In-line AI vision for the raw casting defect

The problem

The defect is discovered after the plant has paid machine hours on it.

In a foundry, scrap is not an anomaly, it is a structural cost line, and late detection multiplies the damage. A rejected raw casting costs metal and energy; the same part rejected after machining also costs machine hours, tooling and the capacity of a cell that is usually the bottleneck of the whole plant. One hundred percent inspection at line speed is not humanly sustainable, and sampling lets exactly the sporadic defect through — which is the one that reaches the customer. The defects themselves are known and recurring: shrinkage, gas porosity, trapped sand, inclusions, mold mismatch, badly cut risers, flash. And there is a second, invisible loss: with no systematic record of the defect, root cause is investigated by intuition and never fully closed.

• At line speed, raw casting inspection is done by eye, shift after shift, with criteria that move from one colleague to the next. Sampling lets exactly the sporadic defect through, which is the one that reaches the customer.
• A rejected raw casting costs metal and energy. The same part rejected after machining also costs machine hours, tooling and the capacity of the cell that is the bottleneck of the plant.
• The defects are known and recurring: shrinkage, gas porosity, trapped sand, inclusions, mold mismatch, badly cut risers, flash. None of them is new; what is new is seeing all of them on every part.
• With no systematic record by heat, root cause is investigated by intuition and never fully closed, so the same shrinkage comes back three months later.

How it fits the IRIS system

Edge — inspection in milliseconds at the shot-blasting outfeed, before machining.

Edge, AI vision running on the line itself:

Catching the defect before machining does two things. It stops the plant spending its scarcest hours on scrap. And it builds the record nobody has: every detection with its image, anchored to the heat number, the furnace and the lab result, so that when trapped sand concentrates in one shift or shrinkage in one chemistry range, the system says so. Inference is local, on the line, with no cloud and no plant network in the loop, and the diverter acts before the next part arrives. Scrap stops being a structural cost line nobody can explain and becomes a set of dated patterns with a heat number attached.

• An industrial camera over the shot-blasting outfeed conveyor, with controlled lighting.
• A neural network trained on real defect images from the part numbers this plant actually produces, not a generic catalog.
• Local inference in milliseconds — no cloud or plant-network dependency — and automatic diversion of the part before it reaches machining.
• Every detection is stored with its image and anchored to the heat number and lab result: when a defect type concentrates in a chemistry range, a furnace or a shift, the system says so.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	Sampled, depends on fatigue	One hundred percent at line speed
Criteria	Shift from one inspector to the next	Stable and traceable
Point of detection	After machining, or at the customer	Before machining
Machine hours on defective parts	Paid	Not spent
Defect record	None	Image, heat, furnace and lab result
Root cause	By intuition	Defect to heat to sand to furnace correlation

from sampled, subjective inspection to continuous inspection with stable criteria · from defects discovered after machining to defects diverted before · from root cause by intuition to a defect ↔ heat ↔ sand ↔ furnace correlation.

Impact estimate

Impact estimate — to be validated against real scrap by stage.

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 6-12 months.
• A 30-50% reduction in machining spent on parts that were defective from the start.
• A drop in the cost of poor quality detected downstream, including at the customer, where a sporadic defect costs the most.
• And root cause that closes, because the defect is finally correlated with heat, sand and furnace.

estimated payback 6 to 12 months, with a 30-50% reduction in machining spent on parts that were defective from the start and a drop in the cost of poor quality detected downstream. *Estimate to validate against real scrap by stage*.

And the fair question from the production manager

“What if it diverts good castings?” — the false reject is the real risk of any vision system, which is why the network is trained on images of good and defective parts of your own part numbers, at that outfeed and with that lighting, not on a generic catalog. Borderline cases are not simply diverted: they are escalated for a person to decide, and every decision feeds back into training. Classifying a known defect against reference images is an anchored task where the best models drop below 1.5% error [1], and the diverted part is still a raw casting: it can be reviewed and returned to the flow.

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