FTI-CT: Flight Trajectory Interpolation with Constrained Transformer

High-resolution flight trajectory data is vital for modern Air Traffic Management (ATM), but inherent irregular sampling and significant data gaps in surveillance pose a challenge to robust interpolation. Existing methods—from geometric fitting to deep learning—suffer from two core limitations: 1) Insufficient dynamic modeling: they struggle to capture complex non-linear motion and long-range temporal dependencies, leading to poor accuracy during extended data outages. 2) Lack of physical constraints: data-driven models prioritize fitting over physics, potentially generating physically infeasible or safety-violating “fake” trajectories. To address this, we propose an innovative solution. We design a Bidirectional Self-Attention-based Encoder to deeply capture the long-range temporal context around missing points. Crucially, we introduce a Physics-Constrained Loss Function that explicitly integrates aircraft kinematics and airspace limits into training, ensuring the physical feasibility and safety of the results. Extensive experiments on real ADS-B data show that our method achieves significantly superior performance in key metrics RMSE and MAE, validating its effectiveness and practical value for ATM.