• Twitter
  • Facebook
  • Google+
  • LinkedIn

Analysis of Strategies used by ATCOs for handling Cognitive Workload

Name: Anudeep Peela

Department: Aerospace Engineering

Program: Dual Degree (4th  year)

Name of supervisor: Prof. Rajkumar Pant



Driven by economic growth,  the world’s air traffic has been increasing rapidly in the last decade. An increase in air traffic resulted in an increased workload on air traffic controllers, compromising safety. Hence our research aims to find out strategies used by expert air traffic controllers which will help design efficient training programs and better interaction tools to support the demands of increased air traffic.

First, we analyse the essential parameters obtained from eye-tracking data, EEG(Brain Waves), and metadata (like age, designation and similar other features.). We further went on to build a machine learning model which predicts the expertise of the air traffic controller from which we obtain essential features.



Air traffic controllers are personnel responsible for the safe, orderly, and expeditious flow of air traffic in the global air traffic control system. Controllers apply separation rules to keep aircraft at a safe distance from each other in their area of responsibility and move all aircraft safely and efficiently through their assigned sector of airspace, as well as on the ground. Because of its significant responsibility, which is on duty and makes countless real-time decisions daily, the ATC profession is consistently regarded around the world as challenging and stressful.

One of the most common causes of aviation accidents is human error, including the pilot, flight crew and air traffic controllers.

Following are types of air traffic controllers errors:


• Fatigue: Exhaustion is a significant issue for ATCs. Their work is often under highly stressful conditions and long shifts. Fatigue will cause slow response time and will contribute to poor decision making.

• Improper staffing and training: Staffing shortages will lead ATCs to extra stressful hours, and also inadequate staff training will lead to unprepared controllers in the event of an emergency.

• Failure to properly communicate: Communication between ATCs and pilots is simple and to the point. There should not be any misunderstandings.

So it is clear that analysing the strategies used by expect controllers will help in minimising the above errors.


Definition of expertise:

The cognitive overload in ATCs is due to excessive demands on perceptual and cognitive resources concerning processing capacities. Furthermore, a very high cognitive workload as observed through the questionnaire suggests the direct impact of increasing air traffic in pushing ATC’s cognitive capabilities. So Peak Hour is considered cognitive workload estimation. The ability to work in peak hour and no compromise on the safety of all the stakeholder indicates the expertise of ATCs.

Accordingly, we defined expertise as the number of sequenced planes that landed successfully, given the same conflicting scenario with an increasing number of flights with time.



Around 60 ATCs of varying background, age, experience and level of ratings are given the same conflict scenario in simulation software. Their channelling, eye movement and physiological conditions are recorded in Real-Time. The combined data set is obtained from the Eye-tracker, EEG and metadata of each controller.


Results of the analysis:

Arrival Rate vs Designation:

Figure 1

The earlier defined metric of expertise is Arrival Rate. Career growth of ATC from junior executive to executive director is as follows Junior Executive -> Assistant Manager ( After 3 years) -> Manager (After 3 years from previous post/direct recruitment) -> Senior Manager ( After 9 years of completion of service) -> Assistant General Manager ( Time-bound/ selection) -> Deputy General Manager -> Joint General Manager -> General Manager -> Executive Director.

From Figure 1, it seems like Arrival Rate does not follow career growth. The designation involves sequential growth and candidates who changed from other departments hence with zero experience. Therefore no conclusions can be made from just the Arrival Rate vs Designation plot.

Figure 2

Arrival Rate vs Stream: Figure 2 plot intuitively shows that expertise is independent of the current stream the ATC is working on.

Arrival Rate vs Blinks

Arrival Rate vs Blinks: It indicates expertise ATCs blink higher on average than novices. This supports the research conducted by Pant, Atish, Amit. It means that, on average expert, controllers are relaxed and have more number blinks measure.


Heat Maps:

A heat map (or heatmap) is a data visualisation technique that shows the magnitude of a phenomenon as colour in two dimensions. Here the magnitude is indicated by the intensity of the colour.

The Figure 3 plot shows that arrival rate and memory deviation are negatively related. Saccades, Dwell Time, and Fixations are highly correlated. Also, age and experience are associated.

Figure 3

Figure 4

Heat map analysis of Eye Tracking Data: Figure 4 plot of Arrival rate < 7 and Arrival rate >=7 shows significantly that Expert ATCs have a high correlation between IA REGRESSION IN and IA REGRESSION OUT FULL. In contrast, novices showed a high negative correlation between IA REGRESSION IN and IA REGRESSION OUT FULL. This indicates that expert ATC retraces back to conflict zone significantly greater than novices. This brings out the awareness or experience component of expert ATCs compared to novices.

Furthermore, we have the T-distributed Stochastic Neighbor Embedding plot. It is a machine learning algorithm for data visualisation. A nonlinear dimensionality reduction technique is well-suited for embedding high-dimensional data for visualisation in a low-dimensional space of two or three dimensions. Precisely, it models each high-dimensional object by a two- or three-dimensional point in such a way that nearby points model similar objects and dissimilar objects are modelled by distant points with high probability.

Figure 5

This plot (Figure 5) is of 2D projection of 120 dimension space such that violet indicates arrival rate <= 4 and yellow indicate arrival rate >=7. So this is used for the threshold calculation of expertise.


Multiple Object Tracking and Tracing:

Below sequence of images(Figure 6) paths of aircraft directed by novices(< 4) and expert(> 7) ATCs. It is visible in the second image that the red path novices divergence from the initial point and also as time goes on paths directed by expert ATCs This shows that forwarding thinking and the ability to detect conflicting areas ahead of time help ATCs handle sequence more aeroplane.

The tracing of the path is done with the help of computer vision algorithms like template matching and deep learning techniques.

Figure 6


Machine Learning Model for the prediction of expertise of ATCs:

Logistic Regression with regularisation is used for this model but with Weight of Evidence transformation. WOE transformation can:

• Handles Outliers

• Handles Missing Values

• Need to create dummy variables


Then Sequential Feature selection Algorithms which are a family of greedy search algorithms used to reduce an initial d-dimensional feature space to a k-dimensional feature subspace where k < d. The motivation behind feature selection algorithms is to automatically select a subset of the most relevant features to the problem. The goal of feature selection is two-fold: We want to improve the computational efficiency and reduce the model’s generalisation error by removing irrelevant features or noise. In a nutshell, SFAs remove or add one feature at a time based on the classifier performance until a feature subset of the desired size k is reached. Using Hyperparameter optimisation, we obtained the best set of parameters for model performance. Below the plot(Figure 7) is the output of SFS.

Figure 7


So the top features of best performing model are

  1. Blinks  
  2. Dwell time - Amount of time participant glances at an interest area
  3. Fixations
  4. IA AREA - Pixel area for the current interest area
  5. IA AVERAGE FIX PUPIL SIZE.- Average pupil size across all fixations in the interest area


These support the analysis from the plot. With a machine learning model, we can achieve an accuracy of above 82.5 per cent in the prediction of expertise.