The defective closure no longer reaches the customer

The problem

At line speed, the eye lets through exactly the drum that opens in transit.

At assembly line speed, the human eye tires and lets through misaligned closures, off-center reinforcement hoops or fiber creases. 100% human visual control is impossible at real line speed, and sample-based inspection leaves gaps — that defective drum reaches a chemical or pharma customer with a risk of accidental opening in transit.

• On the assembly line the human eye tires and lets through misaligned closures, off-center reinforcement hoops or fiber creases in the drum body. Each of them can turn a sound kraft body into a drum that opens or leaks.
• 100% visual control by a person is impossible at real line speed. Nobody keeps the same attention on the thousandth drum as on the first.
• Sample-based inspection leaves gaps between samples, and the defective drum falls into one of them. The sample passes, the batch ships, and the defect travels with it.
• That drum reaches a chemical or pharma customer with a risk of accidental opening in transit — with a dangerous good inside. A single claim like that weighs more than a year of sampling costs.

How it fits the IRIS system

Edge — a local CNN classifies every drum in milliseconds and ejects it.

A camera over the line, with a local CNN trained on the customer's specific closures and hoops, infers in milliseconds per piece and ejects the defective drum before it reaches testing and shipping.

The model is trained on your closures and your hoops, under your line's lighting, not on a generic catalog. That is what keeps false rejects low enough for the line to trust the ejector. And every ejected drum keeps its image, so quality can check any decision the camera made.

See the full IRIS architecture →

Before and after

Sampling the closure today versus inspecting every drum

Aspect	Visual sampling	With iLEAN Edge
Drums inspected	A sample	100%, in line
Misaligned spring-latch closure	Seen if sampled	Classified and ejected
Off-center steel hoop	Easy to miss when tired	Detected on every drum
Fiber creases in the body	Depends on who looks	Same criterion, every shift
Where the defect is stopped	At the customer, sometimes	Before testing and shipping
Where inference runs in fiber drum making	—	Locally, in milliseconds

Before: partial sampling, defects slipping through. After: 100% drum control, in line.

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 5-12 months, depending on your current reject ratio. The higher it is today, the faster the camera pays back.
• It comes from reduced customer claims for drums with closure or hoop defects. A claim from a dangerous-goods customer costs far more than the drum itself.
• From partial sampling with defects slipping through to 100% drum control, in line. The same criterion applies on the night shift as on the morning one. Fatigue stops being a variable in what ships.
• And every ejected drum keeps its image, which turns claims into a defect history by closure type and line. Maintenance gets data to act on instead of a complaint to defend.

Estimated payback 5-12 months depending on current reject ratio, from reduced customer claims. Estimate to be validated.

And the fair question from the production manager

“What if it starts ejecting good drums?” — that is the real risk of any vision system, so the model is trained on good and bad drums from your own line. Telling a seated closure from a misaligned one against reference images is an anchored task, where the best models drop below 1.5% error [1], and doubtful cases go to a person rather than being rejected blindly. Every human decision feeds back into the model. The longer it runs on your line, the fewer cases it needs to escalate.

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