91
consecutive days
Field Data · A Data Story
What a monitor actually sees when it reads four colonies every five minutes — and never blinks for three months.
Most of what happens inside a beehive happens in the dark, between inspections. A beekeeper opens a hive every week or two, reads the frames for a few minutes, closes it, and infers the other thirteen days. It works — beekeeping has run on it for centuries — but it means the moment a colony starts to slip usually isn't the moment anyone is looking.
From April 20 to July 21, 2026, we did the opposite. Four honey bee colonies at a northern New Mexico apiary each carried an ApisNode monitor that woke every five minutes to measure temperature, humidity, pressure, tilt, a 10-band acoustic spectrum, a 6-band vibration spectrum, and a 10-step gas fingerprint — 30 metrics in all — and relayed them over LoRa radio to the cloud. Every morning, an AI analysis fused the last 24 hours against peer-reviewed bee science and wrote a plain-language health report for each colony. Every one of those 91 reports is public, and every chart below is drawn from the raw telemetry behind them. This is what three months of never looking away actually showed.
Start high and wide. The chart below is the apiary's daily average health score — a 0–100 summary the AI assigns each colony, here averaged across the active hives. You can read the season in it: a jittery, uneven spring where colonies were still building; a strong, settled early summer; and two sharp notches where something went wrong fast.
Zoom all the way in. Here is a single colony — "Iris," the apiary's steadiest — across one ordinary day, June 10, at full five-minute resolution. Every sensor family the monitor carries, on one screen. None of these lines is dramatic. That's the point: this is what healthy looks like, and you can only recognize the abnormal against a dense, boring baseline like this.
One colony on one day is a detail. The value of monitoring shows up when you pull back and put the whole apiary — plus the world outside it — on a single multi-day timeline. That is what our wall display does, and it is where the patterns jump out. Here is a real week, June 8–14, rebuilt from the raw data.
The wall display also renders each colony's spectrograms — sound, vibration, and the gas heater-scan — as scrolling heatmaps. The gas view is the most alien and the most revealing: every reading, the sensor sweeps its heater through ten temperatures (100–400°C) and records the air's response at each, building a chemical fingerprint. Stack those fingerprints across days and a rhythm appears in the air itself.
Sound and vibration get the same treatment. The acoustic spectrogram splits the microphone signal into ten frequency bands (100 Hz to 3.6 kHz); the vibration spectrogram splits the accelerometer into six bands (5–200 Hz). Read top-to-bottom they're a fingerprint of what the colony is doing — and read left-to-right, the daily pulse of activity is visible.
One day is an anecdote. Average every day of a stable three-week stretch by the hour it was recorded, and a rhythm sharpens out of the noise. A honey bee colony has a daily cycle — brood-tending, foraging, and fanning all ebb and flow with the sun — and the brood-nest temperature carries its signature: coolest before dawn, warmest in the early afternoon.
The metrics aren't independent, either. When you plot brood-nest temperature against in-hive humidity for all 5,549 paired readings that June, they fall along a clear downward slope: as the nest warms, it dries — warmer air holds more moisture as vapor, so relative humidity falls. It's a textbook physical relationship, and the sensors recover it cleanly from real hive air.
On May 17, one colony — Hive #3 — fell off a cliff. Its health score dropped from the 70s into the 20s and 30s and stayed there for a week, the brood-nest temperature sliding while its gas readings climbed. On paper, that is the profile of a dying colony. A weekly inspection catching it mid-slide might well have read it as terminal.
But the monitor wasn't reading one number — it was reading thirty, every five minutes. And one of them refused to fit the collapse story: the acoustic spectrum stayed uniformly stable across all ten frequency bands. A colony that is actually dying goes quiet unevenly as its population thins. This colony's sound signature was intact. On May 24 the analysis reversed the call — what looked like collapse was the brood cluster migrating upward through the hive, a normal spring reorganization, with the population fully intact. By May 25 the whole apiary was back in healthy territory.
The temperature said dying. The sound said thriving. Only watching both, continuously, told the difference.
The second notch, in late June, hit differently: all four colonies dipped at once. When every hive in an apiary moves together, the cause usually isn't inside any one of them — and this time the monitors could point straight at it. On the morning of June 24, the McCauley Springs Fire ignited in the Jemez Ranger District, less than 25 miles upwind. It grew from 30 acres to more than 700, forcing evacuations near Jemez Springs, and for four days it filled the airshed over the apiary with smoke.
The gateway that relays the hives' data also carries its own air-quality sensor, pointed at the outdoor air. It watched the airshed collapse: outdoor gas resistance — high when the air is clean — fell more than tenfold as the fire grew, bottoming out in the small hours of June 27, exactly as the fire reached its peak size. And in lockstep, every colony's brood temperature sagged together, the apiary average sliding from a healthy 34°C to just over 30°C on the same night, before all four recovered as the smoke cleared and the fire came under control.
The colonies were stressed but adapting; within days most had recovered, and no hive-side intervention would have helped. Knowing that — that the threat was in the sky, not the frames — is its own kind of useful. There was also an honest complication running underneath: for part of June the microphones across all four hives showed suppressed acoustics that the AI correctly flagged as a likely sensor artifact — propolis or debris on the mic — rather than four simultaneous colony failures. Distinguishing a fouled sensor from a real signal is exactly the judgment continuous data makes possible; a single reading can't.
Two of the three scares looked, in isolation, like colony death. Continuous multi-sensor monitoring didn't just raise alarms — it lowered two of them correctly, and named the real cause of the third as a wildfire 25 miles away. Fewer needless inspections; the right worry at the right time.
None of these signals mean anything unless they match what bee scientists have already measured. Every daily report is written by an AI that is grounded in a curated library of peer-reviewed apidology — a retrieval-augmented (RAG) knowledge base of findings on thermoregulation, hive acoustics, vibration, and colony volatiles, injected into the analysis so conclusions trace back to published evidence rather than a language model's guesswork.
The strongest validation of a monitoring system isn't a novel claim — it's reproducing established results. Across 91 days, our raw telemetry independently recovered findings from the literature the analysis is built on. A few of the clearest:
| Published finding | What ApisNode measured |
|---|---|
| Brood nest is held at 34–36 °C within ±0.5 °C, independent of outside weather.Jones et al. 2004 · Human et al. 2006 | Apiary brood temperature averaged 33.7 °C (±0.65 °C) across 91 days while outdoor temperature swung from 5 °C to 36 °C — the tight regulation the literature describes, seen in the whole-apiary chart above. |
| Healthy colonies show a small (<3 °C) diurnal brood-temperature cycle, tightly clamped and peaking in early afternoon.van Dooremalen et al. 2024 (B-GOOD) · Stalidzans et al. 2017 | Hour-of-day averaging recovered a clean ~1.1 °C cycle, coolest before dawn and peaking ~1 PM — well under the 3 °C healthy bound, exactly the shape and phase reported. |
| Flight and fanning activity track light and temperature, peaking in early afternoon and dropping at night.Stalidzans et al. 2017 · Hrncir et al. 2011 | In-hive vibration energy across all six bands — including the 80–120 Hz wing-fanning band — rose through the day and peaked mid-afternoon, falling toward night, matching the published activity rhythm. |
| As the brood nest warms, relative humidity falls — an inverse microclimate coupling.Human et al. 2006 · Ellis & Delaplane 2008 | Across 5,549 paired readings, temperature and humidity correlated at r = −0.44 — the inverse relationship, recovered directly from hive air (see the correlation plot above). |
| Environmental VOC events (e.g. wildfire smoke) shift MOX gas-sensor baselines; the hive acts as an air sampler for its surroundings, and an outdoor reference distinguishes environmental from colony causes.Frizzera et al. 2020 · Cepero et al. 2023 | During the McCauley Springs Fire, the outdoor gas sensor fell >10× and every colony's brood temperature dipped together — an environmental event, correctly separated from colony failure by the outdoor reference, precisely as the method prescribes. |
| Fusing multiple sensors cuts false-positive anomalies by ~30% versus single-sensor thresholds.B-GOOD Consortium 2024 · Kulyukin et al. 2021 | Twice a single metric (falling brood temperature) read as colony death; each time a second modality (stable acoustic structure) overruled it. Fusion turned two would-be false alarms into correct calls. |
Citations are drawn from the same reference library the analysis pipeline uses to ground each report. Reproducing them from an independent apiary is evidence the sensors — and the science reading them — are measuring what they claim to.
This is one apiary, four colonies, one season, at altitude in northern New Mexico — not a clinical trial. We're not claiming a monitor replaces a beekeeper's eyes on the frames; it doesn't. But three months of five-minute data made a few things concrete:
Continuous beats periodic for catching the fast stuff. Both temperature crises developed and resolved inside a single week — the interval between many inspections. More sensors beat one sensor. Every time a single metric told a scary story, a second metric was needed to know whether to believe it. And context beats the hive alone — the outdoor air sensor is what turned "four failing colonies" into "a wildfire upwind," a call no in-hive reading could make by itself.
Ninety-one days, 379 analyses, three scares, zero colonies lost. The point was never the alarms. It was knowing which ones to trust.
Want to read the primary source? Every daily report behind this story — all 91 of them, per colony, in full — is public in the ApisNode report archive. This is the same analysis our monitored beekeepers wake up to each morning.
ApisNode is giving free monitor kits to selected US beekeepers through The Open Apiary Project, in exchange for contributing anonymized data to pollinator research. Your colony gets its own 90-day story.