Rejects double the morning a new coil or resin lot goes on

The material did not necessarily get worse. It got different, and an inspection system trained on how the old lot looked is measuring departure from that appearance, not conformance to a drawing. A coil 5 gloss units brighter, a resin with a fractionally different pigment load, a plated layer 2 micrometres thicker — each of these moves the whole population of scores at once, and a fixed threshold that sat comfortably above the old distribution now cuts through the middle of the new one. Every part is affected simultaneously, which is why the pattern is a step from 1.2 percent to 5 percent at 06:00 rather than a drift over a week.

That is the likely explanation, not the certain one. The awkward possibility is that the lot really is defective and the system is doing exactly what it was bought for. Both hypotheses produce the same headline number, and the two checks below separate them in a single shift.

The overlay that turns 'it got worse' into a timestamp

1. Pull hourly reject rate per part number for the last 60 days. Per part number matters: mixing part families averages away the step you are looking for.
2. Pull goods-receipt records for the input materials over the same period, with the timestamp at which each lot was first issued to the line — not the date it arrived in the yard.
3. Plot them on one time axis. You are looking for a step, not a slope, and for whether the step edge lands within an hour or two of a changeover.
4. Check the counter-example. If the previous three changeovers produced no step, the material is not automatically the variable — it may be this lot in particular, which is a supplier conversation rather than a model one.
5. Note what else happened that morning. Shift start, a lamp replacement, a fixture change and a lot changeover often coincide, and only one of them is the cause.

If the step does not align with a lot boundary, look elsewhere before you go further. A reject rate that follows the clock rather than the goods-receipt log is an illumination problem, diagnosed in the model that finds the scratch on day shift and misses it after dark.

Worse, or merely different: the 30-part retest

Take 30 consecutive rejects from the new lot to the inspection bench and have them judged against the written class list by someone who does not know why they are being asked. Count how many the human confirms as genuinely non-conforming. That single number decides everything that follows, and it takes under an hour.

• Nearly all confirmed. The lot is worse. The system found it, the escape risk was real, and the conversation is with purchasing and the supplier — not with engineering.
• Nearly none confirmed. The lot looks different and remains good. This is overkill, it is expensive, and the fix is revalidation of the inspection against the new appearance.
• A genuine mix. Usually two things at once: a real quality shift plus an appearance shift. Split the rejects by defect class before deciding, because one class may be real and the rest cosmetic.
• Disagreement between two inspectors on the same parts. Stop and fix that first — a class list two people read differently cannot arbitrate anything, as labelling defect images so two inspectors agree sets out.

The four properties that move a score distribution

The second row is the one that surprises people most. A monochrome sensor's spectral response is not the eye's, and neither is the station's illumination, so two lots a colour inspector signs off as identical can land 20 or 30 grey levels apart, out of 255, in the image the model actually receives. If your defect judgement rests on absolute brightness anywhere in the pipeline, colour drift will find it. An 8-bit pipeline has no headroom to hide that in.

A model trained on one lot has learned what that lot looks like, and a new lot is a new question. Widening the definition of normal until it accommodates both is not robustness, it is a slow surrender of the detection you paid for.

Lot-aware thresholds, and when they are the wrong answer

The tempting fix is a per-lot threshold: recalibrate on the first parts of each lot and carry on. It works, in a bounded set of circumstances, and it is dangerous outside them. Recalibrating against the incoming population assumes that population is mostly good; if the lot carries a systematic defect, an auto-tuned threshold quietly learns to accept it and the escape rate goes up while the dashboard looks better than ever.

The alternative to per-lot thresholds is per-lot revalidation with a fixed threshold — slower, more disruptive, and much easier to defend when a customer asks how the accept criterion was set on the day their part was made.

The standing control: a revalidation run on first use of a lot

1. Trigger it from goods receipt, not from a person noticing. The first issue of a new lot to a line should raise the run automatically — the kind of small event-driven automation work that is worth building once and forgetting.
2. Run 50 parts under normal conditions and capture every score, not just the dispositions.
3. Compare the score distribution against the running baseline for that part number: the median, the spread, and the 95th percentile against the threshold.
4. Hand-inspect every reject in the run, and a random sample of 10 passes. Record agreement between the human and the system as a number.
5. Check the illumination proxy at the same time — the background grey-level statistic — so a lighting change is not misread as a material change.
6. Exit with one of three states: release, release under an operator gate with elevated sampling, or quarantine the lot for engineering. Write the state, the evidence and the owner into the lot record.

Fifty parts and an hour is the price. It is small compared with a shift of overkill, and much smaller than the alternative of finding out from a customer. Where a lot fails the run, resist the reflex to retrain immediately: capture the images, keep the lot separate, and decide whether the correct answer is a lighting change, a fixture change, a specification conversation with the supplier, or genuinely new training data. Presentation is often the cheapest lever, which is why fixturing a part so the camera sees the same thing twice is worth reading before the training set is opened.

The traceability that makes any of this possible

Every method on this page assumes one thing: the lot identifier is attached to the inspection record at capture. If it is not — and in most plants it is not — you cannot overlay anything, because the images know the time and the camera but not the material. The lot number lives on a printed traveller, gets copied by hand into a spreadsheet at the end of the shift, and never reaches the system that took the picture. That gap is the subject of why the floor keeps printing the paper copy, and closing it is a prerequisite rather than an improvement.

• Stamp lot, supplier, goods-receipt reference and issue time onto every image record, alongside model version and threshold version.
• Keep the score even when you discard the image. Score histories are tiny and they are what the overlay is built from.
• Retain enough per-lot history to establish a baseline. 1 previous lot is an anecdote; 6 give you a spread you can quote a tolerance against.
• Keep the corpus where the plant can use it. Revalidation and retraining both need images, scores and the model in the same place, which is why plants that will not send video off site reach the same conclusion as teams weighing running models on their own infrastructure.
• If the answer to a lot change is a second camera or a higher resolution, check the timing budget before committing — the constraints in it keeps up on the bench and falls behind at line speed apply to every added view.
• And rule out acquisition first if images look wrong rather than merely different, using half a part in frame: trigger and encoder faults.

Treated properly, incoming material becomes a first-class variable in the inspection system rather than an unmodelled input that occasionally ruins a Tuesday. The rest of the decisions around it sit in visual inspection and defect detection, part of our manufacturing and industrial vision work.

• material variation
• inspection
• revalidation
• quality