The problem with solder inspection is not finding defects — it is not crying wolf. A populated board has thousands of joints, most cosmetically imperfect and functionally fine. An inspection that flags every deviation drowns the operator in false calls and gets ignored; the engineering is telling a real defect from harmless variation.
Automated optical inspection of a populated board looks at thousands of solder joints and components per panel. Bridges, insufficient solder, tombstones, misplacements and lifted leads all matter; but so does the fact that most joints are cosmetically variable and perfectly acceptable.
The failure mode of AOI is not missed defects — it is the false-call rate. A system tuned hot flags harmless fillet variation as a fault, an operator has to clear each one, and within a shift the operators stop trusting it. A calibrated system is judged on how few good joints it questions.
So the inspection is built to classify rather than merely detect: multi-angle imaging that captures joint shape, models trained on the real distribution of acceptable joints, and a decision weighted by the true cost of a false call versus an escape.
Stated before any of them had an answer.
Detecting defects is easy; the value is a low false-call rate, because an operator swamped with false alarms stops trusting and clearing them.
Acceptable joints vary enormously in fillet shape and sheen; a model that has not learned that range flags good boards as defective.
A single top-down image hides a lifted lead or a hollow joint; capturing the three-dimensional shape needs multi-angle or structured illumination.
Some defects are common, others rare; training has to handle the imbalance rather than optimise for the frequent case and miss the dangerous one.
AOI is judged on false calls, not detections. The work is a model that knows the full range of a good joint, so it questions only the ones that matter.
The scope of the work, rather than a capability list.
Rarely the subsystem that sounds difficult.
The entire value of AOI is trust; getting the false-call rate low enough that operators keep using it is the hard, unglamorous work.
A good joint has enormous acceptable variation; capturing that range in training is what stops the system from flagging healthy boards.
Lifted leads and hollow joints do not show top-down; multi-angle or structured imaging is needed, which complicates capture and speed.
Some of the worst defects are the least common; the training has to weight them by consequence, not by how often they appear.
A system operators trust because it questions good joints rarely, not one they mute by shift two.
Bridges, insufficient solder, tombstones and lifts named, not just flagged as anomalies.
Lifted leads and hollow joints seen through multi-angle imaging.
Operator verdicts folded back so the false-call rate keeps falling.
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 missed defects are rare and false calls are constant. A populated board has thousands of joints, most with harmless cosmetic variation, and a system tuned to catch everything flags a flood of good joints. Operators have to clear each false call, and once the rate is high they stop trusting the machine and start rubber-stamping it — so the engineering target is a low false-call rate at an acceptable escape rate, not maximum sensitivity.
By teaching the model the full range of what an acceptable joint looks like, which is wide, so that only genuine deviations stand out; by imaging the joint from multiple angles so its actual shape, not just its top-down appearance, drives the decision; and by classifying the specific defect type rather than flagging a generic anomaly. The threshold is then set on the real cost of a false call versus an escape.
Because some of the most important defects are three-dimensional. A lifted lead, a hollow joint or insufficient solder can look acceptable from directly above and only reveal itself from an angle or under structured light. Multi-angle capture is what lets the system see joint shape, at the cost of a more complex and carefully timed imaging setup.
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.