I am commissioning a line
Sensing requirements, connector mapping, edge install, shadow-mode validation and the promotion gates. Read sections 02–05 in order.
Everything needed to take a line from first connector to bounded autonomy — written for rolling engineers and OT teams, not for a marketing page.
Docs used during commissioning at design-partner sites
The four paths people actually take through the documentation.
Sensing requirements, connector mapping, edge install, shadow-mode validation and the promotion gates. Read sections 02–05 in order.
The write-back contract, OPC UA node layout, webhook payloads and SDK reference. Start at section 03 and the API reference.
OT architecture, segmentation, control-path guards, tenant isolation and the control matrix. Section 09, plus the security package under NDA.
Golden datasets, evaluation harnesses, CI gates and the promotion record. Sections 06–07.
Audit export, replay tooling and the claims bundle format. Section 08 and 10.
Fleet management, rollout rings, rollback and multi-site model registry. Sections 04 and 07.
A signed container image, a GPU node and a profile for the line type. Nothing else.
# 1. verify the signed image before anything else
cosign verify --key steelira.pub steelira/edge:2.4.1
# 2. install for a line profile (tandem-cold, hot-strip, cgl, plate)
steelira edge install \
--line tcm-2 \
--profile tandem-cold \
--historian ip21://plant-hist:10014 \
--level2 opcua://l2-tcm2:4840 \
--mode shadow
# 3. confirm every required signal is present
steelira edge probe --strict
# 4. start read-only. no writes are possible in shadow mode.
steelira edge start --mode shadow
# 5. after the promotion gate, and only then:
steelira edge promote --loop roll_gauge --to assist --approver a.kohler
What each profile needs before a model can be trusted. Missing signals degrade specific models rather than blocking deployment.
| Profile | Required | Recommended | Degrades without |
|---|---|---|---|
| Hot strip | Exit gauge, roll force, looper tension, pyrometers | Shape-meter, surface cameras, accelerometers | Chatter, surface |
| Tandem cold | Exit gauge, roll force, inter-stand tension, speed | Shape-meter, accelerometers, oil-film sensing | Chatter, shape |
| Reversing cold | Exit gauge, roll force, tension | Shape-meter, accelerometers | Chatter, shape |
| Plate | Thickness, roll force, pyrometers | Flatness measurement, surface cameras | Flatness, surface |
| Anneal / CAL | Line speed, zone temperatures, strip temperature | Elongation, tension, surface cameras | Property prediction |
| Coating / CGL | Line speed, pot temperature, coating weight gauge | Air-knife position, surface cameras | Coating control |
See docs section 02 for sensor specifications, mounting requirements and acceptable sampling rates.
Every safety-relevant parameter lives in version control, reviewed like any other mill change.
line: tcm-2
profile: tandem-cold
envelopes:
# per setpoint, per grade family
default:
stand.*.roll_gap_mm: [-0.05, 0.05]
stand.*.tension_mpa: [4.0, 12.0]
stand.*.bending_kn: [-120, 120]
line.speed_mpm: [0, 1250]
DX54D: # tighter for deep-drawing grades
stand.*.roll_gap_mm: [-0.03, 0.03]
approval_above_mm: 0.015
autonomy:
roll_gauge: auto
roll_chatter: auto # veto rights, cannot be disabled
temper_coat: assist
fallback:
watchdog_ms: 250
on_timeout: release_to_level2
on_model_confidence_below: 0.85 -> demote_to_assist
Nobody promotes autonomy on a feeling. Section 07 defines each gate and the evidence required.
6–8 weeks of shadow, model accuracy versus measured outcome above the threshold for the loop.
≥90% recommendation acceptance over 4 weeks, zero envelope breaches, ROI confirmed against baseline.
Sustained performance across grade changes and campaigns, signed by rolling, quality and OT leads.
The same evidence, repeated on the next loop or line. No gate is ever skipped because a previous one passed.
Section 10, summarised. All of these have happened on a real line.
The single most common cause of misleading shadow results. Always run edge probe --strict and reconcile against a known coil before trusting anything.
Gauge and shape sensors drift. Steelira detects divergence between predicted and measured beyond a threshold and raises it rather than silently learning the drift.
Models trained on steady state degrade across grade changes. Confidence drops, autonomy demotes to assist automatically, and the transient is flagged for labelling.
If your Level-2 acknowledges slowly under load, the watchdog will release control. Section 10 covers tuning and the diagnostics to run first.
Surface inspection degrades with lens contamination long before it fails. Steelira scores image quality per camera and alerts on trend, not on failure.
When a replay does not reproduce a production decision, the model version or the envelope changed. The audit trail pins both.
Each connector has its own page covering tag mapping, sampling, failure behaviour and required permissions.
Steelira writes to mill equipment, so security and safety are product requirements, not paperwork.
The complete documentation set, including the OT architecture pack, is provided during evaluation. Offline copies are available for air-gapped sites.