SPC for custom parts: control charts in low-volume production

Why SPC feels awkward on custom parts

Classical statistical process control assumes a long run of one part, enough data to estimate a stable mean and spread, and a subgroup of four or five pieces taken close together. A custom part shop has the opposite conditions. A run may be fifty pieces, the next order is a different part, and the same machine makes five families in a week.

That does not make SPC useless. It makes the textbook version a poor fit. The principle still holds: a process has a natural variation, and a chart shows when the variation stops looking natural. What changes is the chart type, the way the subgroup is formed and the honesty about how much the data can support.

What a control chart actually does

A control chart plots a measured value or a subgroup statistic against time, with a centre line and limits calculated from the process itself. The limits describe the range in which the process normally runs. A point outside them, or a run of points on one side, is a signal that something changed: a tool wore, a coil lot shifted, a gauge drifted, an operator changed a setting.

The limits are not the specification. A process can be in control and still produce out of tolerance parts, because control describes consistency, not capability. A process can also be capable and out of control, because a special cause pushed it off its usual level. Keep the specification limits on the chart as a separate reference, and never calculate control limits from them.

Choosing the chart

For a measured value, the choice is between a chart of individual readings and a chart of subgroup averages. Individuals and moving range suits a slow process where each part takes minutes and a subgroup is impractical. An average and range chart suits a run where four or five consecutive parts can be treated as one snapshot of the process.

For a pass or fail characteristic, an attribute chart is simpler and often more honest. A p chart tracks the proportion nonconforming when the lot size changes. An np chart suits a constant lot size. Counts of defects per part use a c or u chart. Attribute charts need more data than variable charts to detect the same shift, which matters when the run is short.

Rational subgrouping at low volume

A subgroup should capture the variation you want to see, and exclude the variation you want to detect. In a low-volume run, the useful subgroup is a set of consecutive parts made under the same conditions. That shows the short-term capability of the setup. Mixing parts from before and after a tool change in the same subgroup hides the change inside the spread.

For a fifty-piece lot, a subgroup of five consecutive parts every ten pieces gives five subgroups. That is enough to start a chart and to see a trend, but not enough to set firm limits. Use the first two or three lots of the same family to build the limits, then apply them to later lots and review the limits when the process changes.

Short run methods for many part numbers

When the same machine makes twenty part numbers with different nominals, a separate chart for each is unworkable. The short run answer is to chart the deviation from nominal, divided by the tolerance or by the historical standard deviation. The values from different parts then sit on a common scale, and one chart can monitor the machine.

The discipline is in the reference. Each part needs a nominal, a tolerance and a reference standard deviation that is relevant to its size and material. A deviation chart that uses the wrong reference will show a false signal or hide a real one. Document the reference for each part and review it when the process or the material changes.

Capability with small samples

A capability index is a ratio of the tolerance to the process spread. With twenty pieces, the estimate of the spread is uncertain, so the index is uncertain too. Reporting a Cpk of 1.67 from a twenty-piece sample implies a precision the data does not have. Report the sample size, the period covered and the confidence interval where you can.

Ppk uses the overall standard deviation and is the usual measure for a first study on a new process. Cpk uses the within-subgroup variation and suits a process in control with rational subgroups. Do not mix the two in one comparison. For a new part, a first study over several lots is more informative than a single number from the first article.

The reaction plan matters more than the chart

A chart with no reaction plan is a wall decoration. Write down what happens when a point falls outside the limit, when seven points run on one side, when a trend of six points climbs, and when the spread suddenly widens. For each rule, name who checks, what they check and what they are allowed to change.

The common failure is overreaction. An operator who adjusts the machine after every small movement adds variation rather than removing it. The rules exist to separate a real signal from noise, and the plan should say explicitly that a single point inside the limits is not a reason to adjust. That one sentence prevents most of the damage.

• One point outside a control limit: stop, contain, check the setup and the gauge.
• Seven or eight points on one side: investigate a shift in material, tool or setting.
• Six points rising or falling: check tool wear, temperature or gauge drift.
• A sudden increase in spread: check clamping, feed, lubrication or part support.
• A single point inside the limits: record it, do not adjust.

What a buyer should ask for

Ask which characteristic is charted, why that one, and how the subgroup was formed. Ask for the limits, the data behind them and the date they were calculated. A chart with limits but no data is a picture. A chart whose limits were set from the specification is a different tool altogether and should be labelled as such.

For a low-volume part, a control chart is one piece of evidence among several. Pair it with the first article, the gauge study and the inspection record. The chart shows whether the process is stable across the run. The other documents show whether the stable process is making the right part.