Can You Use SPC on Machined Parts Made in Lots of a Few Hundred?

Why does the textbook control chart struggle with small lots?

Because it needs many in-control subgroups to set trustworthy limits, and a lot-based supplier produces a few short runs separated by restarts. Limits built from too little data do not behave the way the chart assumes.

The NIST/SEMATECH handbook quotes Shewhart's rule of thumb for setting up an X-bar chart: a sequence of not less than twenty-five samples of size four that are in control (NIST, variables control charts). The same page warns that when limits are not computed from a large amount of data, the actual properties might be quite different from what is assumed.

A monthly lot breaks both halves of that. The count is short: a run of a few hundred parts gives few subgroups, and sampling four parts at a time uses up a large fraction of the lot. The stream is also interrupted: each restart brings a new setup, new tool offsets, possibly a new material heat lot and a different operator, so the variation between lots can be larger than the variation within one. A chart that restarts its limits for every lot learns nothing from the last one, and the limits it draws from a few points are loose in unpredictable ways.

The design of SPC goes back to high-volume production, and a 2024 Durham University thesis on low-volume SPC notes that methods for low-volume settings have been adapted since the 1970s but that a comprehensive guideline covering all short-run cases is still missing (Medina-Garcia, Durham University). That is an honest picture of the field: there are workable tools, but no single method that fits every lot size, so the agreement with the supplier has to say which one is used.

How does an individuals chart work with one measurement at a time?

It plots each measurement in the order the parts were made and uses the difference between consecutive readings, the moving range, to estimate the process spread. Limits sit three sigma either side of the center line.

NIST describes the individuals chart for sample size one: the moving range of two successive observations measures the variability, and the control limits are the average plus or minus three times the average moving range divided by 1.128, the constant for pairs. The page also says the average can be replaced by a standard or target when one is available (NIST, individuals control charts).

In practice, this means measuring a feature on the first part after setup and then at fixed intervals through the lot, and recording each reading in sequence with the lot number, machine and setup. Because the chart is built on single readings, the measurement system matters more than it does for averaged subgroups: gauge noise shows up directly as chart noise. The fixed intervals matter too, because readings taken only when an operator feels like it will hide drift.

The practical point for a monthly supplier is where the chart lives. Keep one chart per key feature and keep adding the next lot's readings to it instead of starting a fresh chart each time. After a few lots, the chart has enough points to say something, which no single lot could do alone.

How should a supplier report Cpk when the data come from several lots?

Report it cumulatively across lots, only after the chart shows the process is stable, and state the sample count and lot list next to every number. A Cpk from one small lot is a rough indicator, not a capability.

NIST defines process capability as comparing the output of an in-control process to the specification limits, so a stable chart comes first and the index second. It describes about 50 independent data values as generally large enough for an estimate, says the approximate confidence-limit formulas are valid only when the sample is at least 25, and notes that the variation in capability indices is not negligible (NIST, process capability). A figure built from a few dozen readings from one lot can therefore be off by a wide margin and should be labeled that way.

The cumulative approach reads like this: after each lot, the supplier adds that lot's readings to the running set for the feature, confirms the chart is still stable, and reports Cpk with n and the number of lots. Early reports are labeled preliminary. If a lot shows a shift, the report says so and splits the data before and after the change, since a shifted process breaks the pooling assumption. The difference between short-term Cpk and long-term Ppk is covered in the Cpk versus Ppk guide, and what to do when the number is low is in the guide to improving Cpk.

If the data are not close to normal, which is common for features with a one-sided limit such as flatness or runout, NIST suggests transforming the data or using nonparametric indices. Ask the supplier which was used.

How do you ask for this in an RFQ without over-specifying?

Mark the key features on the drawing, say you expect charted data on those features across lots, and ask the supplier to describe what they can supply. Let the quote answer before the contract demands it.

A request that every dimension carry a Cpk is read by most suppliers as a reason to raise the price or decline. A request that names three or four features and asks how readings will be recorded and shared gets a real answer. Ask also how the supplier handles lot-to-lot change, such as a new heat of material, because that is where small-lot processes shift.

When you send a drawing for CNC turning or CNC milling to Saola, mark the key features and the reporting you expect in the quote request, so the reply can say what data will come with each lot. The same marking tells any supplier what you will check at incoming inspection.