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