Noise robust aircraft trajectory prediction via autoregressive transformers with hybrid positional encoding

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Computer Science

Background:

• Aircraft trajectory prediction is crucial for air traffic management and safety.
• Existing models lack robustness in noisy and dynamic environments.
• Challenges include capturing complex spatio-temporal dynamics and positional information.

Purpose of the Study:

• To introduce a novel, noise-robust model for aircraft trajectory prediction.
• To enhance the reliability and accuracy of trajectory predictions.
• To address limitations of current models in handling noisy data.

Main Methods:

• Developed the Noise-Robust Autoregressive Transformer (NAR-T).
• Integrated noise-regularized embeddings and multi-head attention with hybrid positional encoding.
• Formulated robust trajectory prediction as a sequence-to-sequence learning problem using an autoregressive approach.

Main Results:

• The NAR-T model demonstrates enhanced prediction reliability in noisy scenarios.
• Achieved improved capture of temporal-spatial relationships and precise positional information.
• Showcased superior long-term prediction accuracy and real-time responsiveness compared to existing methods.

Conclusions:

• The proposed NAR-T model offers a significant advancement in robust aircraft trajectory prediction.
• Effective for complex, dynamic, and noisy air traffic environments.
• Paves the way for safer and more efficient air travel through improved prediction capabilities.