Designing a Coating Pilot Under Line Constraints
A powder-coating line runs adhesion failures at 7.3% of parts against a 2.0% requirement, and the customer has issued a formal corrective-action request. The team's fishbone and FMEA narrowed the field to four candidates: oven cure temperature, line speed, pretreatment concentration and plant humidity. The pilot window allows 12 runs in total. Engineering will not permit oven temperature above 400°F because of a part specification, so the usable range is 370–395°F. Line speed realistically runs at three settings but the conveyor drive is unreliable at the fastest, and pretreatment concentration has been run at two levels for years. Humidity cannot be controlled and has ranged 41–66% RH during production. The line runs two shifts with different supervisors, and the failure is detected only by a destructive tape-adhesion test performed on a sample from each run. Each run takes about 90 minutes including set-up.
1. Recommend a design for the 12 runs and justify the factor and level choices, including how you treat humidity and shift.
Read the worked answer
Design recommendation: 2³ full factorial with four center points = 12 runs
Factors and levels
- Oven temperature: 370°F and 395°F. These are the practical extremes of the permitted range, not the full physical range. State clearly in the report that conclusions apply only inside 370–395°F and that 400°F+ was never tested. This is a constraint you document, not a defect in the design.
- Pretreatment concentration: low and high, the two levels already proven on the line. Using familiar levels reduces risk of a pilot failure that is itself an artifact.
- Line speed: the two reliable settings. Exclude the unreliable fastest setting - a factor level the equipment cannot hold introduces variation you will misread as an effect.
Why a full factorial is the right family here With three factors at two levels, all eight runs fit, so you do not need to fractionate and you give up nothing. Four center points give you a curvature check and an independent estimate of pure error. The two-factor interactions remain cleanly estimable, which matters because cure temperature × line speed (residence time in the oven) is a plausible interaction and a main-effects-only design would miss it.
Humidity Do not treat it as a factor - you cannot set it. Record RH for every run as a covariate and include it in the analysis. If the plant's humidity swings across the pilot window, the covariate will show whether it matters.
Shift Block by shift: two blocks of six, with runs randomized within each block and centre points distributed across both. This prevents a supervisor effect from masquerading as a factor effect.
Response Adhesion pass/fail on a defined sample of parts per run, plus the ordinal adhesion grade if your test method provides one. Define the sample size per run up front.
2. What acceptance criteria should the team set before the pilot, and how will the result be confirmed in production?
Read the worked answer
Criteria set before the first run - written down and signed off
- Statistical: a factor effect is accepted only if it is significant at α = 0.05 with the block and humidity covariate in the model, and only if the effect size is materially larger than the run-to-run noise estimated from centre points. A statistically significant 0.3-point change is not worth a line change.
- Practical: the optimum must predict an adhesion failure rate at or below 2.0%, the customer requirement. Work back from the model to the predicted rate; do not accept a mean shift and hope.
- Operational: the optimum must sit inside the permitted ranges with margin, so ordinary line variation does not push it out of spec. A setting at 395°F when the ceiling is 400°F leaves no room - if the model points there, choose the next-best robust setting.
- Curvature: if the four centre points sit well off the plane through the factorial points, a linear model is inadequate and you must not extrapolate the optimum from it.
Confirmation
- Confirmation runs at the predicted optimum - at least three, on both shifts, under normal supervision, before any production commitment.
- A 30-day production trial at the recommended settings with the failure rate tracked on a control chart, stratified by shift. Pre-define what "success" looks like and what would force a return to the pilot.
- Hold the old settings documented so the line can be reverted if the trial shows the pilot conditions were unrepresentative - for example, if the trial month happens to be dry and the pilot month humid.
- Hand over to Control only after the trial meets the 2.0% requirement with a demonstrated capable process, not just a better average.
3. Midway through the pilot, two runs fail for reasons unrelated to the factors and one run's adhesion samples were tested by a different technician. How do you handle this without invalidating the pilot?
Read the worked answer
Do not silently delete runs. The instinct to drop inconvenient data is the most common way a Green Belt loses the analysis and the team's trust.
Step 1: classify each problem run
- Runs that failed for unrelated reasons (for example a conveyor stop, a bad powder batch, a power interruption): these are not results of the experimental treatment. Document the cause, mark the run as invalid, and re-run it in the same block. Re-running is legitimate; deleting is not. If you cannot re-run within the window, you have a design with missing cells and must analyze it as such, not pretend the design was balanced.
- The run tested by a different technician: this is a measurement question, not a process question. Before deciding anything, ask whether the two technicians agree. If you have no agreement data, run a quick side-by-side on retained samples. If the technicians differ systematically, include technician as a covariate or as a block; if they agree, the run stands and you note it.
Step 2: protect the error estimate Centre points are your independent estimate of pure error. Do not sacrifice them to make room for re-runs. If you must trade, drop a factorial corner and re-analyze with the reduced design, reporting the aliasing you have accepted.
Step 3: adjust the schedule, not the analysis Twelve runs at 90 minutes is a full pilot window with no slack - that was the design flaw. Ask for two additional run slots now rather than after the data is in. Documenting why you need them, with the failed runs attached, is a far easier conversation than explaining a compromised analysis later.
Step 4: report transparently List every run, its status, the reason for exclusion, and what was substituted. A reviewer should be able to reconstruct the analysis from your log.