Application of text mining and causal analysis for risk identification in air traffic control safety operations

Abstract

Risk assessment in air traffic control (ATC) operations is a core element in ensuring flight safety. Addressing the limitations of current assessment techniques in deeply identifying critical risks within ATC procedures, this study proposes a method that integrates text mining with causal analysis for comprehensive risk identification. First, the BERTopic model is applied to conduct deep semantic analysis of ATC-related safety occurrence reports, extracting key risk themes and their representative keywords with enhanced semantic coherence and diversity. Second, by incorporating domain-specific operational knowledge and extracted keywords, a structured Risk Factor Representation Framework is established. Event chains are annotated to construct a Causal Network for ATC Operational Safety. Finally, topological analysis (including PageRank and betweenness centrality) and vulnerability analysis are employed to identify critical risk nodes within the network. Using incident reports from China’s civil aviation ATC operations in 2021 as a case study, the results demonstrate that the proposed method significantly improves the accuracy of risk theme identification and highlights the prominent role of nodes with high PageRank and centrality in system safety. By transforming fragmented textual information into a structured risk association network, this study enables a shift from scattered factor identification to systematic causal analysis. It offers a novel approach for precise hazard detection and timely risk mitigation, contributing to enhanced civil aviation safety governance.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-25142-w.

Introduction

The continued rapid growth of global air transportation, coupled with the increasing complexity of the airspace operational environment, has posed severe challenges to aviation safety¹. As a large-scale, multi-stage system with tightly coupled processes, civil aviation operations are vulnerable to unsafe events at any stage, which may result in serious consequences such as aircraft accidents, casualties, and disruptions to flight operations¹, as evidenced by several major crash incidents in recent years². These challenges underscore the pressing need to strengthen safety risk management across the sector. To address this issue, the International Civil Aviation Organization (ICAO) has promoted the establishment of a continuously improving safety governance framework, within which air traffic control (ATC) plays a central role in ensuring operational safety^3,4.

China, as one of the fastest-growing countries in global air transportation, has maintained a historically high operational intensity in its ATC system. However, the frequent occurrence of ATC-related safety incidents reveals that the current risk assessment mechanisms are no longer sufficient to meet the demands of safety assurance in such a complex operational environment. This is particularly true for the foundational stage of risk identification^5,6. Although vast amounts of unstructured risk-related data (e.g., incident reports and operational logs) have been accumulated in ATC operations, the multi-source and semantically complex nature of these data creates significant challenges for extracting and analyzing key risk information⁷, thereby limiting the accuracy and comprehensiveness of risk identification and evaluation.

Text mining techniques offer powerful tools to address these challenges^8–10. Their core value lies in transforming unstructured texts into structured information, enabling the extraction of latent risk factors and their interrelations^11–14. In particular, text mining methods that use deep semantic modeling, such as the BERTopic model, perform exceptionally well in analyzing complex ATC-related texts. By integrating the semantic power of BERT-based word embeddings with clustering algorithms, BERTopic is well-suited for identifying domain-specific terminology and constructing risk themes from semantically entangled, jargon-heavy reports^11,13–15.

Nevertheless, existing approaches face two major limitations: (1) pure text mining often identifies only fragmented or isolated risk themes, failing to uncover the complex associations between them; and (2) traditional linear analytical methods are inadequate for capturing the networked and interactive structure of risk elements, thus constraining the depth and scope of risk analysis.

To overcome these limitations, this study proposes an integrated method that combines text mining and causal analysis to systematically identify key risk factors in ATC operations. The core idea is to use deep text mining to extract foundational risk elements and then apply causal analysis to construct a structured risk association network. Causal analysis facilitates the identification of critical nodes and structural patterns, helping to reveal the underlying mechanisms of risk evolution^16–20.

The main contributions of this study are as follows:

1. A refined methodological framework for risk identification and analysis in ATC-related safety occurrences is proposed. This framework systematically integrates risk topic extraction, factor representation, network construction, and critical node analysis, enabling the transformation of fragmented textual information into a structured risk network.

2. An optimized BERTopic-based semantic modeling strategy is developed to accommodate the textual characteristics of ATC operations. By addressing challenges such as terminology density and semantic overlap, the risk topic extraction process is improved, enhancing the accuracy of risk information identification under complex linguistic conditions.

3. A causal relationship network for civil aviation ATC operations is constructed and thoroughly analyzed. Building upon an enriched risk factor representation system, topological analysis is applied to uncover deeply embedded key risk factors and their influence pathways within the system.

Literature review

Risk Identification

Research on risk identification in civil aviation has primarily centered on the structuring of safety reports and the development of corresponding risk models. Sun²¹ integrated holographic modeling with risk filtering, rating and management (RFRM) theory to establish a model for identifying and assessing risks in flight training, enhancing the capability to detect critical risk factors. Li²² proposed a dynamic risk assessment model for runway excursions based on quick access recorder (QAR) flight parameter data, improving the identification and prediction of high-risk flights during the landing phase.

Compared to traditional methods that rely on structured data, text mining techniques have emerged as a powerful tool for extracting semantic information and domain knowledge from unstructured aviation safety narratives. Rose²³ employed the Structural Topic Model (STM) to uncover latent risk themes from Aviation Safety Reporting System (ASRS) and National Transportation Safety Board (NTSB) reports. Xiong¹⁶ applied a Bi-directional Long-Short Term Memory and Conditional Random Field (BERT-BiLSTM-CRF) model to extract six categories of aviation safety-related entities and constructed a civil aviation hazard knowledge graph through named entity recognition. Abuzayed²⁴ developed a BERTopic-based model and experimentally demonstrated its superior performance in text mining compared to traditional Non-negative Matrix Factorization (NMF) models. Farea²⁵ empirically showed that, under the same number of topics K, BERTopic generates more stable topic structures and smoother temporal evolution sequences. Bhuvaneswari²⁶ proposed a utility function incorporating multiple normalization and regularization parameters, in which weights are assigned to perplexity and topic coherence metrics to determine the optimal number of topics for BERTopic. These studies highlight the enhanced adaptability and application potential of text mining techniques in aviation safety data processing, semantic recognition, and knowledge modeling.

These studies underscore the strong adaptability and effectiveness of text mining techniques in aviation safety data processing, semantic understanding, and knowledge representation, laying a solid foundation for advanced risk identification in complex operational scenarios.

Causal analysis

Causal analysis methods for civil aviation safety occurrences have become increasingly diversified, primarily encompassing three categories: structural models, coupling analysis, and causal networks.Structural modeling decomposes accident processes through systematic logic to identify root causes and evolutionary paths, making it suitable for attribution analysis in contexts involving intertwined causal factors.Coupling analysis focuses on the dependencies and interactions among contributing factors, uncovering potential risk accumulation and triggering mechanisms within the system. This method is particularly effective for assessing latent risks under multidimensional factor interactions.Causal network approaches construct causality-based graphs to reveal inter-event relationships and causal chains, offering a systematic framework for modeling risks and identifying propagation pathways in complex scenarios.

In terms of structural modeling, Chen²⁷ integrated event tree and fault tree analysis methods to develop a departmental-level safety performance indicator system for airports, enhancing their operational risk assessment capabilities. Che²⁸ proposed a human factor risk analysis framework that combines the Multiple Resource Theory and fault tree analysis to model mental workload overload, and introduced the mental workload overload (MWLOL) gate to trace the root causes of human errors. For coupling analysis, Jiao²⁹ applied a hybrid approach combining Decision Making Trial and Evaluation Laboratory (DEMATEL) and Bayesian networks to evaluate multi-factor coupling and associated risks in flight operations, identifying potential risk contributors. Xiong³⁰ employed text mining and coupling measurement techniques to identify high-risk interactions among pilot operations, aircraft systems, and engines, uncovering the accumulation and triggering mechanisms of systemic risks.In the domain of causal network modeling, Kim³¹ integrated Systems-Theoretic Accident Model and Processes (STAMP), Human Factors Analysis and Classification System (HFACS), and Functional Resonance Analysis Method (FRAM) models to perform a comprehensive system-level analysis of aviation accidents, identifying potential evolutionary trends. Zhang³² constructed a Bayesian network based on NTSB accident data to analyze causal chains, while Nadine³³ combined improved Human Factors Data Domain (HFDD) and HFACS models to examine the latent impacts of human factors on ground operation safety.

In summary, substantial progress has been made in the risk identification and analysis of civil aviation safety occurrences, and a relatively systematic analytical framework has been established. However, existing studies remain largely focused on the flight operation phase, with limited attention to ATC scenarios characterized by strong human–machine collaboration and complex system coupling. In particular, gaps persist in the areas of topic-based semantic extraction and risk chain identification. Moreover, traditional risk assessment methods fall short of meeting the requirements of ATC safety, which demands early identification, rapid warning, and effective intervention. Therefore, there is an urgent need to introduce risk analysis mechanisms with enhanced capabilities in semantic understanding and causal inference.

The subsequent chapters of this study are organized as follows: Chapter 3 introduces the data sources and the theoretical foundation on which the study is based. Chapter 4 presents the empirical analysis, which identifies and analyzes the main risk topics and key risk factors in ATC reports and proposes corresponding risk management strategies. Chapter 5 summarizes the key findings and outlines directions for future research.

Methodology

Overall research framework

Figure 1 presents the overall research framework of this study, which consists of the following four key stages:

1. Data Preprocessing. Air traffic control safety occurrence reports are collected as the primary data source. The raw data undergo cleaning to remove irrelevant and duplicate entries. Abbreviations are expanded to their full forms. In addition, a domain-specific stopword list and customized tokenization dictionary are developed to support accurate processing of risk-related text.

2. Text Mining. The BERTopic model is employed to extract risk topics from the unstructured text of civil aviation air traffic control safety reports.

3. Causal Analysis. Building on previous research and extracted topic keywords, a structured representation system of risk factors is developed. Event chains are annotated from the incident narratives, based on which a causal risk network is constructed. Topological and vulnerability analyses are conducted on this network to identify high-risk nodes, defined as critical risk factors that strongly influence risk propagation within the system.

4. Development of Risk Control Strategies. Drawing upon the structure of the causal risk network and the characteristics of the identified critical risk factors, targeted risk control strategies are formulated.

Fig. 1.

Research process flowchart.

Data source

The dataset employed in this study was sourced from the CAAC East China Regional Administration. To ensure data reliability and analytical validity, the dataset underwent rigorous filtering and verification procedures. It comprises a total of 1427 air traffic safety occurrence reports from the East China region, all recorded during the year 2021. Each report contains information including the time and location of the incident, relevant keywords, and a detailed narrative description of the event. The overall structure of the dataset is illustrated in Fig. 2.

Fig. 2.

Data structure diagram.

Theoretical foundation

Topic modeling method

BERTopic is a four-stage topic modeling approach grounded in the BERT language model. It demonstrates strong generalizability and a high degree of automation, enabling effective performance without the need for extensive parameter tuning. The BERTopic framework comprises the following four key components:

1. BERT Document Embedder. This module encodes raw textual input into high-dimensional semantic vectors using a pre-trained BERT model, capturing rich contextual information.

2. UMAP Reducer. Utilizing a nonlinear dimensionality reduction algorithm, this module compresses the high-dimensional embeddings into a lower-dimensional space, yielding more stable and cluster-friendly representations¹².

3. HDBSCAN Clustering Engine. This unsupervised clustering algorithm groups the reduced embeddings into discrete topic clusters.

4. C-TF-IDF Representer. Within each identified topic cluster, this module calculates the relative importance of terms to generate a representative set of topic keywords.

The overall topic modeling workflow is illustrated in Fig. 3.

Fig. 3.

Diagram of potential risk information mined by BERTopic model.

Causal network analysis method

A causal network is a system modeling approach grounded in complex network theory. It represents elements within a system and their interrelationships through a node–edge structure. In the context of constructing the causal network for air traffic control safety, nodes are defined as operational risk factors, while edges denote the causal relationships among these factors. The causal network analysis method aims to identify critical links in the propagation of risks by examining the structural characteristics and resilience of the network. This method comprises two major components: topological analysis and vulnerability analysis.

1. Topological analysis

Topological characteristics are quantitative metrics that describe risk nodes and their causal linkages within the air traffic control safety causal network. These metrics capture the connection patterns, propagation efficiency, and structural configurations of risk interactions. Through topological analysis, the structural role of each node within an incident chain can be elucidated, facilitating the identification of key risk nodes and enhancing the targeted effectiveness of risk prevention strategies. The following introduces the topological indicators adopted in this study and their applications in risk identification.

1. Node degree

Node degree is a fundamental metric indicating the number of direct connections a node has with other nodes in the network. Given that the air traffic control safety causal network is a directed and weighted graph, node degree is examined from three perspectives: total degree, in-degree, and out-degree. Total degree reflects the sum of all incoming and outgoing connections associated with a node. In-degree denotes the number of directed edges pointing toward a node, indicating how many other risk factors influence it. Out-degree represents the number of edges originating from the node, reflecting its influence on other factors. A higher node degree suggests a greater number of associations with other risk factors, implying a higher likelihood of the node serving as a hub within the accident chain. The formulas for these metrics are presented as follows:

1

2

3

where represents the in-degree of a node; represents the out-degree; represents the total degree. n is the total number of nodes in the network. denotes the number of edges from node i to node j; if nodes and are connected, then ; otherwise, .

• (2)Average degree

The average degree quantifies the mean number of connections per node in the network, serving as an indicator of the overall connectivity density. A higher average degree suggests a denser web of causal relationships among risk factors, increasing the likelihood of multi-path propagation structures emerging within the network. The calculation formula is as follows:

4

where represents the average degree of the nodes.

• (3)Average path length

Average path length measures the overall efficiency of risk transmission across the network. A shorter average path implies faster diffusion of risk, thereby raising the probability of cascading unsafe events occurring within the system. The calculation formula is as follows:

5

where L represents the average path length, and denotes the shortest path length between node i and node j.

• (4)Clustering coefficient

The clustering coefficient characterizes the degree to which neighboring nodes of a risk factor are interlinked through causal relationships. A higher clustering coefficient indicates tighter local connectivity, which facilitates the formation of closed-loop structures. These localized loops can amplify the circulation and localized diffusion of risk. The calculation formula is as follows:

6

where is the clustering coefficient of node i, is the number of actual edges between its neighbors.

• (5)Betweenness centrality

Betweenness centrality evaluates the extent to which a particular risk factor acts as an intermediary on the causal paths between other nodes. A risk factor with high betweenness centrality serves as a critical bridge for risk propagation across different subsystems or regions, and plays a key role in chain reactions of risk events. The calculation formula is as follows:

7

where denotes the Betweenness centrality of node i, denotes the number of shortest paths from node i to node j that pass through node v, and represents the total number of shortest paths between node i and node j.

• (6)PageRank value

The PageRank value estimates the likelihood that a risk factor is targeted by other high-importance nodes. A high PageRank score implies that the node is a convergence point in multiple critical causal paths. Such nodes often represent the endpoints of accident chains or act as common causative factors, making them key indicators for identifying systemic high-risk zones. The calculation formula is as follows:

8

where denotes the PageRank value of node i, d is the damping factor, typically set to 0.85, represents the set of nodes that point to node i, and denotes the number of outbound links from node j.

• 2.Vulnerability analysis

Vulnerability in the air traffic control causal network refers to the degree of performance degradation resulting from random failures or intentional attacks. Vulnerability analysis provides a robust means to assess the resilience of the network. By simulating the failure of individual nodes and evaluating its impact on overall network performance, this method identifies structural weaknesses and supports data-driven risk assessment and early warning mechanisms.

Text mining and causal analysis method

Text mining

This study used text mining methods, including text preprocessing and BERTopic-based topic modeling, to extract and identify the main risk themes in air traffic control (ATC) safety incidents.

1. Text preprocessing

The primary goal of text preprocessing is to convert raw, unstructured textual data into structured, machine-processable input suitable for topic modeling. In Chinese-language processing, conventional pipelines typically involve three fundamental steps: word segmentation, stopword removal, and synonym consolidation. However, when applied to safety incident reports in the ATC domain, these standard methods encounter several domain-specific challenges:

1. Inadequate coverage of domain-specific terminology: Generic text processing tools and lexicons often fail to recognize specialized aviation-related phrases such as “空中交通管制” (“Air Traffic Control”) and “自动飞行控制系统” (“Automatic Flight Control System”). These domain-specific terms frequently serve as semantically critical elements in incident narratives.

2. Extensive use of abbreviations and mixed expressions: The reports contain a wide variety of Chinese-English abbreviations, such as “左发” (short for “左发动机,” meaning “left engine”) and “MSAW告警” (referring to “Minimum Safe Altitude Warning”). These abbreviated forms are often interwoven with technical jargon and vary substantially in usage across individuals, significantly complicating the accurate vectorization and semantic interpretation of the text.

To address the challenges in the preprocessing workflow, this study proposes a customized procedure that incorporates domain-specific semantics from air traffic control. This approach helps reduce information loss during traditional preprocessing and provides a clearer semantic foundation for topic modeling in the context of ATC safety. Figure 4 presents a comparative visualization of word clouds before and after text preprocessing. The detailed steps of the processing workflow are as follows:

1. Text deduplication: Remove semantically redundant information to ensure the uniqueness and clarity of the corpus.

2. Construction of an ATC-specific stopword list: Eliminate non-informative tokens (e.g., general terms like “aircraft” or placeholders such as “Flight A”) and filter out irrelevant punctuation marks.

3. Abbreviation normalization: Expand frequently used abbreviations into their standard full forms based on aviation regulatory standards and expert-curated glossaries.

4. Development of a domain-specific segmentation dictionary: Incorporate ATC-relevant phrases and commonly co-occurring compound terms to improve segmentation precision and enhance the accuracy of named entity recognition.

Fig. 4.

Comparative word cloud visualization of the data before and after preprocessing.

• 2.BERTopic topic modeling

This study employs the BERTopic model to extract and represent risk-related themes from ATC safety occurrence reports. The modeling procedure includes the following steps:

1. Document embedding: The bert-base-chinese model is used to encode the preprocessed incident texts into semantic embeddings. Newly generated embedding vectors are compared with existing ones using cosine similarity, enabling topic assignment for new documents.

2. Dimensionality reduction and clustering: The UMAP algorithm reduces the dimensionality of the embeddings while preserving semantic structure, improving clustering efficiency. HDBSCAN is then applied to perform unsupervised clustering, identifying semantically coherent groups of texts. Each cluster represents a potential risk topic.

3. Topic representation: Based on clustering results, the C-TF-IDF (Class-based Term Frequency–Inverse Document Frequency) algorithm is used to calculate the importance of keywords within each topic. Representative high-frequency terms are selected to characterize the core semantics of each topic, enabling structured representation of safety occurrences in ATC operations.

As BERTopic is an unsupervised learning method, it does not intrinsically determine the optimal number of topics. To address this limitation, evaluation metrics such as Topic Coherence (TC) and Topic Diversity (TD) are utilized to guide the optimization of both the number and content of topics.

Topic Coherence (TC) measures the semantic consistency of keywords within a topic. It is commonly evaluated using Normalized Pointwise Mutual Information (NPMI), as defined in Eqs. (9) and (10). Topic Diversity (TD) reflects the degree of distinction between topics, and is calculated using Eq. (11). Higher values of TC and TD indicate stronger internal coherence and clearer separation among topics, enabling more accurate and comprehensive identification of potential risk characteristics in ATC operations.

9

10

11

where, represents the probability that words and co-occur in the same document; while and denote the probabilities of and appearing in the document, respectively. T denotes the number of topics; k represents the number of keywords per topic; refers to the number of unique keywords across all topics.

Causal analysis

Potential causal relationships often exist among the risk topics identified in ATC operations. To explore the underlying mechanisms behind ATC-related safety occurrences and to identify key risk factors, this study conducts causal analysis by constructing a causal network of ATC operational safety, followed by topological and vulnerability analysis.

1. Construction of the causal network for ATC operational safety

1. Risk factor representation framework

To comprehensively capture the sources of risk in ATC-related safety occurrences, we construct a multi-layered risk factor representation framework. This framework integrates: topic keywords derived from BERTopic modeling, representative incident cases, risk factor classifications from prior research^28,31,32. The risk factors are categorized into five major types: human factors, equipment-related factors, environmental factors, managerial factors, and stochastic event factors. Among them, the stochastic event factors category encompasses atypical or abnormal operational states, such as taxi-back procedures or deviations from assigned taxi routes. Although these irregularities may not directly result in severe outcomes, they often serve as critical intermediate or outcome indicators in the evolution of potential risk chains.

• (2)Causal network model for ATC operations

Based on the established risk factor representation framework, this study extracts causal relationships from 1349 ATC safety occurrence reports, identifying numerous representative incident chains. These chains are then integrated into a directed and weighted causal network, which represents the propagation and interaction of risks across different incidents. The construction methodology is illustrated in Fig. 5.

Fig. 5.

Toy example of directed and weighted causal network construction.

In Fig. 5, network nodes represent the risk factor indicators related to ATC operations. Directed edges indicate the direction of causality between events, reflecting the potential pathways of risk propagation. Edge weights denote the strength of connection between nodes.

The nodes in the network are classified into three types: initial nodes, contributing nodes, and outcome nodes. These respectively represent the original cause, intermediate contributing or aggravating factors, and final operational outcomes of the unsafe event.

During the construction of the causal network, multiple event chains may share overlapping nodes and duplicate edges. To simplify the network topology and better reflect the strength of causal relationships, this study merges duplicate edges by assigning edge weights based on conditional probabilities between nodes. A weighted directed graph is thus formed. \* MERGEFORMAT Table 1 presents the weighted adjacency matrix corresponding to \* MERGEFORMAT Fig. 5. Node I1 points to nodes C1 and C2 with conditional probabilities of 0.33 and 0.67, respectively, indicating the proportion of its connections to each event type across three event chains.

Table 1.

Weighted matrix.

Number	I1	C1	C2	C3	C4	A1
I1	0	0.33	0.67	0	0	0
C1	0	0	0	1	0	0
C2	0	0	0	0.5	0.5	0
C3	0	0	0	0	0	1
C4	0	0	0	0	0	1
A1	0	0	0	0	0	0

Therefore, a directed weighted network m can be constructed based on the weighted adjacency matrix CN, where N represents the number of nodes. The formulation is as follows:

12

where, represents the weight of the edge from node i to node j, and Edges denotes the set of edges.

• 2.Network topology analysis

Following the construction of the causal network, a topological analysis is conducted to examine both the structural properties of the network and the importance of individual nodes. This analysis aims to uncover the underlying causal mechanisms and identify critical risk factors contributing to ATC-related safety occurrences.

1. Evaluation of Network Structure. First, the average path length and clustering coefficient are calculated to assess the efficiency of information propagation and local aggregation among risk factors. Then, average node degree and cumulative degree distribution are analyzed to reveal the connectivity patterns and distributional characteristics of the network.

2. Evaluation of Node Importance. First, total degree is computed for each node. Nodes falling within the top 15% in terms of total degree are selected to examine their strong connectivity and structural influence. Total degree is chosen as it comprehensively reflects a node’s integration and interaction within the network. Second, Betweenness centrality is computed to identify nodes with high values, which typically serve as “bridges” in the transmission of risk across different causal chains. Finally, PageRank scores are computed to assess the global influence of nodes. Nodes with high PageRank values are examined for their aggregation capacity and overall prominence within the causal structure.

• 3.Vulnerability analysis

To evaluate the robustness and structural resilience of the causal network, this study conducts a series of simulated network attack experiments by progressively removing 1 to 10 nodes under different attack strategies. The change in network efficiency before and after node removal is used as an indicator of the network’s vulnerability. The following four attack strategies are employed:

1. Random attack. Nodes are removed from the network randomly.

2. Degree-based attack. Nodes are removed in descending order based on their degree, targeting the most highly connected nodes.

3. Betweenness-based attack. Nodes are removed in order of decreasing Betweenness centrality, prioritizing those that act as bridges in causal paths.

4. PageRank-based attack. Nodes are removed according to their PageRank values, targeting those with the highest overall influence.

Application and results

Risk theme extraction

Topic modeling results and analysis

To identify potential risk factors in ATC operations, this study applied the BERTopic model to 1328 preprocessed safety incident reports, resulting in the extraction of 14 distinct risk themes, as illustrated in Fig. 6.

Fig. 6.

Topic-based word cloud visualization of ATC-related safety occurrence reports.

As an unsupervised learning method, BERTopic requires manual intervention to assign meaningful semantic labels to the generated topics. In this study, each topic label was determined by integrating contextual analysis, domain-specific knowledge of ATC operations, and the interpretation of the top four keywords associated with each topic. To illustrate the labeling process, we use Topic 1 as an example. The top four keywords for this topic are “bird strike,” “airborne holding,” “windshield,” and “radome.” These keywords reflect two defining characteristics of bird strike incidents: (1) the primary impact areas are typically the windshield and radome, and (2) such events can lead to airborne holding patterns for nearby aircraft. As shown in Fig. 7, “bird strike” is the most frequently occurring keyword, further confirming the theme’s relevance to bird strike–related risks. Accordingly, this topic was labeled “Bird Strike.” Table 2 presents the semantic labels and associated keywords for all 14 topics.

Fig. 7.

Topic–word probability distribution.

Table 2.

Table of topic labels and keywords.

No	Topic label	Keywords
T1	Bird strike	Bird strike, airborne holding, windshield, radome
T2	Passenger illness-return to stand	Return to stand, passenger, cardiac, discomfort
T3	Weather conditions-return to stand	Return to stand, weather, destination weather, refueling
T4	Weather observation system failure	Automated meteorological observation system, weather, radar, electrical failure
T5	FOD–fuselage damage	Foreign object damage, fuselage damage, wing impact, component damage
T6	FOD–engine damage	Foreign object damage, engine damage, horizontal stabilizer damage, APU cowling damage
T7	Taxiway deviation	Taxiway deviation, incorrect taxiing entry, pilot operational error, wrong entry
T8	Equipment Malfunction–Command	Return to stand, engine system failure, control deviation, braking
T9	Equipment factors-altitude deviation	Altitude deviation, radar, CLAM alert, turbulence
T10	Unauthorized ground operation	Operational error, pilot violation, safety protocol, jet bridge
T11	Route deviation	Route deviation, incorrect flight execution, onboard equipment, weather
T12	Air-ground communication issue	ATC–pilot air-ground communication issue, continuous call, failure to promptly, equipment failure
T13	Weather conditions-altitude deviation due to weather	Altitude deviation, CLAM warning, MSAW alert, weather
T14	Tire puncture	Tire puncture, stone, screw from apron operations, foreign object

Figure 8 displays the proportional distribution of the identified themes. Among them, the “Bird Strike” topic accounts for the largest share (28.8%), followed by the theme related to “Passenger-Induced Return to Stand” events. These results suggest that such incidents occur with relatively high frequency in ATC operations and warrant close monitoring as key safety risks. Empirically, bird strike incidents are not only frequent but also pose serious consequences, representing a typical operational hazard in global ATC systems. According to data from the Federal Aviation Administration (FAA), approximately 280,000 bird strikes were reported in the United States between 1990 and 2023, with 18,400 incidents occurring in 2023 alone. Common points of impact include engines and windshields, and in some cases, bird strikes have led to engine failure, forced returns, or emergency landings³⁴. Return-to-stand events, by contrast, are typically triggered by a diverse range of factors and are managed through relatively standardized operational procedures. As a routine contingency measure in flight operations, return-to-stand events may be prompted by passenger-related emergencies, sudden weather changes, or technical malfunctions. Due to considerations of safety redundancy, airlines often adopt this measure as a conservative response, even when the initiating risk is low, which contributes to the high frequency of such events in operational records.

Fig. 8.

Topic distribution chart.

Overall, the distribution of risk themes provides insight into the types of recurring and high-frequency safety incidents in civil aviation operations. These insights can help safety managers prioritize risk categories, allocate monitoring resources more effectively, and develop targeted interventions and personnel training programs tailored to specific operational scenarios. However, while topic proportions offer a quantitative overview, they do not reveal the semantic structure or internal consistency of the themes. To address this limitation, we further examined the clustering of topics in semantic space, as depicted in Fig. 9. In this figure, each data point represents an individual incident report, with color indicating topic assignment and spatial proximity reflecting semantic similarity. The topics demonstrate clear boundaries and well-defined clusters, suggesting a high degree of separability and coherence in the semantic embedding space. This confirms the robustness of the topic modeling approach and lays a foundation for further semantic relationship analysis among topics.

Fig. 9.

Visualization of topic clusters.

To explore the semantic relationships between different topics, this study used UMAP to extract topic embedding vectors, then applied cosine similarity and hierarchical clustering to generate a topic association map, as shown in Fig. 10. The figure reveals clear inter-topic connections, indicating the existence of underlying causal linkages and risk propagation mechanisms across different safety incident types.

Fig. 10.

Topic correlation–hierarchical clustering map.

Some topics exhibit particularly strong semantic correlations due to shared risk origins. For instance, T5 (Foreign Object–Fuselage Impact) and T6 (Foreign Object–Engine Impact) have a correlation coefficient of 0.87, indicating a strong overlap in the semantic space. This high degree of similarity stems from their common causal factor—foreign object intrusion causing structural damage. Although the affected aircraft components differ between the two topics, their shared causality leads to localized semantic clustering and structural correlation. Moreover, high semantic proximity between certain topics can also arise from keyword-level interactions. For example, T12 (Air-Ground Communication Issues) and T7 (Taxiway Deviation) may appear unrelated at first glance. However, one of T7’s key terms, “crew operational error,” is often influenced by communication failures, which are a defining characteristic of Topic 12. Such disruptions can lead to flight conflicts, loss of communication, and other unsafe scenarios, thereby establishing a latent causal relationship between the two topics.

Reliability analysis

To evaluate the quality of topics generated by the BERTopic model, this study adopts two metrics: topic coherence (TC) and topic diversity (TD), and compares the results with those of the classical LDA model. The results show that BERTopic achieves TC and TD scores of 0.68 and 0.57, respectively, significantly outperforming LDA’s scores of 0.31 and 0.50. These findings suggest that BERTopic provides more coherent and diverse topics, indicating superior stability and accuracy in topic extraction.

Specifically, the risk topics identified in this study not only cover traditional factors widely discussed in the literature, such as weather conditions and human factors, particularly pilot operations, but also uncover previously underexplored themes³¹. These include risks arising from disruptive passenger behavior, bird strike–induced operational hazards, and systemic flaws in human–machine interaction. The inclusion of such themes underscores the model’s capability to uncover latent risk factors embedded within complex ATC safety events.

Causal analysis

Causal network of ATC operational safety

Risk factor representation framework

Building upon existing indicators, this study constructs a multi-level risk indicator system based on real cases extracted from ATC-related safety occurrence reports, as well as the identified risk topics and their associated keywords. A total of 90 micro-level indicators were constructed and mapped to five macro-level categories. These indicators encompass critical causal factors such as human errors, system malfunctions, adverse weather conditions, and managerial shortcomings. A partial illustration of the proposed multi-level indicator system is presented in Table 3, while full definitions and detailed explanations are provided in Appendix A.

Table 3.

Risk factor representation framework.

Macro indicators	Micro indicators
	No	Causative factors
Human factors	H1	Insufficient pilot experience or capability
	H2	Misjudgment of operational situation by the pilot
	H3	Pilot ignoring or misinterpreting instructions
	H4	Pilot stress, fatigue, or distraction
	H5	Issues in pilot instruction readback
	…
Equipment factors	F1	Malfunction of automatic flight control system
	F2	Malfunction of onboard radar
	F3	Malfunction of flight instruments
	F4	Malfunction of aircraft lighting system
	F5	Malfunction of landing gear system
	…
Environmental factors	E1	Low cloud and low visibility weather
	E2	Heavy precipitation
	E3	Foggy weather
	E4	Strong wind conditions
	E5	Thunderstorm conditions
	…
Managerial factors	M1	Issues in pilot training, supervision, and management
	M2	Pilot safety awareness management issues
	M3	ATC training, supervision, and management issues
	M4	Training and supervision issues for other personnel (e.g., ground handling, maintenance)
	M5	Scheduled inspection management issues for machinery, components, and systems
	…
Stochastic event factors	T1	Route deviation
	T2	Altitude deviation
	T3	Aircraft crossed designated holding position
	T4	Aircraft taxiway deviation
	T5	Aircraft return to stand
	…

Causal network of ATC operational safety

An example of a specific event chain extraction and the resulting overall causal network for ATC operational safety are presented in Table 4 and Fig. 11, respectively. In the network, nodes represent risk factors and edges denote causal relationships between events. Node size reflects the PageRank value, with larger nodes indicating greater influence within the network. Edge thickness corresponds to the conditional probability between events, with thicker edges indicating stronger causal ties. Node color differentiates risk categories—e.g., green for human factors and purple for equipment issues. Arrow direction represents the direction of causal propagation.

Table 4.

Illustrative table of event and event chain annotations.

Event

Event chain

(5) On May 3, a communication failure conflict incident occurred between ZH8801 and NX205 in the Guangzhou area Keywords: Radar warning delay ZH8801/B738 was flying from Chengdu to Guangzhou at a flight altitude of 9500 m with an estimated arrival time of 09:15; NX205/A321 was flying from Macau to Xiamen at an altitude of 8900 m with an estimated encounter time of 09:26 At 09:23:22, due to a communication system malfunction, ZH8801 did not receive the controller’s instruction to maintain its altitude and continued descending to 8300 m, forming an intersecting flight path with NX205. At 09:25:10, after the radar warning was triggered, the controller urgently instructed NX205 to climb to 10,100 m to evade According to the recorded radar video, when the minimum vertical separation between the two aircraft reached 120 m, the horizontal distance was 4.2 km

M5 → F29 → H23 → T2

Fig. 11.

Causal relationship network of ATC operational safety.

Network topological analysis

Structural characteristics of the network

1. Network properties

The average path length, a measure of separation and transmission efficiency within the network, is calculated to be 3.951. This suggests that risk events in ATC operations are tightly interconnected, with risk propagation requiring only about four steps on average. However, the average path length primarily reflects global connectivity and does not capture local clustering behavior. To address this, the clustering coefficient of each node was calculated, as illustrated in Fig. 12. Nodes such as E11, M11, and F35 exhibit high clustering coefficients, indicating strong local aggregation and pronounced neighborhood clustering.

Fig. 12.

Clustering coefficient distribution (clustering coefficient > 0.25).

The network exhibits both a high average path length and a high clustering coefficient, characteristics indicative of a small-world network. Small-world networks are typified by the coexistence of short global distances and tightly clustered local neighborhoods. In such networks, disruptions at a few critical nodes can rapidly propagate through the system, triggering cascading failures and accelerating systemic risk evolution. To verify the small-world properties of the constructed causal network, three random networks with identical node and edge counts were generated using Gephi. Comparative results for clustering coefficient and average path length are shown in Table 5.

Table 5.

Comparison of topological properties between the causal relationship network and random networks.

Network model	Average path length	Clustering coefficient
Causal relationship network of ATC operational safety	3.951	0.551
Random network 1	5.212	0.256
Random network 2	4.530	0.392
Random network 3	4.315	0.310

The causal network’s clustering coefficient is significantly higher and its average path length shorter than those of the random networks, confirming its small-world nature. This finding underscores the importance of early intervention on high-risk nodes occupying structurally central positions in order to improve system resilience and prevent local disruptions from escalating.

• 2.Network parameter analysis

The average degree of the network is 4.057, indicating that each risk factor is connected to approximately four other factors on average. However, average degree only reflects the overall connectivity of the network and does not reveal the degree of association for individual nodes. To further explore the heterogeneity of node connectivity, a cumulative degree distribution was plotted, as shown in Fig. 13. The distribution follows a power-law pattern, suggesting that the network exhibits scale-free characteristics. In such networks, most nodes have few connections, while a few “hub” nodes are highly connected. Disruption of these hub nodes can lead to disproportionate system-wide consequences. Therefore, accident prevention strategies should prioritize the identification and real-time monitoring of these critical nodes to enhance system robustness and adaptability.

Fig. 13.

Cumulative degree distribution.

Node importance analysis

1. Node degree

Figure 14 presents the distribution of total node degree among the top 15% of nodes. Human-related factors (category H) demonstrate the highest aggregate degree, indicating their centrality and high interconnectivity within the causal network. This prominence reflects ATC’s heavy reliance on human judgment and communication. Human performance is inherently variable, influenced by cognitive abilities, experience, and psychological states. Human factors not only act as direct contributors to unsafe events but also frequently interact with equipment failures, management flaws, and environmental changes. As a result, human-related nodes appear across multiple event chains, serving as key drivers of risk transmission and interaction.

Fig. 14.

Distribution of node degrees (total degree > 20).

In addition to total degree, it is particularly important to identify nodes with both high in-degree and high out-degree. These bidirectionally connected nodes can amplify and propagate risk, acting as high-risk interaction hubs. For example, H6 (pilot violation of operating rules) is a representative human factor, often shaped by limited situational awareness or insufficient safety training, and may trigger a sequence of adverse events such as flight conflicts and route deviations.

• 2.Betweenness centrality

The distribution of betweenness centrality is shown in Fig. 15 Node H23 has the highest Betweenness centrality. On one hand, air–ground communication is a critical link in ATC operations that heavily relies on human–machine interaction and is often affected by factors such as communication equipment failures and pilot distraction. On the other hand, it serves as the sole channel through which control instructions are transmitted to pilots. If this channel is compromised, the likelihood of critical events, such as flight deviations or in-air conflicts, increases dramatically.

Fig. 15.

Betweenness centrality distribution.

In addition, nodes such as H21 (controller misjudgment of operational situation), H14 (airfield personnel violations), and E12 (airspace closure or restriction) also exhibit high Betweenness centrality. This indicates that these seemingly localized issues can propagate downstream through the event network and trigger large-scale cascading effects. Although T2 (deviation from assigned altitude) often appears at the end of the causal chains, it demonstrates high connectivity and a central position in the network, linking multiple event paths. This “bridging” role suggests that even outcome-type risk nodes may serve as new origins of risk propagation and should therefore not be overlooked.

• 3.PageRank values

The distribution of PageRank values is shown in Fig. 16. As illustrated, node T5 (aircraft taxi-back) ranks highest because it frequently results from multiple converging risk factors, including pilot misjudgment, mechanical failure, and deteriorating weather conditions. This makes it a typical outcome of multi-causal influence. E12 (airspace closure) also exhibits a relatively high PageRank, reflecting its potential to be triggered across diverse operational scenarios and thus acting as a common-cause node. Similarly, T8 (go-around, diversion, or rerouting) has a high PageRank value, underscoring its frequent activation during unstable operational states and its pivotal role in system response mechanisms.

Fig. 16.

PageRank value distribution.

Vulnerability analysis

The results of multiple attack simulations on the causal network for ATC safety are presented in Fig. 17. Compared to random node removals, the three targeted attack strategies show markedly higher destructive effects, reducing network efficiency by over 68%. Under targeted attacks, network efficiency declines with the removal of additional nodes but the decline rate stabilizes, indicating that the network is particularly reliant on nodes with high topological importance. Among the metrics evaluated, attacks targeting nodes with high PageRank and betweenness centrality caused the most severe efficiency drops, demonstrating their utility in risk assessment. Notably, the most significant declines occurred during the removal of the first four nodes, emphasizing their structural importance. Based on these findings, the study identifies six key high-risk nodes: H23, H21, E12, T2, T5, and T8.

Fig. 17.

Network vulnerability analysis under different attack strategies.

Optimization strategies for mitigating ATC-related safety occurrences

Building on the preceding analysis of the topological structure of the causal network of ATC-related safety occurrences, this study identifies the functional roles of various risk factors within the civil aviation air traffic control system. Accordingly, targeted optimization strategies are proposed for critical nodes with high transmission potential or strong systemic influence.

1. Optimization of “single high-centrality nodes”

Priority should be given to mitigating “bridge-type” nodes with high betweenness centrality, such as H23 (ATCO–pilot air–ground communication issues) and H21 (ATCO misjudgment of the operational situation). These nodes frequently appear along multiple causal pathways and function as key intermediaries in systemic risk transmission. Any deviation occurring at these points is highly likely to trigger cascading effects, thereby intensifying the consequences of an incident. For node H23, communication mechanisms between controllers and pilots should be strengthened by optimizing speech recognition and message confirmation procedures, enhancing readback verification protocols, and establishing a bidirectional semantic validation system. For node H21, improvements should focus on enhancing situational awareness through intelligent alerting and decision support systems, thereby facilitating proactive risk detection and preventing cognitive bias from developing into operational errors.

Additionally, risk convergence nodes with high PageRank values require enhanced early warning and proactive intervention. Representative nodes include T5 (aircraft taxi-back), E12 (airspace closure), and T8 (go-around/diversion), which frequently serve as consequence nodes resulting from multiple upstream risk factors or system-wide common causes. To address these, root cause analysis should be employed to identify contributing upstream factors, and integrated cross-system prevention and control mechanisms should be established. For instance, in response to taxi-back events, a multi-source coordination system involving equipment status monitoring, pilot behavior recognition, and weather alert services should be implemented to facilitate early detection and timely intervention.

• 2.Optimization of “dual high-centrality nodes”

Operational control should be strengthened for “dual high-centrality nodes” that exhibit both high betweenness centrality and high PageRank values. Nodes such as T2 (deviation from assigned altitude) and T5 (aircraft taxi-back) not only represent high-risk outcomes but also serve as hubs across multiple risk propagation pathways, exhibiting strong risk accumulation and amplification effects. For these nodes, intelligent behavior recognition and risk trend prediction models should be deployed within operational monitoring systems to enable real-time detection and alerting of deviations. In addition, decision support systems should be utilized to facilitate rapid intervention and prevent further escalation along the risk chain.

• 3.Optimization of overall network structure

The causal network of ATC operations exhibits small-world and scale-free properties, reflecting both efficient transmission pathways and a centralized structural configuration. Under such a network topology, core nodes situated at the hubs of risk propagation are capable of triggering rapid, multi-path diffusion of risks, resulting in exponential and cascading effects. Accordingly, risk mitigation strategies should be focused rather than evenly distributed. Emphasis should be placed on nodes with high centrality and densely clustered substructures. Structural decoupling approaches should be adopted to reduce the structural coupling within the network and enhance the robustness and resilience of the ATC system. Such approaches include incorporating procedural redundancies, establishing information verification channels, and defining hierarchical accountability.

Conclusion

This study proposes and applies a risk identification method that integrates text mining and causal analysis to investigate ATC-related safety occurrences in civil aviation. The approach systematically reveals the distribution of risk topics, causal mechanisms, and key contributing factors. The main conclusions are as follows:

1. Topic Extraction: by applying the BERTopic model to 1,349 ATC-related safety occurrence reports, this study extracts 14 distinct and well-defined risk topics, including categories such as bird strikes and taxi-back to stand. BERTopic demonstrates greater stability and interpretability compared to traditional topic modeling approaches, particularly when processing texts characterized by dense technical terminology and frequent abbreviations. It effectively reveals latent risk themes of practical significance.

2. Risk Factor Identification: a comprehensive framework is constructed to represent safety risk factors in ATC operations by integrating insights from prior research and keywords from the extracted topics. The resulting causal network exhibits small-world and scale-free properties, reflecting a systemic structure in which a limited number of core nodes play a disproportionately critical role in risk propagation. Nodes with high PageRank and high betweenness centrality exert considerable influence over the network. Based on this, several key risk factors are identified. Strengthening the management of these critical factors enables proactive intervention and real-time monitoring of key risk points, thereby effectively preventing the escalation of unsafe events, reducing safety management workload, and significantly enhancing operational safety in civil aviation.

3. System-Level Insights: the analysis reveals that nodes with high betweenness centrality (e.g., H23 and H21) serve as bridges for systemic risk transmission, while nodes with high PageRank (e.g., aircraft taxi-back and airspace closure) function as convergence points for multiple risk pathways. Furthermore, dual high-centrality nodes (e.g., deviation from assigned altitude) simultaneously act as initiators and outcomes in multiple risk chains, thereby amplifying systemic vulnerabilities.

In summary, the integrated method of text mining and causal analysis proposed in this study enables a more complete and quantitative representation of ATC-related risks and facilitates the efficient identification of critical risk factors. This contributes valuable insights for enhancing existing risk assessment models. Additionally, the findings offer practical support for aviation safety managers in accurately detecting potential hazards, implementing timely countermeasures, and improving the overall safety level of ATC operations.

Despite these contributions, several limitations remain. First, the dataset is relatively limited in size and diversity, which constrains the generalizability of findings to various airport contexts, operational models, or adverse weather scenarios. Second, the extraction of causal relationships still relies on manual annotation, introducing subjectivity and limiting scalability for larger datasets. Lastly, the identified insights have not yet been integrated into real-time early warning or decision-making systems. Future work should explore technical solutions for embedding the analytical models into operational workflows, thereby enhancing their applicability and translational value in real-world civil aviation safety management.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yuhan Wang: Writing—original draft, Visualization, Investigation, Funding acquisition, Datacuration, Conceptualization. Zhongye Wang: Writing—review & editing, Supervision, Resources, Funding acquisition. Zongbei Shi: Writing—review & editing, Supervision, Conceptualization. Honghai Zhang: Writing—review & editing, Resources, Funding acquisition. Xiaocheng Liu: Writing—review & editing.

Funding

National Natural Science Foundation of China, Chinese Special Research Project for Civil Aircraft.

Data availability

To protect the privacy of the research participants, our data are confidential. The data that support the findings of this study are available from the corresponding author, Yuhan Wang, upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.