Skip to content

README.md

Latest commit

 

History

History
133 lines (105 loc) · 3.4 KB

README.md

File metadata and controls

133 lines (105 loc) · 3.4 KB
Raw

SilentScout Firmware

This is the ESP32-S3 firmware for SilentScout. It captures audio from an INMP441 mic, runs Edge Impulse inference to detect chainsaw/mining sounds, and sends LoRa alerts when a threat is confirmed.

Everything runs offline - no WiFi, no cloud. Just the mic, the ML model, and LoRa.

What you need

Part Spec
MCU ESP32-S3 DevKitC-1 N16R8 (16MB Flash, 8MB PSRAM)
Mic INMP441 I2S MEMS Microphone
Radio Ra-02 SX1278 433MHz LoRa module
Power Solar panel + 18650 battery on 3.3V rail

Wiring

INMP441 → ESP32-S3

INMP441 GPIO
WS 1
SD 2
SCK 42
L/R GND
VDD 3.3V
GND GND

SX1278 → ESP32-S3

SX1278 GPIO
SCK 18
MISO 19
MOSI 23
NSS 5
RST 14
DIO0 26
VCC 3.3V
GND GND

Building

You need PlatformIO (CLI or the VS Code extension).

Setting up the Edge Impulse model

  1. Go to Edge Impulse, train your model, and export as Arduino library
  2. Drop the exported folder into firmware/lib/: 
    firmware/lib/project-1_inferencing/
  3. Make sure the model is trained on mic input with these classes: chainsaw, mining, ambient

Flash it

cd firmware
pio run                        # just build
pio run --target upload        # build + flash
pio device monitor --baud 115200  # serial monitor

Config

Everything's in src/config.h - change it before flashing:

What Default Notes
NODE_ID "N1" Change this per node
NODE_LAT / NODE_LNG 6.9270 / 79.8610 Hardcoded for now (no GPS)
CONFIDENCE_THRESHOLD 85% Min confidence to count a detection
CONSECUTIVE_REQUIRED 3 Has to detect same thing 3x in a row
ALERT_COOLDOWN_MS 10s Wait time between LoRa sends
HEARTBEAT_INTERVAL_MS 20s How often to send "I'm alive" packet
LORA_FREQUENCY 433 MHz ISM band
LORA_TX_POWER 17 dBm Can go up to 20 but uses more battery

LoRa packet format

We use a simple CSV string over LoRa. Keeps it compact.

Alert:

N1,alert,chainsaw,0.95,3,123456789,6.9270,79.8610

Heartbeat:

N1,data,ambient,0.00,0,123456789,6.9270,79.8610

Fields: nodeId, type, class, confidence, consecutiveCount, timestamp_ms, lat, lng

The receiver node will parse this and convert to JSON for the desktop app.

How the detection works

  INMP441 mic
      │
      ▼
  I2S capture (16kHz 16-bit)
      │
      ▼
  Edge Impulse inference
      │
      ├── ambient? → do nothing, reset counter
      │
      ├── chainsaw/mining + confidence >= 85%?
      │       │
      │       ▼
      │   same class 3x in a row?
      │       │
      │       YES → send LoRa alert
      │
      ▼
  light sleep → repeat

Power saving stuff

  • WiFi + BT turned off at boot (we don't need them)
  • LoRa module sleeps except when actually transmitting
  • CPU goes into light sleep between inference cycles
  • Watchdog timer set to 30s in case something hangs

Files

File What it does
src/main.cpp Main loop - capture, infer, alert, sleep, repeat
src/config.h Pin defs, thresholds, timing constants
src/audio_capture.h/.cpp I2S driver setup and buffer capture
src/lora_handler.h/.cpp LoRa init, send alert/heartbeat, sleep/wake