systems can exchange and repurpose information seamlessly, which is a core principle of KGs. An ontology delineates the intrinsic attributes of entities and their interrelations through a structured set of classes. To minimize information redundancy and complexity in knowledge formats, external data are typically adapted to fit the predefined ontological structure. For example, Wang et al. [53] integrated risk management data from an existing ontology library; Zhu and Luo [54] merged a comprehensive public KG into their specifically designed safety rule KG to enhance information accuracy; and Yang and Liao [38] incorporated The People’s Daily Corpus in the form of semi-structured text data to broaden the knowledge base.

Following the ontological templates, data extraction focuses on formatting the information into nodes and edges, the fundamental elements of KGs. When two triplets share a common node, they form preliminary KGs through node–edge–node relationships. However, overly large and sparse KGs can lead to inefficient use of storage resources and other issues. Common problems include repetitive triplets that convey the same meaning in various forms and a lack of connections between nodes, leading to disconnected segments within the graph. The process of completion checking is crucial for identifying such issues. If problems are detected, graph completion techniques are applied; if not, the KG construction is deemed complete.

5.2. Methods Used in Developing KGs

5.2.1. Scope Identification

In defining the scope for KG development, it is essential to consider the problems that the KG aims to solve and the needs of stakeholders involved in the study, given the limited resources and priority of issues. The domain of safety management inherently encompasses extensive data due to its multifaceted nature [27]. The safety knowledge base in a construction project includes written and graphic documents, as well as raw data from equipment or devices [6]. The vast quantity of data and the varied authority of data sources necessitate filtering to establish a valid foundation [26]. Therefore, articulating the problem statement and prioritizing safety knowledge information based on specific needs are crucial steps in clearly defining the scope of analysis [53].

5.2.2. Ontological Template Development

Developing a safety ontology can follow a top-down approach; for instance, Fang et al. [55] defined classes such as “time, space, event, thing, part, attribute, and attribute-value” based on reports of hazardous events, while Simone et al. [56] conceptualized classes like “event, apparatus, substance, activity, barrier, and people” from near-miss reports. This categorization takes into account relevant standards, specifications, codes, and regulations. The next step involves defining subjects, relationships (predicates), and objects to construct a basic ontology model for each class.

Alternatively, an ontology can be developed from a bottom-up perspective. Leveraging big data and advanced data-processing techniques, valuable insights can be extracted from unstructured data, allowing for the construction of ontologies based on the attributes of safety information. For example, researchers attempted to automate the development of a railway safety risk management ontology using statistical methods and k-means clustering [53]. Additionally, tools like the Jena inference engine assist in ontology construction through various algorithms [38].

As ontological theory in construction safety management evolves, the development of ontologies increasingly adopts hybrid approaches. Chen and Bria’s review [57] indicates that ontology-building methods and their effectiveness have diversified, providing a robust foundation for advancing knowledge graph development.

5.2.3. Data Extraction

In the data extraction phase of ontology development, the process involves navigating through data and filtering out triplets, which can be approached in two ways: extracting entities and relationships separately or using models designed to identify both simultaneously. The choice of approach largely depends on the format of the safety data source. For structured formats like JSON, CSV, RDF, etc. from safety organizations such as OSHA, KOSHA, and SWA, data can be straightforwardly transformed into triplets using data processing tools or packages. For instance, construction accident fatality CSV files from OSHA can be processed into triplets like “<accident_name><caused><number_of_deaths>” after appropriate filtering.

However, unstructured data sources such as news articles, regulations, field notes, and semi-structured sources like Wikipedia necessitate additional steps for processing. For these types of data, the foundational approach involves labeling both entities and relationships to extract them in a manner that aligns with the ontologies. To facilitate this, researchers have utilized a variety of statistical tools and AI models, encompassing machine learning and deep learning models.

Table 2. Data extraction details from publications from the last 3 years.

Publication Year	Publication Title	Raw Data	Extraction Approach	Models
2023	A study on a KG construction method of safety reports for process ndustries [58]	Safety reports	Entities & Relationships	BERT-BiLSTM-CRF-TFIDF
2023	Deep learning-based relation extraction and knowledge graph-based representation of construction safety requirements [59]	OSHA sections	Entities & Relationships	Attention-based CNN-perceptron + Attention-based BiLSTM-perceptron
2023	Industrial safety management in the digital era: Constructing a knowledge graph from near misses [56]	Near missing reports	Entities & Relationships	did not specified
2023	Building a knowledge graph for operational hazard management of utility tunnels [60]	Normative documents + Hazard description text + Control measure description text	Entities	BiLSTM-CRF
2023	Automatic construction hazard identification integrating on-site scene graphs with information extraction in outfield test [61]	On-site scene graphs + Chinese safety regulation dataset	Entities & Relationships	BERT-BIEO-Mask RCNN
2023	A text mining-based approach for understanding Chinese railway incidents caused by electromagnetic interference [22]	Chinese railway incidents reports	Entities	CNN-BiLSTM-BERT
2022	Automatic construction site hazard identification, integrating construction scene graphs with BERT based domain knowledge [62]	On-site scene graphs	Entities & Relationships	Mask RCNN-Transformer-C BERT
2022	Vision-based method for semantic information extraction in construction by integrating deep learning object detection and image captioning [63]	On-site scene graphs + Safety regulations	Entities & Relationships	Mask RCNN- Attention-based LSTM
2022	A novel method for constructing knowledge graph of railway safety risk [53]	Railway safety documents + Railway safety risk text documents + Public works data	Entities & Relationships	CNN
2022	Identification of accident-injury type and bodypart factors from construction accident reports: A graph-based deep learning framework [64]	Accident reports	Entities & Relationships	GCN + Co-occurrence network
2022	Using text mining to establish knowledge graph from accident/incident reports in risk assessment [48]	Accident/incident reports	Entities	HMM-BiLSTM-CRF
2022	Computer vision-based hazard identification of construction site using visual relationship detection and ontology [65]	Visual Relationship Dataset	Entities & Relationships	Visual Translation Embedding
2021	Combining computer vision with semantic reasoning for on-site safety management in construction [66]	On-site scene graphs	Entities & Relationships	Mask RCNN
2020	Knowledge graph for identifying hazards on construction sites: Integrating computer vision with ontology [55]	Safety hazard reports + related images	Entities & Relationships	Mask RCNN

summarizes the common categories of extraction and models employed in the recent literature over the past three years.

As indicated in Table 2, raw data in the construction safety management domain often encompass both documents and pictures, necessitating the use of both computer vision (CV) and natural language processing (NLP) techniques for data extraction. Models such as Mask RCNN are frequently utilized in CV to simultaneously identify entities and spatial relationships, effectively meeting ontological needs. However, deep learning models from the Recurrent Neural Network (RNN) family, including LSTM and GRU, which are typically employed for NLP tasks, might only capture entities or relationships. In instances where entities are missing, they can be supplemented through integration with CV models [59]. Additionally, missing relationships can be annotated based on the attributes of entities. For example, missing comparative, inclusive, or judgmental statements (“higher”, “includes”, “is”, etc.), can be indicated in ontological templates [22,48].

The use of deep learning models for extracting safety information has grown significantly with the advancement of AI technology. LSTM (Long Short-Term Memory), an enhanced version of the RNN model, is adept at capturing information across lengthy sequences of input words or graphical elements [67]. Conversely, Mask RCNN (Mask Region Convolutional Neural Network) is capable of segmenting images, detecting objects, and assessing spatial distances, making it ideal for extracting safety triplets. The introduction of the transformer model has further optimized and facilitated the integration of text and visual data [68]. For capturing contextualized representations of words and sentences, the application of Bidirectional Encoder Representations from Transformers (BERT) has set new benchmarks in NLP [50]. Moreover, the potential of Graph Convolutional Networks (GCN) in analyzing complex relationships within graph-based data suggests promising applications in conjunction with KGs [64].

5.2.4. Knowledge Graph Completion

Extracted triplets are uploaded to KG storage or a Database Management System (DBMS) to generate a preliminary KG. Automatically extracted triplets may include redundant, incomplete, or erroneous elements. Consequently, before utilizing the KG, it is essential to refine, augment, and adjust the initial content. The following four strategies are commonly employed to enhance the completeness of KGs:

(1)

Manual Verification: Common-sense and logical errors are identified and corrected through expert review. This method ensures accuracy by addressing errors that automated systems might overlook [68];

(2)

Filtering: Using filter functions or conditional operations available in most graph-based DBMSs, such as the “MATCH” function in Neo4j, allows users to identify and remove redundant or conflicting entries effectively;

(3)

Terminology Consolidation: Yin et al. [58] designed an entity library to standardize terminology, thereby preventing conflicts and duplications. This library provides a unique representation for each concept and ensures that multiple entities within the same sentence are separated and linked individually;

(4)

Reasoning-Based Graph Completion: This method involves generating new links or connections. For instance, from the relationship <A><BELONG_TO><B><CAUSE><C>, a new link <A><CAUSE><C> can be derived. For more intricate scenarios, graph algorithms are utilized to dynamically predict relationships [38]. Additionally, in the construction safety management domain, various sophisticated approaches are employed for KG reasoning and inference, including logical inference, the development of entity and relation embeddings, and the use of statistical relational learning to refine the network [30].

6. Discussion—Issues and Potential Solutions

6.1. Main Issues with Applying KGs in the Construction Safety Management Domain

A KG integrates data from multiple sources, often with varying formats and schemas that are not as standardized as IFC files from BIM software. For instance, on-site project log documents may employ terminologies that differ from those used in construction standard documents. Additionally, establishing links between on-site CCTV-based image or video data and construction accident records can be problematic due to the absence of a unified data schema, which would include data definitions, labels, and relations. These discrepancies significantly complicate the construction of high-quality KGs. Extracting meaningful information from unstructured data sources is a challenging task, markedly more so than from structured data. Although advances in natural language processing and image processing have opened new possibilities for effectively extracting information from text and images, defining conditions to filter out noisy data remains complex. Consequently, automating the data cleaning process in safety management is both crucial and challenging.

The scientometric analysis of this research shows that regulation-related documents constitute a significant portion of safety management data and serve as crucial references in management practices. The typical language structure of clauses in safety management is often imperative, advising or specifying certain actions within specific contexts. The absence of subjects and the complexity of activation conditions in these clauses compromise the accuracy of automated information extraction, thus complicating the formation of effective rule triplets and the construction of a clear KG. For instance, the OSHA Standard [69] states, “The employer shall assess the workplace to determine if hazards are present, which necessitate the use of personal protective equipment (PPE)”. The complex conditions specified make it challenging to construct effective triplets from this standard.

The foundational elements of a KG are triplets, yet these may sometimes inadequately convey the full meaning of safety management semantics, particularly when generated through automatic extraction. There are two primary limitations in the representational capability of triplets. First, the standard <subject-predicate-object> format often fails to capture all essential elements of safety information. For instance, the statement “workers cannot access the restraint area” appears straightforward but requires additional details to specify the type of workers, effective duration, and exceptions such as “unless…” to clearly define safety actions under specific conditions. Second, ontological models may struggle with complex relationships [57]. For example, a construction project might be associated with multiple safety plans, each containing various procedures. This can lead to complex many-to-many relationships, posing challenges in establishing clear, direct relationships between ontological elements.

Furthermore, professional jargon within the construction domain not only affects text extraction accuracy, but can also lead to KG misunderstandings. For example, “Retrofit” might be used as a relationship in a KG. In construction safety management, it refers to the process of updating existing buildings to meet new safety standards, whereas, in a general context, it simply means adding new components to an existing structure. Overcoming this ambiguity requires precise labeling of nodes and relationships to avoid misinterpretation. Currently, there are limited automated methods for achieving this clarity. Moreover, most KGs use a duplication removal function based on syntax rather than semantics, which can lead to the erroneous deletion of ambiguous nodes and edges.

The applications of KGs in the construction safety management domain necessitate timely additions, deletions, and updates of project information to the graph’s content due to the dynamic nature of construction works and their environments. Additionally, the continuous accumulation of safety information requires ongoing updates in the KGs. However, the development of a dynamic KG that is capable of adapting to data updates is still in its early stages. While some studies have used multiple KGs to represent the progression of events, these graphs are constructed independently based on information from different time points and lack continuity [70].

6.2. Proposed Solutions Based on Current Technologies

In order to address the abovementioned issues, firstly, the data preprocessing step can be optimized through knowledge standardization and/or the use of large language models. For structured data from various sources, researchers and project managers could establish knowledge-sharing standards that regulate terminology, data transmission length, format, and the granularity of information transmission. This would dictate what knowledge should be shared within a construction project and how. Additionally, large language models have shown exceptional capabilities in information extraction and filtering. By inputting domain-specific data, construction safety researchers and practitioners could fine-tune a model to help filter out irrelevant information and use prompting engineering to customize outputs into Resource Description Framework (RDF) or JavaScript Object Notation (JSON) formats for input into a KG.

To address the challenge of inadequate KG representation in complex conditions, developing multiple KGs for different aspects of a task could be an approach. If representing the entire safety management process of a project is challenging, it could be segmented into subdivisions and sub-items, with multiple KGs developed for each. This allows each KG to contain a more focused body of information, functioning effectively as a module. Conflicting terms or triplets could be localized within these modules to prevent misrepresentation of the entire project information. For complicated sentences, converting them into multiple triplets within a KG and organizing them into a hierarchical KG could be a solution.

Generative models for graphs offer a promising approach to facilitate evolution within construction safety management KGs. By developing a deep learning model based on existing graph data, these models can generate a new graph that closely resembles the actual data structure found in construction projects. This capability is expected to be particularly valuable in environments where safety conditions and regulatory requirements frequently change, requiring KGs to adapt and update continuously. Furthermore, by comparing the generated KG with one constructed from real-world construction data, researchers can enhance model performance, ensuring that the KG remains effective in representing complex safety management knowledge and solutions. This approach not only mitigates the challenges of outdated information but also simplifies the management of extensive datasets typically encountered in construction safety management, thereby maintaining the integrity and utility of the safety management system.

7. Conclusions

This paper provides a systematic literature review on the topic of the creation and utilization of KGs for construction safety management. A total of 139 relevant publications were sourced from the Scopus and Web of Science databases. The research involved co-occurrence and cluster analysis of keywords to identify prevalent topics. The research also involved categorizing KGs used in the construction safety management domain such as static versus dynamic KGs, single versus multiple source KGs, and domain-specific versus general KGs. The development process of KGs was also examined. The paper detailed the methods used in each phase of KG development, such as scope identification, ontological model construction, data extraction, and completion. The main contributions of this study are as follows:

(1)

Identification of research trends: Charted the evolution and current state of KG applications in the construction safety management domain, highlighting predominant themes, technological integrations, and methodological approaches;

(2)

Methodological framework: Developed a structured methodology for creating KGs tailored to construction safety, covering essential stages like scope identification, ontological modeling, data extraction, and KG completion;

(3)

Categorization of KGs: Identified and categorized various types of KGs in the domain, such as static vs. dynamic KGs and single vs. multiple source KGs, providing clarity on their applicability and utility;

(4)

Integration techniques: Discussed the integration of KGs with advanced data science techniques such as graph-based learning algorithms, which enhance the depth and effectiveness of safety data analysis.

The application of KGs in construction safety management is rapidly evolving. Beyond risk factor identification, safety recommendation systems, and risk assessments, new application directions are emerging. However, the application and development of KGs in construction safety management face several challenges, including the complexity of data cleaning, difficulty of extracting graph structures from unstructured data sources such as texts, the insufficient representativeness of ontology models, handling language ambiguities, and the high demand for dynamicity of KGs. This paper indicates that various data science techniques offer promising solutions to these issues, suggesting that KG applications could benefit from integrating large language models, a common data environment, and graph deep learning models. Based on the analysis results, three future research directions are identified as follows and these future directions will be pursued in the authors’ future research:

(1)

Integration of KGs with graph-based learning algorithms: Graph theory has not been fully exploited within the safety domain. Graph-based deep learning models, such as Graph Convolutional Networks and Graph Neural Networks, could be instrumental in conducting deeper, more effective analyses of construction safety knowledge data. These models are well-suited for analyzing complex interdependencies within safety management systems, potentially leading to more sophisticated predictive safety models;

(2)

Integration of KGs into a common data environment for construction projects: The common data environment in the construction industry serves as a centralized platform for managing and sharing data across all stages of a construction project. Integrating KGs into this environment can enhance the manageability and analysis of construction data, supporting dynamic updates and information sharing. This integration can help in maintaining up-to-date safety protocols and ensuring compliance with safety regulations, thus enhancing overall project safety management;

(3)

Dynamic KGs for real-time safety management: The construction industry requires KGs that can adapt to rapid changes in the environment and project conditions. Developing KGs that incorporate time-series data and that can evolve in real time could significantly improve the responsiveness of safety management systems. Future research should explore methods to integrate real-time data updates into KGs, allowing for continuous adaptation and learning. This could involve developing new frameworks or algorithms that enable KGs to dynamically update and predict potential safety issues based on ongoing construction activities.