Flight Log Anomaly Detector (Hampel + optional STL)

Flags weird events in UAV telemetry: battery voltage spikes, GPS drift/jumps, IMU outliers.

What it does

• Hampel filter (rolling median + MAD) on chosen numeric columns.
• Optional STL decomposition to remove trend/seasonality before Hampel (per-column).
• Optional GPS jump detector using lat/lon vs reported speed.
• Outputs an anomalies JSON list with timestamps, columns, values, and reasons.

MVP model

• For each selected column, apply a windowed Hampel filter:
  • window size k (neighbors on each side, so total window = 2k+1)
  • threshold n_sigma (default 3.0) on MAD-scaled deviation
• If you pass --stl-cols, we run STL (if available) and apply Hampel to the residuals.
• If lat, lon, and gps_speed_mps exist, we can also flag GPS jumps where the per-sample movement exceeds a plausible distance given speed.

This is deliberately simple, fast, and explainable. Treat it as triage; you can review flagged rows.

Quick start

1. CSV columns you can use (example below):

   • timestamp (ISO)
   • batt_voltage_v
   • gps_speed_mps
   • lat, lon (optional, for GPS jump detector)
   • imu_ax, imu_ay, imu_az (optional)

2. Detect anomalies and write JSON:

python flight_log_anomaly_detector.py detect   --csv example_log.csv   --time-col timestamp   --cols batt_voltage_v,gps_speed_mps,imu_ax,imu_ay,imu_az   --out anomalies.json   --k 7 --n-sigma 3.0   --stl-cols batt_voltage_v --stl-period 60   --gps-jump 1 --gps-alpha 2.0 --gps-margin-m 30

3. Output: anomalies.json like:

[
  {"index": 120, "timestamp": "2025-01-01T12:00:10Z", "column": "batt_voltage_v", "value": 15.9, "reason": "hampel", "residual": -0.8},
  {"index": 245, "timestamp": "2025-01-01T12:02:05Z", "column": "gps", "value": 185.3, "reason": "gps_jump", "distance_m": 185.3, "allowed_m": 70.1}
]

Arguments

detect
  --csv <path.csv>                Input telemetry CSV
  --out <anomalies.json>          Output JSON path
  --time-col <name>               Timestamp column (ISO or parseable)
  --cols <c1,c2,...>              Numeric columns to Hampel-scan
  --k <int>                       Half-window size (neighbors). Default 7
  --n-sigma <float>               MAD threshold. Default 3.0
  --stl-cols <c1,c2,...>          Columns to detrend/seasonal-remove with STL before Hampel
  --stl-period <int>              Seasonal period (samples). Example: 60 for 1 Hz * 1 minute
  --gps-jump <0|1>                Enable GPS jump detector (requires lat,lon,gps_speed_mps)
  --gps-alpha <float>             Allowed multiplier on speed*dt (default 2.0)
  --gps-margin-m <float>          Extra fixed margin in meters (default 30)

Notes

• If statsmodels isn't installed, STL is skipped with a warning.
• Time deltas are inferred from --time-col differences; if missing/invalid, GPS jump is skipped.
• Hampel flags are independent per column; review context in your log viewer.

Example data

example_log.csv contains a synthetic 10-minute flight at 1 Hz with planted anomalies.

License

MIT