What is a false reject, and how do you count it
A false reject is a good piece marked as a defect. A miss is a real defect that was not marked. Count false rejects over all alarms, and over the number of good pieces. Without both numbers you cannot decide to run. Fewer false rejects do not hide misses.
A false reject is a good piece marked as a defect. A miss is a real defect that was not marked. Count them apart: false rejects divided by all alarms, and false rejects divided by the number of good pieces. Saying the system is sharp without those two numbers does not tell you whether to switch it on. A miss is not hidden by reporting fewer false rejects.
Two denominators
In one hour there are 100 alarms, and 40 of them open as good. False rejects are 40% of alarms. If that hour also has 20,000 good labels, false rejects as a share of good pieces are much smaller. The first number tells you whether the person will be drowned. The second tells you whether many good pieces will be wasted. You need both. A miss is counted from known defects you inserted that did not alarm. The denominator is those known defects, not the whole production.
With no defect samples put in on purpose, there is no miss data. Do not write the miss rate as zero.
How to record it
| Name | Numerator | Denominator |
|---|---|---|
| False rejects among alarms | Times a good piece alarmed | All alarms |
| False rejects among good pieces | The same count | Good pieces in that stretch |
| Misses | Known defects that did not alarm | Known defects put in |
Check before the job runs
- Run a stretch of known-good material and count the red boxes.
- Put in defect samples and count how many did not alarm.
- Decide the threshold only when both columns have numbers. Do not optimize only one column.
