Turning defect events into process diagnoses, settling-time predictions and operator actions, on every web process in the plant.
Every web process in a flexible-packaging plant produces defects with the same handful of spatial and temporal shapes. Those shapes map to the same classes of cause on every line, so the pattern detectors can be built once and only the cause-to-action tables differ by process. This article sets out that process layer: what it needs, what it detects, and what it gives the operator.
The vision model stays a replaceable component. It answers "what is this blob?" from the image alone. The process layer answers "what is the line telling us, and what should the operator do?" from the stream of classified events plus machine-direction (MD) position, cross-direction (CD) position, time and the plant's event feed. Keeping the two apart avoids shortcut learning in the classifier, keeps retraining cheap, and makes every diagnosis explainable to a process engineer.
| Pattern | Signature on the roll | What it points to |
|---|---|---|
| Streak | Fixed cross-direction position, continuous along the roll | A stationary part at that CD: die lip deposit, doctor-blade nick, blocked dryer nozzle, scored guide roll |
| Repeat | Fixed machine-direction period, any CD | A rotating part with that circumference: chill roll, nip, embossing roll, plate or cylinder |
| Burst | Rate spike at a moment in time, then decays | A transient from a logged event: screen change, job start, adhesive batch change |
| Drift | Slow trend in rate or severity over hours or days | An ageing consumable or slow thermal shift: screen pack, blade wear, viscosity |
| Step | Sudden, sustained change in level | A discrete process change: resin lot, tension setpoint, new substrate roll |
| Edge | Concentrated in the outer CD bands | Edge handling or web geometry: neck-in, edge bead, trim knife, bubble instability |
| Scatter | Random in MD and CD at a stable rate | Process noise or environment: dust, resin contamination, airborne particles |
Two compound patterns matter in a multi-line plant. Propagated: a pattern at a downstream line whose MD period or CD position matches an upstream line's parts, meaning the defect was inherited with the roll. Correlated: two defect classes whose rates move together, usually one cause, as when a tension excursion produces both wrinkles and misregister.
Every line moves through the same states: startup, settling after a change, steady, planned transition, unplanned upset, shutdown. Most false alarms and most missed diagnoses come from reading steady-state rules during a transition. The state is derived from a small set of events the plant already logs: job start and end, speed setpoint changes, roll changes and splices, consumable changes, setpoint changes, stops and restarts.
For each line, product family and event type the layer learns the expected trajectory of each defect class after the event: the median curve, a tolerance band, and the time or metres to return to steady. A live run is scored against its template, which yields three things a fixed threshold cannot: a predicted time to spec ("register will be within tolerance in about 180 m"), an abnormal-transient flag ("this startup is taking 2.3x the usual length"), and a clean separation between startup artefacts and genuine steady-state deviations.
On a gravure press, colour variance and misregister at job start are expected and their template is known per job. The layer reports startup waste in metres against the template, flags startups that settle slowly, and treats any colour drift that begins after settling as a new event with its own cause search: blade wear, ink viscosity drift, substrate roll change.
The process-specific content is a table per process type keyed by (pattern, defect class, state), with likely causes in rank order and the first action. A few rows from extrusion show the shape:
Printing is where the state model earns most, because the two most common events, misregister and colour variance, are expected at job start and only meaningful once the template says settling should be over. Converting is the last web process and the first place most inherited defects are caught, so its most valuable rule is "not ours": recognise the upstream pattern and route it to the line that made it. The tables are a starting draft to be corrected by plant process engineers and then tuned by operator feedback; the layer advises, and setpoint changes stay with the operator and the existing controllers.
Each roll ID records its parent rolls, the line and run that made it, and the MD transform between parent and child: direction (reversed on every rewind), MD offset (trim at start and end), stretch (lamination and printing stretch the web by a measured percentage) and CD lane. With that transform an event on the child roll projects back onto the parent roll and the parent line. That enables attribution to the right line, earlier detection of faint defects the upstream camera missed, a downstream warning when a roll with a known map is loaded, and a claims trail from finished product back to every line that touched it.
Nothing exotic, but each must be logged with the defect events from the start because it cannot be reconstructed later: defect events with class, confidence, size, MD, CD and timestamp; MD position from the encoder; CD from camera calibration; line speed; the process events above; key setpoints from the historian; roll IDs in and out with trim and splice positions; a circumference list per line; and operator responses to advice. Most inspection systems already export per-defect records with MD and CD; the layer consumes that export rather than the raw images, which keeps it vendor-neutral and off the inspection PC. Latency needs are seconds, not milliseconds; the hard real-time path stays with the inspection system and Roll Tagger.
In One Box, this layer is Roll Grader; Roll Tagger supplies the roll identity, encoder truth and event feed it depends on; ImageFlow or any existing inspection system supplies the defect events.