CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Predicción del riesgo de go-around en el aeropuerto de Barcelona-El Prat mediante aprendizaje automático

Estudiants que han llegit aquest projecte:

Director/a: REYES MUÑOZ, M. ANGÉLICA

Departament: DAC

Títol: Predicción del riesgo de go-around en el aeropuerto de Barcelona-El Prat mediante aprendizaje automático

Data inici oferta: 22-01-2026     Data finalització oferta: 22-07-2026


Estudis d'assignació del projecte:
    GR ENG SIST AEROESP
Tipus: Individual
 
Lloc de realització: EETAC
 
Segon director/a extern: Juan Antonio Guerrero Ibáñez
 
Paraules clau:
Machine learning, ADS-B, GRU, Transformer, go-around
 
Descripció del contingut i pla d'activitats:
Este proyecto aplicará técnicas de machine learning al ámbito de la Gestión del Tráfico Aéreo para la predicción de go-arounds. Para ello, se emplearán datos reales de trayectorias de aeronaves obtenidas a partir de sistemas ADS-B, junto con información meteorológica operacional (por ejemplo, METAR), con el fin de caracterizar el entorno operativo durante las aproximaciones al aeropuerto.

El proyecto pretende aportar una herramienta de apoyo a la toma de decisiones que contribuya a una gestión más proactiva, segura y eficiente del tráfico aéreo.
 
Overview (resum en anglès):
This project develops and evaluates machine-learning models to estimate the risk that an approach to Josep Tarradellas Barcelona-El Prat Airport will end in a go-around. The main aim is to assess whether patterns related to this type of manoeuvre can be identified using information available during the approach, such as ADS-B trajectories and METAR weather reports. In this way, the model could provide useful information as a support tool for air traffic management.

To carry out the study, more than 100,000 historical arrival trajectories to LEBL were downloaded and processed. An algorithm was developed from the altitude data to automatically detect and label go-arounds. The algorithm is mainly based on identifying an aircraft descent followed by a sustained climb, indicating that the approach has been discontinued. In addition to the original trajectory data, variables related to runway geometry and meteorological conditions were included.

Two types of datasets were created. On the one hand, a tabular dataset was built, in which each flight is represented by a single point of the approach, close to 700 m of altitude. On the other hand, a sequential dataset was created to preserve the evolution of the approach between 3,000 m and 700 m, representing each flight through 40 time points. A second version of this sequential dataset was also produced by adding the time separation from the previous approach assigned to the same runway.

First, tabular models based on logistic regression, Random Forest and XGBoost were tested. Afterwards, a GRU network and two Transformer Encoder-based models were trained: one without the operational separation variable and another including it. Since normal approaches were substantially more frequent than go-arounds, class weights were used to prevent the models from focusing only on the majority class.

The tabular models did not achieve sufficiently good results, as they were unable to detect go-arounds without generating a high number of false alerts. The best-performing model was the Transformer Encoder with operational separation. In the temporal test set, it detected 15 out of 35 go-arounds, achieving a precision of 0.789, a recall of 0.429 and a PR-AUC of 0.433. In addition, it generated only four false alerts among 10,622 normal approaches.

Overall, the results show that analysing the temporal evolution of the trajectory and considering the separation between aircraft can help estimate the risk of a go-around. Nevertheless, the model does not detect all events and should therefore be regarded as a proof of concept rather than an autonomous alert system. Finally, an interactive simulator was developed to reproduce trajectories and display the risk score generated by the model.

© CBLTIC Campus del Baix Llobregat - UPC