Brochures lead with megapixels and pieces per minute. Neither survives a production line. In fastener, connector and molded rubber work, yield is settled before the purchase order, when a resolution figure is accepted without conversion into microns.
Choosing an optical sorting machine means ranking nine specifications by how much they move yield: resolution in microns per pixel, camera count and view coverage, illumination, feeding architecture, loaded throughput, accuracy split into escape and false reject, ejection performance, software and changeover, and mechanical stability. The first four decide what the machine can see, the next three decide what it does with what it sees, and the last two decide whether either survives a year of production.
The sections below follow the order in which those decisions should be made, because a mistake at resolution cannot be corrected by a later upgrade.
1. Optical Resolution: Microns per Pixel, Not Megapixels
Resolution should be written as the physical area one pixel covers at the part surface, with the smallest defect you must detect spanning at least three pixels and any dimension you must measure spanning five or more.
Pixel equivalent equals field width divided by horizontal pixel count. A five-megapixel sensor with 2448 horizontal pixels imaging a 24 mm field delivers about 9.8 microns per pixel, enough to detect near 30 microns and measure near 50. The same sensor on a 60 mm field delivers 24.5 microns per pixel and loses every defect below roughly 75 microns.
| Defect size to detect | Resolution and hardware that fits |
| 100 microns and above, such as mixed parts | 2 MP sensor, wide field, fixed-focal lens |
| 30 to 50 microns, such as burrs and plating gaps | 5 MP sensor, field under 25 mm, low distortion |
| 10 to 20 microns, such as burrs on 0.1 mm terminals | 12 MP sensor or telecentric optics, field under 10 mm |
| Below 10 microns | Narrow-field stations, telecentric optics, isolated frame |
Work backwards from the defect: its smallest dimension times three, divided into the widest field the part allows. Processing cannot recover detail the optics never sampled, so tightening the field and the fixture buys more than a larger sensor. Dimensional work needs telecentric optics; surface inspection tolerates a lens with distortion under 0.1 percent.
2. Camera Count and View Coverage
Camera count follows from the surfaces that carry defects, not from the throughput target, and the most common failure is a machine that images the top face perfectly while never seeing the thread flank, the underside of a screw head or the cut edge of a stamped contact.
Shade every surface where a defect has ever been found, then ask which of them one camera sees at a single orientation. A part resting on a glass disc shows one face; a part rolling past a lens shows an unpredictable mix. Four cameras imaging a rotating part at four moments describe four different parts.
| Part type | Views that decide acceptance |
| Screws, bolts and nuts | Head top, head underside, thread flank, tip, one oblique for bend |
| Stamped contacts and terminals | Top, bottom, both side edges, cut edge where burrs form |
| Molded rubber seals and O-rings | Outer diameter, inner diameter, both sealing faces |
| Injection-molded housings | Top, bottom, four side walls, cavity-side marks |
| Crystal and glass components | Transmitted silhouette plus two reflective views |
Platforms in the Optik Ayıklama Makinesi category differ less in sensor count than in how many views they hold at once while keeping the part stable. Bores and thread roots usually need an axis-aligned station, since no oblique angle reaches the bottom of a hole.
3. Illumination Design
Illumination is the only specification that manufactures contrast, and a defect with no contrast under the chosen light cannot be found by any algorithm, so samples should be lit and reviewed before resolution is frozen.
Geometry, not brightness, does the work. Coaxial light returns from flat, mirror-like surfaces and turns a dent into a dark spot. A dome flattens the reflections that hide short shots in black plastic. Low-angle dark field skims the surface so burrs glow while the flat face stays dark. Backlight leaves a silhouette that is excellent for dimensions and useless for scratches, and crossed polarizers remove the glare that machined parts throw back.
| Material and defect | Light geometry that usually works |
| Machined or plated metal, scratches and burrs | Low-angle dark field, crossed polarization, one fill |
| Dark molded plastic, short shots and flow marks | Dome diffuse with high-angle fill into cavities |
| Transparent rubber and silicone | Backlight for edges, coaxial for contamination |
| Glass and crystal, chips and bubbles | Dark field for chips, transmitted light for inclusions |
| Assembly with a missing component | Coaxial plus ring in two colors, so the gap changes hue |
At a few hundred parts per minute the exposure window is tens of microseconds, so the light must strobe with the trigger. Continuous output heats the housing, and a tenth of a millimeter of growth softens a 10-micron edge.
4. Feeding Mechanism and Part Handling
The feeding architecture decides which part geometries the machine can present, how much handling damage it adds and how long a changeover takes, which is why it belongs in the specification before conveyor speed is discussed.
Loose parts must be singulated, oriented and held still long enough to be imaged from several angles. A glass turntable carries parts on a transparent disc so top and bottom cameras share one surface, gently enough for plated threads. An index plate seats each part in a machined pocket, buying repeatability for dimensional work. A belt carries flat parts continuously but offers one dominant viewing side, and a vibratory track feeds the smallest parts fastest with the least orientation control.
| Feeding architecture | Best fit and trade-offs |
| Cam döner tabla | Parts under about 30 mm, gentle handling, rim coverage; disc scratches |
| İndeks plakası | Fixed pocket for repeatable orientation; tooling per part number |
| Belt conveyor | Flat or elongated parts, inline flow; weak on side and underside |
| Vibratory track | Very small parts at high rate, lowest cost; overlap causes false rejects |
A part with no stable resting face needs a custom nest rather than a different machine, and that tooling should be quoted with the equipment. A platform such as the Cam Döner Tabla Optik Ayıklama Makinesi suits parts whose two faces and outer diameter must be judged in one pass. Tooling wears silently: a pocket worn by 0.05 mm moves the part out of the depth of field.
5. Loaded Throughput, Not Nameplate Speed
Nameplate speed is measured with one inspection item on a clean sample stream, and the figure that governs planning is loaded throughput, counted while every item is active and rejects are being ejected.
At 600 parts per minute each part owns 100 milliseconds, and that budget must cover transfer, acquisition across all views, processing, decision fusion, ejector scheduling and confirmation that the reject left the stream. A machine reaching 1,200 parts per minute on a presence check may hold 500 once ten dimensional and appearance items run on a 12-megapixel image.
| Cycle-time item | Typical contribution at 600 parts per minute |
| Transfer, settle and trigger | 25 to 35 ms per station |
| Acquisition across four views | 15 to 25 ms |
| Processing per view | 8 to 20 ms |
| Decision fusion and reject scheduling | 3 to 8 ms |
| Ejection and departure confirmation | 5 to 10 ms |
If the per-station total exceeds the available cycle, the line needs a parallel station, fewer items on the fastest recipe, or a smaller field. Ask for loaded data with the reject rate stated, and compare the result with the slowest station upstream.
6. Accuracy Metrics That Survive an Audit
A single accuracy figure cannot be accepted, because it conceals the two errors that cost money: defects that escape to the customer and good parts rejected into the waste bin, so escape rate, false reject rate and carry-over ratio should be specified separately, each with its own sampling method.
What matters is not the share of correctly classified images in a laboratory set but the count of defective parts leaving in the good bin during a shift, and that count depends on severity, orientation and lighting stability.
| Metric | Definition and how to verify it |
| Kaçış oranı | Undetected defects divided by defects presented; seed 200 to 500 pieces |
| Yanlış reddetme oranı | Good parts rejected divided by good parts presented, at speed |
| Carry-over ratio | Good parts ejected per defective part ejected; becomes material loss |
| Tekrarlanabilirlik | Thirty passes of one part must give identical verdicts |
| Gauge repeatability | Repeated measurement against a reference, as tolerance consumed |
Acceptance sampling should follow a recognized scheme rather than an argument. The AQL system in ISO 2859-1 provides lot-by-lot plans indexed by acceptance quality limit, turning an agreed escape rate into thresholds both parties can apply to shipped lots. Programmes where one escaped burr stops a customer line add severity classes with separate limits, as applied in this Otomotiv Bağlantı Elemanları Optik Ayıklama Çözümü. Raising sensitivity cuts escapes and raises false rejects, so the balance depends on defect cost downstream.
7. Ejection System: Nozzle Density, Response Time, Air Demand
Ejection performance decides whether a correct decision becomes a correct action, so it should be specified as nozzle spacing against part pitch, valve response time and sustained air consumption at the real cycle rate.
A machine can classify every part correctly and still ship defects if the jet fires late, wide or not at all. Nozzles set 10 mm apart cannot isolate a 4 mm part without disturbing its neighbours. A 10 millisecond delay on a part moving at 300 mm per second becomes 3 mm of positional error, and sustained air consumption decides whether the existing compressor can carry the machine.
| Ejection parameter | What to ask and why it matters |
| Nozzle pitch and count | Spacing against smallest part pitch; too wide misses neighbours |
| Valf tepki süresi | Slow valves smear the reject window at line speed |
| Sustained air consumption | Cubic meters per minute with reject rate included, not peak |
| Air quality | Filtration for water and oil; contamination sticks valves |
| Departure confirmation | A sensor proving the reject left, so a blocked nozzle cannot hide |
8. Software, Recipes and Changeover
Software decides how fast a validated recipe can be reproduced on the next shift and the next part number, so changeover time, recipe versioning and data export belong in the specification alongside detection features.
Detection features show up in a demonstration; recipe discipline does not, and it separates a machine running at validated capability from one that depends on a single experienced operator. A recipe should capture lighting per channel, exposure, regions of interest, thresholds, ejection delay and the item list, and it should be versioned so a night shift change can be traced. Ask for a measured changeover, and confirm that a proven recipe can be copied to a second machine.
Data output is part of acceptance wherever the customer audits the process and not only the product. The machine should expose a documented interface, with handshake signals for start and stop and a standard protocol for richer data. OPC UA, maintained by the OPC Foundation, is the common route into supervisory systems, while EtherNet/IP, PROFINET and Modbus TCP cover controller links.
9. Mechanical Stability, Environment and Service Access
Mechanical stability decides whether the accuracy measured at commissioning survives thermal swings, vibration from nearby presses and a year of production, so frame rigidity, isolation and calibration access belong in the same specification as optics.
A frame that flexes by microns between optics and part destroys repeatability, and the only symptom is a drifting reject rate. At 0.01 mm and below, a cast or granite base is a stiffness decision, and it needs isolation from stamping presses, since a press fifty meters away can blur a 5-micron edge. A workshop swinging from 12 to 35 degrees Celsius moves focus on any machine that does not compensate, so a stated range of 20 plus or minus 5 degrees is realistic.
Service access is the least glamorous item and the one that decides long-term yield. Lenses need cleaning on a schedule measured in shifts, calibration needs a reference artefact, and nozzles need inspection at intervals tied to operating hours. Ask for mean time to repair, spare part lead times and remote diagnostics, because a machine waiting three weeks for a camera module produces no yield.
10. How to Verify All Nine Specifications Before You Sign
Verification means running your own parts, with every inspection item enabled, on the exact configuration being quoted, and writing the results into the acceptance clause as measured values rather than catalogue values.
The acceptance test is where a specification becomes enforceable. Send three production lots of at least 500 pieces, seeded with each defect type at the severity classes you intend to control, and require escape and false reject counts per type.
| Acceptance test | What to record |
| Sample set | Three lots of 500 pieces with seeded defects |
| Loaded speed | Sustained parts per minute over 30 minutes, reject rate stated |
| Escape and false reject | Counted per defect type, never aggregated |
| Tekrarlanabilirlik | Thirty passes of one part with identical verdicts |
| Changeover | Timed swap including recipe load and first article check |
| Documentation | Defect images, recipe backup, calibration certificate, spares list |
Judge cost over the payback period rather than on the invoice. Labor replacement is easy to quantify at two to four inspectors per shift, while yield recovery from fewer false rejects is often larger and harder to see. Payback of 12 to 30 months is common, and applications in the Yarı İletken Endüstrisi Optik Ayıklama Çözümü area justify higher capital cost through escape cost alone.
SSS
The three questions below address the boundaries buyers probe last: what optics cannot see, how this equipment differs from board-level inspection, and how long a new part number takes to release.
Can an optical sorting machine detect defects below the surface?
No. Imaging works on light returned from or transmitted through the part, so internal voids, inclusions, cracks that never reach a surface and hardness variation stay invisible at any resolution. Those classes need radiography, ultrasonic testing or eddy current, which means a second inspection stage sized to the risk. If a drawing lists internal porosity as a reject criterion, assign it to another process step.
What is the difference between an optical sorting machine and AOI?
Automated optical inspection images boards held in a fixed position, checking placement, polarity and solder joints against a reference, and usually reports rather than removes. An optical sorting machine handles loose parts in bulk, singulates and orients them mechanically, inspects several faces at production speed and ejects rejects. The imaging overlaps; the handling and throughput economics do not.
How long does it take to validate a new part number?
For a part that fits an existing platform and tooling, allow one to three weeks: two to five days for a sample test, then recipe build, first article approval and a capability run. A part needing new pockets or an added viewing angle takes four to six weeks, most of it tooling manufacture. Spare stations and a documented recipe template shorten every later release.