High-Altitude Fall Accidents in Construction: A Text Mining Analysis of Causal Factors and COVID-19 Impact

Abstract

The construction industry remains one of the most hazardous sectors despite its economic importance, with high-altitude fall accidents being the most prevalent and deadly type of incident. This paper aimed to study and analyze the accident data of the past accident cases in China and find out the key causes and rules of the accidents. This research analyzed 1223 Chinese accident reports (2014–2023) using Latent Dirichlet Allocation topic modeling to identify causal factors, followed by Apriori algorithm correlation analysis to reveal accident causation patterns. This study comprehensively uses topic model, association rules and visualization methods to systematically analyze the causes of high-altitude fall accidents. The research identified 24 distinct accident cause topics across personnel, equipment, management, and environmental dimensions. Key findings revealed that incorrect use of labor protective equipment, inadequate safety inspections, and failure to implement safety management protocols were persistent issues throughout the study period. Notably, the post COVID-19 pandemic introduced new safety challenges, with the intensity of topics related to “subject of responsibility for safety production has not been implemented” showing significant post-pandemic increases. These findings highlight the evolving nature of construction safety challenges and the need for targeted interventions to address persistent and emerging risks.

1. Introduction

The construction industry has long been recognized as one of the most hazardous sectors across various industries [1,2]. Annually, hundreds of construction accidents result in considerable casualties, underscoring the persistent urgency of improving safety management within the sector [3,4]. Among these incidents, high-altitude fall accidents stand out as the most prevalent and deadly type of accident, accounting for a substantial proportion of fatalities and serious injuries [5]. Therefore, it has become an urgent task to learn from accidents to formulate targeted prevention and control measures, reduce the accident rate and mitigate the related human losses.

The evolution of digital technology has introduced transformative potential to construction safety management [6]. Natural language processing enables efficient extraction of risk factors and accident causes from textual reports [7,8]. Existing scholarship has explored multiple methodologies for analyzing construction accidents, including hybrid frameworks incorporating fuzzy logic techniques [9], text mining combined with association rules [10], and integrated models utilizing Bayesian networks [11]. However, a notable gap persists in the comparative analysis of construction accidents across different temporal contexts, particularly regarding changes in accident patterns before and after the COVID-19 pandemic, despite its acknowledged impact on construction safety [12,13].

Building upon these technological foundations, this study focuses specifically on high-altitude fall accidents, aiming to uncover underlying causal factors and patterns using Latent Dirichlet Allocation (LDA) topic modeling, and analyze the correlation of the causes of the accident. Accident investigation reports between 2014 and 2023 from China were used for the analysis, especially for the consideration of COVID-19. By extracting key information elements from accident reports using text mining techniques, this research seeks to enhance the efficiency and accuracy of accident analysis, ultimately contributing to more informed safety management practices within the construction industry.

2. Literature Review

As a major death factor in the construction industry, high-altitude fall accidents have received extensive attention from the academic community in recent years, and the research perspectives and methods show a diversified trend. Khan [14] summarized the main risk factors through a systematic literature review, pointed out the limitations of single technical means, and emphasized the need to integrate multiple technical systems to improve the prevention effect. Oliveira [5] analyzed accident data in the United States based on the HFACS model, and revealed the high risk of roofers in specific age groups and periods, which provided a basis for targeted intervention. Zermane [15] combined statistical analysis and fault tree modeling to find that the significant factor in accidents in Malaysia was not wearing personal protective equipment.

With the progress of technology, data-driven methods have gradually become a research hotspot. Piao [16] proposed a dynamic Bayesian network and computer vision fusion framework to realize dynamic risk assessment by real-time monitoring of worker posture and environmental risk factors. Zermane [17] further tested the random forest model and found that management factors and individual characteristics had the highest prediction accuracy for fatal accidents in Malaysia, confirming the potential of data-driven methods in safety management. Qi [18] used an ensemble learning model to automatically extract causal factors from accident reports by optimizing text mining technology. The results show that the ensemble model is more stable than a single algorithm, which provides an efficient tool for large-scale accident data analysis.

In terms of risk coupling and systematic analysis, Niu [19] constructed a risk coupling analysis model to quantify the interaction of five factors of human, machine, material, management and environment based on the N-K model. It was found that multi-factor coupling significantly increased the probability of accidents. Peng [20] combined grey DEMATEL with the explanatory structure model to identify low safety awareness as the key driving factor, and constructed a Bayesian network to quantify the contribution of each factor, revealing the nonlinear characteristics of the causal chain of the accident. Based on 20 years of data in the United States, Halabi [21] uses frequency analysis to analyze the data to obtain the trend of fall accidents, the correlation analysis between accident factors and injury degree, and uses logistic regression analysis to establish a prediction model that can diagnose fatal and nonfatal accidents.

In recent years, research on high-altitude fall accidents in construction mainly focuses on the following aspects: identification of risk factors, coupling of risk factors, worker factors research, safety management research, accident prediction research and innovation of research methods. Some representative studies are shown in

Table 1. Research summary of high-altitude fall accidents in construction.

Author	Research Methods	Key Findings/Contributions
Oliveira et al. [5]	HFACS model	Poor organizational management is the main cause; Roofers aged 31 to 44 have a high incidence of accidents
Khan et al. [14]	Systematic review of literature	Missing guardrails and lack of personal protective equipment are the main risk factors
Zermane et al. [15]	Statistical analysis, Fault tree analysis	Failure to wear Personal Protective Equipment is the main reason
Piao et al. [16]	Computer vision, DBN	It can automatically detect risk factors and dynamically assess workers’ fall risk
Zermane et al. [17]	Machine learning model, Dataset testing	Management factors and individual characteristics factors recorded the highest accuracy
Qi et al. [18]	Machine learning model, Ensemble learning model	Ensemble learning models provide more stable predictions and can identify causal factors efficiently
Niu et al. [19]	Risk coupling analysis model, K-means clustering analysis	The higher the risk coupling value and the higher the risk of accidents
Peng et al. [20]	Grey-DEMATEL, ISM	Low-security awareness is the most significant factor contributing to falls from heights
Halabi et al. [21]	Correlation analysis, Logistic regression analysis	The usage of fall protection has no considerable improvement; The prediction model could correctly diagnose the injury degree outcome by 77.7%
Duan et al. [22]	Personalized stability monitoring framework	The study provides a practical reference for active safety monitoring for workers with high fall-from-height risk
Tehrani et al. [23]	Experimental study, Quantitative analysis	Working at height increases mental fatigue; Working-at-height group had higher mental fatigue levels
Wong et al. [24]	Structural equation modelling, Mediation analysis	The safety management practices were influential in attitude towards using PPE; PU and PEOU were crucial determinants
Robson et al. [25]	Longitudinal survey, Quasi-experiment	The incidence rate of lost-time claim injuries attributed to falls targeted by the training declined

. Although many achievements have been made in the research of high-altitude fall accidents in construction, there are still some shortcomings. The analysis of the cause of the accident mostly focuses on the surface factors, and lacks systematic theories and methods. Most studies focus on specific countries and do not consider differences in regional norms, worker behavior and environment. While these studies have significantly advanced our understanding of risk factors, a critical gap remains in the dynamic, temporal analysis of accident causation, particularly concerning how major disruptive events reshape safety landscapes. The COVID-19 pandemic, as a global exogenous shock, provides a unique natural experiment to examine the evolution and resilience of safety management systems. Few studies have systematically compared accident patterns before and after the pandemic to identify newly emerging safety hazards and changes in existing safety hazards. This analysis helps identify the changing characteristics of risk factors, which is crucial for formulating robust safety policies.

To fill this research gap, this study adopts a research method that combines LDA model and Apriori algorithms. Although predictive models such as random forests and deep learning provide strong classification capabilities, the goal of this study is interpretive rather than predictive. LDA model can inductively discover potential topics from unstructured text without a pre-set classification framework, offering a comprehensive and objective perspective on the causes of accidents. Subsequently, the Apriori algorithm is used to reveal the co-occurrence relationships and association rules among these topics, uncovering the interactions between different factors. This combined approach constructs a transparent and interpretable analytical framework that can elucidate individual causes as well as reveal the complex interrelationships of causative factors. This paper adopts LDA topic modeling and Apriori association rule analysis, combined with visualization technology, having more comprehensive advantages in data-driven, dynamic adaptability and management guidance.

3. Materials and Methods

The research and analysis were carried out according to the methodology shown in Figure 1.

3.1. Text Mining Method

In this paper, LDA was selected as the text mining method [26]. Compared with other text mining methods, LDA shows high scalability and robustness when processing large-scale text data. LDA model is a text topic analysis method based on a probabilistic graphical model. It was first proposed by Blei et al. [27] in 2003, aiming to automatically discover the hidden topic structure of text data by analyzing it. The flow of the LDA algorithm is shown in Figure 2.

The LDA model structure diagram is shown in the Figure 3. Among them, $\alpha$ and $\beta$ are Dirichlet’s prior parameters, m represents the m text, ${\theta }_m$ represents the topic distribution of text m, n represents the number of words in text m, $z_{m,n}$ represents the topic corresponding to the nth term of text m, $w_{m,n}$ represents the nth term of text m, ${\varphi }_{z_{m,n}}$ represents the distribution of words of the topic, K represents the total number of topics, M represents the total number of texts, k represents the kth topic, $N_{\mathrm{m}}$ represents the total number of words in the m document, Dir represents the Dirichlet distribution, and Mult represents the multinomial distribution.

The detailed generation process of the topic is as follows:

①

Generate the topic distribution ${\theta }_{\mathrm{m}}$, ${\theta }_{\mathrm{m}}|\alpha ~Dir(\alpha )$ of document $\mathrm{m}\in (1,M)$ by sampling from the Dirichlet distribution $\alpha$.

②

For each term $w_{m,n}$ in document m, where $1\in (1,N_m)$. First, sample the topic $z_{m,n}|{\theta }_m~Mult({\theta }_m)$, of the term $w_{m,n}$ from the multinomial distribution ${\theta }_{\mathrm{m}}$. Then, sample the term $w_{m,n}|{\varphi }_{z_{m,n}}~Mult({\varphi }_{z_{m,n}})$ from the term distribution ${\varphi }_{z_{m,n}}$, where ${\phi }_{z_{m,n}}|\beta ~Dir(\beta )$. Given parameters $\alpha$ and $\beta$, the joint probability distribution of the model is:

$$P(\theta ,\varphi ,z,w | \alpha ,\beta )=({\displaystyle \prod _{k=1}^{K}P({\varphi }_k | \beta ))}P(\theta | \alpha ){\displaystyle \prod _{n=1}^{N}P(z_n | \theta )P(w_n} | z_n,{\varphi }_{z_n})$$

3.1.1. Data Acquisition and Preprocessing

Accident investigation report is an important data source for safety risk accident analysis [28]. Compared with other documents, the accident investigation report is more accurate and abundant in describing accidents. Therefore, this study used the accident investigation report as the corpus source for text mining. This study collected relevant accident investigation reports from the administrative departments of various provinces and cities. The accident investigation reports were screened, excluding those that were too brief or lacked a description of the accident causes. A total of 1223 investigation reports of high-altitude fall accidents during construction in China from 2014 to 2023 were finally selected as research data. The number of reports is relatively sufficient, which could ensure the objectivity in analyzing the causes of high-altitude fall accidents through text mining. There may be deviations and inconsistencies in the survey reports from different provinces and cities. To reduce such issues and ensure the quality and consistency of the data to support robust topic modeling, this study established a synonym replacement table to standardize different expressions of the same concept. Additionally, during the context filtering process to remove stop words, special attention was given to retaining words that have safety significance, such as “failure” and “lack”. These measures are aimed at minimizing interference factors and enhancing the reliability and stability of topic identification.

Data preprocessing mainly includes Jieba word segmentation [29], synonym substitution and removal of stop words. Word segmentation is the process of converting a continuous sequence of words into words. This paper uses Jieba word segmentation, which supports the import of custom professional vocabulary, to segment the text. Because the compilation of investigation reports on high-altitude fall accidents in construction has not yet formed a unified standard, the same type of accident investigation report may have different expressions, so it is necessary to add a synonym replacement table to replace such words. A list of synonyms is shown in

Table 2. List of synonym.

Origin	New
Safety hazard	Potential accident dangers
Hidden danger	Potential accident dangers
Safety accident hidden danger	Potential accident dangers
Protective measures	Safety protection measures
Protective equipment	Labor protective equipment
Education	Safety education
Protection	Safety protection

. Stop words are terms that often appear in texts, such as punctuation, conjunctions and pronouns, which are meaningless to the text analysis and even interference with the subsequent data analysis. Therefore, is necessary to filter stop words. This study is based on the Harbin Institute of Technology’s stop word list and manually adds high-frequency meaningless words related to construction accidents such as “accident, occurrence, cause, death, injury”, etc. Finally, the vocabulary set of each document is obtained, which is used as the input feature of the LDA model.

3.1.2. Determine the Number of Topics

At present, there are three common methods to determine the number of topics by calculation: Perplexity, Bayesian statistical model and Hierarchical Dirichlet Processes. This study employs a dual criterion method, combining perplexity and coherence, to determine the optimal number of topics for high-altitude fall accidents topic mining. Perplexity measures the uncertainty of a text being identified as a certain topic by the model after training. Therefore, the perplexity value is smaller, which means that the topic recognition ability of the LDA topic model is better. The calculation formula of perplexity is shown as follows:

$$perplexity=\exp \left(-{\displaystyle \frac{{\displaystyle \sum _{\mathrm{d}}{\displaystyle \sum _i^{N_d}\mathrm{lnp}(w_{d,j})}}}{{\displaystyle \sum _{\mathrm{d}}{N}_d}}}\right)$$

Among them, $N_d$ represents the word frequency in document d, $w_{d,j}$ represents the jth word in document d, and $P(w_{d,j})$ represents the probability of each word.

Perplexity does not always reflect the interpretability of the topic. Therefore, coherence scores are introduced to evaluate the semantic consistency of the topic by calculating the semantic similarity between high-frequency words within each topic. The higher the coherence, the more relevant the semantics of the vocabulary within the topic, making the topic easier to interpret.

The perplexity and coherence curve of the topic model is drawn by using the collected text data. Figure 4 shows that when the number of cluster topics for high-altitude fall accidents is K = 24, the model has a good generalization ability at this time, and the topic has good semantic interpretability. Therefore, K = 24 is determined as the number of topics clustered in the cause of high-altitude fall accidents.

3.1.3. Identify the Influencing Factors

Dirichlet priors α and β were set to 0.1 and 0.01 respectively, following common practices in short-text topic modeling. The accident document set and the number of topics were input, and the LDA model was constructed by python coding to extract the cause topics of the high-altitude fall accidents. The LDA topic model was used to obtain the corresponding topic feature words under each topic.

3.2. Visualization of Cause Topic Association Rules

In order to more efficiently and accurately identify and obtain the related information of cause topics, so as to take measures. A mining model of cause topics association rules based on Apriori algorithm is constructed [30,31], and the strong cause topics association rules are mined by setting the minimum support and minimum confidence parameters. The Apriori algorithm was applied with a minimum support threshold of 0.01, minimum confidence of 0.7, and a lift value greater than 1.5 to ensure meaningful and strong association rules. The formulas for support and confidence are as follows. And the visual network diagram of cause topic association is drawn.

$$\mathrm{support}(A)={\displaystyle \frac{\sigma (A)}{N}}$$

$$\mathrm{confidence}(A\to B)={\displaystyle \frac{\sigma (A\cup B)}{\sigma (A)}}$$

$$\mathrm{lift}(A\to B)={\displaystyle \frac{\sigma (A\cup B)}{\sigma (A)\cdot \sigma (B)}}$$

$\sigma (A)$ is the collection where event A occurred, $\sigma (B)$ is the collection where event B occured, N is the total collection, and $\sigma (A\cup B)$ is the collection where event A or event B occurred.

Data Mining of Association Rules was proposed by American scholar Agrawal in 1993, which is to find out the potential association from a large number of data or objects [32], and find frequent patterns or association rules based on feature words from large-scale text sets. The hidden relationship in the text information is mined for law analysis. The association rule X → Y represents an association such that a document $D_i$ contains all the terms in itemset X, and then the document $D_i$ is also associated with the terms in itemset Y [33].

3.3. Visualization of Spatio-Temporal Correlation of Accident Causes

Construction is a complex and dynamic process where the interplay of various factors can lead to safety accidents [34]. To analyze these accidents, we examine the spatio-temporal characteristics of their causes, helping us understand how they vary across different times and locations.

To reveal how the causes of high-altitude fall accidents in construction change yearly, we set a one-year time window, calculated the topic intensity for each window, and displayed the trend with a line chart. To compare the intensity changes of these accidents’ causes before and after the COVID-19 pandemic, we divided the study period into two four-year intervals. We calculated the intensity distribution and transfer characteristics for each period and used a Sankey diagram to show the differences and evolution trends.

To understand the quarterly distribution of these accidents’ causes, we conducted word frequency statistics based on accident characteristic words and drew a radar chart of these words. To explore the correlation between accident causes and spatial locations, we used the locations and cause labels from the text data. We selected the top 60 risk feature words for each construction location to form a “construction location-cause feature word” matrix. Then, we used Gephi 0.10.1 to visualize the network diagram of accident cause associations related to construction locations by analyzing factors like average path length, nodes, edges, and degrees [35].

4. Data Analysis

4.1. Results of Text Mining

From the perspective of topic causation analysis, the words with ambiguous meanings and large semantic ranges under each topic were eliminated, and the feature words in line with the causation process of the accident were retained. The final extracted accident causation topic feature words are shown in

Table 3. The topics of the causes of high-altitude fall accidents in construction.

Serial Number	Topic Words Included in the Topic	Topic Summary
0	Wear, labor protective equipment, safety helmet, possess, safety belt, use, measures, provide, education and training	Incorrect use of labor protective equipment
1	Platform, set up, operation, rise and fall, regulations, specification, safety protection, safety technology, construction, erection	Operating platform is not standard
2	Dismantle, tower crane, demolish, operation, unlicensed, lifting machinery, begin to work, audit, illegal operation, hoisting	Illegal disassembly of equipment
3	Regulations, violation, possess, work qualification, safety production, safety management, contract out, management, qualification, rule	Violation of safety management regulations
4	Risk, safety management, identify, safety awareness, implementation, deficiency, weak, safety education, safety training, possess	Safety management is not implemented
5	Edge, balustrade, eliminate, demolish, set up, potential accident dangers, protection facilities, balcony, adopt, find out in time	Failure to set up edge protection facilities
6	Hanging baskets, exterior wall, use, safety rope, wire rope, hang, fracture, inspection, lifting tool, fix	Failure to use specification lifting tool
7	Hole, advance reservation, potential accident dangers, measures, cover plate, adopt, eliminate, find out in time, template, technology	Failure to eliminate potential accident dangers in time
8	Lifter, use, install, hoisting, demolish, maintenance, hoisting cage, detection, operation, force	Lifter maintenance is not in place
9	Regulations, violation, specification, safety technology, construction, hoisting, requirements, use, by rule, safety production law	Violation of safety technical regulations
10	Hoister, operation, climb, springboard, protective door, window, fix, without authorization, safety awareness, directive rule	Illegal operation of hoister
11	Failure to, find out in time, arrest, safety awareness, training, on-site safety management, weak, safety management, eliminate, safety education	Weak safety awareness
12	Glass, awning, install, plank, holes, ceiling, wear, safety belt, set up, sign	Failure to wear safety belt
13	Get out of line, weak, safety awareness, neglect, adventure work, perform, adventure, safety protection measures, safety rope, management	Adventure work
14	Elevator shaft, safety protection, set up, demolish, requirements, safety warning signs, protection facilities, lack, guard rail, risk factors	Failure to set up safety warning signs
15	Install, special operations, training, operation certificate, safety education, disclosure, light well, begin to work, work qualification, management	No special operation qualification
16	Steel structure, roofing, demolish, roof, color steel plate, safety protection measures, disclosure, organization, safety belt, color steel tile	Safety protection measures are not in place
17	Supervise and urge, wear, use, inspection, eliminate, work safety, safety belt, safety operation rules, safety production rules and regulations, labor protective equipment	Inadequate safety inspection
18	Elevator, install, illegal operation, perform, eliminate, find out in time, duty, management, adopt, use	Illegal operation
19	Safety belt, training, disclosure, wear, safety education, safety technology, safety helmet, correct, use, implementation	Failure to disclose safety technology
20	Work qualification, illegal, get out of line, contract out, contract for, organization, by rule, unload, supervise, simple	Contracting out without qualification
21	Supervision, rectify and reform, perform, implementation, supervise and urge, subcontract, discovery, management, responsibility, requirements	Rectification is not implemented
22	Safety production, implementation, responsibility, subject, check, system, training, governance, safety production responsibility system, education	Subject of responsibility for safety production has not been implemented
23	Scaffold, steel tube, hand board, erection, safety net, use, standard, safety belt, fastener, set up	Scaffolding erection is not qualified

.

The excavated topic feature words described the cause topic of the accident investigation report text in detail. The cause accident topics are summarized from four aspects: personnel, environment, equipment, and management. The summary of the accident cause topics is as follows.

Among the cause topics related to personnel, topics 0 and 12 focus on the use of safety protective equipment by staff, mainly referring to the improper use of “safety helmet” and “safety belt”. Topics 3, 9, 11, 13 and 18 refer to “violation of safety management regulations”, “violation of safety technical regulations”, “weak safety awareness”, “adventure work” and “illegal operation” respectively. Topic 11 focuses on unconscious mistakes of people, such as “failure to”, “find out in time” and “weak”, mainly referring to errors caused by negligence of staff. Topic 3, 9, 13 and 18 focus more on the mistakes made by staff in violation of relevant regulations, which are subjective intentional mistakes. Topic 2 and topic 10 focus on the illegal use of equipment, mainly referring to the use of “tower crane”, “lifting machinery”, “hoister”, etc., involving the specific links of “dismantle”, “demolish”, “hoisting”, “operation” and so on. Topic 15 focuses on the professional qualifications of special operators, such as “operation certificate” and “work qualification”, reflecting that the staff made mistakes due to insufficient qualifications.

Topics 1, 6, 8 and 23 indicate the equipment-related causes, which are “operating platform is not standard”, “failure to use specification lifting tool”, “lifter maintenance is not in place” and “scaffolding erection is not qualified”. The operating platform is not standard specifically means that the structure of the platform itself does not meet the safety technical regulations, and the safety protection measures of the platform are not in place. Lifting tool use is not standard mainly refers to the use of “hanging baskets” “safety rope” “wire rope”, etc. The use of equipment that does not meet the requirements of the specification may lead to fracture or fixation instability during use, endangering the safety of operators. Lifter maintenance is not in place includs “install”, “hoist”, “demolish” and other links. The lifter maintenance is not in place may lead to uneven force on the cage, resulting in safety accidents. Scaffolding erection is not qualified mainly includes the nonstandard use of facilities such as “steel tube”, “hand board”, “safety net”, and “fastener”, which may lead to the instability of the scaffold structure and affect the construction safety.

Topic 4, 17, 19, 20, 21 and 22 indicate the causes related to management, which are “safety management is not implemented”, “inadequate safety inspection”, “failure to disclose safety technology”, “contracting out without qualification”, “rectification is not implemented” and “subject of responsibility for safety production has not been implemented”. Topic 4 mainly involves the lack of safety awareness, safety education and safety training. Topic 17 mainly refers to the lack of attention to safety operation rules and safety production rules and regulations, and the inadequate inspection of the use of labor protective equipment. Topic 19 can cause employees to work without understanding the potential risks and correct safety operation, which can lead to accidents. Topic 20 focuses on the existence of unqualified contracting, which mainly refers to the contracting of construction to an organization without corresponding qualifications, in which case the quality of construction cannot be guaranteed and accidents are prone to occur. Topic 21 is concerned about the failure to implement rectification, which mainly involves the neglect of supervision and performance. Topic 22 mainly refers to the imperfect production safety responsibility system leading to the failure of responsibility subject to to be implemented, which may lead to the chaos of safety management, thereby increasing the probability of high-altitude fall accidents.

Topics 5, 7, 14 and 16 are the causative topics related to the construction environment, including “failure to set up edge protection facilities”, “failure to eliminate potential accident dangers in time”, “failure to set up safety warning signs” and “safety protection measures are not in place”. Topic 5 focuses on the safety protection of the production site, the main protection locations are “edge” and “balcony”, etc., and the main manifestation is that protective railings are not set up in time. Topic 7 mainly refers to the safety problems caused by the failure to eliminate the potential accident dangers in time, “find out in time” reflects the negligence of the on-site inspection. Topic 14 mainly indicates that there is no warning for dangerous areas, mainly after the removal of elevator shaft protection, resulting in the occurrence of safety accidents. Topic 16 refers to the improper use of safety belts during roof operations, which leads to the occurrence of safety accidents.

4.2. Results and Analyses

4.2.1. Cause Topic Correlation Results

By using the Apriori algorithm to mine the data set, the visualized network diagram of the cause topic association of the high-altitude fall accidents in construction is obtained, as shown in Figure 5.

It can be found from Figure 5 that many cause topics are associated with each other. The size of the node represents the number of the association relationship, the larger the node; the more associations it has with other nodes. The thickness of the edge represents the strength of the association rule; the thicker the edge, the stronger the association rulers between the two nodes. It can be found that “rectification is not implemented”, “subject of responsibility for safety production has not been implemented” and “violation of safety technical regulations” are associated with more, indicating that in the construction at high altitude, it is necessary to focus on prevention and strengthen monitoring for these key causes, so as to reduce the probability of other causes. The causal association rules can be clearly obtained by network diagram. It can be found that “safety protection measures are not in place” and “failure to wear safety belt”, “incorrect use of labor protective equipment” and “violation of safety management regulations”, “illegal disassembly of equipment” and “rectification is not implemented”, “safety protection measures are not in place” and “adventure work”, “illegal operation” and “violation of safety technical regulations” have a strong relationship. The network diagram can be used to efficiently and accurately identify and obtain the key cause association, so as to take targeted measures to strengthen the safety risk management of site construction.

4.2.2. Cause Spatio-Temporal Correlation Results

After extracting LDA topics from the text data of the investigation reports of falling accidents from 2014 to 2023, a line chart of topic intensity variation with the year was established, as shown in Figure 6.

According to the statistics of the evolution trend of 24 topics, it was found that four topics, Topic 0, Topic 17, Topic 18 and Topic 22, showed an increasing trend. There were five decaying topics, including Topic 2, Topic 7, Topic 11, Topic 12, and Topic 20. The stable topics are Topic 1, Topic 5, Topic 6, Topic 8, Topic 10, Topic 14, Topic 15 and Topic 16. There are seven zigzagging topics, namely Topic 3, Topic 4, Topic 9, Topic 13, Topic 19, Topic 21 and Topic 23. As can be seen from Figure 3, Topic 4 and topic 18 almost did not appear in 2014. The topic intensity of Topic 4, Topic 11 and Topic 19 in 2015 is relatively large. The topic intensity of Topic 4 in 2016 is relatively large. From 2017 to 2020, the change in the intensity of the topics was relatively slow. The topic intensity of Topic 0, 19, 21 and 22 in 2021 is relatively large. The topic intensity of topic 0, 21 and 22 in 2022 is relatively large. The topic intensity of topic 0, 4, 19 and 22 in 2023 is relatively large. The main problems in recent years are similar, indicating that safety management has played a certain role, but the main problems have not been solved, and should be continuously concerned in the future. For the causes of common accidents, safety managers need to strengthen safety prevention and control and governance from the aspects of technology, equipment, daily inspection, education and training, and improvement of management system, so as to prevent and control hidden dangers early when there are no hidden dangers, and discover and manage hidden dangers early when there are hidden dangers. For the unique causes of accidents in some years cannot be paralyzed. Managers should carefully analyze the accident causes, and effectively control the recurrence of accidents to avoid them evolving into frequent causes of accidents.

According to the evolution rule of topic intensity before and after the epidemic, Sankey diagram of the evolution of topic intensity of accident causes is drawn, as shown in the Figure 7.

Based on the analysis of the topic intensity of the accident causes before and after the COVID-19 pandemic, it can be found that the accident causes show the following characteristics. On the one hand, some key causes have long-term stability and continue to affect the safety of construction. On the other hand, the intensity and manifestation of the cause of the accident are affected by the change of the external environment, showing dynamic volatility. Safety management is not implemented, failure to disclose safety technology, and incorrect use of labor protective equipment are the main problems before and after the COVID-19 pandemic, which indicates that these factors are the long-term weak links in the safety management of construction, and it is necessary to continue to strengthen supervision and improvement measures. The change of the cause topic before and after the epidemic also reflects the different challenges faced by the safety management of construction. After the epidemic, the implementation of the main body of responsibility for safety production has become the primary cause, which indicates that there are shortcomings in the implementation level of construction safety management after the epidemic, and it is necessary to further strengthen the supervision and implementation mechanism of hidden danger rectification.

The Sankey diagram not only visualizes the flux of topic intensity but also reveals the specific impact of the COVID-19 pandemic on safety management. The significant increase in the intensity of “subject of responsibility for safety production has not been implemented” after the pandemic is primarily attributed to chain reactions caused by supply chain interruptions and halts in work due to the pandemic. In this context, the clear definition and enforcement of safety responsibilities may have been overlooked or face execution difficulties, thus forming potential safety hazards. Qualitative analysis of representative accident reports from the post-pandemic period revealed that disruptions in safety supervision and reduced on-site inspections due to mobility restrictions contributed to the lax implementation of responsibility systems. For instance, multiple reports cited “subject of responsibility for safety production has not been implemented” and “safety management is not implemented”. These factors were often coupled with increases in topics such as “rectification is not implemented” and “failure to disclose safety technology” suggesting a cascading effect of weakened oversight after the pandemic.

According to the statistical results of accident causation, the radar chart of the quarterly distribution of accident causation is drawn, as shown in Figure 8.

Through the radar chart of the cause characteristic words of accidents in four quarters, clear seasonal patterns in accident causation emerge, influenced by environmental, operational, and behavioral factors. In spring, the issues of safety belts usage and regulations inplementation are more prominent, which may due to insufficient rebuilding of safety protocols after the slowdown in winter, as well as potential gaps in workforce restructuring and retraining. In summer, the issues of safety management and safety belts usage are more pronounced. Heat fatigue, accelerated work pace, and heightened production pressures my contribute this situation. Autumn faces similar challenges as summer, with training execution and safety belts usage being the main issues. This may be due to the pressure of needing to complete the project on time, as well as the fatigue that arises after a long construction season. During the winter season, the issuse of safety belts usage is most ptominent. This may be due to wearing heavy clothing, which reduces flexibility, and the weather conditions affecting the workers’ awareness. From a macro point of view, it can be analyzed that the use of safety belt, the implementation of regulations, the implementation of management, education and training will lead to a high probability of causing accidents in each quarter. It can be seen that the common problems existing in the construction safety management in each season, as well as the unique cause characteristics of each season, can take targeted safety management measures for each season to reduce the occurrence of safety accidents.

According to the top 60 risk characteristic words of each construction location, the visual network diagram of the correlation between the cause of the accident and the construction location is drawn, as shown in Figure 9.

Figure 9 reveals pronounced correlations between accident causes and specific construction locations, each characterized by distinct risk mechanisms. Accidents in hole operations mainly occur in elevator shafts and light wells. This is because during dynamic construction phase, protective covers often change or are missing, creating transient hazard areas. Accidents in edge operations usually occur on balconies, mainly related to improper safety belt use, often caused by insufficient anchor points and workers taking risk behavior under time constraints. The main reason for the accident in hanging operation is the illegal or irregular use of equipment such as hanging baskets and hoists. The direct cause of the accidents is the fracture of safety ropes or safety belts. The accidents of the platform operation mainly occurred during scaffolding operation. Scaffolding accidents often involve structural instability or missing components, highlighting the importance of strict installation standards and inspections. The main reason for roofing operation accidents is the improper use of labor protection equipment. When working on the roof, due to the special working environment, the failure of protective measures can easily lead to accidents. And the occurrence of other accidents is mostly because of lax management and supervision. Focusing on the most likely hazards in each specific construction location is crucial for accurately preventing the recurrence of accidents.

5. Discussion

The study makes three theoretical contributions through analysis on high-aititude fall accidents of construction in China. First, this study contributes to the theoretical understanding of construction safety by comparing accident causes before and after the COVID-19 pandemic in China. While previous studies have highlighted the disruption caused by COVID-19 on construction safety management [12,13], there is little empirical research examining its specific impacts on construction safety management. Our study reveals a significant post-pandemic increase in the topic intensity of “subject of responsibility for safety production has not been implemented,” a factor less emphasized in pre-pandemic literature. This shift suggests that the pandemic may have exacerbated systemic weaknesses in safety governance, particularly in accountability mechanisms. Furthermore, this result empirically supports the theoretical view that risk attributes are dynamic and may undergo fundamental changes due to exogenous shocks.

Second. the study advances the theoretical framework by demonstrating the dynamic and evolving nature of accident causes in the construction industry. The temporal analysis shows that certain causes, such as “incorrect use of labor protective equipment” and “inadequate safety inspection,” have increased over time, while others like “illegal disassembly of equipment” and “failure to eliminate potential accident dangers in time” have decreased. This evolution suggests that safety management efforts have had differential impacts on various causes. The persistence of some causes indicates deep-rooted issues in safety culture and regulatory enforcement, while the decrease in others highlights the effectiveness of targeted interventions. This dynamic perspective, supported by the work of Mohandes [9] and Shen [11], enriches the understanding of how accident causation patterns shift with industry practices, technological advancements, and regulatory changes. This emphasizes the importance of adopting a dynamic systems perspective in security research, shifting from the identification of static factors to modeling how factors change as the environment evolves.

Third, the application of advanced text analysis techniques, particularly the LDA approach, represents a significant theoretical contribution. This study showcases how this method can uncover hidden patterns and causal factors in unstructured accident report data. The LDA model effectively identified 24 distinct accident cause topics, providing a granular understanding of the multifaceted nature of high-altitude fall accidents. Furthermore, the Apriori algorithm is used to mine the cause topic association, which improves the understanding depth of the accident causality. This is in line with the view of Li that text mining technology should be used comprehensively to systematically analyze accident data [36]. By establishing a robust methodological framework, this study paves the way for more sophisticated and automated safety risk assessments in the construction industry. And at the methodological level, it supports the view that accidents are caused by the interaction of interconnected factors.

Last but not least, the findings have direct practical implications for improving construction safety management. Safety managers can prioritize interventions based on the identified high-risk factors, such as enhancing supervision of labor protective equipment use and strengthening safety inspections. The analysis of cause topic association also suggests the key direction of safety management. Strengthening prevention and monitoring for key cause factors can reduce the probability of other cause factors. The seasonal distribution of accident causes suggests the need for targeted safety campaigns aligned with specific seasonal risks. The correlation between accident causes and construction locations highlights the importance of location-specific hazard inspections. Implementing these strategies can lead to more effective risk control and a reduction in high-altitude fall accidents.

6. Conclusions

This study has provided valuable insights into the complex landscape of high-altitude fall accidents in the construction industry based on the context of Chinese construction industry. Through advanced text mining, we have identified key causal factors across personnel, equipment, management, and environmental dimensions. The analysis revealed that issues such as incorrect use of protective equipment, inadequate safety inspections, and failure to implement safety management protocols persist as major contributors to accidents. The temporal evolution of these factors showed both improving trends in some areas and concerning increases in others, highlighting the dynamic nature of construction safety challenges. These findings emphasize the need for targeted interventions, improved safety culture, and enhanced regulatory compliance to reduce the incidence of high-altitude fall accidents in construction. Therefore, we propose the following recommendations. Implement a specialized training program to focus on addressing the long-standing issue of the proper use of equipment such as safety belts. Establish a strict review mechanism to investigate and rectify instances of “inadequate inspections” and “failure to implement rectifications.” Strengthen the accountability system for safety management to address shortcomings during the implementation of safety production responsibilities.

Despite the comprehensive analysis presented, this study has several limitations that should be acknowledged. First, the data relied on accident investigation reports from different provinces and cities, which may contain reporting biases or incomplete information. For example, the mismanagement of underreporting or overemphasis on worker behavior. This may lead to an undervaluation of systemic issues. Furthermore, the incompleteness of certain reports may affect the comprehensiveness of topic modeling results. Second, while the LDA topic modeling and machine learning approaches provided valuable insights, they may not capture all nuances of accident causation, particularly in cases with multiple interacting factors. Third, the study focused primarily on the Chinese construction context, which may limit the generalizability of findings to other regions with different regulatory frameworks and industry practices. Future research could address these limitations by incorporating multiple data sources, refining analytical models to better handle complex causal relationships, and conducting cross-national comparative studies to enhance the applicability of safety management strategies.