AOI Programming: How the Golden Board Sets Inspection Limits
An AOI programming workflow usually begins with a physical sample rather than a CAD file, and that sample is the golden board. It is a known-good assembly that captures the process in its nominal condition, and every limit in the inspection recipe is referenced back to it. The golden board does not describe what should be acceptable in theory; it records what the line produces when every parameter is centred.
Automated optical inspection compares images instead of measuring absolute values, so the reference defines the centre of the acceptance window. A program written without one drifts toward whatever the first production panels happen to look like, which is a weaker and less repeatable standard. The reference is what makes a recipe portable between shifts, machines and sites.
What the Golden Board Is
The board is a fully assembled unit drawn from a run that has already passed functional test, X-ray inspection and electrical verification. Its joint shapes, component positions and fillet profiles have been confirmed by independent methods, so the AOI recipe can treat it as ground truth. Every deviation detected later is measured against that confirmed state rather than against an assumption.
A golden board is not the same item as a first-article sample. A first article shows that a new product can be built at all, while a golden board records what normal looks like to the inspection system. The two are produced at different stages of qualification and are approved separately. At gopcb they are tracked as distinct release records with their own revision history.
Building the Reference Sample
The sample should come from a run built with production tooling, production stencils and a qualified reflow profile. Boards from an engineering pilot often carry small differences in paste volume and placement offset that shift the whole acceptance band. A reference taken from such a run will encode an abnormal baseline for the life of the program.
Between 12 and 20 candidates are normally inspected and the closest to nominal is chosen, rather than simply taking the first board that passes test. The selected unit is then photographed, serialised and stored. The serial number is recorded in the program header so the reference board used at release can always be identified.
How the Recipe Uses the Reference
The program stores the reference image as a template for each inspection window. During production the system aligns each panel to its fiducials, warps the template onto the actual image and then compares local features such as fillet area, centroid position and grey-level spread. The output is a score for each feature rather than a binary pass or fail.
Those scores are converted into pass, review and reject bands. The centres of the bands come from the reference, and their widths come from process capability measured on a real run. A recipe with bands copied from another product will either miss defects or flood the review queue with work that has no value.
Threshold Setting and Tolerance Bands
An inspection threshold is set from measured spread rather than from a supplier default. A common approach is to measure the same feature across 30 to 50 production panels, calculate the standard deviation, and place the review limit at roughly three standard deviations from the reference mean. The reject limit is placed further out, where a genuine defect is the only plausible cause.
Bands are usually tighter for fine-pitch packages and wider for large passive components, because optical contrast and solder volume differ between them. A single global tolerance applied across the whole board is one of the most common causes of an unbalanced recipe, with heavy review loading in one area and no detection in another.
Reducing False Calls
A false call is a report that consumes operator time without finding a defect. Most are generated by benign variation, such as a slight solder mask colour change between panels, flux residue, or a component body that sits a fraction of a millimetre differently from the reference. Widening a tolerance removes the call but also removes sensitivity, so the trade must be made deliberately.
The better route is to improve the reference and the feature set. Replacing a whole-body comparison with a local joint measurement removes much of the noise that comes from body colour. Adding a second golden board that represents the opposite edge of normal variation also helps, because the recipe can then be tuned to accept both valid states.
Qualification and Sign-Off
Before release, the program is run against a set of known defects and known-good units. The defect set should include shifted components, insufficient solder, bridging, missing parts and polarity errors, and it should be produced on the same line with the same materials. The measured escape rate and the false-call rate together decide whether the recipe may be released.
Sign-off normally requires the process engineer, the quality function and the customer where a contractual requirement exists. The record includes the reference serial number, the threshold table, the software revision and the defect-set results. Without that record, a later change cannot be distinguished from ordinary process drift.
Storage and Maintenance
The reference itself degrades. Solder surfaces oxidise, and a board handled repeatedly accumulates contamination that changes its appearance under the inspection lighting. Storage in a dry cabinet with low particulate exposure slows this, and handling should be reduced to the minimum that the calibration schedule requires.
Many programs re-verify against the retained reference every one to three months, or after any change to lighting, camera or software. When the score drifts beyond a defined band, the board is retired and a new one is built from a current production run following the original procedure.
Re-Qualification After Change
A material change, a stencil revision, a reflow profile adjustment or a new paste lot can all move the process enough to invalidate the bands. The trigger for re-qualification should be written into the control plan rather than decided case by case, since the effect of a small change is rarely obvious at the time it is made.
Where the change is permanent, a new reference is built and the old one is retained for comparison. Where it is temporary, such as a short production deviation, the recipe is not altered and the deviation is handled through concession instead. Keeping those two paths separate protects the long-term meaning of the program.
Program Change Control
Every edit to the recipe should be versioned with the reason, the author and the supporting data. Passing a program between machines also requires verification, because lens ageing and illumination differences shift the effective thresholds even when the file itself is identical.
A practical rule is that no program may be edited during a production shift. Changes are made off-line, verified against the defect set and the reference, and released under the same approval as the original. That discipline is what keeps the golden board meaningful after several years of use.
FAQ
How often should a golden board be replaced? Typically every six to twelve months, or sooner if the reference score drifts outside its band or if a permanent process change has been made.
Can one golden board serve several products? No. Each product needs its own reference, because component mix, solder volume and mask colour all change the image the camera sees.
Does a golden board replace first-article inspection? No. First-article inspection proves the process can build the product, while the golden board only defines the inspection baseline.
