Skip to main content

Manoeuvre detection (CUSUM and detectors from frozen models)

Source: cusumDetect, disagreementDetect and courseRateDetect in packages/classical/src/detect.ts; scoring in packages/eval/src/gate-g6.ts. CLIs: tma-detect (scores) and tma-g6 (calibration and gate).

Requirements: SRD Annex C §C.12, spec 017.

Detectors​

Each detector gives a per-timestep score from zig_index + 1. An alarm is a score at or above the detector's threshold.

  • CUSUM on the reference IMM's innovations (the baseline), restarting at 0 when the tracker re-initialises. There are two kinds:

    • nis (variance): g ← max(0, g + min(NIS, clip) − c);
    • mean (mean shift, two-sided, on e = ν/√S clipped to ±clip): g± ← max(0, g± ± e − c), with score max(g+, g−).

    The clip matters. The reference IMM's NIS is heavy-tailed (quiet median 0.06, mean 0.83, 0.2% above 10), and one outlier would otherwise hold an alarm for hours.

  • Course disagreement: |course_T2 − course_IMM| in degrees.

  • Course rate: T2's own course change over the last w samples, in degrees.

Scoring (G6)​

  • Events. Target course changes on σ = 1° scenarios. Each window runs 10 min from the start, cut at the next manoeuvre's start.
  • Delay. The time to the first row in the window at or above the threshold. A detector already alarming on the row before the start is credited nothing.
  • False alarms. Alarm onsets on quiet (settled) rows, per hour.
  • Thresholds. Each detector's is calibrated on validation to the same false-alarm rate, 0.5 per hour.

Result: phase 5 stopped at validation (seed 211)​

Detected within 5 min0.5 false alarms/h (declared)1/h2/h5/h
Best CUSUM (mean, c 0.5, clip 3)28.2%38.0%47.8%62.6%
T2's course change over 1 min19.2%39.0%58.1%71.7%
T2 against the IMM, course disagreement5.9%12.8%22.9%47.9%

At the declared operating point the CUSUM wins (−8.95 pp [−10.16, −7.73]). The learned detector was not carried to a test split. From about 1 false alarm per hour it catches more, as a reported observation only.

The record is experiments/20261105-p5-research/validation.md.