AOI false calls: what inspection review records are worth

What is an AOI false call?#

An AOI false call is a defect flagged by an automated optical inspection machine on a good solder joint or component, which an operator then has to review and reject. False calls cost review time, and when they pile up, operators trust the machine less and real defects can slip through with the noise.

Every flagged location goes to a review station, where an operator looks at the image and decides whether it is a real defect or a false call. The AOI software or the MES usually stores that verdict with the panel serial, the reference designator and the defect type the machine assigned.

False call reduction normally starts with program and library tuning: thresholds, component models and library entries. Some EMS teams add machine learning classifiers that pre-sort calls before an operator sees them. Both approaches need the same input, a history of calls with verdicts that can be trusted.

Why inspection review records are labeled data#

Inspection review records are labeled data because each one pairs an image with a machine prediction and a trained person's verdict. The operator's verdict is the label, the machine's call is a baseline prediction, and the repair or test outcome checks whether the label was right.

That structure is what model developers usually pay annotation teams to create. An EMS provider that kept its review history produced it as a by-product of daily work, with defect types grounded in real production rather than staged samples.

The records also span a long tail of components, packages, solder finishes and board materials, because every customer program brings its own mix. That variety is hard to assemble in one place, and it is also why customer rights need care.

The fields that make a review record useful#

A useful review record captures the board, the component, the machine's call, the operator's verdict and what happened next. The table lists the fields that matter most and what each one adds.

What raises or lowers the value of review history#

Review history gains value when verdicts can be trusted and traced, and loses it when they cannot. Internal teams and outside buyers ask the same questions: were verdicts made by trained people, were they checked, and do the records connect to outcomes?

Bulk clearing is a frequent problem. When an operator passes every call on a panel at once, the verdicts look like labels but carry no judgment, and they should be flagged or dropped before anyone trains on them.

How to make review records trustworthy#

Trustworthy review records come from a handful of configuration and discipline changes at the review station and in the MES. Most cost little once they are set up.

The payoff comes twice. Lower false call rates reduce review load on the line, and an archive with logged library changes and linked outcomes can stand up to an outside reviewer who asks how each label was produced.

• 1Store the verdict, the reclassified defect type and the reviewer role for every call, not only the panel's final pass or fail.
• 2Keep images, or at least crops around each call, for as long as customer agreements and storage allow.
• 3Log every program and library change with the date and the reason.
• 4Link review records to rework, in-circuit and functional test results by board serial.
• 5Audit a sample of verdicts on a regular schedule and record the result.
• 6Flag panels where every call was cleared at once.

Whose images are they?#

AOI images are records the EMS provider created, but they show the customer's design, so customer agreements decide how they can be used outside the company. A tight crop of one solder joint reveals little, while a full-board image can reveal layout, part choices and product identity.

Read the customer agreement for how it defines confidential information. Some definitions cover anything that reveals the customer's product, which would include images; others cover only documents the customer supplied, such as Gerber files and drawings. Counsel makes that call program by program.

A practical approach treats images in three tiers: verdict records without images, which are usually the easiest to clear; component-level crops, cleared program by program; and full-board images, which normally stay inside the company. Defense and export-controlled programs stay out entirely, and operator names become role codes in any shared version.

Illustrative: an EMS provider reviews its AOI archive#

Illustrative: Copperline Assembly is a fictional EMS provider building industrial and instrumentation boards on several SMT lines. Its AOI software stored calls and verdicts, but images were purged on a rolling basis, and repair actions were logged in the MES under a different board identifier.

The quality manager mapped the two identifiers, linked verdicts to rework and in-circuit results, and found a group of programs where false calls clustered on one package type after a library change. Tuning that library entry cut review load on those programs.

The same work made the archive reviewable. Copperline changed its retention to keep crops, excluded its defense programs and customers whose agreements restrict any use, and asked for a metadata-only fit check on the remaining verdict history.

How SourceX approaches inspection review records#

SourceX approaches inspection review records as labeled workflow data. Under the SourceX Enterprise Data Value Framework, human-generated signal from trained operators, domain expertise, scale across programs, recency and data cleanliness raise value, while reproducibility, preparation cost and privacy burden pull it down.

In the SourceX five-step transaction, Rights separates programs by customer agreement, Preparation removes operator names and customer identifiers, and the EMS provider approves every file before Delivery. The records are licensed, not sold, and the company keeps ownership.