Machine vision assisted manual assembly: verifying steps in place

High-variant, batch-size-1 assembly is where automation usually gives up: too many variants for fixed tooling, too few pieces per variant to justify purpose-built fixtures. The classic fallback is a double check, two employees watching each other, plus end-of-line tests that find errors only after the value is added. A single camera over the bench replaces the second watcher. It reads the workpiece code, confirms each step the operator performs, and reports the step to the manufacturing execution system (MES). This walk-through gives the mount height, lens, focus budget, lighting and handshake numbers for such a station, plus the training counts and the cases where the camera adds nothing.

What can the camera verify during manual assembly?

Presence, count, position, orientation and colour of everything visible from above - a seal in its groove, a clip closed, four fasteners in four holes - while torque, clamp force and anything inside a closed housing stay with the tool or instrument that produced them. The camera is a witness, not a replacement for the person. The operator assembles, the image confirms, and a failed confirmation is corrected in place before the next step adds cost.

Split the checks the way you would on any inspection station. Presence, count and position of well-lit features are rule-based work - blob counts, edge positions, colour thresholds. Parts that vary in finish, reflections that move, seals that sit at slightly different angles across hundreds of variants, are where semantic segmentation and object detection earn their keep: segmentation marks the error pixels, detection returns a box around each recognised component. The design order does not change: stable 2D vision inspection systems still start with lighting, lens and exposure, and the model is the last thing you buy. For a picture of assembly verification running at rate, see the SMT assembly verification station at 3.3 boards per minute; the bench version runs the same logic slower, with a person in the loop.

Where does the camera mount at the assembly bench?

Overhead on the bench's upper rail, lens axis vertical, ring light around the lens, 700-800 mm above the fixture. That height is high enough for hands and parts to move freely, and low enough that a 25 mm lens frames the whole workpiece. Vertical is a measurement decision, not a preference. An angled camera turns part-height variation into image displacement: at 45 degrees, every 10 mm the operator lifts the part slides its features 10 mm sideways (10 mm x tan 45°), which breaks any position check. Overhead, the same 10 mm only defocuses slightly, and focus is a budget you can plan.

Three mounting rules:

1. Bolt the camera to the bench frame, not to a height-adjustable table, or every table move changes the field of view.
2. Keep the operator's normal working posture out of the shadow of camera and light; the station must not change how the person works.
3. Put the instruction screen at eye level beside the field of view, so the step text and the confirmation are read without turning.

The screen is part of the station, not an afterthought - serving drawings and schematics to the bench is its own workflow, and the vision result appears on it as an overlay on the last captured image.

What lens and working distance cover the workpiece?

For a 180 mm workpiece on a 1/1.7" sensor, a 25 mm lens at 750 mm working distance frames about 220 mm across at roughly 0.055 mm per pixel. That is comfortable for a 1.5 mm seal cord, which then spans about 27 pixels. Work the numbers in this order:

• Target field of view: the 180 mm workpiece plus 20 mm of fixture margin each side = 220 mm.
• Sensor: a 1/1.7" format is about 7.6 mm wide, so m = 7.6 mm / 220 mm = 0.0345.
• Working distance: 750 mm lens-to-fixture, which leaves hand clearance under the ring light.
• Focal length: f = 750 mm x 0.0345 / 1.0345 = 25 mm, a stock 25 mm C-mount lens. Reversing the check, m = 25 / 725 = 0.0345, so the field of view is 7.6 / 0.0345 = 220 mm.
• Scale: 220 mm / 4000 pixels = 0.055 mm per pixel. A 1.5 mm seal cord covers 27 pixels, a 0.5 mm screw head 9 pixels. Rule: the smallest feature you must confirm needs at least 5 pixels across.

A 12 MP camera is the right class for a 220 mm field. For the camera, start here:

The LS-CMR-0166 is a 12 MP USB3 camera with a 1/1.7" sensor at 4000 x 3000 and 30 fps; the rolling shutter is harmless because capture is gated on hands-clear. Browse the area scan cameras or the camera finder for other mounts and interfaces. When a product family exceeds the 220 mm field, move up in sensor before moving up in distance. The MV-CS200-10UM is a 20 MP camera with a 1" sensor over USB3. A 1" format is about 13 mm wide, so the same 25 mm lens at 750 mm frames roughly 380 mm across, about 0.07 mm per pixel at 20 MP.

Keep every check feature at least 10 % inside the field-of-view edge; the 10 % field rule exists because placement variation eats margins from the edges inward.

How do you keep focus when part height varies by 50 mm?

Stop the aperture down to f/11 and you hold about plus or minus 38 mm of part-height range at 750 mm. Beyond roughly plus or minus 40 mm per variant family, drive a liquid lens, which refocuses in milliseconds. The focus budget is arithmetic, not trial and error:

Take the circle of confusion as two pixels: 2 x 1.9 µm = 3.8 µm for the sensor above. At 750 mm with the 25 mm lens, (WD / f)² = (750 / 25)² = 900, so:

• f/8: 2 x 8 x 0.0038 mm x 900 = 54.7 mm, call it 55 mm total, plus or minus 27 mm around the focal plane.
• f/11: 2 x 11 x 0.0038 mm x 900 = 75.2 mm, call it 75 mm total, plus or minus 38 mm.

Aperture costs light: f/2.8 to f/11 is four stops, 16x less light on the sensor, so either the exposure grows 16x or the gain does, and gain is noise. Decide by variant geometry:

• All variants seat their check features within about 30 mm of one plane: fix focus at f/8 to f/11 and stop there.
• Height spread of 50 mm or more, or short exposures for fast steps: a liquid lens driven from the recipe holds f/2.8 to f/4 and the short exposure, refocusing in milliseconds. Key the focal command to the barcode so the focus travels while the operator still has parts in hand.
• A laser distance sensor reading part height can gate both the focus command and the capture, which keeps the recipe honest when a fixture is swapped.

Which light keeps seals and metal readable without glare?

A diffuse white ring light around the lens, 100-120 mm across, run continuously. The diffuser flattens the specular hotspot off machined metal, and black elastomer seals read as clean dark shapes against it. Lighting decides this station more than any other choice. Black rubber on bright metal is its own subject - the black O-ring lighting guide covers the geometry that makes a seal readable - and when the housing is polished, a cross-polarizer pair on lens and light removes the last specular streak at the cost of about two stops.

Set exposure with a hands-clear gate: the recipe captures after the operator's hands leave the frame, so motion is residual drift, not assembly. At 10 ms exposure a 5 mm/s drift blurs 0.05 mm, one pixel at the scale above. Raise light output before gain; every decibel of gain is noise you then ask the classifier to ignore.

For the ring over the bench, the 100 mm diffused pattern is the stock answer:

The LS-RNG-0143 runs on 24 V with a 30° beam and a 3-pin flat connector. The ring lights range and the light finder carry the same pattern in other diameters and colours.

How does a passed check release the next MES step?

Through a five-step handshake - scan the workpiece code, MES validation of step and station and user, instruction display, a background recipe run on the camera, result report - and the screen advances only on a pass. In order:

1. The operator scans the code on the workpiece. The code resolves to a variant and a build state; the MES confirms the part belongs at this station, at this step, with this logged-in user.
2. The screen shows the step: images of the assembled state, the parts to fetch, the details that matter.
3. The operator assembles. The camera watches the same field of view with the recipe loaded and the focus already set.
4. On hands-clear, the recipe runs and returns the result with the image.
5. Pass: the step is reported complete, the UI advances, and the image is archived against the serial number. Fail: the failing region is drawn on the image, the operator corrects in place, the check repeats.

The entry scan is a code-reading job like any other; if the codes are hand-marked or sit on curved metal, 1D barcode preprocessing for rotation, blur and glare covers the failure modes.

The recipe runs either on a bench PC or on the camera itself. On a PC, object detection without a GPU is realistic at this image size and rate, which keeps the station to a small fanless box under the bench. With hundreds of variants and segmentation in the recipe, size the PC around inference time, not around the camera link.

When the checks stay at presence, count and position and the variant count is modest, a smart camera replaces the PC and the software seat - the choice between AI and rule-based inspection on a smart camera decides it. The smart cameras line is built for exactly that:

For the one-device version:

The LS-SMC-0063 is a 1.6 MP smart camera at 60 fps with a 16 mm lens and polarized white light over GigE.

How many images train a reliable seal-presence check?

Collect 100-200 labelled images per class to start, failure states included, and grow the set from what production logs. With batch size 1 you train on the feature across variants, never on one variant alone. The counting logic is worth reading in full - how many samples deep-learning defect inspection really needs - because the classes here are asymmetric: one large "seal correct" class and several small, precious failure classes.

What varies in production must vary in training:

• Part height per fixture and per variant, each shifting focus and scale slightly.
• Ambient light: day shift against window light, night shift against high-bay lamps.
• The failure states themselves - missing, tilted, wrong seal, double seal. If a failure has never occurred, photograph staged examples; if you cannot stage it, that check stays rule-based or human.

Validate per condition, not as one averaged score - per-condition generalization metrics - or a model that only ever saw daytime images passes review and fails on the night shift.

When is a camera at the bench the wrong answer?

When the critical step leaves no visible trace - torque values, clamp force, leak tests, internals - or when the feature sits under the operator's hands at check time, the camera cannot witness it and the station becomes theatre. An apparently correct fastener image does not establish the applied torque; that stays with the torque tool, which logs its own values. The camera's job is the visible consequence: the washer under the head, the marked fastener, the clamp fully home.

Three more walls:

• Takt: the check and handshake must fit inside the step. Budget about one second; the frame arithmetic is in the questions below.
• Hidden geometry: if the failure mode is height, tilt or gap rather than presence, greyscale 2D will not carry it - that is the boundary where 3D beats 2D machine vision.
• Data policy: assistance platforms that insist on cloud image storage collide with on-premises quality systems; standard vision software on your own hardware avoids that negotiation entirely.

What should you measure before ordering the parts?

Measure the workpieces and the bench before buying. The largest part sets the field of view, the height spread sets the focus budget, and the smallest feature must land at five pixels or more. Write these down before the first part number:

• Largest workpiece plus fixture margin, and the rail height you actually have - together they fix the focal length.
• Height spread of check features across all variants: inside the f/11 budget (plus or minus 38 mm at these numbers) or beyond it.
• Smallest feature divided by (field of view divided by pixels): at least 5.
• Bench lighting - windows, high-bay lamps - and a plan for a shroud or more light power; lock white balance.
• Network: a drop for the camera and a path to the MES server.
• Storage and retention for images per serial number.
• Operator clearance: 700-800 mm under the camera, screen at eye level.

Prove the checks before they gate production - verifying a presence system without calibration is a workable acceptance test for exactly this class of station.

Then decide by case:

• One field of view, height inside plus or minus 38 mm, presence-and-count checks on tens of variants: fixed-focus camera at f/11, or a smart camera with built-in light.
• Height spread beyond about plus or minus 40 mm or short exposures: an area scan camera with a liquid lens.
• Hundreds of variants, segmentation and detection recipes, OCR to come: an area scan camera plus a bench PC.
• Failure modes about height, tilt or gap: a 3D camera instead of, or beside, the 2D one.
• Steps with no visible trace: keep the instrument as the witness and point the camera at the consequences.

Run your numbers through the station configurator, or talk the variant list through with us - bring the largest workpiece, the smallest feature and the height spread; those three numbers decide the hardware.