CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Development of an Artificial Intelligence Assistant for Air Traffic Control based on Speech Recognition and Machine Learning

Estudiants que han llegit aquest projecte:

Director/a: BARRADO MUXÍ, CRISTINA

Departament: DAC

Títol: Development of an Artificial Intelligence Assistant for Air Traffic Control based on Speech Recognition and Machine Learning

Data inici oferta: 04-12-2025     Data finalització oferta: 04-07-2026


Estudis d'assignació del projecte:
    GR ENG SIST AEROESP
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
Air Traffic Control, Automatic Speech Recognition, Natural Language Understanding, Machine Learning, Whisper, Named Entity Recognition, Human-in-the-Loop.
 
Descripció del contingut i pla d'activitats:
L'objectiu principal del treball és desenvolupar un 'bot' capaç de respondre utilitzant la fraseologia estàndard de control de trànsit aeri. El projecte es basarà en l'ús de dades reals (àudios i transcripcions de converses pilot-controlador) per entrenar un model de ML i LLM. La finalitat és classificar els missatges i generar una resposta automàtica que sigui coherent i correcta segons la normativa, servint com a eina de simulació o entrenament.
El pla d'activitats inclou:
- recerca de conjunts de dades oberts i etiquetats
- recerca de models de veu-a-text i validació amb el vocabulari de control aeri
- classificació de missatges
- regles de resposta per missatge
- creació del pipeline de prova
- validació a dades no vistes
- redaccio del treball
L'estudiant ha de tenir coneixament de Pandas i de SciKit-Learn.
 
Overview (resum en anglès):
Air Traffic Control (ATC) is a safety-critical domain where human operators manage aircraft flow through Very High Frequency (VHF) voice communications. A significant portion of these communications consists of routine, predictable interactions (such as initial contacts) that contribute to controller cognitive workload without requiring complex decision-making.
This project aims to design, develop, and evaluate a low-latency automated ATCo assistant capable of handling the specific initial contact subset of communications. To meet civil aviation safety requirements, a deterministic modular architecture is proposed, integrating Automatic Speech Recognition (ASR) models with lightweight Machine Learning (ML) classifiers and rule-based extraction algorithms.
Ten ASR architectures were evaluated, with a fine-tuned Whisper Large model achieving 92% accuracy on real ATC audio. Three Natural Language Understanding (NLU) approaches were compared; the selected modular architecture, designed specifically for this project, achieved 100% accuracy on speaker and intent classification and 84.6% callsign extraction accuracy, with a mean processing time of 4.6 milliseconds per sequence. A functional Client-Server prototype was developed to validate the complete system in a simulated operational environment, featuring a graphical interface with Text-to-Speech (TTS) response capabilities.
Results confirm that deterministic, non-generative architectures can reliably automate initial ATC contacts while maintaining the safety and latency standards required by civil aviation. The remaining challenge is infrastructural: achieving full real-time operation requires GPU-accelerated hardware deployment beyond the scope of this project.

© CBLTIC Campus del Baix Llobregat - UPC