EN
News
Two Errors, Two Costs: Balancing False Calls Against Misses
Overview
Every inspection station makes two kinds of error. It misses a defect that is there, and it flags a part that is good. The two are produced by the same threshold, and moving that threshold trades one against the other. A specification that asks for both to be very low is not asking for a better machine; it is asking for a machine that does not exist.
What makes the trade manageable is that the two errors rarely cost the same. A missed defect on a safety component can cost a recall; a false call usually costs a few minutes of re-inspection. When the costs differ by orders of magnitude, the correct setting is obvious, and the specification should say so rather than demand both numbers be small.
This article sets out how the two rates interact, what each costs in practice, and how to write a specification that states which error the line can absorb. Big Bird Industrial (formerly Shimada Big Bird Industrial) has been building cleaning and deburring systems for automotive powertrain components since 2001, and the detection figures used here come from production stations.
1. The Two Errors
The terminology is worth fixing, because the two errors are often described with the same words by different people.
A Miss
A defect present on the part is not detected, and the part continues down the line. The miss rate is the fraction of defective parts that pass, and production stations commonly specify it in the region of one part in a thousand, with the tightest applications reaching one in ten thousand.
A False Call
A conforming part is flagged, and it leaves the line for review, rework or scrap. False-call rates on real surfaces sit between about one and seven percent depending on the defect class and the surface, and a station that reports zero is usually measuring something easier than the requirement.
2. Why They Move Together
A detector decides by comparing a signal against a threshold. Lowering the threshold catches more defects and flags more good parts, because the population of good parts overlaps the population of defects in whatever feature the detector measures.
The Overlap Is Physical
The overlap is not a modelling weakness that a better algorithm removes. A shallow scratch and a tool mark produce similar contrast; a small pore and a dust particle produce similar darkness. Where the two populations genuinely overlap, no threshold separates them perfectly, and the only way to improve both is to change what is measured rather than where the line is drawn.
Changing What Is Measured
Adding a second illumination, a second view or an air pulse can separate populations that a single measurement cannot. That is the legitimate way to improve both rates, and it is a hardware change rather than a tuning exercise.
3. What Each Error Costs
The costs are asymmetric, and quantifying them is what turns the trade into a decision.
The Cost Of A Miss
A missed defect travels. On a machined component it may be found at assembly, where the cost is a line stoppage; on a safety component it may reach the field, where the cost is a recall. The further it travels, the more the cost grows, which is why the miss rate is specified at the operation where the defect is cheapest to catch.
The Cost Of A False Call
A false call consumes a review, a re-inspection or a rework. It also consumes trust: an operator who sees good parts rejected repeatedly learns to bypass the station, and a bypassed station provides no protection at all. That second cost is the one that is never in the calculation and always in the outcome.
| Application | Typical miss rate | Typical false-call rate | Bias |
|---|---|---|---|
| Casting surface on a head | About 0.1% | 3 to 5% | Reject-leaning |
| Transmission housing | About 0.1% | About 5% | Reject-leaning |
| Crankshaft journals | About 0.1% | 5 to 7% | Reject-leaning, variable surface |
| Painted exterior panels | About 0.2% | 2 to 5% | Balanced |
| Bearing cap sequence | About 0.01% | About 1% | Strongly reject-leaning |
4. Where The Balance Should Sit
The correct balance follows from the ratio of the two costs, and the ratio is usually large.
When The Miss Is Catastrophic
On a sequence check, a safety surface or a hydraulic passage, a miss can become a field failure and a recall. The station should be biased heavily towards rejecting, accepting a higher false-call rate as the price. A bearing cap station running at one miss in ten thousand and one false call in a hundred is correctly biased.
When The False Call Is The Bigger Cost
On a cosmetic surface inspected at high volume, a false call sends a part to polishing on every shift, and the accumulated cost can exceed the value of the defects caught. Those stations are tuned closer to balance, which is why a paint station runs at a lower miss rate and a higher false-call rate than a safety check.
5. Measuring Both
A station is characterised by two numbers, and a supplier who quotes one is describing half of it.
The Validation Set Has To Contain Both
A test set of defective parts measures the miss rate. A test set of conforming parts measures the false-call rate. Without the second, a station that rejects everything scores perfectly on the first, which is why the acceptance test has to include normal production parts.
Measuring In Operation Rather Than At Handover
Both rates move after commissioning, as lamps age and as the part population changes. Tracking flags per hundred parts and confirmed defects per thousand is what shows the drift before the station is bypassed.
6. The Organisational Half
The technical trade is only half of the problem. The other half is what the organisation does with the flags.
A station whose flags are reviewed by someone with authority to release a part will be trusted. A station whose flags go into a queue that nobody clears will be overridden, and once overridden the false-call rate stops being a cost and becomes a habit. The review position is therefore part of the error budget, not an administrative detail.
7. Writing It Into A Specification
A specification that names both rates, states which one is allowed to move, and defines the test sets that measure them is a specification a supplier can design against.
A specification that demands both be minimal will be answered with a number that describes a laboratory rather than a line. The question worth asking is not what the lowest achievable miss rate is, but what miss rate the application needs and what false-call rate the line can absorb while maintaining it.
8. The Cost Of A Station Nobody Uses
The largest cost in the error budget is neither the miss nor the false call. It is the station that stops being used, because at that point the miss rate becomes whatever the process produces without inspection.
How A Station Stops Being Used
It happens in three steps. A false-call rate rises above what the line can absorb, so operators begin to expect rejections on good parts. The review queue grows because nobody has authority to release a flagged part quickly. Someone discovers the override, and thereafter every ambiguous part is passed manually. None of those steps is a decision; each is a response to pressure.
The countermeasure is organisational rather than technical: a named reviewer with authority, a target review time, and a monthly figure for how often the station was bypassed. A station that is reviewed and trusted can run at a false-call rate the line tolerates. One that is bypassed has no error rate worth measuring.
Frequently Asked Questions
Can A Station Have Both Rates Low?
Only where the defect and the background do not overlap optically. Where they do, improving both requires adding information, such as another illumination or view, rather than tuning a threshold.
Which Rate Should I Specify First?
The miss rate, because it is set by what the defect can cause downstream. The false-call rate follows from what the line can absorb while achieving it.
Why Do False-Call Rates Seem High?
Because they are measured on real production surfaces with real variation. A rate in the low single digits on a casting or a painted panel is normal, and a rate reported as zero usually means the test set was too easy.
What If Operators Ignore The Station?
That is the real failure mode. It is caused by an unreviewed flag queue or a false-call rate the line cannot absorb, and it turns a working station into a record-keeping exercise.
Conclusion
A miss and a false call are produced by the same threshold and cost very different amounts. The useful specification states which one the line can afford, rather than asking for both to be small.
Improving both at once is a hardware question, not a tuning question. And the rate that ultimately matters is the one measured in operation, because a station nobody trusts has a miss rate of one hundred percent.
