Automated inspection stations: false rejects, escapes and what happens to the reject

Prove the measurement before you believe the result

An inspection station produces numbers or verdicts that look authoritative from the moment it is switched on. Before relying on them, run a capability study: present the same parts repeatedly to see how consistent the station is with itself, present known good and known bad parts to see whether it separates them, and compare against an independent measurement traceable to a recognised reference. Do it across the range of conditions the station will meet, not once at commissioning on a warm afternoon. Metrology institutes exist precisely because measuring the measurement is a discipline, and a station never checked this way is an expensive source of confident nonsense.

The two errors and the honest trade between them

Tightening limits reduces escapes and increases false rejects. Loosening does the reverse. There is no setting that removes both, and pretending otherwise leads to a limit chosen by whoever complained most recently. Make the trade explicitly: what does an escape cost in this application, in warranty, recall exposure, customer credibility or safety, and what does a false reject cost in scrapped material and disrupted flow? Safety-related characteristics justify a high false reject rate; cosmetic ones rarely do. Write the reasoning down with the limit, because the pressure to loosen arrives during a bad week and needs something to push back against.

Overrides are the failure mode that hides itself

When a station rejects too much, people find ways round it: a bypass switch, a habit of re-presenting the part until it passes, an inspection turned off during a rush and never turned back on. Each of these is rational from the operator's position and destroys the station's value silently. Two countermeasures work. First, treat a high false reject rate as an engineering defect to fix rather than an operating nuisance to absorb. Second, make bypass a deliberate, logged, authorised act with a defined containment plan, so that using it produces a record somebody reviews rather than a quiet gap in the evidence.

Rejected material has to go somewhere defined

A surprising number of stations reject correctly into a bin that then gets tipped back into the line. Design the physical path: a reject route the part cannot leave without a decision, a container that cannot be confused with good stock, and a documented rule for what happens next — scrapped, reworked and re-inspected, or reviewed by a competent person. Where rework is permitted, the re-inspection must be as rigorous as the original, since a part that has been through the station twice and passed on the second attempt is either genuinely fixed or evidence that the station is marginal on that feature.

Rejects are data about the process, if anyone looks

Every rejected part is a measurement of something that went wrong upstream, yet most stations record only a count. Capturing which characteristic failed, by how much and when, converts inspection from a filter into a diagnostic instrument: a drift appearing after each changeover, a failure concentrated on one cavity or one supplier batch, a cluster following a tool change. That is the difference between paying for a machine that removes bad product and paying for one that reduces the making of it. It requires that the failure reason be recorded in a structured way at the moment of rejection, not reconstructed from a bin at the end of a shift.

Frequently asked questions

Can automated inspection replace human inspectors entirely?

Only for characteristics the equipment genuinely measures well, and there are usually fewer of those than a project assumes. People bring context: they notice the unfamiliar defect nobody programmed for, they hear the machine sounding wrong, and they connect a mark to something that happened upstream. A common structure keeps automated inspection on the defined, repeatable checks at full rate, and retains skilled human review for sampling, for new products and for anything the station flags as borderline.

How often should an inspection station be verified?

Frequently enough that you can bound how much production is affected if it has drifted. Many plants run a known reference part at each shift start and after any adjustment or maintenance, with a fuller study at defined intervals or after any change to lighting, optics, fixturing or software. Record the result rather than just acting on it, because the record is what lets you scope containment when a check finally fails and shows how long the drift had been developing.

What should we do when the station rejects far more than the process is producing as defects?

Treat it as a fault in the station, not as cautious behaviour. Investigate presentation first, since inconsistent positioning is the usual cause, then illumination or sensor degradation, then whether the limit reflects a real requirement or an arbitrary tightening. Loosening the limit without understanding why should be a last resort and needs the same authorisation as any other change to a quality control, because it moves risk to the customer rather than removing it.