Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)
Independent multi-flyover development fixtures
Historical fixture: the aircraft-type selection described below predates the current non-jet-only search policy.
These cases exercise the requirement to use three, four and more independent flyovers under one recording clock and location. They are synthetic development fixtures, not the real field corpus or sealed holdout requested in issue #53.
The generator at 1f0342594bdc7fc7089a075b6dbf26aca9342711 does not import the
production acoustic, Doppler or search implementation. It transports source
oscillators from emission time to receiver time using WGS84 slant distance and
343 m/s still-air sound speed. Curved trajectories have varying headings, speeds
and heights. Noise, stationary tones and second harmonics are present. Model
velocity is independently differentiated and expressed in local ENU coordinates.
The synthetic terrain and geoid establish a declared common height datum.
Every case includes the complete finite synthetic traffic population with opaque IDs. Two remote regions contain distractor trajectories. Their first two passes are similar to the true region; later passes differ in timing and geometry. Jets and unknown types are retained. No receiver coordinate, manual event window, frequency band or correct flight ID is present in the search request.
Reproduce or download
.venv/bin/python scripts/make_multi_event_case.py --events 4 --output data/new-four-event-case
.venv/bin/python scripts/discover_audio.py data/new-four-event-case/inputs/recording.mp3 --output results/new-four-event-discovery
Choose --events 3, 4 or 6; the default seed is 20261103. Existing output
directories are refused. The generated WAV preserves the renderer output; the
MP3 is encoded with FFmpeg/libmp3lame. The manifest records exact media/code
hashes and encoder command/version.
The checksummed fixture release
contains all three generated cases. Verify the bundles against SHA256SUMS, then
each extracted file against manifest.json. The source YouTube MP4 is excluded.
For a blinded product run, expose only inputs/: MP3, trajectory Parquet,
terrain/geoid rasters, coverage declaration and request with a known synthetic
clock. Build the flight index from that complete Parquet and coverage declaration.
Do not expose truth/, renderer/, generation.json or the generator source to
the search process. Join truth/labels.json after predictions are frozen. The
initial discovery checks below did not supply labels, but did not assert a
filesystem isolation boundary.
Frozen discovery results
The first runs used unchanged default settings, including contrast 2.5, with no retuning after generation or after observing results. Labels were read afterward for interval matching at IoU 0.3; harmonics count once.
| Independent flyovers | Duration | Groups found | Hits / misses / extra groups | Runtime |
|---|---|---|---|---|
| 3 | 740 s | 3 | 3 / 0 / 0 | 8.26 s |
| 4 | 960 s | 4 | 4 / 0 / 0 | 9.30 s |
| 6 | 1,400 s | 6 | 6 / 0 / 0 | 13.09 s |
Complete measurements, plots and receipts retain every generated case. Recorded peak RSS was 343,792 / 1,648,608 / 398,472 KiB respectively; the four-event outlier is preserved without an established causal explanation. The generator's analytic tests check both approaching and receding constant-velocity Doppler frequency, propagation delay and invalid arrival ordering. The combined clean Python 3.14 suite passed 125 tests in 36.60 seconds, with one existing pandas future-dtype warning.
These results establish event discovery on these fixtures. Flight associations, location error, ambiguity and actual accepted/inconclusive decisions require the subsequent full pipeline run, reporting 100/500/1,000 m thresholds. No location success is claimed from detection counts or from known synthetic terrain alone.