AOI False Calls: 7 Ways to Cut Them Without Missing Defects

AOI false calls are the flagged images that turn out to be acceptable when a person looks at them. They slow the line, consume the attention of the review station and, in the worst case, train operators to pass defects without looking properly. The goal is not zero false calls, which would mean a blind machine, but a rate the review team can handle without fatigue.

Reducing them safely means understanding why they happen. Most come from the inspection program, the lighting, or the board itself, and only a minority come from the machine being unable to see a genuine defect. Treating the program as a fixed asset is what keeps the rate high.

What a False Call Actually Costs

The direct cost is review time, which is easy to measure. The larger cost is attention. When six out of ten flagged images are acceptable, the operator learns to expect acceptable images, and the one genuine defect in the batch is more likely to be released. That is the real price of a high false call rate.

There is also a throughput effect. A line that flags heavily is a line that queues at the review station, and the queue encourages faster decisions. Before long the escape rate rises and the inspection program that was supposed to protect the customer has become a formality that is passed rather than used.

Programming Choices That Set the Baseline

The inspection program is built from CAD data, a golden board or both. CAD-derived programs know where every pad should be, which makes them sensitive to placement and paste variation, while golden-board programs learn what the process actually produces and are more tolerant of normal variation.

Starting from the wrong source is the most expensive programming error. A CAD program run on a process that has not yet stabilised will flag everything, while a golden board captured from a marginal build teaches the machine to accept marginal joints. The program should be built from CAD and then taught with boards that passed a manual inspection.

<img src="https://www.gopcba.com/wp-content/uploads/2026/05/smart-energy-PCBA-1.jpg" alt="Operator at an AOI review station judging a flagged solder joint image” />

Lighting and Camera Settings per Joint Type

Different joint types need different illumination. A chip resistor fillet shows best under angled low light, a fine pitch lead needs a controlled top light, and a shiny lead-free fillet can saturate a camera that was calibrated for tin-lead. One lighting recipe for the whole board guarantees a compromise.

Camera gain, exposure and colour balance work with the lighting, and all of them drift with lamp ageing. Lamps dim over thousands of hours, so a program that was tuned on new lamps will slowly become less stable. Recalibration on a schedule, with a reference target, keeps the images consistent.

Libraries, Golden Boards and Tolerance Bands

Component libraries define how a part should look, including the acceptable range of body position, lead presence and fillet shape. Libraries built for one package type and copied to another are a common source of false calls, because the tolerance band no longer matches the part being inspected.

Tolerance bands should come from measured process capability, not from a guess. If placement varies by 0.1 mm across a shift, a band of 0.05 mm will flag half the boards for nothing. Measuring the real distribution and setting the band outside it removes those calls without hiding a trend that matters.

Thresholds: Balancing False Calls and Escapes

Every threshold is a trade. Lower it and the machine catches more subtle defects along with more good joints; raise it and the line runs smoothly while a few real defects pass. The right setting is a value the process can support, reviewed with data rather than adjusted when the queue grows.

The escape rate should be measured with the same rigour as the false call rate, using verification on a sample of passed boards. Without that number, any reduction in false calls looks like an improvement, even when it has been achieved by making the machine less sensitive.

The Review Station and Operator Judgement

The review station is where the decision is made, so it deserves as much attention as the machine. A clear image, the correct reference picture beside it and a defined accept or reject criterion turn a judgement call into a comparison. Without those, two operators will decide the same image differently.

Boundary samples help more than any instruction. Keeping physical examples of an acceptable fillet, a marginal one and a reject, and reviewing them at the start of a shift, aligns the team with the IPC-A-610 criteria the program is written against.

Repair Loop Feedback Into the Program

Every call that reaches the repair station produces information. Recording the defect type and the outcome lets the programming team see which calls were real and which were noise, and that data is what justifies a change to a threshold or a library.

Without that feedback the program never improves. Calls are reviewed, decisions are made, and the information is lost at the end of the shift. A simple log of defect type, board number and operator decision costs minutes and turns the review station into a source of program improvements.

Measuring Performance Properly

Track false calls per board and per thousand joints, along with escapes found at a later stage. Both numbers move together, and a program should be judged on the pair. A falling false call rate with a rising escape rate is a regression dressed up as progress.

Trend the results against the product mix and the shift. A rate that changes when a different board runs is a programming issue for that board, while a rate that drifts over weeks on the same board points at lighting, calibration or a process change upstream.

When the Machine Is Not the Problem

Some high call rates come from the process rather than the inspection setup. Paste volume that varies across the panel produces joints that genuinely differ, and no threshold can separate them cleanly. In that case the fix belongs upstream at the printer.

Others come from the board itself, such as a mask colour that reduces contrast, a finish that reflects unevenly or a layout that puts a tall part behind a low one. Where machine vision cannot see a joint at all, X-ray inspection covers the joints that AOI cannot, and the two are best planned as one inspection strategy.

Related reading: our fabrication notes, board quality and design release notes cover the same ground.

FAQ

What is an acceptable false call rate? There is no universal figure, because it depends on joint count and board complexity, but a rate the review team can handle without fatigue is the practical limit. Once operators begin passing images without proper scrutiny, the program has become unsafe whatever the number says.

Should thresholds be tightened to catch more defects? Only with evidence. Tightening without measuring the escape rate usually moves the failure from the inspection station to the customer, and it adds review labour. Verify on a sample of passed boards first, then adjust with data.

How often should the AOI program be re-validated? After any change to lighting, camera, software, stencil, paste or board layout, and on a periodic schedule even when nothing has changed. Lamps and mechanics age, so a program that was correct at commissioning will not stay correct on its own.