Abstract:
Unmanned Aerial Vehicles (UAVs) are increasingly employed in diverseapplications-from surveillance and inspection to delivery-generating vastamounts of structured telemetry data. To enable accurate, interpretable missionprofiling, we propose a dual- stage TabNet framework trained entirely onsynthetic UAV telemetry. First, we simulate 12,000 missions (Surveillance,Delivery, and Inspection) in Microsoft AirSim under varied environmentalconditions, logging 10 Hz multivariate data (GPS, velocities, battery status,event flags). Segments of 10-15 s are aggregated into 20 statistical and spectralfeatures. Stage I classifies each segment into one of three super-classes, achieving80.36% accuracy, while Stage II refines predictions into five sub-classes,achieving 76.04% accuracy. Both TabNet models leverage entmax attentionmasks for per-instance feature selection, enabling insight into the most salienttelemetry variables. We compare our work to ten related studies on tabular deeplearning, synthetic data generation, and hierarchical classification,demonstrating that our approach uniquely combines hierarchical modeling,simulation-driven dataset creation, and interpretability. Our open-source codeand synthetic dataset are publicly available. These results validate the viabilityof simulation-first, interpretable deep tabular learning for UAV mission analysisand pave the way for real- world deployment via sim-to-real adaptation.
Page(s):
110-110
DOI:
DOI not available
Published:
Journal: 4th International Conference of Sciences “Revamped Scientific Outlook of 21st Century, 2025” , November 12,2025, Volume: 1, Issue: 1, Year: 2025
Keywords:
intrusion detection
,
Cyberphysical systems
,
Fault Detection
,
UAV security
,
tabnet