BeeGuardian AI

Stack

Edge AI, end-to-end.

Every layer of BeeGuardian AI — sensor, model, transport, UI — was picked to be fast at the edge, cheap in the field and delightful in the browser.

ESP32 firmware

Arduino IDE, modular functions, deep-sleep aware. Reads BME280, HX711, INMP441 and runs on-device inference.

  • · WiFi + MQTT
  • · Deep sleep < 100 µA
  • · Serial + OTA debug

Edge Impulse TinyML

MFCC signal processing + neural network trained on Normal and Stress bee sounds. Deployed as Arduino library.

  • · Quantized int8
  • · 148 ms inference
  • · Multi-class output

Sensor fusion

Weighted score across audio confidence (35%), temp (25%), humidity (20%), weight (20%). Continuous 0–100.

  • · Ideal 34–36 °C
  • · Ideal 50–65% RH
  • · AI-weighted

MQTT + HiveMQ

Telemetry published to structured topics. Remote commands accepted for reboot, calibration and sleep intervals.

  • · beeguardian/hive-01/state
  • · TLS-ready
  • · QoS 1

Realtime dashboard

React + TanStack Start + Recharts. Streams telemetry, computes trends and surfaces alerts with beautiful visuals.

  • · Live charts
  • · Dark & light
  • · Mobile-first

Solar power system

5 W panel → TP4056 → 18650 → 3.3 V rail. Runs weeks between full-charge cycles thanks to deep sleep.

  • · Solar input meter
  • · Battery gauge
  • · Brownout safe

Edge Impulse workflow

  1. Step 01
    Collect

    Field recordings of normal and stressed hives (queenless, overheated, intruder).

  2. Step 02
    Label

    Two classes: Normal, Stress. Noise added as third class for robustness.

  3. Step 03
    MFCC

    Signal processing block extracts mel-frequency cepstral coefficients.

  4. Step 04
    Train NN

    Small 1D CNN. Data augmentation with gain, noise, time shift.

  5. Step 05
    Quantize

    Post-training int8 quantization for fast ESP32 inference.

  6. Step 06
    Deploy

    Export Arduino library, drop into firmware, call run_classifier().