Fatigue Detection of Air Traffic Controllers Through Their Eye Movements

Abstract

Eye movement patterns have become an essential element in modern approaches for identifying air traffic controller fatigue. By observing eye movements among various individuals and environments, researchers have discovered correlations with multiple physiological metrics and cognitive processing abilities. This study involved human-in-the-loop simulations to collect eye movement and fatigue data from air traffic controllers and students. The eye movements were classified into three main types: fixation, saccade, and blink. Statistical analyses were performed to determine the most important indicators. Using support vector machine and random forest models for training and prediction, it was found that the fixation characteristic is significantly important for monitoring air traffic controller fatigue. The implementation of this model has the potential to identify forthcoming instances of controller fatigue during their shifts, thereby helping to avert the possibility of unsafe situations.

1. Introduction

In the modern air transport sector, various factors must be taken into account to ensure efficient and safe operations, with air traffic control (ATC) being of crucial importance. During the operation of a flight, an aircraft faces numerous constraints or challenges, such as weather conditions, terrain, and its own limitations. Air traffic controllers (ATCOs), or controllers, are responsible for maintaining a safe and orderly air traffic flow, ensuring that each aircraft reaches its destination [1]. Until now, researchers have conducted numerous studies and improvements in the ATC process, including using standard air-ground communication, optimal design of airspace, and improving intelligent decision support systems [2]. Despite advances in intelligence and automation, humans remain central to the whole air traffic control process. Consequently, investigating ATCOs and human factors is essential. To date, research on ATCOs encompasses various aspects, such as work experience and voice recognition [3]. In this paper, our aim is to analyze the fatigue states of controllers within their working conditions to develop a fatigue state detection model.

In the context of fatigue detection for air traffic controllers (ATCOs) or other high-stress aviation roles, we could divide fatigue into “physical” and “mental” fatigue. Physical fatigue refers to muscle soreness or discomfort due to prolonged sitting, repetitive movements, or poor ergonomics. It can also include eye strain or discomfort due to prolonged screen usage. Mental fatigue means cognitive overload from continuous monitoring, decision-making, and multitasking. It is the depletion of attentional resources over time, leading to reduced alertness and slower reaction times [4]. For operators, fatigue causes easily distracted minds and difficulty focusing, which reduces work efficiency and can lead to accidents [5]. Thus, there is a strong link between fatigue and safety.

Methods for detecting fatigue can be categorized into subjective and objective techniques [6]. Subjective detection methods commonly involve questionnaires that assign a specific score to indicate fatigue level. Nevertheless, the precision of these questionnaire outcomes can be influenced by how aware individuals are of their own fatigue, which is a very subjective factor. On the other hand, using objective detection techniques can successfully circumvent the subjective biases [7], present in questionnaires, thus improving the trustworthiness of the information. Objective techniques can be divided into physiological measurements [8], which often require controllers wearing various testing devices, potentially interfering with their routine tasks, and behavioral analysis, which evaluates fatigue levels through observations of the controller’s eye movements, actions, and voice [9]. Specifically, eye movement behavior analysis is low-impact and highly precise, making it a widely used approach in monitoring controller fatigue [10]. During work, controllers must continuously observe variations in flight status on radar screens; thus, tracking eye movements can effectively indicate an individual’s fatigue level to some degree.

Taking into account the characteristics of air traffic control tasks, this study will use eye movement behavior to develop a model to identify controller fatigue. Throughout the experiment, eye movement data sampling devices were used to record the fixation and saccade attributes of the subject’s eyes in real-time. Subsequently, we analyzed the data to determine the correlation between variations in eye movement behavior and fatigue levels.

To summarize, this paper focuses on exploratory research. Previous studies have mainly used eye movement behavior as a supplementary tool, with few studies establishing a direct link to controller fatigue. Therefore, this study aims to identify the most effective indicators of eye movement behaviors and choose the appropriate algorithm for model development. This study achieved high accuracy in detecting fatigue through eye movement indicators, filling a gap in previous research. In addition, in the final data processing phase, algorithm-based filtering is used to discard insignificant indicators, optimizing the model. Through this research, we reveal the relationship between eye behaviors and air traffic controller fatigue, particularly with regard to sleepiness.

2. Related Work

Research on eye movement dates back to the mid-20th century [11]. The behavior of eye movement is typically examined in relation to two types of images: static and moving. Studies on static images usually occur when the head is still and the observer is focusing on non-moving objects, primarily analyzing eye saccades and fixations. Saccades refer to quick eye movements that shift the central foveal field of view from one point to another, while fixation is the act of maintaining the central foveal field of view on a target for enough time to capture detailed visual information [12,13]. Practically, eye movement behavior is significantly related to physiological indicators and the ability to process information from individuals [14]. Monitoring eye movement indicators during air traffic control tasks, in conjunction with other physiological data, helps to explore the specific relationships between various eye movement behaviors and fatigue.

In 2000, van Orden et al. investigated the average duration and frequency of blinks, as well as the duration and frequency of fixations, and the diameter of the pupil, using nonlinear regression and artificial neural networks for analysis [15]. The results showed that information from multiple ocular measurements can be combined to produce accurate personalized real-time estimates of sub-minute scale performance changes during a sustained task. In 2001, Sheridan et al. designed two experiments in which subjects were required to anticipate an aircraft conflict to investigate the effect of predictive warnings of varying reliability with varying duration on eye movements [16]. In 2006, Ahlstrom et al. used eye movement data related to cognitive work when investigating the effect of weather displays on controller operations [17]. The authors found that controllers had significantly larger mean pupil diameters at times of high workload, suggesting that eye movement measurements could provide a more sophisticated assessment of workload, thus allowing to better assess controller workload variability. In 2010, Paubel et al. evaluated ERASMUS and found that it had a large effect on the amplitude of controller saccades, the duration of fixations, and the distribution of attention across the visual scene [18]. In 2010, Stasi et al. used Wickens’ multiple resource model and found that peak saccade speed decreases with increasing cognitive load, in agreement with subjective test scores and performance data, demonstrating that eye-saccade speeds are quite sensitive to changes in brain workload [19]. Yet, the conclusion might be debatable. When experiencing a high cognitive load, the functional field of view shrinks, leading saccades to traverse a reduced distance and, consequently, not reaching their peak ballistic speed. In 2013, Imbert et al. evaluated the ability of five notification designs to attract attention during an air traffic control task and found that controllers’ eyes were more likely to be attracted to motion compared to color or even animation [20]. This aligns with the fact that our visual system prioritizes responding to motion rather than color. First, attention should be directed toward the colored object to ensure it is noticed, which implies that it must fall within a retinal area sensitive to color. Subsequently, we can distinguish the object’s color. In 2013, Di et al. simulated an ATC task to investigate the effects of task duration and task complexity on eye-saccades and fixation movements and found that micro-saccades and peak saccade speeds decreased with decreasing task duration, while task difficulty did not affect eye movements, and changes in eye movements over task duration were correlated with activation of the brain’s sleep center and mental fatigue [21]. In 2014, Imants et al. found that search strategies varied between air traffic controllers in the same task scenario [22]. Kang et al. found promising results by applying the expert controllers’ eye scanning path to improve training effectiveness and reduce training time by applying it to the novice active learning process, which is beneficial for improving novice performance in a conflict detection task [23]. In 2016, Marchitto et al. investigated air traffic controller eye movement behavior in a conflict monitoring task using eye movement recordings to assess the effect of complexity on cognitive workload in a simulated air traffic control conflict detection task [24]. Conflict trials were more complex and time-consuming than conflict-free trials and required more frequent eye movements. In addition, the bursts of large saccades decreased as the complexity of the task increased. An increased complexity due to trials involving conflicts may increase cognitive load, causing the functional field of view to narrow. Consequently, this limits the field to objects nearer to the current fixation during the planning of the next saccade.

In 2017, Kearney et al. concluded that eye movements are closely related to visual attention and can be analyzed to explore the shift of attention during surveillance tasks [25]. In 2017, Yoshida et al. found that flight placard color salience also affects controllers’ eye movement behaviors during control tasks [26]. In 2019, Di Flumeri et al. proposed an attention recognition system based on eye tracking and EEG to monitor when the attention of an air traffic controller will be in an inattentive state [27]. In 2021, Eisma et al. measured visual attention distribution by designing an aircraft conflict prediction experiment, and found that there was a significant increase in the conflict monitoring rate of the controllers after using enhanced visual feedback [28]. Wang et al. revealed that the professional status of controllers significantly influences eye movement patterns by comparing metrics such as fixation, saccades, and gaze entropy. Eye movements are vital indicators of information search actions, providing a valuable understanding of critical cognitive strategies for decision-making [29]. In 2022, Li Qinbiao et al. used EEG and eye tracking to identify controllers while conducting radar map monitoring, and developed a situation awareness recognition model based on EEG-ET data while considering workload [30]. John et al. found that the controller’s pupil size dilated with increasing workload after collecting data through a multi-object tracking task and a collision prediction task [31]. The blink frequency decreased with workload. Li et al. found that when controllers have greater situation awareness, the number of fixation counts, as well as the duration of the fixation, increases accordingly [32].

Existing research has fully recognized the critical role air traffic controllers play in the operation of air traffic, and further realized that understanding the fatigue state of controllers is crucial for the safe and efficient operation of flights. In terms of subjective assessment, there is already a considerable foundation for the detection of controller fatigue [33]. As for eye movement behavior, many previous studies have integrated it with facial information, such as huffing and other mouth features when collecting data [34]. This study will use the research methodologies described in the current literature to select and implement eye movement behavior as a means of assessing fatigue and designing experiments.

3. Experimental Design

We designed two different experiments to collect real-time data on eye movements and fatigue levels of air traffic controllers. The first experiment used a simulation environment created with Python. In the second experiment, an approach control simulation was conducted using a radar control simulator. Both experiments used Tobii eye trackers to monitor participants’ eye movements and gathered Karolinska Sleepiness Scale (KSS) ratings to denote their present fatigue states.

3.1. Experiment I: Conflict Monitoring and Warning

The main task of this simulation is to monitor and detect air traffic conflicts. The total duration of one simulation exercise is 40 min. The simulation system can provide a certain level of accuracy in monitoring conflicts and warning alerts. As shown in Figure 1, aircraft enter the simulation experiment platform randomly from one of the four waypoints, with each aircraft adhering to its predetermined flight path. On the right side of the simulation platform, a series of experimental parameters are shown, such as the aircraft’s call sign, speed, simulation duration, and the number of successfully coordinated flights. The established flight paths are designed to intersect and can be modified as necessary. Participants must ensure that all aircraft (either 9 or 18) throughout the experiment remain conflict-free and reach their destinations. At the end of the experiment, participants are required to complete the workload and fatigue scale. Each individual performs two tests with 9 and 18 aircraft, respectively, as illustrated in Figure 1. In this experiment, an aircraft conflict is defined as two planes being closer than the minimum safe distance of 20 km.

3.2. Experiment II: Simulation of Approach Control

In addition to conflict monitoring and alerts, this study collects eye movement data from air traffic controllers in a simulated approach control experiment and employs the KSS scale to evaluate fatigue levels. Throughout the simulation, air traffic controllers are tasked with methodically coordinating the sequencing of flights in the approach zone and guaranteeing their landing on the assigned runway, while aiming to prevent or warn of conflicts and effectively managing horizontal speed and vertical descent. Furthermore, during the approach, air traffic controllers must monitor and manage departing aircraft to ensure that they reach the departure altitude before exiting the control area, thus avoiding conflicts with incoming aircraft, as shown in Figure 2.

3.3. Subjects

Eight participants were involved in Experiment I, all male, 23 years old, majoring in air traffic control at Nanjing University of Aeronautics and Astronautics (NUAA). In the second experiment, three air traffic controllers from the Southwest Air Traffic Management Bureau were involved.

3.4. Simulation Procedure

3.4.1. Experiment I: Conflict Monitoring and Warning

The processes of simulation in Experiment I are described as follows:

• Step 1: Prepare the simulation environment to minimize the effects of light, particles, and other variables, and explain the experiment’s goal and procedures to the participants.
• Step 2: Let the participants become accustomed to the platform and simulation tasks; adjust their seats and other equipment.
• Step 3: Set-up and calibrate eye-tracking device.
• Step 4: Simulation starts, and the participants evaluate the traffic scenarios as instructed and complete the workload form.

To measure participants’ workload, we used standard NASA−TLX workload form. This scale is a widely used subjective workload assessment tool, designed to measure the workload perceived by operators during task performance and to explore the relationship between workload and fatigue. For controllers, the NASA−TLX scale can effectively evaluate their workload levels while using automated systems, thereby aiding in the improvement of system design and operational procedures and enhancing overall work efficiency and safety. The scale measures workload across six dimensions: mental demand, physical demand, temporal demand, task performance, effort, and frustration level. Once the controllers have completed the specified tasks, the system automatically presents a self-assessment questionnaire for each dimension, and scoring is done according to the controllers’ selections, as illustrated in Figure 3.

3.4.2. Experiment II: Simulation of Approach Control

The processes of the second experiment were described as follows:

• Step 1: Prepare the simulation environment to minimize the effects of light, particles, and other variables, and explain the experiment’s goal and procedures to the participants.
• Step 2: Let air traffic controllers become acquainted with the airspace map, understand the requirements for entering/leaving altitude, runway direction, and pre-adjust radar screen size.
• Step 3: Set-up and calibrate eye-tracking device.
• Step 4: Simulation starts, and air traffic controllers provide air traffic control service to the flights, while periodically completing the KSS scale.

3.5. Algorithms for Eye Movement Data Processing

3.5.1. Determination of Fixation and Saccade

In the fixation phase, the variability in the air traffic controller’s gaze leads to slight shifts in the point of gaze. The K-means clustering algorithm can classify these fixation points into different clusters. By counting the fixation points, saccade data can be obtained on the basis of the distance and time between consecutive fixations. The K-means clustering algorithm process is illustrated as follows:

• Step 1: Identify the number of clusters, K, i.e., classifying the fixation points into K clusters.
• Step 2: Select K points at random to serve as initial centroids from the set of fixation drop points.
• Step 3: Measure the distance from each point to each centroid and assign each point to the cluster associated with the nearest centroid.
• Step 4: Once each centroid has a group of points assigned, update the algorithm to choose new centroids (for each cluster, compute its average position to determine the new centroid). This might not effectively identify glissades, which are scenarios in which the centroid gradually changes due to the fatigue of the extraocular muscles. Determining a fixation based on its distance from the centroid could result in several brief fixations instead of one prolonged fixation. This method requires careful consideration when applied.
• Step 5: Repeat Steps 3 and 4 iteratively until the termination condition is met (e.g., the clustering outcome no longer changes).

Following the aforementioned steps, we obtain K clusters of fixation points. The intervals between the first and last fixation points within each cluster are measured and compared with the previous and subsequent clusters, enabling the determination of saccade frequency and duration. Ultimately, through a series of calculations, we determine the total number of fixation points and differentiate them from saccades.

3.5.2. Statistics on Blink

When analyzing metrics related to blinking, it is important to consider whether the eyes are closed. To determine if the air traffic controller’s eyes are closed, the eye aspect ratio (EAR) formula is used. First, the locations of the eyes must be identified, and six points are selected to represent the eyes, as illustrated in Figure 4. Based on these 6 points, the EAR is calculated as Equation (1).

$$EAR={\displaystyle \frac{\Vert P_2-P_6\Vert +\Vert P_3-P_5\Vert }{2\Vert P_1-P_4\Vert }}$$

(1)

3.6. Fatigue State Assessment

Several established fatigue measurement metrics have been introduced and validated in prior studies, including the Multidimensional Fatigue Inventory (MFI) and the Fatigue Severity Scale (FSS). Within the air traffic operational environment, only a limited number of measures are feasible. Subjective fatigue assessment tools include the Visual Analogue Scales (VAS), the Samn–Perelli seven-point Fatigue Scale (SPS), and the Karolinska Sleepiness Scale (KSS) [35]. Although the KSS is primarily designed to assess sleepiness, it is widely used to measure the fatigue levels of aviation operators, especially pilots and air traffic controllers [35,36,37,38]. This could be due to (i) the potential impact of both sleepiness and fatigue on aviation safety, making it challenging to distinguish between the two, leading managers to equate sleepiness with fatigue; (ii) the convenience of using KSS compared to other measurement methods [35]. In air traffic control studies, it is typically difficult for controllers to reach an extreme level of fatigue [39]. As a result, controllers rarely fall asleep. When rated using the KSS, fatigue levels generally stay within the first five stages. Despite the subjective nature of the sleepiness scale, it offers convenience for analyzing and summarizing experimental data and collections. However, when considering the fatigue of air traffic controllers, the KSS may only capture sleepiness as a component. Hence, this study will use the KSS to examine fatigue, potentially enhancing the accuracy of the findings.

4. Results

4.1. Single Eye Movement Indicator to Detect Fatigue

Before combining all the eye movement indicators chosen in this study, it is essential to individually link these data with the fatigue state. In the second experiment, six types of eye movement data were selected and collected, namely saccade velocity, saccade duration, saccade distance, pupil diameter, fixation area diameter, and fixation duration. The highest fatigue level observed is level 5 according to the self-assessments of the KSS sleepiness scale by the subjects. Specifically, there are 31 samples at fatigue level 1, 30 samples at level 2, 27 samples at level 3, and 9 and 8 samples at levels 4 and 5, respectively. For analysis purposes, the eye movement data are averaged across each fatigue level.

4.1.1. Saccade Velocity

The average eye-saccade velocities of air traffic controllers in 105 samples at various fatigue levels are depicted in Figure 5. At low fatigue levels (KSS levels 1–3), it is clear that a reduction in alertness leads to an increase in eye cascade velocity in the subjects. However, as alertness drops further, there is a significant decrease in average eye-saccade velocities.

4.1.2. Saccade Duration

Saccade velocity, duration, and travel distance are naturally interrelated and may exhibit a high degree of correlation. Variations in the duration of the eye cascade among subjects under different fatigue conditions are illustrated in Figure 6. As the KSS scale indicates higher fatigue levels, there is an overall trend of reduced saccade duration. This implies that, in approach control, air traffic controllers gradually spend less time on target tracking, possibly signifying diminished attention.

4.1.3. Saccade Distance

As depicted in Figure 7, the mean saccade distance of the controller exhibits a variable pattern with the fatigue condition. Given the unpredictability of the control task and the intricate nature of saccade behavior, isolating the saccade distance as a factor to correlate with the fatigue state proves challenging.

4.1.4. Pupil Diameter

Figure 8 demonstrates the variations in average pupil size in relation to fatigue levels; as the pupil size enlarges, the KSS level declines. This implies that with the rise in fatigue levels among controllers, the average pupil size tends to grow, signifying that pupil size can act as a physiological indicator of fatigue status.

4.1.5. Fixation Area Diameter

Figure 9 illustrates how the diameter of the fixed area of the subjects changes with the level of fatigue. Similarly to the diameter of the pupil, the diameters of the fixed area of air traffic controllers generally increase when their alertness, measured by the KSS self-test, decreases. When air traffic controllers are fatigued, their attention tends to vary. This fatigue leads to increased muscle tremors, resulting in less consistent focus, which is reflected in a larger and more dispersed fixation area [40].

4.1.6. Fixation Duration

In the analysis of fixation duration, a threshold of 120 milliseconds was set for fixation duration based on recommendations for handling short-duration fixations, aiming to exclude instantaneous fixations that may not provide meaningful information [41]. The results are shown in Figure 10. Unlike the diameter of the fixation area, the duration of fixations generally decreases as the alertness level measured by the KSS declines. When performing air traffic control tasks, controllers tend to exhibit shorter average fixation durations as their fatigue level increases. This could be attributed to the difficulty in maintaining steady fixation caused by fatigue, which disrupts information processing and potentially shortens fixation duration.

4.2. Evaluating the Importance of Eye Movement Metrics

Support vector machine (SVM) and random forest (RF) are selected for their complementary strengths in handling small datasets, noise, and analyzing the importance of features. SVM is particularly adept in situations with little data, as it increases the classification margin to achieve high accuracy, even with limited samples [6]. Its emphasis on support vectors (data points close to the classification boundary) ensures robust performance while minimizing sensitivity to a few outliers, which is often crucial in eye movement research where data can vary. In contrast, random forest helps overcome overfitting through an ensemble method that involves building and averaging multiple decision trees [42]. This makes RF exceptionally capable of dealing with noise in eye movement data, a common issue in these analyses. In addition, RF highlights the importance of the features, helping researchers determine which eye movement metrics, such as fixation duration or saccade speed, are the most impactful in classification or prediction tasks. This improves the interpretability of the results and helps in model refinement.

Given the benefits of SVM and RF for classifying and analyzing eye movement data, this research selected these techniques to detect and analyze fatigue levels.

4.2.1. Support Vector Machine

Support vector machines (SVMs) offer significant benefits and show great potential for detecting fatigue levels using eye tracking data. Their proficiency in handling high-dimensional and nonlinear datasets allows them to map intricate eye movement features into a high-dimensional space, facilitating the determination of the optimal classification hyperplane. This capability ensures the precise differentiation of various fatigue states, enabling real-time monitoring.

Eye movements generally exhibit complex nonlinear characteristics; thus, a Radial Basis Function (RBF) kernel is used to identify these intricate patterns, and the penalty parameter C is optimized through cross-validation. Based on cross-validation outcomes, the kernel type, C, parameter, and other hyperparameters are fine-tuned to ascertain the best model setup. The study utilized denoised data to construct a feature dataset for analysis. The data features included the following key indicators: fixation duration, fixation area diameter, pupil diameter, saccade velocity, saccade duration, and saccade distance. A total of 108 sample datasets were collected and proportionally divided into a training set and a test set, with the test set accounting for 25%. The training set was used to build and train the model, while the test set was used to evaluate the model’s performance. The aim was to explore the role and performance of these features in assessing controller fatigue.

The initial training and prediction using the support vector machine yielded an accuracy of approximately 63%. As illustrated in Figure 11, the red color represents the predicted values, while the blue color indicates the original data. The X-axis represents the number of samples in the test data sets, while the Y-axis represents the level of KSS. The contribution of individual eye tracking data to the prediction is examined during model evaluation, as depicted in Figure 12.

These results suggest that saccade distance has minimal positive influence on the model’s predictive performance. Consequently, excluding the saccade distance can lead to an improved model accuracy of approximately 67%, as demonstrated in Figure 13. Figure 14 illustrates the contribution of each eye-tracking metric to the prediction.

Based on the support vector machine analysis, it is clear that in the approach control experiment performed in this research, the size of the fixation area strongly affects the selected eye tracking indicators for estimating the controller’s fatigue state. The size of the pupil is the next most crucial factor, followed by the length of the fixation period as the third important measure.

4.2.2. Random Forest

Random forest is an ensemble method that constructs multiple decision trees using random subsets of data and features during the training phase. Due to its randomness and the deployment of multiple trees, it is resistant to noise and can effectively assess the significance of each feature for the selection and understanding of the model.

Eye movement behavior data typically have a high dimensionality, including fixation positions, pupil diameters, and other metrics. When employing random forests for analysis, it is not necessary to perform additional dimensionality reduction. For the specific analysis, parameters such as the number of trees, their maximum depth, and the maximum number of features per tree are configured. The data features include the following key indicators: fixation time, fixation area diameter, pupil diameter, saccade speed, saccade duration, and saccade distance. A total of 108 samples were collected and proportionally divided into training and test sets, with the test set accounting for 15%. The built-in feature importance function is then used to rank the influence of each eye-tracking data type on fatigue. Based on the outcomes of model evaluation, performance can be enhanced by fine-tuning the parameters.

The random forest model achieves an accuracy of approximately 59% for both training and prediction. As illustrated in Figure 15, the red indicates the predictions while the blue represents the original data. The X-axis represents the number of samples in the test data sets, while the Y-axis represents the level of KSS. Using the traits of the random forest algorithm, the significance of each feature for the target can be determined during the prediction process, as illustrated in Figure 16.

The figures demonstrate that the random forest algorithm’s evaluation reveals the diameter of the fixation area as the most influential predictor, followed by the pupil diameter, fixation duration, saccade distance, saccade speed, and saccade duration. After reordering based on significance, the optimized model achieved an accuracy of about 0.65, as shown in Figure 17. textcolorblueThe X-axis again represents the number of subjects, while the Y-axis represents the level of KSS. The contributions of each variable at this stage are illustrated in Figure 18.

By applying the random forest algorithm, the simulated approach control experiment indicates that the size of the subject’s fixation area most significantly influences the results, with the diameter of the pupil being the next most important factor, and the duration of fixation ranking third. These fixation properties are very useful for creating a fatigue model for controllers.

4.2.3. Summary

Support vector machines and random forests were used to analyze and predict indicators of eye movement behavior and fatigue status. Given their unique attributes, support vector machines are highly sensitive to parameters, making their prediction accuracy highly dependent on parameter selection, and they have difficulties with managing non-linear, large-scale data. In contrast, random forests are not suitable for data sets with many missing values or sparse data, and they have a high model complexity. This study illustrated that both algorithms demonstrated the significant contribution of various eye movement metrics, such as the diameter of the fixation area, the diameter of the pupil, and the duration of the fixation, to fatigue levels.

5. Discussion

Based on the analysis of changes in eye movement metrics presented in this paper, we concluded that with increasing fatigue in air traffic controllers, their visual system adapts. When fatigue is minimal, controllers maintain task attention by increasing saccade speed. However, as fatigue increases, pupil dilation becomes larger, saccade speed decreases, and saccade range contracts. This indicates that their attention becomes more restricted, increasing the risk of tunnel vision—a well-known phenomenon among controllers that they work diligently to prevent. Furthermore, controllers’ fixation patterns diversify with greater fatigue, as fixation areas expand and fixation durations tend to shorten. In contrast to earlier research, this study investigates and confirms the relationship between pupil size, saccade characteristics, and fixation patterns as markers of fatigue. Furthermore, it highlights the differences among multiple fatigue detection approaches. Whereas most prior studies have concentrated on evaluating a single detection method, this research offers a comprehensive perspective by contrasting SVM and random forest techniques.

Future investigations should aim to expand the sample size to improve the applicability of the findings. Furthermore, it is crucial to categorize the eye movement data according to the levels of experience of the participants to understand how different levels of expertise affect fatigue detection. Incoming research should also investigate the combination of other physiological and behavioral indicators with eye movement data to improve the precision of fatigue detection. In conclusion, this study offers a new point of view on fatigue detection for controllers and proposes avenues for future research.

6. Conclusions

As air traffic controllers continue to play a crucial role in maintaining the safe operation of air traffic, it is essential to promptly detect their fatigue. Several eye movement metrics have been suggested to predict fatigue levels, such as fixation characteristics and pupil diameter. In the approach simulation task, when the controller’s fatigue levels increase, the pupil tends to dilate, and the fixation becomes more dispersed, resulting in a larger fixation area diameter. As the alertness level, measured by the KSS, decreases gradually, the subjects’ eye-saccade characteristics will also change; specifically, when the controller’s alertness drops to a certain threshold, both the velocity and duration of the eye-saccade substantially decrease. Moreover, comparing the results obtained from the support vector machine and random forest algorithms shows that in both cases the parameters ranked from most to least critical are the diameter of the fixation area, the diameter of the pupil, and the fixation duration, respectively. This implies that these three metrics may have significant value in future research on eye movement behavior and fatigue.

In contrast to prior research, this study examines the relation between air traffic controller fatigue and eye movement indicators. Very few investigations have looked into this specific subject. Consequently, this paper has the potential to prompt further study. However, there are some limitations. First, only 11 participants were involved in this study, leading to a limited sample size for more extensive analysis. Second, the duration of our experimental sessions was not long. Typically, in actual air traffic control, controllers manage flights for as long as 2 h before having a break. Furthermore, individuals with different levels of control expertise may exhibit different patterns in eye movement data during tasks. In addition, sleepiness may represent just one aspect of fatigue. Therefore, future research should aim to encompass more aspects of fatigue. It would be beneficial to classify participants and review the detection data of individuals with different levels of experience to develop more effective fatigue detection strategies.