Automotive plastic injection with AI — the defect is not discovered, it is anticipated before the scrap piles up.

The problem

The defect shows up when the scrap bin is already half full.

On an automotive plastic injection line everybody knows the pattern: the press starts drifting mid-shift — the mold runs hot, material viscosity shifts with the moisture in the pellets, the process technician is busy with another press — and nobody notices until the end-of-line operator sees the pile of bumpers with flash along the edge and raises the alarm.

1. The defect is upstream — in the cycle, in the mold, in the material. But it is discovered downstream, at visual review.
2. The parameter is on the press's PC — Engel, Arburg, Krauss-Maffei, Battenfeld… every brand has its own screen, and the data is almost never cross-referenced with the outcome of the part.
3. There is only one process technician — and six or eight presses. He cannot watch them all at once. The veteran's intuition covers 95% of cases; the other 5% is scrap.

The familiar irony: the data that would explain the defect (actual holding pressure, mold temperature on the previous cycle, effective back pressure) exists, the press stores it, and nobody cross-references it with the part that came out of that cycle.

How it fits the IRIS system

iLEAN does not change your press — it puts vision where there was none and connects the press data with the outcome of the part.

The classic injection problem is that the information is already there, scattered across islands: the press knows its own story, the MES knows the work order, the operator knows the defect at the end of the shift. What does not exist is the cross-reference. iLEAN acts as the putty that fills that joint without asking you to throw out the press, the MES or the ERP.

Edge looks at the part as it leaves the mold. Connect reads the press wherever it is. The agent cross-references part ↔ cycle ↔ history and warns before the drift becomes scrap. The person decides what to change.

The three iLEAN pieces applied to automotive plastic injection:

• Edge — a terminal with machine vision (CNN) installed at the mold exit (over the take-out robot or over the discharge conveyor). It reads the freshly demolded part in milliseconds, recognizes the typical defects (flash, sink mark, visible weld line, burn mark, silver streak, short shot) and fires the actuator to divert the part into the bin if it does not pass. It works with no network: if the plant loses WiFi, Edge keeps inspecting and rejecting, because what is critical cannot depend on connectivity.
• Connect — captures the press parameters through three graduated modes (manual with a photo of the panel, intermediate over the old isolated local PC, integrated via Euromap 63/77 where available). It also captures what arrives from outside: the pellet supplier's notice of a batch change, the WhatsApp from the shift lead saying the temperature control unit is behaving oddly, the customer email with a PPAP change.
• Agent — cross-references the image of the part with the cycle parameters, the material batch, the mold history and the MES work order. If it detects drift (not just an isolated defect, but a trend), it does not send an email at 10 p.m.: it alerts the process technician on whatever channel they use, with the part, the cycle parameters and a root-cause hypothesis. The person decides what to correct on the press; the system does not touch critical actuators without a signature.

See the full IRIS architecture →

Before and after

End-of-line inspection vs. in-cycle control with iLEAN

Aspect	Visual inspection + SPC on screen	With iLEAN Edge + Connect + Agent
When it is detected	End of line or shift change	Part by part, at the mold exit, in milliseconds
Typical size of the scrap pile	20-50 parts before the alarm is raised	1-3 parts before holding and alerting the technician
Link between defect ↔ cycle parameter	Manual, rebuilt hours later	Automatic, at second zero, with a root-cause hypothesis
Older press with no interface	Not integrated, data lost	Connect captures manual / intermediate / integrated
Operation without network	n/a	Edge keeps running on cabinet power
File for PPAP/IATF 16949	Rebuilt by hand, batch by batch	Dossier per part, automatic, with image and parameters

Impact estimate

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

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Injection plant with 6-10 presses (200-2,500 T range), a mix of class A cosmetic parts (bumpers, door panels, instrument panels) and functional parts (structural components, brackets).
• Edge pilot on 1-2 critical presses (camera at the mold exit + diverting actuator + integration with the press through whichever Connect mode applies). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the current frequency of scrap incidents, the average part weight and the cost per minute of recovered cycle.
• Hard levers: scrap avoided on large parts made of premium material (PP+EPDM-TD, ABS-PC) + cycle minutes recovered + an IATF 16949 audit with an automatic dossier per part, not per batch.
• The automotive quality standard sits in the order of 25 PPM (parts per million) ^[1] — a margin like that cannot be held with visual inspection alone.

And the process technician's reasonable doubt

“What if the AI mistakes a harmless cosmetic defect for a critical one and stops good parts?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely classifies an image against patterns trained on your own parts, the best models brought error below 1.5% ^[2]. And even so, what is critical is never decided alone: the system holds the part and raises the alert, the process technician decides. iLEAN's three safety rings are there precisely for this.

^[1] Symestic — automotive quality standard ~25 PPM.

^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

Related processes: powder coating defects · body panel sheet metal defects · jidoka in robotic welding.