Gallery
Description
The problem
Pão de Rio Maior makes frozen par-baked bread. Every box passes an X-ray for foreign bodies. This is a critical control point.
The machine recorded why each box was rejected. The reason stayed in USB exports nobody read.
From 31 Aug to 26 Oct: 16,109 rejects of 143,130 boxes (11.3%). A reject is good bread thrown away, or checked again.
Each reject stopped the belt. Up to 609 stops a day. In three weeks, 60 hours still.
Single-detector rejects: 47% at most 5% over the limit, and 73% within 10%. That fits denser bread, or loaves stacked.
Limits changed on 73 product-and-detector pairs in eight weeks. No record of who, or why. LV3 and LV8 are AI detectors and rejected boxes while still learning.
Our solution
An operator's screen, and a panel for the operations room. Each reject is explained from the machine's numbers. Nothing is sent back. Fixed rules, not a model.
A person still runs the X-ray. Quality still sets the limits.
On screen: detectors, value, limit, how far over, and the image. Strong: two or more detectors at least 10% over, or one at least 25% over. Weak: one detector, at most 10% over. Otherwise moderate. LV3, LV8, or LEARNING: the machine's own learning detectors.
By shift or by production day. The panel shows reject rate against the usual rate, stops, and time lost restarting.
Signals by hour. Rejects by product and by detector. A limit change at the first reject with the new limit. Filters: signal, product, detector, borderline.
One click: replay or live. Last 15 days, at ×1, ×10, ×60 and ×300. History rebuilt three times a day, each day checked. Live reads the share every second, read-only, on the machine clock, latest images pre-loaded.