High-Speed Industrial Checkweigher (Raspberry Pi 4)

Author: Bibin VR
License: MIT — see LICENSE
Hardware: Raspberry Pi 4 · Thames Side XT1000 · T16 Load Cell · IR + Motion GPIO sensors

Dynamic, in-motion checkweighing for a 3-section conveyor using a Thames Side XT1000 indicator (RS232), a T16 fluid-damped load cell, and an IR photoelectric trigger — no encoder required, products never stop.

Targets 300 products/min (5/s) by capturing the full weight waveform as each product crosses the weigh conveyor and estimating the static weight from its shape (plateau / peak / area / rise-fall), refined by a lightweight ML model.

XT1000 ─RS232─▶ SerialReader ─┐
                              ├─▶ RingBuffer ─▶ TrackingEngine ─▶ SQLite + HMI
IR sensor ─GPIO─▶ IRTrigger ──┘     (5 s)      (segment→weigh→classify)

pip install -r requirements.txt          # numpy/pyyaml/pyserial required; rest optional

# Run with NO hardware (synthetic line) — proves the whole pipeline:
python -m checkweigher.main --simulate

# Real hardware:
python -m checkweigher.main --config config/config.yaml

# Dashboard (default): http://<pi-ip>:8090   (8080 is used by another app here)

Console prints live stats; the HMI shows the live waveform, counts, throughput, accept/reject/lost, belt speed and a recent-products table.

checkweigher/
├── config/config.yaml          # tune to your physical line
├── checkweigher/
│   ├── main.py                 # thread orchestration + lifecycle
│   ├── config.py               # typed config
│   ├── metrics.py              # live HMI metrics (thread-safe)
│   ├── core/                   # clock, ring_buffer, events, simulator
│   ├── acquisition/            # serial_reader, ir_trigger
│   ├── tracking/               # product, tracking_engine (state machine)
│   ├── dsp/                    # filters, vibration, waveform
│   ├── estimation/             # classical, ml_model
│   ├── position/               # encoder_interface (future-proof)
│   ├── storage/                # logger (SQLite WAL, batched)
│   └── dashboard/              # Flask HMI + template
├── scripts/                    # train_model.py, replay.py
├── tests/                      # pytest, runs without hardware
└── docs/                       # ARCHITECTURE / diagrams / performance

• docs/ARCHITECTURE.md — full software architecture, thread model, pipeline.
• docs/diagrams.md — data flow, state machine, class & sequence diagrams.
• docs/performance.md — single-occupancy constraint, sample-rate sizing, latency budget, throughput math, tuning checklist.

• XT1000 RS232: set the indicator to continuous output at its max stable rate (≥ 200 Hz). Adjust serial.frame_format parsing in acquisition/serial_reader.py:parse_frame to your exact print string.
• IR sensor: wire to a 3.3 V-safe GPIO (default BCM 17). Uses libgpiod hardware timestamps when available, else RPi.GPIO.
• Level shifting: the Pi UART is 3.3 V — use a MAX3232 (or USB-RS232) for the XT1000’s RS232 levels.
• Low jitter: run with sudo (or CAP_SYS_NICE) so the serial thread can take SCHED_FIFO; consider isolcpus/taskset to pin it on a busy line.

python -m pytest tests/ -q          # filters, waveform, ring buffer, config
python scripts/replay.py --n 5      # compare filters on logged waveforms
python scripts/train_model.py       # train the LightGBM estimator

Implement QuadratureEncoderSource in position/encoder_interface.py (edge counting → counts/position/speed) and return it from make_speed_source. The tracking engine, geometry math and windowing are unchanged — windows become exact even if the belt accelerates.