Цель
Production phase переводит Entrance Observer из field prototype в повторяемый Gratheon kit. Финальный product должен выбираться по measured model, energy, video and field data, а не по удобству Jetson developer kit.
Production product must be:
- accurate enough for beekeeper decisions;
- power-aware enough for target installation mode;
- weatherproof and serviceable;
- secure and remotely supportable;
- manufacturable from stable suppliers;
- integrated with Gratheon telemetry, video and device-management flows.
Recommended product strategy
Не стоит заставлять один hardware SKU решать все пасеки. Entrance video имеет разные ограничения в urban powered apiary и remote solar apiary.
| SKU |
Target site |
Video behavior |
Connectivity |
Likely compute direction |
Product promise |
| Urban WiFi/video |
Apiary with mains/PoE and good internet |
On-demand live plus selected clips |
Ethernet, PoE or WiFi |
Jetson production SOM, Raspberry Pi + Hailo or RK3588 |
Best live inspection and model QA. |
| Field telemetry-first |
Remote apiary with limited power/data |
Movement telemetry by default, rare low-res clips |
LTE gateway, local gateway or store-and-forward |
Pi + Hailo, RK3588 or smart camera only if energy fits |
Bee traffic trends with low bandwidth. |
| Research/dev kit |
Gratheon team and advanced contributors |
Flexible recording and overlays |
Lab network |
Jetson Orin Nano dev kit |
Dataset and model development, not customer default. |
Production architecture
flowchart LR
subgraph CameraHead[Camera head]
lens[Fixed locked lens]
sensor[Camera sensor]
window[Optical window and hood]
bracket[Repeatable entrance bracket]
end
subgraph Device[Entrance Observer device]
compute[Production compute module]
accel[AI accelerator or GPU/NPU]
encoder[Video encoder path]
storage[Industrial local storage]
power[Power and protection]
watchdog[Watchdog and health telemetry]
connector[Sealed connectors]
end
subgraph Gateway[Optional apiary gateway]
localradio[Local radio or Ethernet/WiFi]
backhaul[LTE/Ethernet backhaul]
end
subgraph Cloud[Gratheon cloud]
telemetry[telemetry-api]
video[gate-video-stream]
graphql[graphql-router]
deviceMgmt[Device registry and updates]
end
lens --> sensor
sensor --> compute
window --> sensor
bracket --> sensor
compute --> accel
compute --> encoder
compute --> storage
power --> compute
watchdog --> telemetry
connector --> compute
compute -->|movement + health| telemetry
compute -->|live/session media| video
compute -->|optional local network| Gateway
Gateway --> telemetry
Gateway --> video
graphql --> telemetry
graphql --> video
graphql --> deviceMgmt
Production component recommendations
| Subsystem |
Recommended direction |
Why |
Defer or avoid |
| Compute |
Benchmark Jetson reference against Raspberry Pi 5 + Hailo AI HAT+ 26 TOPS and one RK3588 board. Choose by count accuracy per watt and serviceability. |
TOPS alone is not product metric; device must run camera, AI, tracking, encode, upload and watchdog together. |
Do not select production board only because it is cheaper or has higher advertised TOPS. |
| AI runtime |
Keep model export paths for TensorRT, HailoRT and one secondary NPU path. |
Avoid locking dataset/model work to one vendor too early. |
Avoid vendor-only smart-camera AI until model quality stable. |
| Camera |
Move from generic USB camera to locked, documented camera head. |
Production install needs repeatable FOV, focus, exposure and cable retention. |
Avoid manual varifocal lenses on customer units unless physically locked/documented. |
| Sensor shutter |
Prefer camera with low motion blur; test global shutter if rolling shutter causes count errors. |
Bees move fast near entrance. |
Do not overpay for global shutter until side-by-side clips show benefit. |
| Video encode |
Prefer platform/camera pipeline with efficient H.264/H.265 encode. |
Live sessions and clips should not starve AI or burn solar energy. |
Avoid software H.265 on low-power CPUs for field SKU. |
| Storage |
Use industrial-rated storage sized to retention policy. |
Customer device should survive power cycles/temperature. |
Avoid storing continuous video by default. |
| Enclosure |
UV-resistant IP65/IP67 enclosure or camera-head + electronics-box assembly. |
Weather failures become support failures. |
Avoid acrylic hobby cases and ad-hoc frames for customer kits. |
| Connectors |
Sealed connectors or rated glands with documented pinouts. |
Cameras, power, antennas and optional sensors must be serviceable. |
Avoid soldered field cables that cannot be replaced. |
| Power |
Split powered/PoE SKU from solar SKU. |
Product promise and BOM differ. |
Do not claim solar autonomy without Wh/day and sunless-day validation. |
| Connectivity |
Ethernet/PoE or WiFi for powered SKU; gateway/LTE for remote SKU. |
Remote apiaries need deliberate data/power model. |
Avoid cellular in every unit until costs and power proven. |
Compute selection matrix
| Candidate |
Strength |
Weakness |
Best production fit |
Decision gate |
| Jetson Orin Nano production module/carrier |
Best ML ecosystem and TensorRT path. |
Higher cost/power; Orin Nano lacks NVENC. |
Premium/dev or powered video SKU. |
Beats alternatives on accuracy or velocity enough to justify watts/cost. |
| Raspberry Pi 5 + Hailo AI HAT+ 26 TOPS |
Good cost/power candidate with strong community and camera ecosystem. |
Requires Hailo model conversion and tracker CPU benchmark. |
Near-term production candidate for WiFi/PoE SKU. |
Same count quality as Jetson with at least 2x better FPS/W or lower total cost. |
| Raspberry Pi 5 + Hailo AI HAT+ 13 TOPS |
Lower cost than 26 TOPS. |
Less headroom for tracker/future models. |
Cost-down after 26 TOPS proven. |
Meets target FPS/accuracy with thermal margin. |
| RK3588 board |
Integrated NPU and often strong media codecs. |
Vendor support, RKNN tooling and OS maintenance vary. |
Alternative if video encode is central or Hailo conversion fails. |
Stable full pipeline with lower energy/cost than Jetson. |
| Coral TPU |
Very low power for supported models. |
Current YOLO-style detector may need redesign. |
Telemetry-only or simplified detector experiments. |
Simplified model reaches count target. |
| Smart camera / vision module |
Integrated sensor, encoder and sometimes AI. |
Vendor lock-in, less control, supply risk. |
Long-term solar field SKU research. |
Vendor pipeline exposes frames/metadata/control needed by Gratheon. |
Camera-head production requirements
| Requirement |
Implementation rule |
| Repeatable FOV |
Define entrance width coverage, working distance, angle and target pixel density. |
| Locked optics |
Use fixed-focus or physically locked lens settings. |
| Lighting robustness |
Test dawn, noon, shade, clouds, rain and backlight. Add hood or constrained exposure profile. |
| Clean optical path |
Window must be cleanable, tilted/hooded and resistant to scratches/condensation. |
| Service replacement |
Camera cable/head should be replaceable without rebuilding whole device. |
| Model metadata |
Store camera model, lens, FOV profile and calibration image in device metadata. |
Power and solar decision rules
Production must define installation mode before sizing power.
| Installation mode |
Power design |
Acceptance rule |
| PoE powered |
PoE splitter or board with Ethernet data and regulated power. |
Runs full daytime profile and live sessions without brownouts. |
| Mains powered |
Outdoor-safe supply and protected low-voltage DC into enclosure. |
Safe installation, no exposed mains in hobby enclosure. |
| Solar telemetry-first |
Battery from measured Wh/day, panel for worst season, aggressive sleep and low-video policy. |
Survives target sunless days and reports low battery before failure. |
| Gateway remote |
Hive nodes/device connect to one powered gateway with LTE/Ethernet. |
Per-hive unit does not need modem unless economics support it. |
Production telemetry and support checklist
| Field |
Required |
Why |
beesIn, beesOut, unknownDirection, netFlow |
Yes |
Core product value. |
confidence, countIntervalSeconds, modelVersion |
Yes |
Explains data quality and model changes. |
fps, droppedFrames, cameraOnline |
Yes |
Detects camera/processing failures. |
deviceTemperature, uptimeSeconds, resetReason |
Yes |
Support and reliability diagnostics. |
diskFreeBytes, queuedTelemetryCount, queuedClipCount |
Yes |
Prevents silent storage/upload failures. |
rssi, uploadLatencyMs, networkType |
Yes |
Explains missing data and weak connectivity. |
inputVoltage, batteryVoltage, batteryPercent |
Required for battery/solar SKUs |
Low-power alerting and warranty support. |
hardwareRevision, cameraProfileId, firmwareVersion |
Yes |
Links field behavior to exact build. |
Production quality gates
- Two or more identical units produce comparable counts on same labelled clip set.
- Outdoor enclosure passes rain, UV, condensation, heat and cable-pull tests appropriate to claimed rating.
- Live session starts, stops, expires and recovers without manual SSH.
- Device can be paired to Gratheon hive/box through web app without manual DB edits.
- Firmware update and rollback plan exists before customer sale.
- Critical parts have at least two suppliers or approved substitute.
- Production test fixture validates camera, AI, network, storage, power, LEDs/buttons and cloud registration.
Exit criteria
- Production candidate selected using measured count accuracy, FPS/W, bandwidth, thermal stability, cost and support complexity.
- Camera head geometry frozen and documented.
- Power SKU boundaries explicit: powered/PoE, field telemetry-first and research/dev kit.
- BOM includes enclosure, connectors, labels, packaging, service tools and QA fixtures, not only compute/camera.
- Gratheon support can identify device state remotely without asking beekeeper to SSH into unit.
Bill of materials
Подробный список находится в Phase 3 - Production BOM. Он включает production compute candidates, camera-head choices, enclosure/connectors, power variants, networking, factory test tools and sourcing rules.
