The press stamps just fine. Its Siemens S5 tells nobody about it.

The problem

The data lives on a screen nobody ever integrated.

An automotive stamping press produces well. It is a machine worth hundreds of thousands of euros that holds up shift after shift. But the data that matters to the MES — and to real OEE, not the shift supervisor's notebook — lives on its Siemens S5 HMI screen:

1. Real strokes per minute, not the theoretical figure on the datasheet.
2. Pressure and force per stroke, with the real curve when the die wears out or when the coil arrives off-spec.
3. Alarms and micro-stoppages, the gap between Monday's report OEE and the real one.

Why it happens: the Siemens S5 does its job, it has been running for decades, and swapping it for an S7 or TIA Portal means stopping the press for weeks. Integration cost more than it was worth, so it never happened. The operator writes down what matters on a sheet at shift change, the shift supervisor types it into the ERP on Monday, and the MES draws an incomplete reality. This is not sloppiness: it was the only reasonable option. That is what changed.

How it fits the IRIS system

iLEAN Connect photographs the screen — filler from the outside, ring 1 stays sacred.

The classic OT customer is rightly suspicious of anything that gets near their PLC. iLEAN respects that boundary by design: the S5 is not touched, no port is opened on it, the OT network is not touched. Connect applies capture from the outside, in the physical world — a camera looks at the screen, just as a person would, and the data crosses into the system from ring 3.

The camera looks at the HMI. The agent recognises the field. The live data enters the MES. The PLC receives nothing.

The iLEAN pieces applied to a press with a Siemens S5:

• Connect (capture gradient, photo option). Camera pointed at the HMI, capture every cycle or on event. A vision agent (CNN + OCR) recognises the active view (process view, alarm view, die recipe view) and extracts strokes, pressure, SPM and active alarm. It works because it is trained on that machine's real screens, not in the abstract.
• Edge (optional). If the press has an additional visual bottleneck (coil feeding, ejection, defect ranking on the stamped part), an Edge terminal with a CNN can watch it in line. Not mandatory to start.
• Plant agents. They cross the data extracted from the screen with the production order, the source coil and the die maintenance plan. Real OEE, traced micro-stoppages, pressure anomalies linked to a specific die. When they spot an out-of-range pattern, they alert the shift supervisor through whichever channel they use.
• Three rings. The S5 lives in ring 1 (critical OT, untouchable). Connect captures from ring 3 (the physical world, the screen). The data passes through ring 2 (validation + signature) before reaching any system that takes a decision. The armoured mailbox only opens from the inside.

See the full IRIS architecture →

Before and after

Blind S5 screen vs. photo capture with iLEAN Connect

Aspect	Operator writing it down by hand	With iLEAN Connect + vision agent
Strokes per minute (SPM)	Datasheet figure, logged per shift	Real, every cycle, into the MES
Pressure per stroke	Only when the operator happens to look	Continuous curve, anomalies flagged
Alarms and micro-stoppages	Reconstructed by hand	Traced instantly, with root cause
Real OEE	"Approximate"	Calculated from live data
Touching the S5 PLC	n/a	Zero, no port opened on it
Die record	The maintenance lead's folder	History per die, automatic

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the actual data from your plant. It is an order of magnitude so the committee can size it; we refine it during the diagnostic.

• Automotive Tier-1 with 3-5 stamping presses running an isolated Siemens S5 HMI and a per-die maintenance plan.
• Connect pilot (camera + vision agent) on one press. First value expected within a few weeks: real SPM and alarms travelling to the MES.
• Indicative payback between 4 and 9 months. Hard levers: reduction of uncounted micro-stoppages (defensible floor ≥30%), die maintenance planning based on real pressure, real OEE to decide whether a fourth press is needed or one is spare.
• Rollout to the other presses done by the plant team itself with their own trained people — without us.

And the reasonable IT/OT objection

"What if the OCR misreads a value and pushes a made-up SPM into the MES?" — hallucination is a problem of free generation, not of anchored tasks. Reading a number off a screen with known digits and pushing it to the MES is the anchored task par excellence: the best models keep the error below 1.5% ^[1]. And even so, ring 2 is there: out-of-range values go through validation before updating OEE or firing an alarm. The S5 never receives anything back; the data flows only from the screen towards the MES.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.