Trajectory-BERT: Trajectory Estimation Based on BERT Trajectory Pre-Training Model and Particle Filter Algorithm.

Area of Science:

• Aviation
• Data Science
• Artificial Intelligence

Background:

• Trajectory data are vital for aviation safety and flight intention analysis.
• Noise and signal interruptions degrade the precision of collected aviation data.
• Existing methods struggle with incomplete or noisy trajectory information.

Purpose of the Study:

• To develop a robust method for estimating missing or noisy aviation trajectory data.
• To improve the accuracy of flight path reconstruction using advanced AI models.
• To enhance aviation safety through more reliable trajectory data.

Main Methods:

• Utilized a Bidirectional Encoder Representations from Transformers (BERT) model for trajectory data pre-training.
• Employed masked trajectory data estimation based on surrounding points during BERT training.
• Integrated a refined particle filter algorithm to generate alternative trajectory sets for noisy data.
• Determined the optimal trajectory by computing maximum posterior probability with the BERT model.

Main Results:

• The BERT trajectory pre-training model successfully learned intricate motion patterns.
• The combined BERT and particle filter approach demonstrated superior performance in trajectory reconstruction.
• Experimental results indicate the proposed model outperforms traditional algorithms in handling noisy data.

Conclusions:

• The BERT-based trajectory estimation method significantly improves data precision and reliability.
• This approach offers a promising solution for enhancing aviation safety and operational efficiency.
• The model's ability to handle noise makes it valuable for real-world aviation applications.