On a cell line the dangerous defect is rare, faint and expensive to miss. Contamination, coating flaws and edge damage are safety-relevant, appear on a fraction of a percent of cells, and hide inside normal surface variation — so the inspection has to catch the anomaly it has barely seen without rejecting good cells it has seen millions of.
A rechargeable-cell line runs many lanes at speed, and the defects that matter — surface contamination, coating voids, edge and tab damage — are both rare and safety-relevant. Missing one is unacceptable; over-rejecting cripples yield. The inspection sits exactly on that tension.
The defects that matter most are the ones there are fewest examples of, which defeats a straightforward supervised classifier. And the surface has legitimate variation — sheen, texture, print — that a naive model reads as a fault, so the false-reject rate climbs the moment sensitivity is raised.
So the system is built around anomaly detection trained on good cells, imaging tuned to make the subtle defects visible in the first place, and a per-lane decision that holds the same accuracy across every lane rather than an average that hides a failing one.
Stated before any of them had an answer.
Contamination and coating voids appear on a fraction of a percent of cells, too few to train a classifier the usual way — yet they are the ones that must never pass.
Sheen, texture and print vary legitimately from cell to cell; a model that has not learned that variation rejects good product as defective.
Several lanes run in parallel with slightly different optics and lighting; one model naively shared across them drifts lane to lane.
An over-sensitive system destroys throughput and gets bypassed by operators, which is worse than a calibrated one that occasionally passes a borderline case for review.
You cannot train on a defect you have barely seen. Anomaly detection on good cells, plus imaging that reveals the flaw, is what catches the rare and safety-critical case.
The scope of the work, rather than a capability list.
Rarely the subsystem that sounds difficult.
The defects that matter most are the ones with fewest examples; framing the problem as anomaly detection rather than classification is what makes them catchable.
Small per-lane differences in lighting and optics quietly shift a shared model; per-lane calibration is unglamorous and essential.
The threshold is set against the real cost of each error, not an accuracy target, because an over-rejecting line gets switched off.
A defect with no contrast under the installed lighting cannot be learned; most of the accuracy is won in the imaging, before training begins.
Safety-relevant contamination and coating flaws detected as anomalies despite few training examples.
Throughput protected by a threshold set on cost, with borderline cells reviewed rather than dumped.
The same accuracy on every lane, not an average that hides a weak one.
A retraining loop that folds newly seen defects back in as the line runs.
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a detail would identify a customer it is omitted rather than approximated. More is available under a non-disclosure agreement, within the limits the customer has agreed.
Every item links to its own page.
Because the defects that matter most on a cell line — contamination, coating voids, edge damage — are rare, sometimes a fraction of a percent of production. There are too few examples to train a reliable classifier, and they are exactly the cases that must never pass. Framing the problem as anomaly detection, trained on the plentiful good cells, lets the system flag deviations it has never seen a labelled example of.
By tuning the imaging so real defects have contrast and legitimate variation does not look like a fault, by training on enough good-cell variation that sheen and texture are understood as normal, and by setting the decision threshold against the measured cost of each error rather than an accuracy score. Borderline cells are routed to review instead of being force-failed.
Because each lane has slightly different optics, lighting and alignment, and a single model applied blindly across them drifts — one lane over-rejects while another misses. Per-lane calibration keeps the accuracy consistent, and per-lane monitoring catches a single failing lane that a line-wide average would hide.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.