Deep divePackaging & Paper
Predicting web breaks and smoothing grade changes on a paper machine
A web break stops saleable production, loads the broke system and leaves the crew re-threading the sheet; a poor grade change makes off-spec reels until the machine settles. Both leave traces in data most mills already store: QCS profiles, wet-end measurements, draws, historian tags and web inspection events. This deep dive explains how prediction models are built from that data, validated on rare events and delivered to operators as alerts with reasons.
On this page
- Why a break costs more than the minutes the reel stops
- Paper machine terms used in break and grade-change models
- From machine signals to an alert operators act on
- Data sources that carry early signs of a break
- Labelling breaks so a model can learn from them
- Modeling rare events without flooding the control room
- Grade changes as a trajectory problem
- Retention drift ahead of a break, traced through the data
- Keeping a break model trustworthy after go-live
- Questions and answers
- Sources
Why a break costs more than the minutes the reel stops
When the sheet breaks, the reel stops producing saleable paper, the broke system absorbs a surge of fiber, and the crew threads the tail back through the machine. Repulped broke shifts wet-end chemistry for a while afterwards, repeated breaks raise the chance of damaging felts and fabrics, and threading is among the more hazardous routine tasks on a machine.
Grade changes cost in a quieter way. Each move between basis weights, furnishes or shades produces paper that misses specification until moisture, weight and caliper settle, and how long that takes varies with crew, sequence and machine condition. ColdAI's packaging and paper work lists paper machine optimization and grade-changeover prediction among its focus areas2, because the signals that precede both problems are usually already recorded.
Paper machine terms used in break and grade-change models
- QCS
- Quality control system: scanning gauges that measure basis weight, moisture, caliper and often ash across the moving sheet and feed the control loops for those properties.
- MD and CD profiles
- Machine-direction trends over time and cross-direction profiles across the width. Streaks and edge deviations in CD profiles often say more about break risk than averages do.
- Wet end
- The approach flow, headbox and forming section, where consistency, retention and drainage chemistry decide how the sheet forms.
- Draw
- The speed difference between sections that keeps the sheet in tension; too much strains a weak sheet, too little lets it flutter.
- Broke
- Off-specification or broken paper returned to pulping, which changes stock properties as it re-enters the system.
- Web monitoring system
- Cameras along the machine that record break sequences and, with web inspection, detect holes, spots and edge cracks.
- Soft sensor
- A model that estimates an unmeasured or lab-measured property, such as strength or retention, from online signals.
From machine signals to an alert operators act on
- Historian and QCS data
Time-aligned tags, scanner profiles and wet-end measurements at a common sampling rate.
- Break and defect events
Sheet-break signals, web inspection detections and the video that explains each break.
- Labelled break windows
Each break tagged with section, cause and the period before it in which an alert is useful.
- Process features
Rolling statistics, rates of change and profile-shape measures per machine section.
- Break-risk model
Scores risk continuously and reports which variables pushed the score up.
- Alert with reasons
Shown on the operator screen with the moving variables and suggested checks.
- Crew feedback
Operators mark alerts as useful, false or acted on, which feeds retraining.
Data sources that carry early signs of a break
| Source | What it can reveal | Common trap |
|---|---|---|
| Historian and DCS tags | Creeping draws, falling press vacuum, steam and consistency swings | Tags renamed after control upgrades; compression hides short spikes |
| QCS scanner profiles | CD moisture streaks, wet edges and weight bands that weaken the sheet locally | The scanner goes off-sheet at a break, so profiles stop exactly when you need them |
| Wet-end measurements | Retention, charge and white-water consistency drifting with furnish | Lab values arrive late and at irregular times |
| Web inspection events | Bursts of holes or edge cracks shortly before a tear | Alarm thresholds tuned for product quality, not prediction |
| Maintenance and shift logs | Clothing changes, wash-ups and doctor blade changes that explain clusters of breaks | Free text and timestamps typed in after the event |
No single source predicts breaks well on its own; the value comes from combining them on one aligned timeline.
Labelling breaks so a model can learn from them
Build a break register
Combine the DCS sheet-break signal with reel and winder downtime so every break has one start time and one restart time.
Align the clocks
Historian, QCS, camera and maintenance systems rarely share a clock. Measure the offsets and correct them before computing any feature.
Classify location and cause
Use the web monitoring video to place each break in a section and give it a cause category, such as wet end, edge crack, hole, felt or mechanical.
Set warning windows
Agree how early an alert must arrive to be acted on, and exclude threading, restarts and planned stops from training data.
Split by time
Validate on later months than you train on. Random splits leak information between neighboring samples and flatter the model.
Modeling rare events without flooding the control room
Breaks are rare compared with hours of normal running, a textbook case of class imbalance. A model that never alerts scores well on accuracy and is useless. Evaluate on events instead: the share of breaks in each category that had an alert inside the useful window, how much warning it gave, and how many false alerts a crew sees per shift.
Start with a transparent baseline such as multivariate statistical monitoring on principal components, which flags unusual combinations of variables without needing labels. Then train supervised models, often gradient-boosted trees on windowed features, per break category. Sequence models can capture long-memory effects but need more breaks to learn from and are harder to explain to a shift leader.
Set thresholds with the people who will receive the alerts. Fewer, earlier, explained alerts usually beat a sensitive model that crews learn to ignore, a trade-off shared by predictive monitoring work across process industries.
Grade changes as a trajectory problem
A grade change moves many setpoints at once: stock flow, machine speed, steam pressures, headbox settings and chemical dosing. Past transitions between the same pair of grades form a natural experiment. Align them by start time, score each by time to on-specification paper and waste produced, and the best ones reveal setpoint sequences worth repeating.
While the scanner is unstable or the lab cannot keep up, soft sensors that estimate weight, moisture or strength from upstream signals show operators where the sheet is heading. Recommendations should stay advisory at first and complement the grade-change automation many QCS and DCS suppliers already offer. Ordering grades sensibly across a campaign also ties this work to trim and order planning.
Retention drift ahead of a break, traced through the data
Keeping a break model trustworthy after go-live
Drift from furnish, clothing and season
Early signalFalse alerts rise, or breaks in one category arrive unannounced.
MitigationTrack event-level performance monthly and retrain on recent data validated on the latest period; see concept drift.
Tag changes after control-system work
Early signalFeatures flatline or jump after an upgrade.
MitigationVersion the tag map and run data-quality checks on every model input.
Exposure of the control network
Early signalRequests to install software on DCS servers or open inbound connections.
MitigationRead from a historian replica in a demilitarized zone, design zones and conduits to ISA/IEC 624431, and keep any setpoint write-back behind management of change.
Questions and answers
How much machine history does a break-prediction model need?
Enough to contain a meaningful number of breaks in each category you want to predict, across the grades, furnishes and seasons the machine normally runs. The deciding factor is break count and label quality rather than calendar time. If history is thin, start with unsupervised monitoring and collect well-labelled breaks deliberately while it runs.
Does break prediction work on older paper machines?
Usually, if the machine has a historian and a reliable sheet-break signal. Older machines may lack fine scanner resolution or a web monitoring system, which limits how precisely breaks can be located and explained. Start with the data already logged, measure which break categories the model can see, and add sensors or cameras only where a category stays invisible.
How does this differ from the analytics our QCS supplier provides?
QCS and web inspection software focuses on controlling measured quality and detecting defects in the sheet. Break prediction combines signals across systems, including DCS tags, wet-end chemistry, inspection events and maintenance logs, and estimates the risk of a future event. The two complement each other: supplier tools keep the sheet in specification, while a break model looks for combinations that precede a tear.
Should a break model change setpoints automatically?
Not at first. Run it as an advisory system, measure how often its alerts and recommendations were right, and agree with operations which actions could ever be automated. Any closed-loop write-back to the control system should pass through management of change, a safety review and the site's industrial cybersecurity controls.
Sources
- ISA/IEC 62443 Series of Standards — International Society of Automation · checked 10 October 2026
- Packaging & Paper: focus areas and delivery approach — ColdAI