TSAN: Temporal Siamese Attention Network for Flight Trajectory Similarity Learning

Similarity computation is fundamental for numerous tasks in flight trajectory data analysis. With the help of the quick access recorder (QAR), an on-board sensor that collects flight parameters during flight, we can obtain a considerable amount of flight trajectory data. However, existing methods primarily focus on 2-D vehicle trajectory similarity computation and struggle to handle flight trajectory similarity in more intricate 3-D air traffic networks. To fill this gap, we propose a novel temporal Siamese attention network for flight trajectory similarity learning, named TSAN. TSAN first creates variants of the input trajectories by trajectory augmentation methods to learn against data noise. Then, after utilizing parallel temporal convolution networks (TCNs) to capture the spatio-temporal features between different trajectory variants, TSAN integrates the Siamese attention mechanism, facilitating interaction among different feature vectors to learn comprehensive representations of flight trajectory. Experimental results on two real-world datasets demonstrate that TSAN substantially outperforms competitors in effectiveness and efficiency while exhibiting low parameter sensitivity. Moreover, a case study demonstrates the proficiency of TSAN in downstream tasks.