CL-DFTC: A deep flight trajectory clustering framework based on self-supervised contrastive learning

Trajectory clustering, as an unsupervised data-driven method, can effectively identify and characterize air traffic flows, playing a key role in airspace planning, air traffic flow management, and decision support tools. However, existing trajectory clustering methods struggle to effectively differentiate trajectories that exhibit high global similarity and significant local differences. Additionally, trajectory errors accumulated at terminal area boundaries can also lead to incorrect clustering. To address these challenges, this paper proposes a deep trajectory clustering framework CL-DFTC based on contrastive learning. First, the model leverages contrastive pretraining to model the relationships between trajectory samples, improving the performance of trajectory embeddings in downstream clustering tasks. Second, three trajectory augmentation techniques tailored to the flight characteristics of the terminal airspace are proposed to guide the model in learning embedding representations that align with terminal area operating patterns. Finally, a trajectory clustering module is designed that integrates multiple clustering losses to jointly optimize both representation learning and clustering performance. Experimental results show that CL-DFTC outperforms existing methods on trajectory datasets from three different airports. Moreover, CL-DFTC can effectively generalize to the terminal airspace of different airports for trajectory pattern recognition, including special patterns like holding patterns, without the need for tedious manual feature engineering.