CPS Project Trashcan – Audio Trigger, Classification and Serial Control

This service continuously listens to audio, captures a 1-second window (200 ms before, 800 ms after) when a sound threshold is exceeded, classifies it with an Edge Impulse model, and drives the mechanical trashcan over a simple serial protocol with state events.

Highlights

• Event-triggered audio capture with a 1 s snapshot (200 ms pre + 800 ms post)
• Edge Impulse inference (model .eim) on captured audio (16 kHz, 1 s, int16)
• Serial API integration with state events and start cycle
• Diagnostics and runtime settings via new serial commands
• Config from .env, no CLI required
• Optional waveform/spectrogram/WAV export for debugging
• systemd unit provided for auto-start on boot

Requirements

• Python 3.10+
• Linux/Raspberry Pi OS or macOS
• A working microphone input and PortAudio (sounddevice)
• Edge Impulse .eim model (placed under model/)

Install dependencies:

cd CPS-Project-Trashcan
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Note for macOS: Grant microphone access to Terminal/Python (System Settings → Privacy & Security → Microphone).

Configure .env

A template .env is included. Adjust at least the serial port and model path.

Keys (defaults in parentheses):

• TRASHCAN_SERIAL_PORT=/dev/ttyACM0 (or /dev/ttyUSB0, /dev/cu.usbserial-XXXX)
• MODEL_EIM_PATH=model/modelmac.eim
• AUDIO_DEVICE_ID= (empty = auto)
• AUDIO_RMS_THRESHOLD=1200 (int16 RMS threshold)
• PRE_MS=200, POST_MS=800 (must sum to 1000 ms)
• TRIGGER_COOLDOWN_S=0.3
• VISUALIZE=0 (1 = save last_segment.png/.wav and spectrogram)
• DIAG_ENABLED=0, DIAG_INTERVAL_S=10 (periodic serial diagnostics)
• BOTTLE_SPEED_MS= (servo step duration in ms)
• TRAY_POS_0/1/2= (steps per type) or TRAY_POS_PLASTIC/GLAS/CAN=

Example:

TRASHCAN_SERIAL_PORT=/dev/ttyACM0
MODEL_EIM_PATH=model/modelmac.eim
AUDIO_RMS_THRESHOLD=1200
PRE_MS=200
POST_MS=800
TRIGGER_COOLDOWN_S=0.3
VISUALIZE=0
DIAG_ENABLED=1
DIAG_INTERVAL_S=10

Run as a daemon (systemd)

A service unit is provided at deploy/trashcan.service. Edit the file to match your paths and user, then:

sudo cp deploy/trashcan.service /etc/systemd/system/trashcan.service
sudo systemctl daemon-reload
sudo systemctl enable --now trashcan.service
journalctl -u trashcan.service -f

Stop/Start:

sudo systemctl stop trashcan.service
sudo systemctl start trashcan.service

Manual run (development)

source .venv/bin/activate
python edgeimpulse/main.py

• Reads .env automatically.
• Logs triggers, classification (top label/score), serial events.

How it works

• Audio stream (sounddevice, mono int16, 16 kHz) feeds a 1 s ring buffer.
• When the RMS threshold is crossed, the service waits until 800 ms of future audio have been recorded.
• A 1 s segment is built from the ring buffer (contains ≥200 ms pre + 800 ms post) and sent to the Edge Impulse runner.
• If a confident label is found (score ≥ 0.7), the type is mapped to PLASTIC/GLAS/CAN and an automatic start:: cycle is initiated over serial.

Type mapping:

• 0=PLASTIC, 1=GLAS, 2=CAN
• Strings are recognized: plastic/plastik → 0, glas/glass → 1, can/dose → 2

Event/state flow (automatic):

• Wait until gState::ack::IDLE
• Send start::
• Process asynchronous events:
  • event::state::CONTAINS_BOTTLE
  • event::state::WAITING_FOR_TRAY
  • event::state::TRAY_IN_POSITION
  • event::state::MOVING_BOTTLE_TO_TRAY
  • event::state::BOTTLE_IN_TRAY
  • event::state::MOVING_TO_IDLE
  • event::state::IDLE (cycle done)

State diagram

The following Mermaid diagram summarizes the automatic state progression and the serial events involved.

stateDiagram-v2
    [*] --> IDLE

    IDLE --> CONTAINS_BOTTLE: start::<type>\n(event::state::CONTAINS_BOTTLE)
    CONTAINS_BOTTLE --> WAITING_FOR_TRAY: event::state::WAITING_FOR_TRAY
    WAITING_FOR_TRAY --> TRAY_IN_POSITION: event::state::TRAY_IN_POSITION
    TRAY_IN_POSITION --> MOVING_BOTTLE_TO_TRAY: event::state::MOVING_BOTTLE_TO_TRAY
    MOVING_BOTTLE_TO_TRAY --> BOTTLE_IN_TRAY: event::state::BOTTLE_IN_TRAY
    BOTTLE_IN_TRAY --> MOVING_TO_IDLE: event::state::MOVING_TO_IDLE
    MOVING_TO_IDLE --> IDLE: event::state::IDLE

Loading

Serial API reference

Commands are sent as a single line: name::value\n. Acks are returned as <cmd>::ack::<payload?>. Asynchronous state events can arrive at any time: event::state::<STATE>.

Core commands:

• start::
  • type: 0|1|2 or plastic|glas|can
  • Ack: start::ack::OK or error (ERR_BAD_TYPE, ERR_BUSY)
• mTray::
  • Manually set target tray and trigger CONTAINS_BOTTLE
  • Ack: mTray::ack::OK or ERR_BAD_TYPE
• mPosBottle::<1|2>
  • 1: moveDrop(), 2: moveInit()
  • Ack: mPosBottle::ack::
• gPosBottle::
  • Ack: gPosBottle::ack::
• gLimitTray::
  • Ack: gLimitTray::ack::PRESSED|RELEASED
• gState::
  • Ack: gState::ack::<STATE_NAME>
• gType::
  • Ack: gType::ack::<PLASTIC|GLAS|CAN>
• estop::
  • Sets error state (EMO_MOOD)
  • Ack: estop::ack::OK
• ping::
  • Ack: ping::ack::pong

Diagnostics and settings:

• gDiagTray::
  • Ack: gDiagTray::ack::pos=,target=,dtg=,speed=,state=
• gDiagBottle::
  • Ack: gDiagBottle::ack::state=
• setTrayPos::=
  • Persist target steps per type in RAM
  • Ack: setTrayPos::ack::OK | ERR_BAD_ARG | ERR_BAD_TYPE | ERR_FAIL
• setBottleSpeed::
  • Set servo step duration (bigger = slower)
  • Ack: setBottleSpeed::ack::OK

The daemon exposes convenience methods and an optional background diagnostic loop controlled by .env:

• DIAG_ENABLED=1 enables periodic logs every DIAG_INTERVAL_S seconds.
• At startup, BOTTLE_SPEED_MS and TRAY_POS_* values are applied once if set.

Tuning

• AUDIO_RMS_THRESHOLD: raise to avoid false triggers, lower to be more sensitive.
• AUDIO_DEVICE_ID: set to a specific input device if the default is wrong.
• PRE_MS/POST_MS: keep sum ≈ 1000 ms; change only if your model differs.
• TRIGGER_COOLDOWN_S: avoid repeated triggers from the same sound.
• Confidence threshold (currently 0.7): adjust if your model requires it.

Troubleshooting

• Serial
  • Verify TRASHCAN_SERIAL_PORT and user permissions (dialout/uucp group on Linux, ls -l /dev/tty*).
  • The firmware must use the same baud rate as the daemon (default 9600).
• Audio
  • No triggers? Increase sensitivity (lower threshold) or verify microphone access.
  • On macOS, allow microphone usage for the terminal/Python process.
• Model
  • MODEL_EIM_PATH must point to a valid .eim, typically 16 kHz, 1 s input.
• Service
  • Adjust paths in deploy/trashcan.service (WorkingDirectory, EnvironmentFile, ExecStart). Check logs with journalctl -u trashcan.service -f.

Project structure (excerpt)

• edgeimpulse/main.py – daemon: audio trigger, classification, serial control, diagnostics
• deploy/trashcan.service – systemd service unit
• .env – configuration
• model/ – Edge Impulse models (.eim)
• arduino/ – firmware/protocol (reference)

License

Internal project. Please clarify license before public release.

🪙 Easter egg

When in doubt, just ping it.

$ echo "ping::<anything>" > /dev/serial
ping::ack::pong

Also, some ASCII trash… can!

    _______
   / _____ \
  / /     \ \
 | |  ___  | |
 | | (___) | |
  \ \_____/ /
   \_______/
     (♻)

Reduce, Reuse, Recycle — and sometimes, Re-route to the right tray.