Engineering resource
AOI false calls: separating programming issues from defects
AOI false calls consume review time, but reducing the number of alarms is not sufficient evidence that an inspection process has improved. A useful change must reduce unnecessary review while preserving detection of the defects the program is intended to find. That requires agreed acceptance criteria, trustworthy review labels and a controlled comparison of the old and new programs.
For a turnkey PCBA project, the buyer does not need to prescribe every camera setting. The buyer and assembler do need a common understanding of what an alarm means, how it is resolved and what evidence supports a program change. This article describes that workflow, with particular attention to separating a programming issue from a genuine manufacturing defect.
Define the outcome categories before calculating a rate
An AOI call is a machine decision requiring the action defined by the inspection process. After review, it may be confirmed as a real defect, classified as an acceptable feature that was incorrectly flagged, or remain unresolved pending further evaluation. Keep those categories separate. An unresolved call does not become a false call simply because the production queue is growing.
Conversely, a defect found after the machine passed the feature is a different event: an escape from the applicable inspection check. Record where and how it was found and whether that defect was within the agreed coverage. A feature the equipment cannot inspect should be visible in the coverage plan rather than counted as though the program had successfully evaluated it.
The definitions must identify the counting unit. A board can contain several component calls, and one component can trigger multiple algorithms. Decide whether the report counts boards, components, joints or individual checks. Without that definition, two teams can report very different false-call rates from the same production run and both believe their arithmetic is correct.
Use a numerical example with an explicit denominator
Suppose a hypothetical run inspects one thousand boards and generates two hundred review calls. Review confirms forty actual defects and classifies the remaining one hundred sixty calls as acceptable features. If there are no unresolved cases in this example, eighty percent of the calls are false calls. The same data represent 0.16 false calls per inspected board. Those figures describe different quantities.
Neither figure is the percentage of boards containing a false call unless the record identifies which calls belong to which board. Several false calls may occur on one board. Nor can the data establish an escape rate, because the example contains no independent examination of machine-passed features. Report what was measured rather than attaching a broad quality claim to a convenient percentage.
When comparing runs, also account for product mix. A run containing more difficult packages may generate a different review workload even when every inspection program is unchanged. Compare the same program and relevant feature groups where possible. If the mix changes, show it in the report rather than implying the entire difference came from a software adjustment.
Check whether the image supports the decision
OMRON’s technical paper on board appearance inspection describes how shadows and secondary reflections can make optical information unavailable or unreliable. It also distinguishes false-positive and false-negative decisions. The practical implication is to inspect the measurement evidence before assuming a threshold change will solve an alarm problem.
For a recurring call, retain the relevant image or measurement view, reference designator, board identifier and program version. Ask whether the feature is clearly represented, whether registration is correct and whether neighboring geometry changes what the camera sees. A measurement from the wrong location cannot be repaired simply by accepting a wider range of values.
Separate an image-acquisition issue from an acceptance-rule issue. If the feature is not reliably visible, the engineering response may require a different supported inspection approach or another agreed check. If the evidence is reliable but the rule does not match the approved acceptance criteria, then a controlled programming correction is the relevant action. Document which situation was identified.
Make the review label trustworthy
A review decision should cite the applicable acceptance requirement and sufficient evidence to support the classification. The operator should not be asked to decide from memory that a particular alarm is always harmless. Repeatedly overriding the same location can conceal a genuine defect when the appearance eventually changes.
Create an escalation path for borderline cases. Preserve the original machine result while an authorized reviewer resolves the question. If additional inspection is required, attach its result to the same board and feature record. That history helps later analysis distinguish a confirmed false call from a case that was merely dismissed without adequate evidence.
Periodically compare reviewer decisions on representative cases using the approved criteria. Disagreement can reveal ambiguous instructions or an image that does not support confident classification. Resolve that issue before using the labels as training or optimization data. A program adjusted from unreliable labels may become more consistent at making the wrong decision.
Group recurring calls without hiding their context
Sort confirmed false calls by product, program revision, reference designator, package and inspection check. Add relevant information such as board position or material lot when it helps explain the pattern. Start with the groups that create the largest verified review burden, but retain smaller groups with potentially significant detection implications.
A hypothetical connector may generate repeated calls at the same reflective terminal across many boards. A different group might appear only after a component-lot change. These patterns suggest different investigations; they do not prove a cause on their own. Compare the images and process records before choosing whether to revise a package library, investigate incoming variation or correct the manufacturing process.
Keep real defects in a separate analysis stream even when they occur at the same reference designator. A location can have both false calls and genuine defects. Treating the entire location as a nuisance and disabling its check removes useful discrimination instead of improving it.
Document a programming change as an engineering change
State the observed issue, proposed adjustment and intended effect. Record which feature, algorithm or library entry is affected and whether the change applies to one program or several products. A shared library change deserves a review of every dependent program that could inherit the new behavior.
Preserve the previous approved program and a clear rollback route. The change record should identify the person who made the adjustment and the person who approved its release under the site’s process. Avoid overwriting the only copy of the baseline before a comparison has been completed.
Do not broaden an acceptance limit merely to reach a target pass rate. The program must implement the applicable product criteria. If those criteria need clarification, obtain that decision explicitly. A lower alarm count created by changing what the factory considers acceptable is not the same improvement as better recognition of already acceptable assemblies.
Challenge the candidate with both acceptable and defective cases
Use a controlled evaluation set that includes the acceptable variation responsible for the false calls and representative known defects within the relevant coverage. Keep the identity and expected disposition of each case. The candidate program should be assessed against both groups, not only against the images that motivated the adjustment.
Where the equipment supports evaluation using stored data, document the scope and limitations of that replay. A stored image comparison may help assess a decision rule, but it does not automatically validate new image acquisition, handling or registration behavior. Include physical inspection trials when the change or the risk requires them.
Report the candidate’s misses and unresolved cases openly. An evaluation set is finite and cannot prove that no future escape is possible. Its value is that it provides explicit evidence for the decision and reveals known tradeoffs before release. Agree further checks or restrictions when the evidence is not sufficient.
Monitor the released program over time
OMRON’s Q-upNavi feature description includes tracking pass rate and false-call rate over time by line or inspection program. Regardless of the software used, that separation is useful: a factory-wide average can hide a deteriorating program or a newly introduced component variant.
After release, compare the same defined metrics with the baseline and retain the observation period. Include confirmed defect counts, review workload and any downstream findings that relate to the changed check. If production conditions change, annotate the record so an apparent trend is not attributed automatically to the program revision.
Define a response when the result differs from the qualification trial. The process may require additional review, investigation or rollback according to the approved change plan. The response should preserve traceability to the boards inspected under each version, making later containment possible if a detection weakness is discovered.
Keep automated review under controlled supervision
Koh Young’s Smart Review description presents automated review and feedback as part of inspection optimization. Such capabilities are equipment-specific. Their availability should not be assumed for every assembly line, and a vendor’s general performance statement should not be quoted as the measured result of your project.
For any automated classification approach, establish who approves the reference labels, how changes are versioned and how uncertain cases are handled. Confirm which records remain retrievable for the actual boards. Automation can reduce repetitive review work, but the project still needs defined evidence for accepting the resulting inspection process.
Discuss the limits alongside the project’s other assembly testing options. A functional test, optical check and other inspection methods answer different questions. Their combined scope should be designed deliberately rather than assuming one favorable metric establishes complete assembly quality.
Ask for a useful inspection report in the quotation
During quality and traceability planning, agree the counting units, disposition categories, program version records and escalation responsibilities. Specify the evidence required for significant inspection changes. This gives both the buyer and assembler a practical basis for discussing improvement without demanding an unsupported promise of zero false calls or zero escapes.
When you compare quotation scope, check whether programming, first-article review and the agreed reporting are included. The useful deliverable is a controlled inspection process with understandable results. A reduced alarm count is valuable only when the records show how it was achieved and what detection evidence was preserved.