Vision guided spot welding alignment for small parts

A micro spot welding station builds microphone capsules, micro speakers and similar electro-acoustic parts one joint at a time, and each joint is placed to a few hundredths of a millimetre. The vision system is what makes that repeatability possible: it images the part, finds the weld point, and corrects the electrode position before the current fires. This guide sizes such a system with three numbers — pixels across the smallest feature, magnification stability as part height moves, and the light that separates a shiny 0.3 mm tip from its surroundings. Worked against a ±0.05 mm placement target, the numbers lead to a telecentric lens, a global shutter camera, a strobed coaxial or dark field light, and an error budget that lands near ±0.025 mm.

What the vision system does before the current fires

The loop is fixed for every part. The fixture clamps the part, the camera takes one image while the head is parked, the software locates the weld point, and the stage moves the electrode to its commanded position plus the measured offset. Only then does the weld fire; the camera is finished before the transformer triggers.

Two coordinate frames meet in that loop. The camera measures in pixels, the electrode moves in stage micrometres, and a calibration map converts between them. A tool offset stores the vector between the imaged weld point and the true contact point of the tip. Both are constants — until the tip is dressed, the lens is moved or the fixture is swapped.

A smart camera closes the loop on the device: locate, transform and correction output on board, over I/O or Ethernet, with no PC in the cell. Locate plus transform takes well under 50 ms there when the search window is small, so vision adds almost nothing to the weld cycle.

Pixel maths: from tip diameter to field of view

Start from the smallest feature the camera must centre on — the electrode tip, not the whole part. The working rule is 4 or more pixels across that feature, which leaves an edge fit at a tenth of a pixel with margin to spare.

Worked for a 0.3 mm tip and a ±0.05 mm target:

1. Required object pixel size: 0.3 mm / 4 pixels = 0.075 mm per pixel — the ceiling.
2. Camera class: 5 MP global shutter in 2/3" format, 2448 pixels wide at a 3.45 µm pitch.
3. At 1x magnification: object pixel size = 3.45 µm / 1 = 0.0035 mm, so the tip spans 0.3 / 0.0035 = 86 pixels and the field is 8.4 × 7.1 mm.
4. At 0.66x: object pixel size = 3.45 µm / 0.66 = 5.2 µm = 0.0052 mm, the tip spans 58 pixels, and the field widens to 8.4 / 0.66 = 12.7 mm.

Both magnifications pass the pixel rule with room to spare, so resolution is not what decides the lens — the field is. A part 6 mm across, placed by a fixture that repeats within ±2 mm, needs a field of at least 6 + 2 × 2 = 10 mm, plus about 20% margin: 12 mm. That is the 0.66x row, not the 1x row.

The same conversion, from magnification to object-space size, is walked through in calculate object space pixel size from magnification, and why sensor size and focal length set the field of view covers the sensor half.

Why the head's ordinary lens fails the accuracy test

The tempting alternative is the wide M12 lens already on the head, run near its focus distance. It passes the pixel test: a 1/2.4"-class sensor is about 5.7 mm wide, so an 8 mm lens about 100 mm away covers 5.7 × 100 / 8 = 71 mm, and at roughly 3,000 pixels across the sensor each pixel is 71 / 3000 = 0.023 mm — the 0.3 mm tip still spans about 13 pixels.

It fails on geometry. An ordinary lens looks down its axis at an angle, so magnification changes with distance: scale change = height change / distance = 0.3 / 100 = 0.3% for a ±0.3 mm stack-height variation. A feature 20 mm off the image centre then moves 20 × 0.003 = 0.06 mm between a short part and a tall one — more than the entire ±0.05 mm budget. The focal length arithmetic behind the 71 mm figure is in the machine vision lens selection calculation.

A telecentric lens removes that term. Its chief rays run parallel on the object side, so magnification holds as the part moves toward or away from the lens — across the whole telecentric range, where the image may soften while the geometry stays true. The blur-limited depth follows an estimate:

At 0.66x, f/4 and a 3.45 µm pixel that is 2 × 4 × 3.45 µm / 0.66² = 63 µm ≈ 0.06 mm; closing to f/8 doubles it to about 0.13 mm. The ±0.3 mm stack variation therefore softens the image at the extremes without shifting the measurement — exactly what an alignment camera needs. The trade-offs are laid out in entocentric vs telecentric lenses.

Working distance decides fit. The upper electrode sits over the tip, so the lens barrel and its folded-in coaxial light must clear it; telecentrics in the 65-110 mm working-distance class do. For the close-in build, the LS-TCL-2118 is a 1x object-side telecentric lens for a 2/3" sensor with an 8-11 mm image circle and a 108 mm working distance on C-mount.

Two shorter-working-distance telecentrics with a built-in coaxial light cover the cramped head and the wider field:

All 2 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Calibration: pixels to stage micrometres

Camera side: image a glass grid plate — a 0.1 mm pitch is standard — and fit scale and distortion. Metrology telecentrics list geometric distortion near 0.01-0.1% on their data sheets, worth 1-13 µm across a 12.7 mm field, and the grid fit absorbs it. What the data sheet numbers mean is in geometric vs TV distortion.

Tool side: the offset between the imaged weld point and the true contact face of the tip. The practical method is a touch-off. Lower the electrode onto the grid plate at five to nine commanded stage positions, mark each touch, re-image the marks, and fit the affine map from commanded positions to pixel positions. The RMS residual of that fit is the honest number for the whole chain — stage, calibration and tip together; a repeating ball-screw stage and a clean fit land under 0.02 mm RMS. A constant offset shared by every point is not error: it is the tool offset itself, and the transform carries it.

Prove the result with a weld-dot test. Fire dots at five positions across the field and measure dot-to-target offset. One shared vector across all five means the tool offset is wrong — enter it and repeat. Scattered offsets mean the stage or the fixture is the problem, and no amount of vision tuning fixes that.

Locating the part when the part varies

The weld point itself is rarely the feature to match — it hides under the electrodes. Match stable part features instead: a punched hole, the lead edge, a moulded notch. Silhouettes drift with oxide, burrs and plating; a hole does not.

Two features, not one, because theta comes from the baseline between them. Angular error is feature error divided by baseline: 0.01 mm of feature error over a 10 mm baseline is 0.001 rad, and a weld point 15 mm from the pair's centre then moves 15 mm × 0.001 = 0.015 mm. The longer the baseline, the cheaper the angular accuracy — so spread the two features as far as the field allows.

Size each search window from the fixture, not the part. A fixture that repeats within ±0.5 mm gets a search window of ±1 mm around the trained position: every part found, matching still fast. Hold a match-score floor around 0.8 and reject the part rather than weld on a weak match. How far a small part may drift from its trained pose before orientation precision falls is bounded by the 10% field rule, and why a fiducial spanning too few pixels fails is in fiducial marker pixel size.

Lighting a shiny 0.3 mm tip

Micro-welding parts are small, shiny and convex — the combination that defeats plain bright field rings. Two geometries cover them.

Coaxial for the tip face. The face is flat and mirror-like, so light folded into the lens axis returns into the lens: the face reads bright, marks on it read dark. That is the geometry described in in-line illumination for specular parts, and it survives a tip recessed between electrodes because the light travels along the optical axis through the gap.

Dark field for the flank and lead edges. Place a spot light at a grazing angle from one side: light on the cylinder wall reflects away, and only edges and scratches scatter back, so the rim reads as a bright line on black. The law of reflection decides the winner — bright field vs dark field lighting walks that choice.

Strobe the exposure: fire the light for 0.2-2 ms while the head is still, before the current pulse. A short pulse also drives the LEDs well above their continuous rating, which buries shop lighting; the weld flash itself saturates any sensor, so the frame must be complete before the transformer triggers.

A 24 mm white spot feeds a coaxial port directly, and the 4 W third-generation spot adds strobe headroom in the same under-50 mm size class:

White is the right default on plated copper and steel; narrow bands earn their place only for spectral separation of marks or plating, as lighting for metal surface defect inspection shows on larger parts. The driver electronics and timing behind the pulse are covered in integrated strobe drivers in machine vision lights.

The limit sits below the image

At 0.005 mm per pixel and tenth-pixel fitting, the image contributes about 0.0005 mm — so the optics are never the deciding term. The mechanics are. Budget each term and add in quadrature:

The sum works out to √(0.00000025 + 0.0001 + 0.0004 + 0.000144) = 0.0254, so ±0.025 mm — half the ±0.05 mm target, which is the margin to keep before a tool wears.

Thermal drift gets too little attention: a 100 mm aluminium path in the head grows about 0.012 mm across a 5 °C morning-to-afternoon swing. A steel base, a warm-up cycle before production, or keeping the camera off the transformer's hot structure each shrink the term.

The stage term usually dominates. The same budgeting method behind vision guided placement accuracy to 40 µm applies directly: list every term, square it, and let the quadrature sum decide whether the target holds.

Decisions for three builds

One 0.3 mm tip, ±0.05 mm target, precision XY stage. 2/3" 5 MP global shutter camera, 0.66x-1x telecentric at a 65-110 mm working distance, coaxial white strobe. Field 12.7 mm, object pixel 0.0052 mm, budget ±0.025 mm — a 2× margin. Pick the camera in the camera finder and the lens in the lens finder.

Two weld points 15-20 mm apart, one image. A 1"-format camera with a 0.5x coaxial telecentric: a 1" class sensor is about 12.8 mm wide, so 0.5x gives a 25.6 × 19.2 mm field — a 20 mm point span plus a ±2 mm fixture error fits with margin. Object pixel at a 5 µm pitch class: 5 µm / 0.5 = 10 µm, so a 0.3 mm tip spans 30 pixels. The LS-SMC-0017 smart camera carries an 8.8 MP sensor in 1" format with global shutter and a C-mount for this lens class.

Cramped head, finest tips. Where the head leaves no room and the tip is 0.15 mm across, the 1.5x lens with built-in coaxial light at a 65 mm working distance is the fit: on a 1/3" class sensor, about 4.8 mm wide, the field is 4.8 / 1.5 = 3.2 mm, and a 3.75 µm pixel class gives 3.75 / 1.5 = 2.5 µm per pixel — a 0.15 mm tip spans 60 pixels. Choose it when the field only needs the tip and its immediate lead; choose the 0.5x row when the whole part must be in view.

The deciding numbers never changed between the three: field from the fixture, magnification from the sensor, working distance from the head.

Checks before ordering

• Field of view covers the worst case: feature span + 2 × fixture error + 20% margin, or the part falls outside the search window.
• Working distance clears the upper electrode at welding height with the coaxial light folded in, or the barrel crashes on the first cycle.
• Global shutter on the camera, so the frame stays true while the stage settles.
• Strobe duration and trigger wiring agreed between the light controller and the weld sequencer: the exposure must close before the current pulse.
• The light's diameter fits the coaxial port or the free space beside the head.
• A touch-off routine exists in the run-up, and its residual is logged after every tip dress.

The light finder sorts spot lights by size, colour and wavelength, and the camera and lens finders above sort by the same fields used in the maths.

Build the chain from the configurator

Most micro-welding cells end up with the same chain: a global shutter camera, a telecentric lens matched to the field, and a strobed spot light driven by a controller that fires sub-millisecond pulses on a trigger from the weld sequencer. The configurator walks that chain part by part; start it at the lens step and the matched lights and cameras follow:

If a number here needs to match your own head — tip diameter, stack height, fixture error — send it through the contact page and we will size the chain with you. The applications guides hub collects the set-up articles for the rest of the cell.