AOI Program Setup and the Defect Library

Automated optical inspection is a camera and a decision. The camera is reliable; the decision depends on how the AOI program was written and on what it was taught to accept. Two machines with identical hardware will find very different numbers of defects on the same board.

Programming is therefore the substance of an AOI installation. It is a judgement about which defects matter, which appearances are acceptable and how much of the decision is handed to an operator, and it has to be revisited as the product and the process change.

What the Machine Can and Cannot See

An AOI system compares each feature on the board with a reference and reports the difference. It sees presence, position, polarity, the shape of a fillet and the texture of a joint, and it cannot see anything that is hidden underneath a package.

The limits are geometric rather than optical. A joint under a ball grid array, a solder fillet on a lead that is shadowed by the body, and a crack that has not yet opened are all invisible to a top down camera, which is why AOI is complemented by X-ray and by electrical test rather than replacing them.

Building the Defect Library

The defect library is the set of images that the program uses to recognise a fault. It begins with the standard defects and grows with the ones the line actually produces, and its quality determines whether the machine finds real problems or reports noise.

A library that is built from a single sample of each defect is fragile, because it recognises that appearance and not the range of appearances. Collecting several examples of each defect, from different batches and different positions on the panel, produces a program that generalises.

Reference Images and Their Pitfalls

The program needs a reference for a good joint as well as for a bad one. Using a single golden board is convenient and risky, because the appearance of a good joint varies with the paste, the finish and the position on the oven.

A better approach is to build the reference from several known good boards, including the ones that represent the edges of the process window. The program then accepts the range that the process actually produces rather than the appearance of one particularly good sample.

False Calls and Their Cost

A false call is a good joint that the machine reports as a defect. Each one costs an operator review and, if the review is not careful, a rework operation on a joint that was sound. The cost is in the labour and in the risk of damaging the board.

A false call rate that is too high destroys the value of the inspection. Operators begin to accept the reports without examining the board, and real defects pass through as a result. Managing the rate is therefore a production decision, not a programming detail.

Tuning Without Losing Sensitivity

The way to reduce false calls is to make the reference and the tolerances match the real process, not to widen the acceptance criteria until the complaints stop. Widening the criteria reduces false calls and stops finding marginal defects, which is the opposite of what the inspection is for.

The distinction is visible in the data. A tolerance that was widened to accept a legitimate variation will not hide a defect that is genuinely outside the process window; a tolerance that was widened to silence the machine will. Tracking which defects are found after the change is the test.

Operator Review and Its Discipline

Every report that reaches an operator has to be reviewed against a written criterion. Where the criterion is a picture with an example of acceptable and unacceptable, the review is quick and consistent; where it is a sentence, the decision varies between operators and shifts.

The review should also feed back. A report that is judged acceptable is evidence that the program needs to change, and a report that is judged a real defect is an example for the library. Recording both makes the program better, and the same feedback loop is what keeps an inspection process honest.

Coverage Compared with Other Tests

AOI covers the visible defects well and the hidden ones not at all. The coverage of the inspection is therefore planned together with the electrical test and the X-ray step, so that every defect class has at least one test that can find it.

Where a class is covered twice, the redundancy should be deliberate: a bridging defect found by both AOI and electrical test is worth catching twice if the escape cost is high. Where a class is covered by nothing, the gap should be documented rather than left to be discovered by a customer.

Maintenance and Program Control

The program is a production asset and should be under change control. A modification to the library or to a tolerance changes what the line accepts, and it must be recorded with a reason and a date, in the same way that any other process change is.

The machine itself needs calibration and lighting maintenance, because the images the program was built on depend on the illumination. A lamp that has aged moves every image slightly, and a program that was tuned on a fresh machine will drift with it. Checking the lighting as part of the routine is one of the measures that belongs with the practice of judging board quality.

Process Control and Verification

Where the requirement is not written down, the shop supplies its own default, and the default is chosen for the process rather than for the design. The measurements that matter are the repeatable ones: conductor width and spacing, annular ring, finished hole size, plating thickness and surface finish are all verifiable on a coupon that travels with the panel.

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

FAQ

Can AOI be used as the only inspection? Only for a product whose defect classes are all visible. Boards with hidden joints need an additional method.

What false call rate is acceptable? The figure should be low enough that operators trust the reports, which in practice means accepting a modest rate rather than aiming for zero.

How often should the program be reviewed? When the product changes, when the process changes, and periodically against a sample of known defects to confirm that they are still found.