Nobody can watch two kilometres of moving steel for a whole shift and stay sharp. Cameras placed five minutes upstream of the inspection desk changed the job from continuous watching to reviewing - and every correction the operator makes retrains the model.
At a European stainless steel producer, a finished coil of stainless steel runs to around two kilometres in length and a metre and a half across. On the finishing line, an operator watches it pass and marks surface defects.
That is the entire job for a shift: watching a continuous ribbon of bright metal go by, waiting for the moment something is wrong with it. The people who do it are experienced and good at it. The task is still one that no human attention system is built to sustain.
This is not a skill problem, and treating it as one is how these projects usually go wrong. It is an attention problem, and it has an expensive tail.
A defect that gets past the desk does not stay flat. It gets rolled into the coil, with a few tonnes of steel wound on top of it. Recovering it means unrolling the coil, which is slow, costly, and consumes line capacity that was meant for making new material.
If it is not caught at all and the coil ships, the arithmetic gets worse. The customer returns it, and now the producer is paying for the freight both ways, for the unrolling, for cutting the defective section out - and for the conversation about why it arrived that way.
The single decision the whole system rests on. The cameras are not at the operator's position - they are five minutes before it, measured in line speed rather than metres. Everything else follows from that gap: it is what converts a detection into something the operator can still act on, because the steel the model just looked at has not reached them yet.
Put the cameras at the desk and you have built an expensive second pair of eyes with no more time to think than the first pair had.
A Faster R-CNN detector runs against the camera feed and raises defects on the operator's screen as bounding boxes, with the position along the coil. Object detection rather than classification, because the useful answer here is not "this coil has a defect" - it is where, and how far it runs. The operator gets five minutes of warning on a specific section rather than an open-ended instruction to concentrate, and stops the line to confirm when that section reaches the desk.
The job has changed shape. It is no longer sustained vigilance, which people are bad at. It is reviewing a flagged candidate, which experienced operators are extremely good at.
The characteristic failure of a box-based detector on a continuous surface is that the box is too short: the model finds the defect but under-calls how far it runs. The operator drags it out to the true extent, and that corrected box is captured as a new training label rather than discarded as a nuisance.
This is the part that compounds. Every shift the operators work, the labelled set grows on precisely the cases the model currently gets wrong - so the model improves fastest exactly where it is weakest, and the volume of work it hands back to the operator falls over time. The people whose workload the system reduces are the people who taught it.
The model is deployed through Azure Machine Learning onto an edge device at the line, and inference happens there. That is not a preference, it is arithmetic: the line does not pause for a network round trip, and a detection that arrives late has spent the five-minute lead time that made it useful in the first place. Inference has to be local to be real.
Training and retraining stay in Azure ML, on the wider data platform the producer was already building across the line - from melting through hot and cold rolling to finishing. So the loop is: infer at the edge, collect corrected labels, retrain centrally, redeploy to the device. Defect detection was one workstream on shared foundations, not a bolt-on carrying its own private infrastructure.
Defects are caught while the material is still flat and still on the line - before the section is wound into a coil that would have to be unrolled to reach it, and well before it could reach a customer as a return.
Detection no longer depends on how far into a shift the operator is. And because every correction is captured, the system gets measurably better at the defect types it initially handled worst, which steadily reduces how much of the coil the operator has to adjudicate at all.
Let's talk about what's achievable for your specific context.
Start the Conversation →