A LoRaWAN beehive monitor that runs for years on three AAA primary lithium cells.
It weighs the hive, reads its temperature — and, with optional Bluetooth sensors, temperature and humidity inside the hive — and reports over LoRaWAN to The Things Network — then sleeps at about 4 µA. Configuration and calibration happen over Bluetooth from a web page, so there is no mobile app to install and nothing to plug in once the lid is closed.
| MCU | Seeed Studio XIAO nRF54LM20A (Standard, non-Sense) |
| Radio | Wio-SX1262 LoRa kit — LoRaWAN Class A, OTAA, EU868 |
| Sensors | NAU7802 24-bit load-cell ADC · DS18B20 temperature · nPM1300 battery gauge · optional: up to 3 SwitchBot Outdoor Meter BLE temperature/humidity sensors (Extended Mode) |
| RTOS / SDK | Zephyr 4.4 via nRF Connect SDK v3.4.0 |
| Sleep current | ~4 µA (System OFF + GRTC wake) |
| Battery life | ~4.6 years on 3× AAA Li/FeS₂ (1200 mAh) at 15 min / 2 h reporting — ~10 y at 60 min / 4 h. With Extended Mode BLE sensors each uplink costs 139 mC instead of 63 mC: ~3.3 y / ~7 y |
| Uplink | 8 bytes — weight (999 kg range), temperature, battery, flags · +5 bytes per BLE sensor heard (max 23) |
A node in the field: the hive stands on the platform, the IP-rated enclosure bolted to it holds the electronics and the battery pack, and the temperature probe runs inside the hive.
📺 Build workshop walkthrough (YouTube) · 🎬 Bench test clip (download — GitHub does not play repo-hosted video inline)
A TanenBase runs under a hive in Bühl, Baden. It reports over LoRaWAN to a gateway on Windeck Castle, about 4 km away, and sends one reading an hour.
| Radio link | ≈ 4 km to the gateway (3.6 km line of sight) |
| Reporting | once an hour |
| Gateway | Kerlink iStation, Windeck Castle, 378 m above sea level |
| Live data | 📈 beelogger dashboard — Bühl apiary (weight and temperature, updated hourly) |
Left to right: the hive on its platform, the enclosure with the red setup button underneath; the gateway tower on Windeck Castle; the 3.6 km path from the gateway to the apiary.
The firmware is a single-pass state machine, not a loop. Every wake is a cold boot from System OFF: it checks why it woke, does one pass, and powers off again.
wake reason
│
┌───────▼────────┐ button / reset
│ SETUP │ BLE config service, sensors live, 180 s window
└───────┬────────┘
│
┌───────▼────────┐ timer
│ MEASUREMENT │ weight + temp + battery, anomaly check
└───────┬────────┘
│ delta exceeded, or heartbeat due?
┌───────▼────────┐ yes
│ TRANSMISSION │ BLE sensor scan (Extended Mode), SX1262 up, join if needed, uplink
└───────┬────────┘
│
┌───────▼────────┐
│ SLEEP │ System OFF, GRTC timed wake
└────────────────┘
The radio is only powered when there is something to send. Between heartbeats the node measures, compares against the last transmitted values, and goes back to sleep without ever touching the SX1262 — which is where most of the battery life comes from.
Full design notes: docs/ARCHITECTURE.md.
A hive scale's real enemy is not resolution, it is thermal drift. Uncompensated, the reference cell reports a 350 g peak-to-peak swing on a constant 22 kg load over a 17.5 K daily cycle — 16 g of phantom weight per kelvin, enough to bury a real nectar flow.
The firmware corrects it with two parameters: one gain constant, plus a first-order lag filter on the probe temperature that models the probe sitting a few minutes behind the aluminium cell body. Residual scatter drops 6.5×, to σ = 15.7 g. The filtered temperature is carried in ZMS across System OFF, and the tare stores the temperature it ran at as the correction's reference point.
Same 22 kg load, 40.7 h, one daily temperature cycle. σ 102.2 g → 15.7 g, peak-to-peak 350 g → 103 g.
Validated in the field, not just on the bench: six days outdoors with a 22.2 kg dead weight and the correction live left +1.10 g/K of temperature sensitivity where the uncompensated cell had −16.6, and σ of 11.6 g against 63.1 g — measured on the same samples, by reconstructing the filter offline and inverting the correction.
Days 4–6 are rain, not drift: +220 g arrived in two hours while the filtered temperature moved 0.4 K, and 506 g of water was still in the wooden platform and the concrete block at the end. Once the drying trend is modelled, the residual thermal term across those days is +0.56 and −1.48 g/K — the correction holds underneath the water.
The cell the shipped coefficient was fitted on — a Bosche H40A-C3-0150 single-point cell bolted between two steel spreader plates. Worth staring at: the report concludes the drift is not coming from the cell but from this mount, which is 8× outside the OIML R60 envelope the cell is certified to.
k is a property of the cell plus its mount, not of the cell alone —
re-characterise after any mechanical change. Which cell is fitted is therefore a
runtime setting, not a build flag: the setup page offers a generic cell (no
correction), the measured H40A-C3-0150, four named-but-uncharacterised cells
(Steinberg SBS-PF-150, Zemic L6E, TAL220, Flintec PC/SB — they run uncorrected
but the unit records which is fitted), or your own measured gain and lag, and
the correction picks its constants up from there.
Profile table and how to characterise a cell:
docs/load_cells/. Method, data, hold-out validation and
the (unflattering) hardware conclusions for the reference cell:
docs/load_cells/H40A-C3-0150/.
The scale tells you that a colony gained or lost weight. It cannot tell you whether the colony is raising brood, holding its winter cluster, or sitting in damp air. Extended Mode adds up to three off-the-shelf SwitchBot Outdoor Meter Bluetooth sensors — at most two inside the hive and one outside — and sends their temperature and humidity in the same uplink as the weight. No cable through the hive wall and no extra gateway: the node's own nRF54LM20A radio listens for them.
Left: a sensor with its Tanen label — the Bluetooth address as text and as a QR code. Right: sensors live on the setup page, two in the hive and one outside.
- Set up from the phone. In the web page's Erweitert tab, scan the label's QR code (or type the address), choose inside the hive or outside, save. Readings appear within seconds, on that tab and on the overview.
- Straight into BEEP and beelogger. The outside sensor becomes
t/TempOut(the DS18B20 moves to BEEP'st_0); the in-hive sensors map tot_i/h_i,TempIn/FeuchteInandTempIn2/FeuchteIn2. Nothing to change on either platform. - Only when it counts. The node scans for the sensors only on wakes that transmit and stops as soon as every sensor has been heard. Each sensor adds 5 bytes to the uplink; one that is not heard costs no airtime and sets a flag instead.
- What it costs. A transmit wake with the scan measures 139 mC instead of
63 mC, so the pack lasts about 3.3 years instead of 4.6 at 15 min / 2 h
(
docs/POWER_BUDGET.md).
Wire format and GATT details:
docs/ARCHITECTURE.md.
| Path | What it is |
|---|---|
src/ |
Firmware — core/ (FSM, power, watchdog), drivers/, features/, config/ |
boards/ |
Application devicetree overlay for the target board |
third_party/ |
Vendored Seeed XIAO nRF54LM20A board definition (Apache-2.0) |
Tanen_Base_pcb/ |
KiCad carrier board — schematic, layout, gerbers |
web/ |
Web Bluetooth configuration page + standalone downlink encoder (no build step) |
ttndecoder/ |
TTN payload formatters — one unified BEEP + beelogger decoder, plus the older per-platform ones |
test_apps/measure_loop/ |
Standalone sensor bring-up app |
docs/ |
Architecture, power budget, sprint log, TRD, open questions, German user manual (BENUTZERHANDBUCH.md + built HTML) |
docs/load_cells/ |
Load-cell characterisation — data, analysis, host replay of the compensation |
media/ |
Build photos and a bench-test clip |
Requires nRF Connect SDK v3.4.0 and its bundled toolchain. The board
definition is out-of-tree and vendored, so no west update of extra manifests is
needed.
git clone https://github.com/hmd83/tanen.git
cd tanen
# LoRaWAN credentials — never commit the real one
Copy-Item prj_credentials.conf.example prj_credentials.conf
# then edit it and set your AppKey
.\build.ps1 debug # logs on uart20 -> on-board CMSIS-DAP VCOM @115200
.\build.ps1 prod # LOG off, size-optimized
.\flash_update.ps1 # flash, preserving stored config + calibrationbuild.ps1 finds the SDK automatically (newest vX.Y.Z under C:\ncs). Override
with $env:NCS_ROOT, $env:NCS_VERSION, or $env:NCS_TOOLCHAIN if your install
lives elsewhere. On a mode change it forces --pristine for you, because
switching in place trips a stale-CMakeCache Kconfig failure.
Raw west equivalent, if you prefer:
west build -b xiao_nrf54lm20a/nrf54lm20a/cpuapp --no-sysbuild -- \
-DBOARD_ROOT=$PWD/third_party/seeed-xiao-nrf54lm20a \
-DOVERLAY_CONFIG="prj_credentials.conf;prj_production.conf"Gotchas worth knowing before your first build — see
docs/PLAN.md Sprint 15 for the full list:
- Keep the build directory path short on Windows (MAX_PATH kills ninja).
- Do not add
-S cdc-acm-console. It repoints the console at the nRF's own USBHS CDC, which never enumerates on this board — you get silence. - pyocd cannot program the nRF54LM20A: its erase works, then its flash algorithm
hard-faults and leaves the chip blank. Use OpenOCD (what
flash_*.ps1do) orwest flash.
- Flash, then press the user button to enter SETUP. The node advertises as
TanenBasefor 180 seconds. - Open
web/index.htmlin Chrome or Edge (Web Bluetooth; Safari and Firefox do not support it) and connect. - Write the DevEUI, JoinEUI, and AppKey from your TTN device registration.
- Tare the empty scale, then calibrate with a known reference weight. The tare also records the temperature it ran at — that becomes the reference point of the thermal correction, so tare at a temperature the hive actually sees.
- Set the measurement and transmit intervals, and the anomaly thresholds.
- Press Done. The node joins TTN and enters its normal cycle.
Everything above is also settable later by LoRaWAN downlink (FPort 10–13) and
persists in ZMS across reboots and firmware updates.
web/encoder.html builds those downlinks — pick the
parameter, type the value in human units, and copy the fPort plus hex/base64 for
the TTN console, or the whole down/push body for the TTN API. It is a separate
page with no Bluetooth at all, so it needs neither the node nor a pairing.
Each physical unit needs its own DevEUI. Leave CONFIG_TANENBASE_DEV_EUI
empty to derive a unique one from the SoC's factory hardware ID.
Register the device as LoRaWAN MAC v1.0.3, OTAA, EU868, then install
ttndecoder/tanen-decoder.js as the uplink
payload formatter (Payload formatters → Uplink → Custom Javascript formatter).
One formatter serves both supported platforms: it parses the frame once (8 bytes,
plus 5 per Bluetooth sensor in Extended Mode) and emits both key sets into the
same decoded_payload.
| Reading | BEEP key | beelogger key | Unit |
|---|---|---|---|
| Weight | weight_kg |
Gewicht |
kg |
| Temperature | t |
TempOut |
°C |
| Battery | bv |
VBatt |
V |
| Outside temp / humidity (BLE) | t / h |
TempOut / FeuchteOut |
°C / %RH |
| Inside #1 temp / humidity (BLE) | t_i / h_i |
TempIn / FeuchteIn |
°C / %RH |
| Inside #2 temp / humidity (BLE) | t_1 / — |
TempIn2 / FeuchteIn2 |
°C / %RH |
Extended Mode adds up to three SwitchBot Outdoor Meters (at most two inside
the hive, one outside), set up in the web page's Erweitert tab by typing the
sensor's MAC or scanning a QR label. The node scans for them only on wakes that
transmit. An outside sensor takes over t / TempOut and the DS18B20 moves to
BEEP's t_0; without one, t / TempOut stay the DS18B20. Wire format in
docs/ARCHITECTURE.md.
Each platform stores the keys it knows and ignores the rest, so no per-platform
formatter juggling is needed. A failed sensor decodes to null on both sides,
never to a sentinel number. flags, anomaly_weight, anomaly_temp and
heartbeat ride along for TTN live-data debugging and are ignored by both
servers. The older single-platform formatters
(beepdecoder.js,
tanen-beelogger-decoder.js) are kept
for reference.
BEEP (beep.nl) — the webhook base URL must carry the device key explicitly:
https://api.beep.nl/api/lora_sensors?key=<DevEUI>
Without it BEEP decodes your payload perfectly and then silently drops it,
because it falls back to a TTN Stack V2 field (payload_fields.hardware_serial)
that V3 webhooks do not send. This one cost real debugging time — see
docs/PLAN.md Sprint 14.
beelogger (beelogger.de) — the community server
takes the Gewicht / TempOut / VBatt names its own payload formatter uses,
which is exactly what the decoder emits. Webhook base URL:
https://community.beelogger.de/<username>/{/devID}/beelogger_log.php?Passwort=<password>&LORA=1
<username>— your community-server account{/devID}— filled in by TTN per uplink, so name the TTN devicesbeelogger1,beelogger2, … to match the stations registered on the server<password>— the same one for every station, which is what lets a single webhook serve the whole TTN application instead of one per nodeLORA=1— tells the server the record arrives via LoRaWAN
Both platforms are confirmed receiving from the unified decoder.
The carrier board is a passive breakout: XIAO socket, Wio-SX1262 socket, screw
terminals for the load cell and the DS18B20, battery input, and the bulk
capacitor. KiCad sources and fabricated gerbers are in
Tanen_Base_pcb/.
Assembled in an IP-rated enclosure — XIAO nRF54LM20A top-left, NAU7802 load-cell amplifier on the right, external 868 MHz antenna, and cable glands for the load cell and temperature probe.
The weighing platform end-on: a single-point cell carries the whole top frame from the centre, so the hive can sit anywhere on it without changing the reading. Enclosure sits on the top deck, cable runs down to the cell.
Underside — the cell between the top plate and the base, load arrow pointing down. Emax = 150 kg, which is oversized for a hive: a 100 kg cell would give ~1.5× better signal-to-noise at the same absolute drift.
Optional Extended Mode sensors: up to three SwitchBot Outdoor Meters (model
WoIOSensorTH), each on its own battery. Nothing to wire — print a label with the
sensor's Bluetooth address as plain text and as a QR code (for example
EB:6B:01:C6:49:95) and stick it on; the setup page reads either.
The KiCad project files are still named XIAO_nRF54L15 V1, while the board itself is silkscreened Tanen Base v0.1. That is deliberate: all XIAO modules share one 21 × 17.5 mm footprint and pinout, so the nRF54L15 and nRF54LM20A are drop-in replacements on this carrier — only the GPIO number behind each header pin changes, and that is handled in the devicetree overlay.
Tanen_Base_pcb/README.mdhas the full pin-by-pin mapping.
Battery: 3× BEVIGOR AAA 1.5 V lithium (Li/FeS₂, 1200 mAh) in series —
4.5 V nominal, 5.2 V measured fresh at room temperature — primary /
non-rechargeable. The firmware
deliberately deletes charging-enable from the nPM1300 charger node; with it
armed, applying USB power would push charge current into non-rechargeable cells.
Do not re-enable it unless you have also changed the chemistry.
docs/POWER_BUDGET.md §5.2 for the chemistry data,
mitigations, the ER14505 AA Li-SOCl₂ cell this replaced, and what the swap bought
(no passivation, −40 °C rating, large pulse margin) and cost (9.3 y → 4.6 y).
Measure cells before assembly and reject any above ~1.80 V; keep the enclosure out of direct sun.
Running in the field — see the Bühl apiary. The core loop — measure, delta-detect, uplink, sleep — is verified end to end on hardware, including ingestion by both BEEP and the beelogger community server from the same uplink. Extended Mode is verified on hardware too (2026-09-13): three sensors live in the setup page, their blocks in the uplink, and the transmit wake measured on the PPK2.
Not yet done: MCUboot/OTA is wired up (sysbuild.conf, pm_static.yml) but
unverified on this module, and no production signing key is provisioned.
Open design questions are tracked in
docs/OPEN_QUESTIONS.md.
See CONTRIBUTING.md. Bug reports and hardware notes from anyone running this on real hives are especially welcome.
Hussein Daj — project owner, product design, hardware design (carrier PCB, sensor selection, power architecture), firmware direction, and all hardware bring-up, PPK2 power measurement, and field validation.
Claude Code (Anthropic) — firmware co-development. Worked alongside the author on the state machine, drivers, BLE GATT service, power-management path, watchdog and fault-tolerance layer, the nRF54L15 → nRF54LM20A port, the load-cell temperature-drift analysis and the compensation formula it produced, and this documentation. Every change was reviewed and hardware-verified by the author before it landed.
Google Gemini — product design input: requirements shaping, use-case analysis, and design-decision review during the concept and specification phase.
- Zephyr RTOS and the nRF Connect SDK — Apache-2.0
- Seeed Studio — XIAO
nRF54LM20A Zephyr board definition, vendored in
third_party/— Apache-2.0 - loramac-node — LoRaWAN MAC
- BEEP — open beehive monitoring platform and data API
- beelogger — open hive-scale project and community server; its field names are mirrored by the uplink decoder
Apache-2.0 — see LICENSE and NOTICE. This covers the firmware, the hardware design files, and the documentation.