Collar telemetry

Source: src/sim/herd.ts (generation), src/sim/types.ts (schema), src/sim/features.ts (feature derivation). Every cow emits the data stream a real collar reports — and the detection layer is only allowed to see this stream, never the simulation's internal state.

The pipeline at a glance

Everything in the "on the collar" box is derived on-device; everything to the right is what this demo implements. The collar reports derived telemetry — a compact 5-minute summary — not raw sensor data, which is what keeps the data volume tiny.

The telemetry sample

Every 5 sim-minutes each collar records a sample:

Field	Type	What it is	How a real collar gets it
t	timestamp	Sample time	RTC / GPS time
x, y	metres	Position	GNSS fix (duty-cycled to save power)
speed	m/s	5-minute average speed	Successive GNSS fixes + accelerometer
behaviour	enum	grazing / walking / resting / ruminating	On-collar activity classifier over IMU data — head position and jaw/ear movement patterns
temperature	°C	Body temperature	Contact thermistor (or ear-tag partner device)
rumination	0–1	Fraction of recent time ruminating	Accumulated from the activity classifier

As JSON (the demo's in-memory form — a real feed would pack this far smaller):

{
  "t": 1834,
  "x": 287.4, "y": 141.2,
  "speed": 0.11,
  "behaviour": "grazing",
  "temperature": 38.7,
  "rumination": 0.31
}

Compact by design

Packed as binary, this is ~20 bytes per sample: position quantised to the paddock (2 × 2 bytes), speed (1), behaviour class (2 bits), temperature offset (1), rumination (1), timestamp delta (2), plus header. At one sample per 5 minutes that is ~6 KB per cow per day, because the collar sends derived telemetry, not raw sensor data. The heavy lifting — activity classification, fence logic — runs on-collar; only the compact summary and alerts leave the device.

How each signal behaves in the sim

• Speed — 5-minute EWMA of true movement. Grazing ≈ 0.12 m/s, walking ≈ 1 m/s, resting/ruminating ≈ 0. Each cow has a persistent ±15% pace factor, so healthy cows genuinely differ — the detector has to cope with that.
• Behaviour — the state machine's current state, i.e. a perfect activity classifier. (A real classifier is ~90–95% accurate; adding label noise here would be a straightforward extension.)
• Temperature — relaxes towards 38.6 °C + a ±0.25 °C diurnal cycle + condition offsets (illness adds +1.4 °C) + a herd-wide heat-stress term (weather). Fevers develop over ~1.5 h, as in practice.
• Rumination — EWMA (~2 h) of time-in-rumination, shown as min/h in the UI. A healthy cow runs ~20 min/h averaged over the day. Illness collapses the underlying behaviour, so the reported rate decays over an hour or two — a leading indicator, and exactly the signal commercial rumination collars are sold on.

The ring buffer

Each cow keeps 24 hours of samples (288). Everything downstream reads this buffer:

• the sidebar's 24-hour sparklines (cow vs. herd mean),
• the detector's 2-hour feature window (recent behaviour),
• the detector's own-baseline window (2–24 h ago) — used to compute self-relative deltas, which separate "naturally slow cow" from "cow that just went slow" (see detection features).

Why synthetic telemetry is the right demo choice

The generator and the detector are strictly separated: conditions are injected into the behaviour model, and the detector must recover them from the telemetry alone. That makes the demo falsifiable — inject a condition, watch whether and how fast the detector finds it — which is a far stronger portfolio claim than a dashboard drawn over canned data. The same separation is what made training a real model possible: the simulator generates unlimited labelled telemetry.