+9.5 points of material yield, found in runs the line had already made.
Rubber & polymer supplier · textile bobbin line · 336 historical production runs. Nothing new collected, no specific data preparation, no client man-days.
The screening is most of the work. Discarding the 967 columns that carry nothing is what makes the remaining ten interpretable by a process engineer.
“Where should the line run so the yield holds?”
- Input
- One existing export: 977 columns of process, quality and material data.
- Target
- Material transformation rate. Starting point 68.6% on average, levers unidentified.
- Output
- A setpoint range for each retained lever, and the yield expected inside each zone.
- Only choice asked
- A process one: which variables are settings, and which are results.
Three zones, and one of them is worse than doing nothing.
The dashed line is the line’s own average across all 336 runs. A zone only counts if it sits above it.
- Zone 1
- Retained. +9.5 points against the line’s own history — the zone the team now produces in.
- Zone 2
- A viable alternative on different settings, useful when Zone 1 is not reachable for material or scheduling reasons.
- Zone 3
- Not recommended. It sits below the line’s own average, so following it would make things worse. We publish it rather than hide it.
- Sample size
- 33, 27 and 17 runs respectively. Small, and we say so — each new run refines the zones.
Seven of ten in one upstream step.
The yield was being decided earlier in the line than the team was looking. That is the finding, and it is the kind of thing a correlation model is good for.
Not every lever is a setpoint
Ambient temperature is observed rather than set. It guides when to produce rather than what to adjust — and the deliverable says which is which.
A side finding, unasked for
Because material data sat in the same export, one of the three material suppliers turned out to be associated with a 1.5× higher chance of exceeding 75% yield.
Two we cannot place
Two of the ten retained levers are not attributed to a process stage in the study. We leave them unplaced rather than guess.
The gain already rests on past runs. The test only confirms it live.
- Method
- Run the next productions inside the recommended zone, and compare their yield to the historical average.
- Cost
- One run is roughly five hours. A week is enough for a first verdict.
- Setup
- Nothing to connect, nothing to install.
- Afterwards
- Each new run refines the zones; the model reloads in one click, by your own team.
Name one line and one target variable.
We read the export you already produce and tell you within a week whether it carries the levers.
