An injection press with no network can still give you data

The problem

The press that drifts most is the one that reports nothing.

Critical machines are, almost by definition, the ones you cannot stop and the ones you cannot touch. Nobody will authorize a control retrofit on a press running a live program, and the machine builder rarely opens the panel's protocol. The result is a data map with holes: the plant has process information from the new machines and none from the old ones, which are precisely the ones that drift most and contribute most to scrap. Switchover pressure, barrel temperature profile, actual cycle time, part counter and alarms are displayed on screen and lost right there. When a systematic defect appears — a run of sink marks, a weight difference between cavities — root cause analysis is done from memory and on the opinion of the most experienced setter. And the only "official" route to data is to budget a machine replacement, which costs hundreds of thousands of euros and weeks of reorganizing the shop floor.

• The presses that mold the biggest parts — instrument panel carriers, door panel substrates — are often the oldest, because they still make good parts. Their panel is proprietary and off the network.
• Switchover pressure, barrel temperature profile, cushion, actual cycle time, part counter and alarms are displayed on that screen and lost the moment the screen changes.
• When a run of sink marks or a weight difference between cavities appears, root cause is done from memory and on the opinion of the most experienced setter.
• The only official route to data is to budget a replacement — several hundred thousand euros and weeks of reorganizing the shop floor around a machine that still works.

How it fits the IRIS system

Connect in photo-over-panel mode — nothing goes inside the machine.

Connect in photo-over-panel mode. A fixed industrial camera with controlled lighting points at the screen; a grounded language model reads each frame by optical recognition and structures it as a time series per part number and per mold; the system cross-checks the live ERP order to know what was being produced at each instant. No cable inside the machine, no change to its software, no recertification. When the panel cycles through screens, the system keeps reading whatever it shows at that moment.

It does not touch the press software, does not open a port and does not need the machine builder to release a protocol. On a press running a live OEM program, that is the difference between a case that gets approved this quarter and a retrofit nobody will ever authorize. The plant keeps the press that still makes good parts and gains the data it never had.

See the full IRIS architecture →

Before and after

The silent press versus the press being read

Aspect	Today	With iLEAN Connect
Molding curve	On the screen, then gone	Time series per part number and mold
Sink marks this week	Argued over from memory	Correlated with the curve of that lot
Alarms	Seen by whoever walks past	Recorded and searchable
Cavity weight drift	Discovered at the scale	Visible in the process trend
Machine software	—	Untouched, no recertification
Replacement CAPEX	In next year's plan	Deferred indefinitely

from zero correlation between process conditions and defect to root cause analysis per lot in minutes · from alarms only seen by whoever walks past to alarms recorded and searchable · from "we have to replace the press in order to measure" to CAPEX deferred indefinitely.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-10 months.
• Avoidance or deferral of replacement CAPEX, in the order of several hundred thousand euros for a large-tonnage press.
• Scrap reduction through curve-to-defect correlation, which is impossible today because the curve does not exist anywhere but on the screen, for the seconds it is displayed.
• And root cause on a systematic defect — sink marks, cavity weight drift, a long cycle — goes from a meeting with opinions to a query per lot in minutes.

estimated payback 5-10 months; avoidance or deferral of machine replacement CAPEX (in the order of several hundred thousand euros for a large press) and scrap reduction through curve-to-defect correlation. *Estimate to be validated.*

And the fair question from the production manager

"Can a camera really read a panel that cycles through screens?" — yes, because the task is anchored: fixed screen, fields in known positions and physical ranges known in advance, where the best models drop below 1.5% error [1]. The system reads whatever screen is showing and discards any frame whose value falls outside the possible range instead of accepting it, which is what would poison the series.

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