Flight-JEPA
Multi-scale JEPA for flight trajectory representation learning.
Key Features
- Event-aware curvature-based segmentation → semantic tokens
- Multi-scale: Local (patch) + Global (segment) JEPA
- VICReg collapse prevention for low-entropy cruise data
- Curriculum adaptive masking (random → maneuver-biased)
- 9-dim features: [x,y,z, u_x,u_y,u_z, r, sin_θ, cos_θ]
Architecture
- Encoder: 3L Transformer, d=256, 8h
- Predictor: 2L bottleneck, d=128
- EMA decay: 0.998
- Params: 4,867,456
Results on ATFMTraj/RKSIa
- Classification ACC: 0.9128
- See metrics.json for full results
Built on
TS-JEPA, Flight2Vec, ATSCC, VICReg, I-JEPA
