SMT assembly verification vision system at 3.3 boards per minute

A slow SMT line — 3.3 boards a minute, boards up to 5 in (127 mm), no more than 20 components — still needs a station that proves every part is present, correct and aligned before the board moves on. The line rate is the best asset in the job: 18.2 seconds per board buys one stopped-board image and a full second of processing, so a single area-scan camera, one C-mount lens and one diffused ring light cover all 20 checks. This guide sizes each part with the arithmetic on the page — pixels per millimetre, focal length, working distance, aperture, exposure — commits to a camera class and a lighting geometry, and marks the four cases where the one-camera answer fails. Nothing here needs the PCI frame grabber the older generation of vision systems was built around.

What 18 seconds per board buys you

Start with the takt: 60 ÷ 3.3 = 18.2 s per board. Twenty checks on one image each run 10-50 ms of template and edge work on an ordinary industrial PC, so the whole pass finishes in 0.2-1.0 s and leaves at least 17 s of slack. That margin decides the architecture: one camera, one image per stopped board. No line scan, no multi-camera rig, no CoaXPress link, no dual-station buffering.

The sequence is equally plain. A photoeye reads the board, a stopper or conveyor gate settles it, the PLC pulses the camera's trigger input, the light strobes, and PASS or FAIL returns on digital lines in milliseconds. The worked example below budgets the whole exchange at about 0.6 s. If the PLC also trends output for OEE, a rolling parts-per-minute average in the PLC is the rate this vision budget has to respect.

The pass line: 0.05 mm per pixel

The measurement rule: put at least 3 pixels across the smallest offset you will alarm on. Post-placement alarms for 0603-class parts sit at 0.15-0.25 mm in most plants, well inside the IPC-A-610 physical limit of 50% side overhang on a chip termination — you alarm early, not at the audit limit. Divide the field of view by that offset and you have the pixel size the camera must beat.

Read the table as a decision, not a menu: order the 12 MP class unless every part on the roadmap is 0805 or larger, because that step costs one lens grade and buys the 0402 row outright. Pixels alone do not measure — the lens must out-resolve the sensor across the whole 140 mm field, and lens contrast and resolution: evaluating fine detail shows how to check that from MTF curves rather than pixel counts. The alignment measurement itself, chip body against its own pads, is the geometry worked through in SMT component orientation measurement for pick-and-place vision.

Worked example: one shot of a 127 mm board

Inputs

• Board width: 127 mm (5 in). Field of view: 140 mm, about 6 mm of margin each side for stop-position wander.
• Camera: 12 MP, 4000 × 3000 pixels, on a 1/1.7-inch sensor about 7.4 mm wide (1.85 µm pixels).
• Alarm threshold: 0.15 mm side offset, the 0402 row of the ladder.
• Height available above the conveyor: 500 mm.
• Light: diffused ring, strobed under 1 ms.

Resolution arithmetic

140 mm ÷ 4000 px = 0.035 mm per pixel. A 0.15 mm offset then spans 0.15 ÷ 0.035 = 4.3 px, clear of the 3 px pass line. A 0.1 mm overhang spans 2.9 px, which is exactly why 0201 parts push this camera off the table. When a larger board appears on the roadmap, rerun the division with its width:

Lens, aperture and perspective

Round to a 25 mm C-mount lens at a 500 mm standoff. The check: field of view = 7.4 × (497 − 25) / 25 = 140 mm, the number we started from. Depth of field is not a constraint here — total DOF ≈ 2 × blur × f-number × (1 + m) / m², and with blur set to 2 pixels (3.7 µm), even f/4 leaves about 11 mm, deeper than any component stack. Spend the aperture on light and edge contrast, not on focus.

Perspective is the honest limit at this magnification: a part h millimetres tall shifts its top edge by h × m on the image, so a 0.6 mm-thick 0603 appears 0.03 mm displaced from its pads. Absorb that into the taught template; a telecentric lens only earns its price below 0.05 mm board-level measurement.

Two housings deliver this frame. The LS-CMR-0166 is a 12 MP USB3 area-scan camera, 4000 × 3000, at up to 30 fps:

Prefer GigE and one frame per board? The LS-CMR-0053 is a 12 MP GigE camera at 9 fps — idle at this line rate, with headroom for a re-shot on doubt.

Timing budget

One 12 MP mono frame is about 12 MB, sent once per board. The 17.6 s of slack absorbs the PLC handshake, a second shot on doubt and the operator's reject bin — the single-camera decision stands.

Camera class: PC-driven area scan or smart camera

The default route pairs an area scan camera with an industrial PC: re-teach a component in minutes, archive an image per board serial, extend later to barcode or trace reads. The full bill of materials for that architecture is in machine vision system components: a buyer's guide.

The no-PC route is a smart camera — sensor, lens, light and logic in one housing, verdict over Ethernet. The constraint is pixels: a 1.6 MP sensor holds 0.05 mm per pixel only over about 72 mm, so a 140 mm board takes two overlapping frames. Two strobes and two rounds of checks still close near half a second against the 18.2 s takt, with no PC, no fan and no Windows update at the line. Two housings fit the job:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The LS-SMC-0063 is a 1.6 MP unit with a 16 mm lens, white polarized light and 60 fps. The MV-SC6050C-08M-XBN is a 5 MP color camera with an 8 mm lens and a multi-color light source, built for polarity marks.

Decision: take the smart camera when the plant will not host a PC at the line and the checks stay at presence plus coarse offset. Take the 12 MP PC camera when traceability images, new check types or 0402 alarms are on the roadmap — that openness is why area scan cameras hold the default. On colour: stay mono unless a colored band is the sole cue, because mono edges are cleaner with no Bayer interpolation splitting the pixels.

Lighting: diffuse the board, don't spotlight it

A post-reflow board is a hostile lighting subject: mirror-flat solder and gold pads sit beside matte black moldings, and the edge you measure lives between them. A bare directional ring throws a hotspot that whites out that edge. The first pick among ring lights is a diffused one around the lens axis — the diffuser turns each LED into a soft patch, so solder reads as stable mid-grey instead of blown white. For a 140 mm field at a 500 mm standoff, the fit is a ring about 100 mm across with a 30° beam. The LS-RNG-0143 is exactly that:

Mount it 150-250 mm above the board, just under the lens. Red light in the 620-660 nm band buys contrast between black bodies and white silkscreen; white light only when a colored mark must be seen — which LED colour gives the best contrast in machine vision works through that choice.

If glare off shiny joints still buries edges, cross-polarize the ring and the lens: cross polarization to cut surface glare shows the set-up and its light cost. For low-contrast raised parts, a low-angle ring (0-15°) outlines component bodies against the board — the geometry trade is covered in machine vision lighting angles: surface or silhouette.

Shop-floor light is the drift your teach will be blamed for. Shroud the station — a blackout card or a cabinet with a sight window — then lock exposure and gain after teaching, so the only variable left is the board.

Trigger, transfer and the PLC handshake

Wiring is three lines each way: a 24 V PLC output into the camera's opto-isolated trigger input, and PASS or FAIL back on opto or relay lines, latched inside 10 ms. Images and verdicts also flow to the PC over the same Ethernet or USB link for archiving against the board serial.

No frame grabber sits in this chain. Modern cameras are GigE Vision and USB3 Vision devices: plug into a standard NIC or port and the frame lands where the software expects it. If you are rebuilding a station around an older card, camera frame grabber compatibility for resolution upgrades tells you which legacy pieces survive the change.

The station needs 3.3 frames per minute, so any camera in these classes is barely working — but measure the delivered rate from image timestamps instead of trusting the front page: camera frame rate measurement from timestamped frames is the method.

Checks before you order

• Largest board on the 12-month roadmap: rerun the pixel division on its width — 12 MP holds 0.05 mm per pixel to a 200 mm field.
• Smallest component: pick the ladder row first, then the camera; never the reverse.
• Collect 20-30 known-good boards and 10 rejects — missing, shifted, rotated, tombstoned, reversed polarity — before teaching.
• Trigger wiring: 24 V PLC output to the camera opto input; PASS/FAIL lines back; both directions labelled and fused.
• Lens coverage: a 25 mm C-mount rated for 2/3-inch or larger format covers a 1/1.7-inch sensor corner to corner.
• Station: 500 mm of standoff clearance, a light shroud, and exposure locked after teach.

When the one-camera answer is wrong

• 0201 parts or alarms under 0.1 mm: the 12 MP full-board shot gives 0.035 mm per pixel and stops at the 3 px line. Step to 20 MP on a 1-inch sensor — 5472 px across 140 mm is 0.026 mm per pixel — and the lens moves to about 50 mm at a 580 mm standoff. The MV-CS200-10UM is a 20 MP USB3 camera on a 1-inch sensor:

• Boards that never stop: rolling shutter smears a moving frame line by line, so take global shutter and strobe — a sub-millisecond pulse freezes any conveyor here to well under a pixel. A global-shutter 12 MP camera carries a 1.1-inch sensor near 14 mm wide, which asks the same 140 mm field for about a 50 mm lens near a 550 mm standoff. The MV-CH120-11UC is that camera: 12 MP, USB3, global shutter.
• Colored polarity bands as the only cue: take the color sensor. The band is large against the pixel size, so the Bayer penalty on edges is affordable; the multi-color-light smart camera above covers the no-PC version.
• Underside inspection: a second station after the board flipper, same math, mirrored.

Run your own numbers in the system configurator, starting at the camera step; if the roadmap is still fuzzy, fill in scoping a machine vision system before you call an integrator first. Either way, the figures above are the ones any proposal has to beat.