What We Got Wrong About Changeover

A plant running nine housing variants on one washer was changing over in about two hours. The production manager asked for twenty minutes, and on paper it looked straightforward: quick-change fixture pads, a program per variant, done.

It took four months, and for two of them the line was measurably worse than before we touched it. This is what happened in each of those months, because the sequence is more useful than the answer.

Month One: We Changed The Fixture

The obvious constraint was the fixture. Nine variants meant nine sets of pads and stops, swapped with bolts, and the bolting was most of the two hours.

We replaced the bolted pads with quick-release pins and added position markings, so a pad could only go in one place. Changeover dropped to about seventy minutes on the first attempt, which everyone considered a success.

It was not. Within three weeks we had two rejections that traced back to a pad in the right position but 12 mm off the variant position it was meant for, and one near-miss where a nozzle box had been left from the previous part. The fixture was faster and the line was less reliable, which is a bad trade if you make it without noticing.

Month Two: It Got Worse

The second month was the worst of the project, and it was our own doing.

Because the fixture change was quicker, operators started changing over more often — splitting batches that used to run together in order to hit delivery dates. That was rational on their side. The problem was that a changeover now cost less time and the same uncertainty, so we had more changeovers with the same error rate, and the error rate became visible.

First-pass yield fell by four points, from 96.2 to 92.1 percent. The cause was not the fixture. It was that the program selection was still manual, the first part was still checked informally if at all, and the operator had less time to notice something was wrong because the change was faster.

This is the part of changeover work that gets skipped. Reducing the mechanical time is easy to see and easy to celebrate. The verification time does not shrink with it, and if you cut the first without touching the second you have simply made a slower process faster at making the same mistakes.

The Option We Rejected

Halfway through the second month, with yield falling, the plant proposed buying a second washer: one machine per variant family, no changeover at all.

We costed it honestly, which meant counting what the second machine would not solve. Two machines still need the right program loaded, still need a first part measured, and still need the fixture cleaned between batches — the three intervals that turned out to be most of the original two hours. What the second machine removes is only the pad swap, which by then was already down to twenty-two minutes.

So the proposal would have bought roughly twenty minutes of changeover for the price of a machine, a second set of utilities and a second maintenance routine, while leaving the forty minutes of first-article waiting untouched. It was rejected on that basis, and the rejection was correct.

I mention it because it is the option most plants reach for first. The second machine is visible, budgetable and easy to justify, while the four intervals that actually make up a changeover are invisible until somebody stands there with a stopwatch.

Month Three: We Found The Real Constraint

We went back and timed the changeover properly, stopwatch in hand, writing down every interval rather than the total.

The fixture swap was 22 minutes. Cleaning the machine between batches — removing the previous part's debris from the fixture base, the drain and the spray shadow — was 31 minutes, and nobody had counted it because it happened while the fixture was being changed and looked like part of the same activity. The first-part check was another 40 minutes, mostly waiting for the measurement.

So the real distribution was: 22 minutes of what we had optimised, 31 minutes of what nobody had named, and 40 minutes of waiting for a result somebody else produced.

Three changes followed. The fixture base was redesigned with a removable tray so cleaning it took four minutes instead of thirty. A variant identification step was added at load, so the program selected itself from the part rather than from the operator. And a first-article gauge was placed at the machine so the check no longer waited for the laboratory.

Month Four: What It Looks Like Now

Changeover now runs at about 24 minutes: 6 for the fixture, 4 to clean, 8 for the program and first part, 6 of margin that we deliberately left in.

First-pass yield recovered and ended slightly better than the starting point, because the variant identification removed a failure mode that had existed before the project began. The line now changes over more often than it did four months earlier, which was the actual objective and had never been stated as one.

The lesson we took from it is not about fixtures. It is that changeover time is usually attributed to the machine and is usually distributed across four owners: the fixture, the cleaning between batches, the program, and the measurement. Optimising the one that belongs to you produces a number that improves and a line that does not.

What I Still Do Not Know

The 6 minutes of margin we left in is a guess. We added it because the process felt tight, and the line has been stable since, but I cannot tell you whether 2 minutes would have been enough or 10 would have been safer.

The honest position is that we traded a measurable risk for an unmeasured one. If somebody has a way of sizing that margin rather than feeling it, that is a conversation I would like to have.