What is in process inspection: in-process inspection explained, from first-off checks to hold points

In process inspection is the checking of work while it is still in production, at defined points between operations, so that a nonconformity is found and corrected before more value is added to the part or before a later operation buries it. The characteristic being checked is chosen because it can drift, and the point at which it is checked is chosen because that is where finding the drift is still cheap.

Written both ways, in process and in-process, and also called process inspection, on-line inspection, or patrol inspection depending on who performs it. It sits between incoming goods inspection at one end of the flow and final inspection at the other.

Where do the check points go?

• First off, after every change. New setup, tool change, material batch change, shift change. In aerospace this is formalised as first article inspection under SAE AS9102, whose stated purpose is to provide objective evidence that all engineering design and specification requirements are understood, accounted for, verified and documented.
• Operator self-check at a defined frequency. Every nth part, recorded, with a defined reaction if it fails.
• Patrol inspection. A quality inspector sampling across several stations on a route, independent of the operator.
• Hold points before concealing or irreversible operations. Before painting, potting, closing a weld, backfilling a trench, plastering a wall. Once the operation is done, the characteristic is unverifiable.
• Statistical process control. Measure a characteristic on a small sample at a fixed interval, plot it on a control chart, and act on the pattern rather than on any single reading. Shewhart's distinction between common cause and special cause variation is what stops operators adjusting a stable process into instability.
• Last off. The final part of a run, retained and measured, to bound the period the run's output can be trusted for.

How the inspection interval sets your containment size

This is the calculation that usually gets skipped. When a check fails, everything produced since the last passing check is suspect. So the interval between checks is not really an inspection decision, it is a decision about how much product you are willing to quarantine when something goes wrong.

A worked example

A CNC cell produces 400 housings a shift on one setup. A cutting tool wears and the bore drifts out of tolerance somewhere in the middle of the run. With an operator check every 20 parts, the worst-case suspect population is the 20 parts since the last good check, plus the part in hand: they get sorted, and the run continues after a tool change. With no check until final inspection, the suspect population is the entire 400, all of which have to be screened, and by then some have moved to the next operation. Same failure, same tool, twenty times the containment. Halving the interval to every 10 parts halves the exposure and doubles the checking labour, which is the trade the process engineer is actually being asked to make.

Why it costs less than catching it later

The standard heuristic is 1-10-100, from Making Quality Work by George Labovitz, Yu Sang Chang and Victor Rosansky (1993): about one unit of cost to prevent a defect, ten to fix it internally, a hundred once the customer finds it. It lines up with the four cost of quality categories ASQ works with, where an in-process check is an appraisal cost that stops an internal failure cost, and an internal failure cost is the cheap version of an external one. The exact multiplier is industry specific and nobody should quote it as a measured figure, but the ordering holds in every cost of quality study: the further downstream detection happens, the more of the escalation you have already paid for.

The mechanism behind the ordering is simple enough. Every operation after the defect was created adds material, labour, machine time and lead time to a part that was already scrap, and each one makes the fault harder to reach.

What an in process check needs to be worth running

• Measurable acceptance criteria at the point of use. A figure and a tolerance, not an adjective. "To a good standard" cannot be passed or failed, only argued about.
• A gauge that can discriminate. Measurement system analysis exists for this: gauge repeatability and reproducibility for measured characteristics, attribute agreement analysis for pass or fail calls. A check with more measurement variation than process variation generates noise and false confidence.
• A written reaction plan. Stop, quarantine back to the last good check, adjust, verify, restart. Without it the operator's only options are to ignore the reading or to stop the line, and one of those happens far more often.
• Records that get trended. A drifting characteristic is visible in a run chart weeks before it is visible in a reject.

Where this fits in modern quality thinking

Deming's third point was to cease dependence on inspection to achieve quality, on the grounds that quality cannot be inspected into a product and has to be built in. In-process inspection is still appraisal rather than prevention, so it does not escape that critique. What it does is place the appraisal close enough to the process, in time and in physical distance, that the result feeds back and changes the next part. That is the difference between control and sorting, and it is why the check every twenty parts earns its keep while the screen at the end mostly buys reassurance.

Related reading: sample inspection, quality assurance inspection, and inspection frequency.