Aerospace SPC has a problem that automotive does not: volume. The methods were built for processes producing thousands of parts a shift, and a great deal of aerospace manufacturing produces eleven parts a month, on a five-axis machine, from a forging that took four months to arrive and is worth more than the car you drove in.
The standard that deals with this is AS9103, Variation Management of Key Characteristics. Its central demand is easy to state and hard to satisfy: the variation of a key characteristic must be managed, which means understood, monitored and reduced — not merely inspected against a tolerance at the end.
A key characteristic on a low-volume line
Part
Limits frozen on the first 20 parts. A fixture reset before part 21 moves the process up by about 0.012 mm; all eight parts made after it are beyond the upper control limit.
Every one of these 28 parts conforms. Part 27, at 6.3748 mm, is 0.0002 mm inside the drawing limit. AS9103 asks for the variation of a key characteristic to be managed, not merely inspected — and the chart called this at part 21, six parts before anyone measuring against the tolerance would have flinched.
Illustrative data — generated to behave like this process, not taken from anyone's plant. The limits, the signals and the capability indices above are computed at page load by the same engine that draws a customer's chart.
Twenty-eight parts. That is a realistic run for a component like this, and it is where the individuals chart (I-MR) earns its place: there is no rational subgroup, because consecutive parts are separated by hours and by setups, so a subgroup would be pooling things that do not belong together.
The tolerance is 6.350 ±0.025 mm. Every part conforms. Part 27, at 6.3748 mm, sits 0.0002 mm inside the drawing limit and still passes. A conformity-only inspection system would report 28 good parts and a perfect month.
What the chart shows is that after a fixture reset before part 21, this is not the same process any more. It moved up by about 0.012 mm and every part since has been outside the limits computed from the twenty parts before it. On material with this lead time, "we found it at part 21 rather than part 27" is not a statistics debate — it is whether you scrap one forging or six.
This is the distinction AS9103 turns on. Inspection asks whether the part is acceptable. Variation management asks whether the process that produced it is predictable, and only the second question can be answered before the parts exist.
What gets charted
- Machined features on rotating parts — blade root and tip dimensions, disc bore and rim geometry, seal diameters, wall thickness on thin-wall components.
- Assembly and joining — fastener torque, interference fit, rivet dimensions, bond line thickness, shimming.
- Special processes — heat treat furnace uniformity, anodise and plating thickness, shot peen intensity (Almen strip), NDT sensitivity.
- Composites — cure temperature and pressure, ply thickness, void content, resin content, autoclave profile.
- Additive manufacture — powder characteristics, melt-pool monitoring, density, build chamber oxygen.
- Engine and system test — flow rates, leak rates, thrust, vibration signature.
The low-volume problem, and what to do about it
The honest difficulty: three-sigma limits from twenty individual readings are not very precise, and with eleven parts a month it takes two months to have even that. Everyone in aerospace hits this. The workable answers:
- Baseline on parts, not on time. Twenty consecutive parts across three setups is a better baseline than six months of one part a week, because it contains the setup-to-setup variation you actually need in the limits.
- Chart the process, not the part number. A single spindle running a family of similar features gives you a usable series where each individual part number gives you five points. Normalise to deviation-from-nominal and chart the machine.
- Include setup variation deliberately. If setups are the dominant source of variation — they usually are on low-volume machining — a baseline that contains one setup produces limits that flag every subsequent setup. That is a chart guaranteed to be ignored within a month.
- Accept wider limits early and re-baseline once. Limits computed from twenty points and frozen are far more useful than limits that shuffle every week, even if they are 20% too wide.
- Say what your limits are based on. A limit set with its baseline window, author and date attached is defensible at an audit; a number in a spreadsheet cell is not.
Why the usual tooling hurts here
Excel. Same four failure modes as everywhere — moving limits, single author, no alerting, one or two rules implemented — with the aerospace addition that the audit trail matters more here than almost anywhere.
The CMM's own software. It reports the part. It rarely reports the process across parts, across setups and across months, which is the thing AS9103 is about.
Enterprise quality suites. The right answer for a prime or a large tier-1 with an integrated quality system. Disproportionate for a 40-person machining house that supplies them and has been asked to demonstrate variation management on nine key characteristics.
Where this platform fits
- Individuals charts as a first-class citizen, not a degenerate case of a subgrouped chart.
- Limit sets are versioned with their baseline window, author and timestamp — the evidence trail that makes a re-baseline a decision instead of an edit.
- Signals get acknowledged, given a cause and closed by a named person, so "what did you do about it" has a written answer.
- Tolerances and capability alongside the chart — Cp, Cpk, Pp and Ppk computed by an engine whose test suite reproduces the published worked examples, so the numbers on your first-article package and the numbers on your screen come from the same place.
- A share link for the customer's SQA who wants to see the last six months of a key characteristic without a licence or a meeting.
What we do not do
We chart the process. Several things around that are a different job, and are worth knowing before you plan around us:
- First-article inspection (AS9102). No FAI report generation, ballooning or bubble-numbered drawings — that stays with your FAI tooling.
- MSA and gauge R&R. AS9103 expects the measurement system to be qualified before a chart means much, and we assume you have done that elsewhere.
- CUSUM and EWMA. We draw Shewhart charts today. For a slow drift on a high-value part those would suit better, and they are on the list.
- Non-normal capability. Position, roundness, flatness and runout are bounded at zero, so a Cpk computed on a normal assumption flatters them. Worth handling in a statistical package until we support it properly.
- Configuration management and serial-number genealogy. Points carry timestamps and labels; full traceability belongs in your ERP or MES.
Where to start
Pick one key characteristic on one machine and export the last two or three setups' worth of measurements — twenty readings is enough to begin. Freeze the limits on that window and let the next month run against them. If nothing else, you will find out whether your setups are your dominant source of variation, which is the single most useful thing a low-volume shop can learn from a chart.
The clause-by-clause reading is on the AS9103 page.