Reduce AOI False Calls Without Hiding Real Defects

A high automated optical inspection call rate can overload reviewers, but a low call rate does not by itself show that inspection improved. A program can appear efficient after its thresholds are relaxed while allowing defects to escape. Evaluate nuisance calls and defect detection as separate outcomes.

Build a reviewed evidence set

Save representative images or measurement records for each recurring call category. Include confirmed acceptable assemblies, confirmed defects and cases that require another inspection method. Record the component, location, population variant and the reviewer decision. Do not use an unreviewed production pass label as ground truth.

For a solder joint call, distinguish appearance variation from an actual acceptance issue. For a missing-component call, check whether the location is intentionally unpopulated on that variant. A wrong variant program is a configuration error; adjusting image thresholds is an ineffective way to correct it.

Look for the common cause

Group calls by package, board position and inspection condition. A cluster near one edge may suggest support or image acquisition issues. Calls concentrated on one reflective component may need a package-specific inspection strategy. First verify focus, lighting, board alignment and the selected library before editing acceptance criteria.

Koh Young’s discussion of AOI classification describes the use of inspection measurements and classification to address false calls. Specific machine features differ, so qualify the actual inspection program and review process rather than assuming an advertised feature guarantees coverage.

Change one defined inspection behavior

State the proposed change and its intended scope. For example, a package-specific adjustment should not silently alter every component using a shared library entry. Identify which defect modes the affected inspection step is expected to detect and which other checks, if any, also cover them.

Run the revised program against the retained evidence set. Then review a controlled production sample using an independent method appropriate to the feature. If the suspected defect is hidden from optical view, AOI images alone cannot establish its absence. Record that limitation instead of forcing a visual classification.

Track reviewer behavior as well

Check whether reviewers consistently resolve the same examples. Excessive nuisance calls can encourage habitual acceptance, while unclear criteria create disagreement. Use a small set of annotated examples to align decisions and route ambiguous cases to engineering.

Keep the previous program, change reason, comparison results and approval record. Monitor both call rate and confirmed defect outcomes after release. A useful improvement reduces avoidable review while retaining demonstrated detection of the faults the inspection step is meant to find. An unexplained reduction in calls is a reason to examine the evidence, not automatically a production success.