Every stamped part inspected at the press, with local CNN vision
At the cadence of a mechanical press the human eye does not catch the fine crack, the new burr from a worn die or the impact mark from a badly torqued shim: parts pass by the minute and the incipient defect slips through where nobody is looking. iLEAN Edge places overhead and side cameras after the press, infers in milliseconds with a local CNN and ejects the defective part before the shipping container reaches the Tier-1. Pure jidoka AI: the process detects its own defect and reacts.
100% human visual control is impossible at real cadence.
In a stamping cell, parts pass by the minute in front of any inspection station. At the cadence of a mechanical press, asking an operator to guarantee the surface of every part is not demanding: it is physically impossible. What exists in practice is a first-off at startup and a per-lot sample every N parts, and sampling has two structural problems.
- It leaves gaps — quality checks a minimal fraction of output. The fine crack in the draw area, the new burr from a worn die and the impact mark from a badly torqued shim slip through exactly in the parts that sit between one inspected part and the next. In Lean vocabulary: there is no poka-yoke, there is statistical hope.
- It travels to the customer — the defect sampling did not catch is found by the Tier-1 at their receiving or, worse, on their assembly line. And it is not one lost part: it is an expensive return — containment of the lot in transit, 100% sorting, an 8D with its PFMEA revised — that triggers the PPM report and erodes IATF 16949 approval as a supplier.
The consequence is the worst muda of all: the defect that consumed coil, press and transport before being discovered. In stamping for seating and suspension Tier-1s, the cost of a defective part leaving the plant runs far above the cost of internal scrap.
iLEAN Edge — overhead and side cameras, a local CNN and ejection before the shipping container.
The problem is not judgment — the quality manager tells a crack from a material slip line perfectly well — it is cadence and sustained attention, which is exactly where machine vision wins. iLEAN Edge replicates the veteran inspector's judgment at press speed, part by part, without fatigue. It is the poka-yoke sampling could never be.
Edge sees every part before the shipping container. The local CNN tells a crack, a burr and an impact mark apart. The actuator ejects the defective part without stopping the press. It works with no cloud and without sending a single image outside the plant — even when the cell sits in a hall with limited connectivity. Every part is recorded with a timestamp, traceable to the production order, the coil heat and the part number.
The specific iLEAN pieces for a stamping cell:
- Edge — a physical terminal with overhead and side cameras installed over the press exit, before the shipping container. It carries a local CNN on an industrial GPU, trained with good and bad examples of that specific part number. It inspects fine cracks in the draw area, new burrs from a worn die and impact marks from a badly torqued shim. Inference is local, with no image sent to the cloud — no latency, no bandwidth cost, and no dependence on connectivity that in many stamping halls is limited exactly at the press cell —: when it detects a defect, it triggers the ejection of the part before the shipping container. Jidoka AI: the process reacts to its own defect in the instant, not at the audit.
- Connect — captures the production order, the coil heat, the part number and the die in use, whether from the ERP or the MES. Every part ejected by Edge is signed with a timestamp and tied to the heat and the part number, so the traceability the Tier-1 demands is automatic, not a manual reconstruction.
- Agent — lives in Central, cross-referencing the Edge history (parts ejected per hour, by defect type, by shift) with the press force curve and the part number from Connect. If the burr always correlates with a specific drop in peak force, it does not send an email at midnight: it presents the already cross-referenced hypothesis — die to be inspected — to the quality manager, who validates and decides. The person supplies the judgment; the system does the gemba walk through the data.
Sampling inspection vs. inspection with iLEAN Edge
| Aspect | Sampling inspection | With iLEAN Edge on the press |
|---|---|---|
| Inspection coverage | A 1-3% sample of output, every N parts | 100% of parts controlled, at real cadence |
| Fine crack in the draw area | Depends on it landing in the sample | A local CNN on every part, without exception |
| New burr from a worn die | Discovered once it grows, with parts already out of spec | Part ejected and die flagged for inspection before the container |
| Impact mark from a badly torqued shim | Appears weeks later, at the Tier-1 receiving | Detected on the line by the overhead camera, before shipping |
| Tier-1 returns for a visual defect | Lot containment + 100% sorting + 8D + approval at risk | Drastic reduction, PPM towards 0 (estimate to be validated) |
| Part certification | Certified on a sample, with no evidence per part | Every lot ships with vision certification of 100% of its parts |
Impact estimate for your plant — to be validated with your own 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.
- Stamping cell with a mechanical press moving parts by the minute and current control by first-off and per-lot sampling every N parts.
- Edge pilot on one press — overhead and side cameras plus an ejection actuator before the shipping container, with no machining work on the machine. First value expected within a few weeks.
- Indicative payback of 5 to 12 months, depending on the current scrap ratio and the weight of Tier-1 returns. Estimate to be validated against your history.
- Expected reduction in returns for visual defects: drastic on moving from a 1-3% sample to 100% part control, with PPM towards 0. (Estimate to be validated against your history.)
- The strategic lever is vision-certified shipping: every lot leaves with 100% of its parts inspected and documented — a growing requirement from seating and suspension Tier-1s towards their stamping suppliers. It enables business, it does not only save scrap.
And the fair question from the quality manager
"What if the AI gets it wrong and lets a crack through?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely classifies an image against known patterns (this part matches the "good" pattern or the "defect" pattern), the best models brought the error below 1.5% [1]. And even then, nothing is decided in a vacuum: the quality manager sees each shift's history, validates false positives in the Edge interface itself, and retraining enters with every model version documented and approved. The press does not stop while training happens. The person supplies the judgment; the machine keeps the cycle turning.
[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.
What people ask about Edge visual control in stamping
Why can the human eye not keep up with the real cadence of a mechanical press?
Because at the cadence of a mechanical press, parts pass by the minute that no operator can review one by one without fatigue. What exists in practice is per-lot sampling — a first-off at startup and checks every N parts — and sampling leaves gaps between one inspected part and the next. The fine crack, the new burr from a worn die or the impact mark from a badly torqued shim slip through exactly those gaps and appear weeks later in the Tier-1 PPM report. With iLEAN Edge, overhead and side cameras over the press exit infer in milliseconds with a local CNN and eject the defective part before the shipping container, so control covers 100% of parts rather than a fraction.
Which specific defects does the local CNN detect on the press exit?
The ones that end in a Tier-1 return. On the surface: a fine crack in the draw or bend area, which at a glance looks like a material slip line. On the edge: a new burr from a worn die, which grows part by part as the punch edge dulls. And on the visible face: an impact mark from a badly torqued shim or from scrap left in the die. The CNN is trained with good and bad examples of that specific part number, so the acceptance criterion is the one in that part's PFMEA, not a generic one: the defective part is ejected before the shipping container bound for the Tier-1.
Does Edge work with no cloud? My press cell has limited connectivity.
Yes — and that is the design case, not the exception. Edge is a physical on-premise terminal with the CNN loaded onto an industrial GPU in the device itself: inference is local and no image leaves for the cloud, so there is no latency, no bandwidth cost and no network dependency. Many stamping halls have limited connectivity exactly at the press cell; Edge keeps inspecting every part and triggering the ejection even with no network. When the connection returns, the signed record of every part uploads to be cross-referenced with the traceability of the production order, the coil heat and the part number.
How does Edge correlate with the press force curve to warn about a worn die?
By cross-referencing the visual defect with the process signal. Every press stroke has its force curve, and when the punch edge starts to dull, peak force drops in a characteristic way and the edge burr grows at the same time. Edge does not wait for the burr to exceed tolerance: on detecting that the burr correlates with that drop in peak force, it presents the quality manager with the already cross-referenced hypothesis — die to be inspected — before the first out-of-spec part reaches the container. That is jidoka AI: the process detects the drift of its own tool and warns in the instant, not at the next shift's audit.
How is the CNN trained on the specific part number on my press?
With good and bad examples of your own part, not with a generic stamping library. The quality manager marks conforming parts and parts with each defect type — crack, burr, impact mark — of that specific part number and that specific material (high strength steel, galvanized, aluminum…). The CNN learns the visual pattern of that part under that lighting from the overhead and side cameras. When a new part number comes in or the die changes, it is retrained with the new samples and every model version is documented, validated against a known part set and approved by quality — within that part's APQP — before entering production, with a record of which version was active in each lot.
Keep the defect out of the container — we will send the estimated ROI of this AI project for your stamping cell within 48h.
We work on your plant's real data, not ours. Assessment with no commitment.
Request estimated ROI within 48h ‹ See all 12 metal stamping cases See automotive