AI Optical Sorting Machine for Small Parts: When It Replaces Manual Full Inspection

Manual 100% inspection on small precision parts fails in a particular way: it is accurate for the first hour and unreliable by the fourth, it cannot be audited, and its cost scales linearly with volume. An optical sorter is bought to replace that variability, not to replace a measuring machine — and the two get confused constantly.

The distinction matters: a measuring machine produces a number against a tolerance, with a calibration chain. A sorter produces a decision — pass, fail, or grade — on many parts per minute, judged against trained criteria. Asking a sorter to report a dimension is asking the wrong instrument, just as asking a measuring machine to sort 800 parts a minute is.

All values below are published POMEAS figures for PMS-PL06. This is a selection method, and no claim is made about throughput on any specific part without sample testing.

What the published specification actually says

The manufacturer states explicitly that actual speed depends on part size, material, feeding condition, the number of inspection items, defect complexity and the acceptance criteria, and that the final figure comes from sample testing. Treat 800 parts per minute as a ceiling to be verified, not a promise.

Throughput arithmetic: what 800 parts per minute means

800 parts per minute is one part every 0.075 s, or 75 ms. That is the entire budget in which the part has to be fed, imaged on six cameras, processed, judged and physically sorted.

Two consequences follow, and they decide most feasibility questions:

- Feeding is usually the binding constraint, not imaging. A sorter cannot run faster than parts can be presented one at a time and in a known orientation. Vibratory bowl feeders, conveyors and guides are part of the specification, not accessories. - The inspection item count eats the budget. Each additional check adds processing time. A part with two checks and a part with nine do not run at the same rate, which is why the published figure is qualified by "number of inspection items".

Why six cameras

POMEAS states the six-camera arrangement images each part from different directions to reduce blind spots from a single viewpoint, improving coverage of complex structures, side features and defects on multiple surfaces.

This is the core difference from a single-camera station: a crack on the side of a part, or a defect visible only at a grazing angle, is not a processing problem — it is a coverage problem, and it is solved optically. The trade is cost and cycle time: six images per part instead of one.

AOI plus AI: what each one contributes

The published architecture separates the two. Traditional algorithms — grey analysis, edge detection, dimensional measurement, contour comparison, position judgement — handle defects that can be described by a rule: a scratch is darker than its surroundings by a threshold, a hole is in the wrong place by n pixels. AI classification handles defects that cannot be described that way: micro-cracks, dark cracks, irregular appearance anomalies and reflective-surface irregularities, where the rule would either miss them or flag good parts.

POMEAS states the AI model is trained on customer-supplied good, defective and borderline samples. That is a project dependency, not a product feature — the quality of the result depends on the samples provided, and the number required is a question to settle before the order, not after installation.

Changeover: the specification that decides whether it gets used

A sorter that takes four hours to change over will not be used on a high-mix line. PMS-PL06 is published with per-product inspection recipes, saved parameters that can be recalled on changeover, and offline adjustment of detection regions, acceptance criteria and the AI model — so the line is not stopped while the recipe is tuned.

For multi-SKU operations this is usually the decisive specification, ahead of throughput.

Four questions that decide whether a sorter is the right answer

1. Is the output a decision or a number? Decision → sorter. Number with a tolerance and a calibration record → measuring machine (one-press image measurement, or a VMM). They are not substitutes.

2. Can the parts be fed reliably? Singulated, oriented, at a stable rate. If not, the feeding problem has to be solved first.

3. Are the defects describable by rules, or do they need training data? Rule-describable defects can run on traditional algorithms; the rest need samples, and enough of them.

4. Is the volume high enough to justify it? Manual inspection cost scales with volume; a sorter's does not. Below a certain volume the case does not close — and the changeover specification is what makes high-mix, lower-volume cases viable.

What this does not cover

It does not quote throughput for any specific part — that requires sample testing. It does not cover the feeding and fixture design, which is usually the largest part of the engineering. It does not cover the lighting selection, which for reflective parts is typically the hardest single problem in the build. And it makes no claim about detection rates on defects that have not been imaged.

FAQ

How fast is it?

PMS-PL06 is published at up to approximately 800 parts per minute — one part every 75 ms. Actual speed depends on part size, material, feeding, the number of inspection items and defect complexity, and is confirmed by sample testing.

Why six cameras?

To cover multiple surfaces and side features. Blind spots from a single viewpoint are an optical problem, not a processing one.

Do I need AI, or is AOI enough?

Rule-describable defects run on traditional algorithms. Micro-cracks, dark cracks, irregular anomalies and reflective-surface defects are the cases POMEAS cites for AI classification, trained on customer-supplied samples.

How many samples are needed to train the model?

The published documentation lists this as a question to settle with the supplier; it depends on defect variety. Settle it before the order.

Can it measure dimensions?

It can apply dimensional measurement, contour comparison and position judgement as inspection items. It is not a calibrated measuring machine and does not replace one where a traceable number is required.

How long does changeover take?

The system stores per-product recipes and supports offline adjustment of regions, criteria and the AI model, so tuning does not stop the line. Actual changeover time is part-dependent.

Next Step

Send part images, material, dimensions, defect types and acceptance criteria, the target throughput and changeover frequency, and POMEAS will arrange sample testing and return an inspection plan with a throughput calculation.