Why the Encoder Comes Before the Algorithm on a Sheet Line
Sheet material is wide and arrives continuously, so the line cannot be stopped for inspection. Manual sampling covered only part of each sheet, and wherever the sample missed, the surface was simply not looked at: a defect went straight downstream. The escape rate sat at about 8 percent. The second loss was quieter. Sampling produced no structured record, so defect type, position and frequency could not be counted, the same problem kept returning, and process engineering had nothing to work from.
Sampling fails for lack of records, not lack of eyes
If the row rate falls behind the strip, the image is already wrong
A line-scan camera builds the surface image row by row, and the fragile point is synchronisation between acquisition and strip motion. If the row rate lags the strip, the image stretches along travel; if it runs fast, the image compresses. Once the geometry changes, defect positions no longer match the real part. The scheme takes the strip displacement signal from an encoder and triggers rows at equal displacement intervals, so row spacing follows motion exactly and speed fluctuation cannot distort the geometry.
Grey contrast for areas, edge analysis for lines
The defect spectrum falls roughly into two families. Area defects such as dents and stains appear as patches of abnormal grey level; line and edge defects such as scratches and missing edges appear as contour or boundary anomalies. Grey-level contrast handles the first, edge analysis the second, and they cover for each other. The same split separates a scratch from a stain: a scratch is a long thin line whose direction relates to the contact that caused it, a stain is a patch of abnormal grey, and the classifier marks them by shape and orientation.
Thresholds graded by defect class
One threshold for every defect forces a binary choice: either light defects generate enough false calls to hold up the line, or serious defects are released. Thresholds here are set per defect class, strict for serious defects and looser for light ones, to sit between escapes and false calls. The precondition is that defects are graded first, and the grading standard comes from process engineering, not from the algorithm deciding for itself.
Thirty seconds is the sheet's window, not the inspection time
The cycle is 30 seconds per sheet, synchronous with the line and without stopping it, subject to site acceptance results. The sheet crosses the inspection zone at speed, so the verdict has to be reached inside that window; a late verdict is an escape. Imaging, detection and archiving therefore run as parallel pipelines rather than capture first and compute later. The system starts and stops with the line, and a new coil or lot opens a new archive by sheet number.
Defects that feed back into the process
Once defects are counted by class, position and batch, process engineers can go after the source: scratches repeating at one position point to a guide roller or a conveyor contact, a rise in stains in one batch points to an upstream operation. Escapes fell from about 8 percent under manual sampling to below 0.5 percent, subject to site acceptance results. The other half of the return is the data trail, which takes inspection from blocking defects to supporting process improvement.
Sheet line questions
Why not a matrix camera for full-width imaging?
A matrix camera needs several stitched shots to cover a wide sheet, and the seams bring both escape risk and registration error. A line-scan camera builds the image from motion, which suits a continuous line.
How is the 0.5 percent reached?
By full-width continuous imaging of every sheet plus defect-specific algorithms and graded thresholds. Real figures depend on material and defect spectrum, subject to site acceptance results.
Does a change of specification need recalibration?
Camera coverage and light layout are rechecked for a new width, and the row rate follows strip speed. Mounting datums usually stay, but trigger and decision parameters are updated.