Five Questions to Ask Before Adopting AOI

AOI (automated optical inspection) projects fail in remarkably consistent ways: the equipment arrives and runs, but the site does not dare switch off manual re-inspection, so the result is "the cost of one more machine, and not one fewer person".

And the problem is usually not the equipment itself. It is that the specification was not pinned down before purchase. This article lists five questions that need clear answers before the contract is signed, each with how to calculate it.

Question 1: what is the definition of a defect?

This is the most basic question, and the one most often skipped. "No scratches on the surface" is not a specification, because it does not answer:

• How long must a mark be to count as a scratch? 0.1 mm or 1 mm?
• How deep? Or is only length considered, not depth?
• Does location matter? Is a scratch in a functional area the same as one on a corner?
• Which is worse: one 2 mm scratch, or three of 0.5 mm each?

If three senior inspectors disagree about the same sample, the machine will not have the answer either.

This is not nitpicking; it is a physical limit. AOI "maps image features to decisions", and if the decision criteria themselves are fuzzy, the system can only learn a fuzzy boundary, which shows up as both false rejects and escapes running high.

Recommended approach: before discussing specifications, have the site pick 30 to 50 real samples, covering definite OK, definite NG, and borderline parts "that even veteran inspectors hesitate over". This set of samples will later become the basis for acceptance, and it is worth far more than the adjectives in a specification.

Question 2: can the defects you need to catch be resolved?

This question is pure calculation; no guessing needed.

Assume a 50 mm field of view and a camera 2448 pixels wide:

GSD = 50 mm ÷ 2448 = 20.4 µm/pixel

Then convert the defects you need to catch into pixels:

If the calculation gives only 2 pixels, the answer is not "a better algorithm" but a change in the optical configuration: a smaller field of view with several shots (longer cycle time), a camera with more pixels (higher cost), or a shorter working distance with a longer focal length (the mechanism has to be redone). All three routes have costs, but all are cheaper than discovering after purchase that the defect cannot be seen.

The full calculation method is in How to Calculate Lens Focal Length and how to configure a vision inspection station.

Question 3: false rejects or escapes, which side do you choose?

This question has a big effect on the outcome, yet it is often skipped at the purchasing stage. With the same optics, lighting and algorithm, the false reject rate and the escape rate cannot both be pushed down: tighten the threshold and escapes fall but false rejects rise; loosen it and the reverse happens.

A concrete example shows the orders of magnitude. 10,000 parts a day with an actual defect rate of 2 % (200 defective, 9,800 good):

So what has to be decided before purchase is which side this line fears more:

• Escapes are extremely costly (automotive, medical devices, aerospace): tighten the threshold, accept high false rejects, and budget for review labor
• High throughput pressure, low unit price: loosen the threshold, and cover escapes with sampling or downstream testing

With the equipment conditions unchanged, "both" is not an option; to improve both at once you have to go back and change the optics, lighting or algorithm. If a specification says "0 % escape rate", that line cannot be proven at acceptance either: with a finite set of samples, all you can prove is "k missed out of N known defective samples". Write escapes in this verifiable form and state an upper limit for the false reject rate at the same time, and the specification can actually be accepted against.

Question 4: is the lighting repeatable enough?

AOI stability depends heavily on lighting, and the lighting question is not "is it bright enough" but "is it equally bright every time".

Factors that change the image:

• Ambient light: day and night shifts, fluorescent lights switching on and off, indicator lights on neighboring machines
• Light source aging: LED luminous flux decays with operating time, so images half a year later differ from those at acceptance
• Part placement: if the fixture's positioning repeatability is worse than the depth of field, imaging conditions differ every time
• Surface condition: the degree of oxidation, oil and adhesive residue within the same batch of material

The questions to ask: does this AOI have a light shield? Does the light source have constant-current drive and brightness feedback? Is there a routine for periodic calibration with a reference part? If any one of these three is missing, changes in yield a few months later become hard to interpret, because you cannot tell whether the process has worsened or the imaging conditions have changed.

Question 5: what is the real cost of cycle time and labor?

The equipment quotation says "0.5 s per part". But the actual cycle time of an inspection station is the sum of five segments, and image capture is only a small part of it: the full cycle time budget breakdown shows that for the stop-and-shoot example on the industrial automation page, moving and settling take 83 %, and exposure only 0.04 %.

The full account should include:

1. Equipment cycle time: whether it can really keep up with the line
2. Review labor: number of false rejects × review time per part (65 minutes a day in the example above)
3. Tuning and changeover: how long does a new part number take? Who is able to tune it?
4. Maintenance: how often the light source and lens are cleaned, and how often the reference part is calibrated

Item 3 is the most often underestimated. If a changeover takes an engineer half a day of parameter tuning and the site changes over three times a week, this AOI will gradually become equipment nobody wants to use. What to confirm at purchase is "can the people on site do the changeover themselves", not "can an engineer tune it to its best".

Summary: five questions, five concrete answers

Closing: AOI is not buying equipment, it is defining a standard

Most of the real work in adopting AOI lies in the first question: turning the judgment criteria in people's heads into rules that can be written down, measured, and give the same conclusion to three different people. Once that is done, choosing the equipment becomes a relatively simple purchasing decision.

Buy the equipment before that is done, and what you usually end up with is a very expensive camera that needs someone standing next to it.