Jan 2025 – Dec 2025 · Graduate Capstone
Predictive Anomaly Detection in Aviation Telemetry
Catching mechanical drift across 64 sensor channels before it becomes a fault — by learning what normal flight looks like and measuring how far reality has moved from it.
Results
- 92.5% accuracy and 0.857 precision on the NAB benchmark (161 windows, 17 labeled anomalies)
- ROC-AUC ≈ 0.69 and PR-AUC ≈ 0.46 — tunable toward high precision (0.86) or high recall (0.76) depending on deployment needs
- Subsequence-level detection of gradual drift, not just point anomalies
- Validated on a benchmark outside the training domain, confirming transfer
Stack
- Python
- PyTorch
- EXAMM
- scikit-learn
- CUDA
- NumPy
- pandas
One flight, before maintenance — sensor correlations (e.g. E1 RPM ↔ IAS at 0.86).
Same aircraft, after maintenance — those same correlations tighten (E1 RPM ↔ IAS rises to 0.91), which is the kind of shift the pipeline is built to catch.
Each window covers 30 seconds of flight data, and a new window starts every 5 seconds — so consecutive windows overlap by 25 seconds. That heavy overlap is deliberate: a single 30-second slice barely moves between windows, so a slow multi-channel drift shows up as a steady change across many windows in a row, instead of getting missed between two coarser, non-overlapping ones.