Risk Topics Discovery and Trend Analysis in Air Traffic Control Operations—Air Traffic Control Incident Reports from 2000 to 2022

Abstract

The safety of air traffic control (ATC) operations is an important cornerstone for the sustainable development of the civil aviation industry. In order to clarify the risk factors in the control operation process and to achieve digital representation of the safety risks of civil aviation control operations, starting from the ATC incident reports, we fully mine the safety risk information and unspoken rules of ATC operations. A risk perception model for air traffic control operations safety based on the Latent Dirichlet Allocation (LDA) topic model and the Semantic Network Based on BERT (BSN) model is suggested. First, 17 risk topics and keywords were found in the incident reports collected using the LDA topic model. These topics included those pertaining to the stage of aircraft operation, human factors in control operation, and the sector or airspace operation status and structure. The findings indicate that while most risk subjects have not changed significantly, they do show an upward tendency. Human factors and operational rules and procedures account for the highest share of all key causes, and they also have a significant impact on how risk topics evolve over time. Finally, the BSN model in the air traffic control field was built based on the keywords of each risk issue in order to highlight any potential correlations between distinct risk topics. The results show that some risk topics have interrelated risk characteristics, and there are regularities of mutual evolution between these risk topics. The relevant research results can better mine air traffic control unsafe information and lay a foundation for accurately perceiving air traffic control operations risks.

1. Introduction

Safety is a paramount concern in the civil aviation industry, and the air traffic control division plays a vital role in ensuring it. With the rapid expansion of aviation, airports and airspace are becoming increasingly congested, putting pressure on air traffic control. This pressure can sometimes lead to unsafe incidents such as the Tenerife accident in March 1997 and the dangerous approaching incident of two aircraft at Shanghai Hongqiao International Airport in October 2016, both caused by ATC operations.

Many recent studies have concentrated on the examination of elements impacting the safety of ATC operations and the investigation of risk evolution rules in an effort to prevent different mishaps. Among them, the analysis of influencing factors includes traditional analysis methods [1,2] and modern text analysis techniques [3,4]. Traditional analysis methods mainly use nonlinear modeling, event vulnerability analysis, Bayesian theory, and fault tree methods, which are difficult to simultaneously mine the potential relationships and risk evolution rules between related events. Modern text analysis techniques primarily use Apriori, FP-growth, and Onto-BN methods to extract control operation risk factors and identify multi-level correlation relationships between different unsafe event attributes. The risk evolution rules mainly analyze runway incursions, heavy landings, communication failures, and other unsafe events and track investigation reports of aviation accidents [5,6]. As a result, the technique of using accident reports for analysis has steadily grown into a crucial instrument for assessing safety performance in various fields. The usefulness of data and text mining methods in aviation safety data analysis was demonstrated by Nazeri et al. [7] by using the association rule mining approach to analyze ASRS reports. In order to obtain risk categories, Figueres-Esteban et al. [8] analyzed railway incident records using network analysis techniques. Robinson et al. [9] identified human factors topics in aviation safety reports through the use of text modeling, while this study provided practical validation of the use of natural language processing in sensemaking and pinpointing trends necessary for prioritizing safety activities and precipitated a change to the way subject matter experts interpreted the narratives and their implications to industry safety.

The progress of text data mining and semantic association analysis has been considerably aided by developments in machine learning and natural language processing up to this point. Topic models have grown in importance as tools in the field of natural language processing. Among these, the LDA topic model [10] is a probabilistic model and an unsupervised learning technique that has strong performance in text data mining and can extract the subject information of documents from a vast amount of text corpus data. In order to better utilize the text information in accident reports to forecast accident duration, Pereira et al. [11] employed a topic model to examine two years’ worth of accident cases. They then presented a machine learning-based duration prediction model that can combine text features with non-text variables. Topic modeling of data from the aviation safety reporting system by Kuhn et al. [12] revealed hidden risk factors in addition to well-known risk issues. The LDA topic model was used by Sun, L. et al. [13] to evaluate 17163 articles from 22 significant transportation journals, resulting in the inference of a total of 50 major themes and the quantification of the similarity between the journals and the countries and regions in the aggregated topic distribution based on the LDA results. Liu, Y. et al. [14] combined text mining techniques and Latent Dirichlet Allocation (LDA) models to analyze standardized accident investigation reports in the Chinese construction industry. This approach helps site managers more quickly and effectively understand the causal factors and key information leading to accidents from incident reports.

The LDA model is a text corpus mining algorithm that is particularly effective at extracting topical information from lengthy texts. However, because the LDA model is based on the bag-of-words model, it only takes word frequency into account and does not take word order or semantic relationships into account. Therefore, when exploring risk topics in the field of ATC, it is difficult to truly reflect the degree of interdependence between topic keywords. However, the BERT language model [15] can combine contextual semantic information and vectorize words to reveal the correlation between keywords. At the same time, combined with semantic networks [16,17], the association relationship between keywords can be explicitly represented, thereby more accurately representing ATC operations safety risk knowledge. Second, although the mining method based on the topic model can discover general topics from ATC incident reports, it will ignore the trend and regularity of ATC risk topics changing over time.

The primary contributions of this paper can be summed up as follows in light of the subject model’s limitations in exploring the safety risk themes of ATC operations.

• This paper created a word segmentation corpus specifically for air traffic control operations safety risk management based on a large number of ATC incident reports, effectively resolving the issue of word segmentation ambiguity and special terms. In parallel, it studied and developed a topic model for air traffic control operations safety risk, showing several possible risk topics and raising awareness of ATC operations safety hazards.
• This study revealed the historical trends and causes of the varying air traffic control risk topics, as well as attaining more precise positioning of the major risk factors within those risk topics.
• The constructed BSN model for air traffic control improved professional understanding of reporting incidents, quantified the evolution of different risk topics’ commonalities, and created a sophisticated unstructured safety management process data analysis mechanism based on natural language processing.
• The risk topics identified in this article provided important reference for the ATC operations risk situation awareness and the comprehensive assessment of ATC security situations

The rest of the paper is organized as follows. Section 2 describes the data source and text preprocessing work. Section 3 discusses the methodology of Latent Dirichlet Allocation and designs the Semantic Network Based on BERT (BSN) model. Section 4 is the results of data analysis. Section 5 concludes the study and provides recommendation for the future study.

2. Data Source

The Aviation Safety Reporting System (ASRS), a public database jointly maintained by the National Aeronautics and Space Administration (NASA) and the Federal Aviation Administration (FAA), is the source of the data utilized in this study. It is a crucial method for gathering data on aviation safety. This database contains a total of 20,174 air-traffic-control-related incident reports from 2000 to 2022. The number, date/time, location, operating environment, aircraft information, personnel information, cause analysis, and incident process are all included in each incident report. Figure 1a shows the sample data of ATC incident reports, and Figure 1b shows the number of ATC incident reports per year.

The number of reports was significantly high from 2010 to 2019, as seen in Figure 1. Due to COVID-19’s effect on civil aviation flights, the number of incident reports decreased between 2020 and 2022. The general trend roughly mirrors the annual variation in air traffic volume in the control region.

This paper used the NLTK (Natural Language Toolkit version 3.6.5) word segmentation tool in Python version 3.9 to segment ATC operations incident reports. In this study, the NLTK python package, one of the best English text segmentation open-source tools, was used to efficiently conduct the above basic NLP tasks. During the word segmentation process, the words were tagged and filtered by part of speech, and then the filtered words were processed. After the above preprocessing, the final corpus of data that could be input into the LDA topic model was generated.

3. Methodology

3.1. Latent Dirichlet Allocation (LDA) Topic Model

The ATC incident reports needed to be arranged into a text set that could be entered before risk subjects were discovered. First, by constructing an English stop words dictionary, words unrelated to ATC operations, such as prepositions, pronouns, and special symbols, were stopped. Second, we continued to add words to the stop words dictionary that frequently appeared in incident reports but could not be used to specifically describe risky ATC operations incidents, like time, runway number, taxiway number, speed unit, altitude unit, and other words. Then, in order to avoid splitting professional vocabulary during word segmentation and further increase the accuracy of semantics, a professional dictionary was constructed based on the professional vocabulary in the fields of ATC and safety and with reference to pertinent regulations, the literature, and existing ATC incident reports.

To extract the topical data from a huge corpus of text and infer the topic distribution of the text, Blei et al. [10] suggested the LDA topic model. A three-layer Bayesian mixture model consisting of documents, topics, and keywords is the foundation of the LDA topic model. In text topic mining [18], sentiment analysis [19], topic trend prediction [20], and other areas, it is often employed. Figure 2 displays its probabilistic graphical representation.

In the LDA topic model in this paper, K is the number of ATC operations risk topics, M is the number of ATC incident reports, α is the prior distribution of θ_m, θ_m is the topic distribution of the m-th incident report, Z_m,n represents the topic of the n-th word in the m-th incident report generated from θ_m, w_m,n represents the final n-th word in the m-th incident report, β is the prior distribution of φ_k, φ_k is the word distribution, and N_m is the total number of words in the m-th incident report. The procedure for creating risk themes for the LDA model using event reports is shown in Figure 3.

Therefore, the risk topic of ATC incident reports in the LDA topic model could be decomposed into two processes. The first process was the process of generating documents, that is, α→θ_m→Z_m,n, which means that when generating the m-th document, the corresponding topic distribution θ_m was extracted from the document–topic distribution α, and then the topic number Z_m,n of the n-th word in the document was generated according to the topic distribution of θ_m. The second process was the process of generating words, that is, β→φ_k→w_m,n | k = Z_m,n, which means selecting the word distribution with the number k = Z_m,n from the K topic–word distributions φ_k and then generating the word w_m,n.

3.2. Semantic Network Based on BERT

Despite showing the relationship between incident reports, risk topics, and keywords, the LDA topic model was unable to show how various keywords were connected to one another. According to a related study [18], there may still be implicit associations between two terms even if they are related to unrelated topics. Therefore, the connect between risk topics may be more clearly described by building a semantic network and connecting keywords to form a network. A semantic network was an idea put forth by Quillian [21]. In a semantic network, knowledge is represented as nodes and edges, where nodes stand in for concepts or objects and edges for the connections among nodes.

It is initially important to digitally represent words before creating a semantic network. Traditional word vector models, such as the Word2vec model proposed by Mikolov, T. et al. [22], include two algorithms: CBOW and Skip-gram. The CBOW (Continuous Bag of Words) is a neural network model used to generate word vectors. The CBOW algorithm generates the word vector of the current keyword through context information, while Skip-gram generates the context word vector through the current keyword. This model can connect context information, but it is unable to address the polysemy issue. The EMLo model [23] focuses on finding a solution to the polysemy issue; however, when compared with the Transformer encoder, its encoder performs rather poorly in feature extraction. As a result, the BERT model is a pre-trained unsupervised natural language processing model [15] that can better comprehend contextual semantics by applying a Transformer feature extractor than conventional word vector models. Figure 4 depicts its structure.

The BERT pre-training model had to be modified since it lacked expertise in specialized fields because it was trained on so much general text. The papers by Sun, C. [24], Hu, X. [25], and Gururangan, S. [26] are cited in this research to modify the BERT pre-training model and improve the BERT model’s comprehension of the air traffic control field. The stages for building a BERT model for the field of air traffic control were as follows:

• Preprocess and clean the incident reports, label each report with the corresponding risk topic category, and divide the dataset.
• Add the professional dictionary constructed in this paper to the BERT vocabulary, and load the pre-trained tokenizer and serialized classifier of the BERT model, where the pre-trained language model used is bert-base-uncased. At the same time, add special tokens (such as [SEP], [CLS], [PAD], [UNK], etc.) to the BERT model.
• Use the training set text corpus of data as input and convert it into an input sequence that meets the BERT model, and train the BERT model at the same time.
• Evaluate the fine-tuned model using the validation set and calculate the performance of the model on various indicators.
• Use the adjusted BERT model to classify risk topics for test set data.

Therefore, the text input is transformed into a one-dimensional word vector using the BERT model for the air traffic control field created by the aforementioned processes. This word vector may completely incorporate contextual semantic information. The correlation between words is then reflected by computing the cosine similarity between word vectors. The stronger the associations between words are, the deeper the information is represented in terms of its semantics. The specific calculation process is as follows: First, the n-dimensional word vectors of keywords A and B calculated by the BERT model in the air traffic control field are A_i = (x₁, x₂, x₃, x₄, x₅, …, x_n), B_i = (x₁, x₂, x₃, x₄, x₅, …, x_n), and then use Equation (1) to calculate the cosine similarity between word vectors.

$$\cos (A,B)=\frac{A\cdot B}{‖A‖‖B‖}=\frac{{\displaystyle \sum _{i=1}^n{A}_i\times B_i}}{\sqrt{{\displaystyle \sum _{i=1}^n{(A_i)}^2}}\times \sqrt{{\displaystyle \sum _{i=1}^n{(B_i)}^2}}}$$

(1)

where cos(A, B) represents the cosine similarity between keyword A and keyword B, A_i represents the i-th component of the word vector of keyword A, and B_i represents the i-th component of the word vector of keyword B.

According to Equation (1), the similarity between keywords can be calculated, and then the similarity matrix S of m keywords can be obtained as shown in Equation (2).

$$S=\left(\begin{array}{ccc}{s}_{11}& \dots & s_{1n}\\ ⋮& \ddots & ⋮\\ s_{m1}& \cdots & s_{mn}\end{array}\right)$$

(2)

where S is the similarity matrix of keywords, and S_ij is the similarity between the i-th keyword and the j-th keyword.

Create a semantic network of the safety risk associated with ATC operations using the similarity matrix. The nodes in this paper’s semantic network were keywords related to various risk topics, while the edges indicated correlations between different keywords. In the semantic network, there was a positive correlation between each node’s size and its average degree value. The larger the node, the more words connected to it. The width of the edge in the semantic network was positively correlated with the similarity between keywords. The wider the edge, the closer the relationship between the two keywords.

4. Results of Data Analysis

4.1. LDA Topic Discovery

Before training the LDA topic model, it was necessary to pre-specify the hyperparameters α, β and the number of topics K. The parameter α affected the sparsity of topics in incident reports. The higher the value of α, the smaller the impact on topic sparsity and the more uniform the topic distribution. That is, the generated documents mixed a large number of topics. The lower the value of α, the greater the impact on topic sparsity and the sparser the topic distribution. That is, the generated documents covered a small number of topics. The parameter β affected the sparsity of keywords under each topic. The higher the value of β, the smaller the impact on keyword sparsity. That is, most of the keywords in the corpus were contained under each generated topic, so these keywords indicated more general topics. The lower the value of β, the greater the impact on keyword sparsity. That is, the keyword distribution under each topic was more uneven and indicated more specific topics. In the process of training LDA topic models, as more and more topics were discovered, it was expected that each generated document would contain fewer and more specific topics. In most previous studies [27,28], α is usually set to α = 50/K, and β is set to β = 0.01. Similarly, in this study, we adopted the above parameter values and found that this model generated better results than other parameter values.

In addition, for the determination of the number of topics K, this study used the perplexity index to measure the optimal number of topics. The smaller the perplexity, the better the effect of the topic model. Its calculation (Equation (3)) is as follows:

$$Perplexity=\exp \left\{-\frac{{\displaystyle \sum _{m=1}^M\log p(w_m)}}{{\displaystyle \sum _{m=1}^M{N}_m}}\right\}$$

(3)

where Perplexity represents perplexity, P(w_m) is the probability of word w_m in document m, and N_m represents the number of words in document m.

At the same time, to investigate the stability and consistency of the topic model results, 20,174 ATC incident reports were first shuffled, and then the perplexity curves of 2500, 5000, 7500, 10,000, 12,500, 15,000, 17,500, and 20,174 reports were calculated, as shown in Figure 5. In these eight control experiments, it could be seen that as the number of ATC incident reports increased, the minimum value of the perplexity curve stabilized at K = 17. As a result, information on a total of 17 topics was gleaned from 20,174 incident reports.

The topic number of the incident reports was obtained according to the topic model, and the number of incident reports under each topic was counted as shown in Figure 6.

The importance of keywords was sorted, and the top 25 keywords were displayed in order to properly display the keyword information for each topic. Figure 7 illustrates the word cloud diagram in drawing form.

The font size and darkness of the words in Figure 7 indicated how significant they were to the topic. The topics determined in the figure revealed the potential risks that existed during ATC operations. Combining the information shown in Figure 6 and Figure 7, for example, topic 16 contained keywords such as radar, ILS (instrument landing system), VOR (very high frequency omnidirectional range), TCAS (traffic collision avoidance system), GPS (global positioning system), localizer, etc. This topic highlighted that the goodness of air traffic control facilities and equipment during ATC operations would directly affect the safety of ATC operations. Topic 04 involved the takeoff and departure stage of the flight. This topic contained keywords such as takeoff, departure, climb, checklist, runway, clearance, cross, vehicle, etc. Topic 07 contained keywords such as approach, ILS, RNAV, runway, RNP, controller, visual, land, etc., which introduced the operation of aircraft during the approach and landing stage. At the same time, topic 12 contained keywords such as arrival, descend, restriction, altitude, speed, pilot, etc. These keywords collectively indicated the operation of aircraft during the descent stage. In Figure 6, the sum of the number of incident reports for topic 04, topic 07, and topic 12 accounted for 23% of all incident reports. As a result, the stages of takeoff, descent, and approach landing were those where air traffic control safety mishaps happened most frequently. According to related studies, civil aircraft must frequently operate manually during the takeoff, descent, and approach landing stages. Untimely ATC instructions and maneuvers can cause incidents to occur [29,30,31].

Topic 03 and topic 17 accounted for a large proportion. Topic 03 involved sectors and airspace. This topic contained keywords such as sector, airspace, traffic, hold, flow, arrival, load, busy, etc. Therefore, this topic reflected the operating conditions of sectors and airspace, factors such as traffic or load. Topic 17 involves the airport ground traffic. This topic contains keywords such as taxi, ground, taxiway, hold, control, runway, vehicle, cross, etc. Therefore, this topic reflects the relevant elements of airport operations.

In addition, some risk topics involve human factors or external environmental factors. For example, topic 15 indicated that during ATC operations, there may have been situations such as unclear semantic expression, slips of the tongue, ambiguity, or noise, which could cause incidents to occur. Therefore, in such risk topics, relevant personnel need to focus on improving their professional skills and concentration to avoid similar events. At the same time, there were also some risk topics that involved the operating conditions of airspace when incidents occurred, airspace structure, or operating rules/procedures [32], etc. From these topics, it was found that unreasonable airspace structure and cumbersome operating rules/procedures would affect the safety of ATC operations.

In summary, the risk topics extracted from 20174 unsafe incident reports revealed potential hazards in the ATC operations process. These risk topics directly or indirectly affected the safety of ATC operations.

4.2. Trend and Reasons for Changes in Air Traffic Control Risk Topics over Time

In the original dataset, each unsafe incident report had corresponding time information. Therefore, in this section, combined with the results of the LDA topic model in Section 4.1, we further explore the trend and evolution of risk topics over time.

Data from 2009 to 2022 were chosen to investigate the trend of risk topics in order to demonstrate the changes in risk topics in recent years. The results are displayed in Figure 8. At the same time, further statistics were analyzed, and we calculated the proportion of various risk topics in each year, to clarify the trend of topics and the relationship between risk topics, as shown in Figure 9. Therefore, combining Figure 8 and Figure 9, it could be seen that some risk topics, such as topic 01, topic 07, and topic 13, which stood for arrival and descent, radar separation, and communication, exhibited a downward tendency over time. Some risk topics showed an upward trend followed by a downward trend over time, such as topic 04, topic 12, topic 16, and topic 17, which represented weather, approach flight, expression, and departure and takeoff, respectively. Most of the remaining risk topics showed significant fluctuations over time in this study. At the same time, due to the impact of the global COVID-19 epidemic and the large-scale reduction in the number of civil aviation flights, the number of all types of risk topics has declined in the past three years.

Additionally, the data were divided into data from 2010 to 2014 and data from 2015 to 2019, and the ring growth rate D_i of each risk issue was computed independently for these two time periods in order to better refine the trend of risk topics. The calculation (Equation (4)) is as follows:

$$D_i=\frac{{\displaystyle \sum _{y=2015}^{2019}{\delta }_{yi}-}{\displaystyle \sum _{y=2010}^{2014}{\delta }_{yi}}}{{\displaystyle \sum _{y=2010}^{2014}{\delta }_{yi}}}$$

(4)

where D_i is the ring growth rate of risk topic i, and δ_yi is the number of risk topic i in year y.

Through the above calculation formula, the ring growth rate D_i of each risk topic is obtained.

Table 1. Ring growth rate of each risk topic.

Topic	Di	Topic	Di
Topic 01	−0.004	Topic 10	0.257
Topic 02	0.963	Topic 11	−0.241
Topic 03	−0.042	Topic 12	0.472
Topic 04	−0.228	Topic 13	−0.131
Topic 05	−0.944	Topic 14	0.268
Topic 06	−0.269	Topic 15	−0.032
Topic 07	0.469	Topic 16	0.336
Topic 08	0.436	Topic 17	0.261
Topic 09	0.681

presents the computation outcomes. Topics 05, 06, and 11 in the table, which dealt with nighttime flights, communication, and radar separation, had the biggest year-over-year declines; topic 02, topic 09, and topic 07 had the largest year-on-year growth rate. Among them, topic 02 mainly involved risk factors related to flight altitude, topic 09 mainly involved risk factors related to operating rules/procedures, and topic 07 mainly involved risk factors related to the approach phase; most of the remaining risk topics had small changes but showed an upward trend.

Figure 10 specifically shows the trends of topic 05, topic 06, topic 11, topic 02, topic 09, and topic 07 over time.

Based on the above results, statistical analysis was performed on the main reasons for topic 05, topic 06, topic 11, topic 02, topic 09, and topic 07. In Figure 11, the main reasons for incidents in topic 05, topic 06, and topic 11 from 2015 to 2019 decreased significantly compared with the previous five years, while topic 02, topic 07, and topic 09 increased significantly. At the same time, it can be seen that among the main reasons for all risk topics, human factors and operating rules/procedures accounted for a large proportion.

According to the results of Table 1 and Figure 11 and combined with specific incident reports, the trend of risk topics over time was largely affected by human factors (such as controller skills or workload and fatigue) and operating rules/procedures (such as unreasonable control transfer procedures or flight procedures design). Therefore, more attention needs to be paid to the impacts of these two aspects in the process of ATC operations. At the same time, with the improvement of the performance and accuracy of communication, navigation, and surveillance systems, the number of ATC incident has been reduced to a certain extent.

4.3. Semantic Network for Air Traffic Control Operations Safety Risk

Based on the results of risk topic identification, the ATC incident reports were input into the BERT language model. Through extensive training, Adam (TensorFlow 1.13.1) was used as the optimizer in this BERT model, Categorical Cross Entropy was used as the loss function, and sparse classification accuracy was used to calculate model performance. In the end, the performance loss of the BERT model in the air traffic control field was reduced by 5% compared with the BERT pre-training model. Therefore, the keyword vectors calculated using the adjusted BERT model would be more in line with the air traffic control field.

Then, the BERT model in the air traffic control field was used to digitally represent the keywords and calculate the word similarity between the keywords through Equation (1). Next, Gephi 0.9.2 software was used to construct a semantic network to visually display the relationship between words. In the semantic network diagram of this article, there were a total of 112 nodes and 2411 edges. At the same time, in order to reveal the potential relationship between various risk topics, this paper used the Louvain (Gephi 0.9.2) clustering algorithm [33] to perform cluster analysis on the air traffic control operations safety risk semantic network. The Louvain clustering algorithm is based on multi-level optimization of modularity for community detection. This algorithm can achieve large-scale network community division with different granularities in a short time without specifying the number of communities. Modularity is a method to measure the quality of a community network division. Its definition is shown in Equation (5).

$$Q=\frac{1}{2m}{\displaystyle \sum _{vw}\left[A_{vw}-\frac{k_v{k}_w}{2m}\right]}\cdot \delta (c_v,c_w)$$

(5)

In the equation, m is the number of connections in the semantic network, and v and w are any two nodes in the semantic network. When there is a connection between them, A_vw = 1; otherwise, it is 0. k_w is the degree of node w, and δ(c_v, c_w) is used to judge whether nodes v and w are in the same community. If they are in the same community, δ(c_v, c_w) = 1; otherwise, it is 0.

Therefore, the Louvain clustering algorithm was used to divide the keywords in the semantic network into different modules and use the same color to represent the keywords in the same module. Finally, it was calculated that there were a total of four modules as shown in Figure 12. The connections between keywords within the same module were relatively dense, while the connections between keywords in different modules were relatively sparse.

The semantic network for ATC operations safety risk not only represented the interconnection between keywords but also revealed the potential relationship between risk topics. In Figure 12, the nodes in the purple area mainly involved the keywords of topic 02, topic 04, and topic 05, and the nodes in the green area mainly involved the keywords of topic 08 and topic 11. At the same time, it could be seen that the risk topics in the same color area had a closer relationship. Based on the risk topic keywords, a statistical analysis was performed on the risk topics contained in each module, as shown in

Table 2. Main risk topics contained in each module.

Module	Main Risk Topics	Risk Topic Keywords
Module 01	Topic 08, topic 09, topic 16	VFR, FR, ILS, TCAS, RNAV, VOR⋯
Module 02	Topic 01, topic 02, topic 04, topic 05, topic 07, topic 10, topic 13	Altitude, airport, climb, traffic, approach, airspace⋯
Module 03	Topic 02, topic 04, topic 06, topic 12, topic 14, topic 15	Frequency, contact, takeoff, MVA, restriction, noise⋯
Module 04	Topic 03, topic 08, topic 11	Controller, conflict, pilot, flow, busy, airspace⋯

.

Combined with Figure 12 and Table 2, the risk topics in the same module had interrelated risk characteristics. For example, module 04 contained topic 03 and topic 08. Among them, topic 03 involved sectors and airspace, reflecting factors such as the operation status, traffic, or load of sectors and airspace. Topic 08 mainly reflected human factors in the control operation process. Therefore, when the flight traffic in the sector or airspace increased, it would inevitably lead to an increase in the control operation load, thereby increasing the probability of human factors.

In summary, the air traffic control operations safety risk semantic network can quantify the degree of association between keywords and reveal the relevance between various risk topics. At the same time, it also has great potential in processing unstructured text data.

5. Conclusions

This paper extracted the risk categories of 20,174 air traffic control operations incident reports through the LDA topic model and found a total of 17 risk topics. These risk topics revealed risk factors during air traffic control operations over the past two decades, enabling data-driven risk management. The results showed that ATC operations risk mainly involved aircraft operation stages, airspace structure, airspace load, and operating rules/procedures, and some risks involved more human factors and external facilities, equipment, or environmental factors. At the same time, combined with specific incident reports to explore the causes of various types of risk incidents, it was shown that in the ATC operations process, unclear semantic expressions by controllers, slips of the tongue, or the issuance of unreasonable ATC instructions caused unsafe events; complex operation environments, unreasonable airspace structure design, and adjustments to some civil aviation transportation policies and ATC rules all affected the safety of ATC operations.

Exploring the trend of each topic over time in recent years, it could be found that some risk topics had a large growth rate, some risk topics had a large decline rate, and most risk topics had a small change but showed an upward trend. Further analysis of the main reasons for incidents showed that human factors and operation rules/procedures accounted for the largest proportion among all main reasons. In addition, research showed that if human factors and operating rules/procedures could be avoided from affecting ATC operations, it would greatly reduce the occurrence of unsafe events. At the same time, a large number of flight accident reports and studies showed that about 70% of aviation accidents were related to people, and most of the human factors were not caused by technical defects but because the crew members had problems in communication, cooperation, decision-making, and other aspects, that is, the crew resource management (CRM) ability. Therefore, the construction quality of “crew resource management (CRM)” must be further improved.

Finally, through the air traffic control operations safety risk semantic network, we further explored the mutual relationship between keywords within the same risk topic and revealed the potential relationship between keywords between different risk topics. The results showed that some risk topics had interrelated risk characteristics. Therefore, there would be mutual evolution rules between these risk topics. Therefore, when preventing and controlling ATC operations risks, attention should also be paid to subsequent risk changes in order to achieve early control of risks.

In summary, this paper’s exploration of air traffic control operations safety risk topics can provide a reference for the precise prevention and control of ATC operations risks. At the same time, relevant research results can also assist air traffic management units in further exploring the causes of incident and optimizing control processes, airspace structure, and personnel management.