Stawamus deck · predictive maintenance

Column health monitoring

Learn what a healthy column looks like on the jig. Then the deck's main PC tells a technician which tile, which column and which part to fix, before it fails.

Standard test movesEmpty column, fixed speeds, up and down. The average tracks friction; half the difference tracks load.

Column vs. itself and its neighboursOwn baseline (first 8 tests), drift and step detectors, comparison with the other columns on the tile.

Alert names the part"Tile 7, column 12: friction rising → clean and re-grease screw and nut; if it returns, replace the nut."

Jig runs on the TMC5240 dev kit. MPS TBMA732 (MA732) on the motor's rear shaft, its A/B/Z outputs into the chip's encoder inputs. No rework, no setup. The chip flags lost steps itself.

1 · Parts

Part	Note
TMC5240 dev kit: NUCLEO-G474RE + TMC5240-BOB	The bench board for every test, from first moves to the encoder (Order 1)
MPS TBMA732-Q-LT-00A encoder board (MA732, 14-bit)	3.3 V only; A/B/Z at 1,024 ppr out of reset. Solder 5 wires, bridge TEST to GND. USD 11.91, 13 at Digi-Key (2026-10-13)
Diametric magnet 6 × 2.5 mm N35SH (Radial 9049)	150 °C max (motor side reaches ≈ 80 °C at 2.0 A)
M2.5 screws: 2 × 5, 2 × 8, 2 × 6	Mount to motor (rear holes only 3 mm deep), tray to mount, board to tray
Printed: encoder_magnet_holder, encoder_mount, encoder_tray	From `print/make_parts.py`; measure the rear shaft and the board thickness first

Prices and order: parts page ↗

2 · Wire it

Wiring and power-up order: bench guide, TMC5240 dev kit section ↗; encoder steps in its part L.

Encoder in the chip	Setting
Input	A/B/Z → BOB ENCA/ENCB/ENCN, 1,024 ppr = 4,096 counts/rev
Scale	Firmware boots at `enc_counts_rev` 4096 (nothing to set) → `ENC_CONST` = 3.125 (binary, `0x00032000`; 12,800 microsteps/rev). The firmware converts to 16,384 counts/rev (1 count = 4), so the analysis is unchanged.
Lag and hysteresis	8–10 µs ≈ 2 counts at 50 rev/s; ABZ hysteresis 0.52° ≈ 6 counts (MA732 datasheet). For the load-angle signal subtract a ≈ 0.26° direction offset (estimate).
Lost-step flag	`ENC_DEVIATION` = 128 microsteps (2 full steps) → `deviation_warn`, read every 1 ms → event `ENC_DEVIATION`
Must stay off	`GCONF.stop_enable` (else ENCA high stops the motor)

3 · Run it

Firmware: firmware/dist/firmware_nucleo_g474re_cube.bin. Add --board nucleo-g474re --no-estop to every bench command (.venv/bin/bench from the repository root) (no E-stop wired on the bench: the supply's output button is the stop).

First run: bench --board nucleo-g474re --no-estop health-run --db jig.sqlite --encoder --ambient 21. Homes, full strokes at 50 and 200 mm/s, ±2 mm reversal probe for nut play. Refuses to start without a valid encoder. Magnet field: MGL and MGH pads both off (meter).
Load ladder: --steps load. 0, 0.5, 1, 2, 3, 4 kg on the weight tower. Gives signal per kg.
Heat soak: --steps soak --soak-min 30. Temperature effect on the readings.
24/7 endurance: --steps stroke,reversal,endurance --cycles 2000 --repeat 0. Standard test every 100 cycles, homing check every 250. Telemetry off (50 Hz = 0.5 GB/day).
Weekly: load ladder, dial-gauge reading of play and height.
Report: deck-health report --db jig.sqlite → HTML with alerts, baselines, a chart per signal.
After a part swap: add --new-unit --unit nut-B2 to restart the baseline.

Stops at the first E-stop, driver fault or unexpected stall. Everything goes into one SQLite file.

4 · Run to failure

Cycle one column until it fails: play above 0.3 mm, more than 5 lost steps per km at 2 kg, or a stall in the standard test. Then look back: which signal moved first, and how many km before. That sets the alarm limits and the production test rate.

Tile-20 board: STM32G431 + 20 × TMC5240, rear-shaft MT6826S encoder per column into the chip. Test moves in quiet moments, a 64-byte summary per column every 10 min over CAN FD, same analysis as the jig on the main PC.

Encoder per column

IC	NOVOSENSE MT6826SGT-AKD-R: A/B = 1,000 ppr (4,000 counts/rev) out of reset, ENC_CONST 3.2 as on the AS5047P. 10 µs delay, 30–1,000 mT, 1–3 mm gap (DS Rev 1.1)
Price	USD 2.38 / 100, 2.12 / 1k at LCSC (2026-10-13). -AKD out of stock; -STD in stock (1 ppr: one SPI write, ABZ_RES = 999, then EEPROM burn)
Placement	Small daughterboard on the motor's rear-face M2.5 pair; 5-wire cable (3V3, GND, A, B, Z) to the Tile-20. Magnet Radial 9049
Supply	22 mA each: 0.44 A at 3.3 V per Tile-20; size the rail for 0.6 A

How it runs

Test moves on an empty column (the camera knows): stroke test hourly, reversal probe daily, homing at least daily. One column at a time.
Counters during work: moves, travel, stalls, zero shift at homing, driver flags, lowest supply, peak temperature, encoder deviations.
Every 10 min: one CAN FD frame per column, 220 per bus. Negligible next to the 10 ms motion traffic.
Main PC (FleetMonitor): baseline per column, drift detectors, peer comparison, symptom rules → alerts on the floor display and API.

What each fault looks like

Fault	Signal	Action in the alert
Friction rising	Average load reading of up and down tests falls	Clean and re-grease screw and nut; if it returns, replace the nut, then the screw
Tight spot	Dip in the load profile at one height	Inspect at that height; replace a bent screw
Nut play	Reversal probe: start of each move gets lighter	Gauge it; replace the nut above 0.3 mm
Coupler slipping	Zero shifts at homing, encoder moved with it	Tighten set screws on the shaft flat; replace the coupler if it slips again
Lost steps	Chip's encoder deviation flag	Look for binding or overload
Something on the cap	Up/down difference rises on an empty column	Remove what rests on or rubs the cap
Supply or cable	Voltage dips; open-coil or short flags	Check connector and motor cable; swap the board
Running hot	Driver temperature rise per test grows	Check current, cooling, friction

What the column does on its own

Event	Reaction	Still working?
A few missed steps	Position reset from the encoder, event logged	Yes
Blocked by a jam	Move stopped and reported; no re-homing	Yes, once cleared
Encoder dies	Runs open-loop; flagged for repair	Degraded
Motor, cable or driver fault	Output off; column sinks to its bottom stop; planner routes around it	No, deck keeps running

Bottom stop 1–3 mm below flush: a dead column leaves a shallow dip, not a pit.

Simulated wear

Real firmware on a simulated 4-column tile: 30 test sessions over 40 km, one column wearing, 8 seeds. Wear models are plausible, not measured: this proves the chain, not the warning times.

Scenario	Fails at	Warned before failure
Healthy columns	—	0 false alarms
Friction rising	3 × new (30 km)	14 km
Nut play	> 0.3 mm (24 km)	4–6.7 km
Tight spot	first lost steps	17–20 km
Coupler slipping	0.5 mm slip	1.3 km (more with more homing)

Open until the jig runs

Noise floors per speed, signal per kg and per newton of friction, temperature effects. Thresholds in the code are placeholders.
Homing repeatability (sets the zero-shift limit) and magnet stability on the printed mount.
The production MT6826S against the jig's MA732: run the first daughterboard beside a TBMA732 on the dev kit. The MT6826S is AMR: the motor's stray field at the rear shaft is not measured.
Which signal moves first before a real failure, and how early.
Play estimate is coarse (0.25 mm steps) until calibrated with the dial gauge.
Load sensing is unreliable below ~1 rev/s and near full speed; tests run in between.

Where it lives

Firmware: firmware/core/ (health.*, machine_health.cpp, machine_tmc5240.cpp), dev-kit hardware layer firmware/board/; design in firmware/docs/HEALTH.txt.
Jig script: software/link/health_run.py (bench health-run).
Analysis and FleetMonitor: software/link/health/ (deck-health report | baseline | calibrate | alerts | simulate).
Printed encoder mount: docs/print/encoder_*.stl.
Methods: NIST e-Handbook (EWMA, CUSUM), NIST/HIWIN ball-screw run-to-failure, PHM Society ball-screw friction from motor current.