PCB AOI Lighting: Coaxial, Low-Angle and RGB in One Cell

A PCB under an AOI camera is not one inspection problem. It is at least three, sitting a few millimetres apart: a copper pad that behaves like a mirror, a solder joint that is a small three-dimensional shape, and a white silkscreen legend printed on green laminate. A single light that flatters one of them usually buries the other two. This is why lighting arguments on AOI lines rarely end with a product name — they end with a sequence.

The order matters more than the hardware. Decide the optical interaction for each feature first, then the geometry, then the wavelength. Skipping to “we need a brighter ring” is how cells end up with four lights and no stable threshold.

Why one light cannot serve the whole board

Reflectance on a PCB spans an enormous range. Bare copper and gold fingers are close to specular at the angles an AOI head sees. Solder mask is semi-matte. Silkscreen is diffuse white. Solder, after reflow, is a curved mirror with a matte oxide skin. If you illuminate all four with the same high-angle source, the specular areas saturate while the diffuse areas sit in the middle of the histogram — the classic “bright pads, grey characters, invisible joints” image.

The reason is geometric, not a matter of power. A high-angle ring sends light down at 60–90°, so flat specular areas return it straight into the lens and clip; the same light arriving on a curved solder fillet leaves at a different angle and never returns. Adding current makes the pads whiter and the joints no more visible. This is the same distinction the standard lighting literature draws between specular, diffuse and transmitted geometries, and it is the first thing to fix before touching exposure or gain.

Pads and bare copper: coaxial, and usually blue

For pad presence, oxide, contamination and mask registration, coaxial illumination is the honest choice. It puts the source on the lens axis through a 45° beam splitter, so only near-normal reflection returns. A flat pad images uniformly bright, an oxidised patch images darker, a missing pad images as laminate. There is no direction, so the result does not change when the board shifts a millimetre.

Blue is usually the better wavelength here. Copper and gold reflect strongly in the blue, while green solder mask absorbs it, which widens the pad-to-mask separation before any threshold is set. Red does the opposite: it lifts the copper but is absorbed less by some masks, and it can make dark laminate look brighter than expected. None of this is universal — mask chemistry varies by fabricator — which is why the wavelength decision belongs in the sample test, not in the specification.

One mechanical constraint decides whether coaxial works at all: the module’s exit aperture must be larger than the field diagonal. If it is not, the corners go dark and the operator blames the light. On a 12.5X coaxial zoom body such as the PMS-LZL-12101D 12.5X coaxial zoom lens, the illumination is inside the barrel, so the constraint disappears; with an external module on a plain zoom, it has to be checked at the widest field setting, not the narrowest.

Solder joints: low angle, because the information is slope

A solder fillet is judged by its profile — insufficient, excess, bridging, voiding at the edge. Profile is encoded in surface slope, and slope is only visible when illumination arrives at a low angle. At 10–30° above the board, the flat mask reflects away from the camera and the fillet flanks scatter light back, so the joint reads as a bright wedge against a dark field. This is dark-field imaging by geometry, not by dimming.

Two practical consequences follow. First, azimuth matters: a joint lit from one side loses the opposite flank, so a single bar is a systematic blind spot. Opposed bars, or a four-quadrant ring, are the minimum. Second, the lower the angle, the longer the shadow — roughly h / tanθ for a feature of height h — and a tall component can throw a shadow across the joint behind it. On dense boards this shadow budget, not brightness, is what limits how low you can go.

Low-angle illumination also interacts with the lens. A telecentric lens rejects most scattered light, so dark field behind a telecentric optic is dim; a zoom lens with a smaller NA at low magnification is dimmer still. Expect to change brightness between magnification steps rather than assuming one setting covers the range. This is the same effect described in why image brightness changes with zoom magnification, and it becomes a lighting problem the moment the AOI recipe switches magnification automatically.

Multicolour and zoned lighting: useful, but not a resolution upgrade

RGB illumination changes contrast by wavelength: illuminate with a colour the feature reflects and it brightens; illuminate with the complementary colour and it darkens. Zoned rings add a second dimension — each sector is an independent channel, so four sectors fired in sequence give four shadow directions from one housing.

What this buys is separability between materials that look alike in greyscale. What it does not buy is optical resolution. A 5 µm defect on a lens with 12 µm object-side sampling stays invisible no matter how many colours are cycled; conversely, a well-lit monochrome setup often beats an elaborately coloured one on a simple presence check. The failure mode is predictable: a cell adds channels, cycle time grows, thresholds multiply, and the defect that motivated the change is still marginal.

A worked reference from a real enquiry makes the arithmetic concrete. A customer inspecting 17.3 × 17.3 mm boards with a 0.25X telecentric lens and a multi-colour source wanted to reduce the field to 12 mm or 8 mm. The relative magnification ratios are 17.3/12 ≈ 1.44 and 17.3/8 ≈ 2.16, which correspond to roughly 0.360X and 0.541X under the stated assumptions about the sensor. Those are geometric ratios, not measured magnifications, and the customer’s exact sensor format is still unconfirmed — but the direction is fixed: object-side sampling improves by the same factor, and image irradiance falls roughly with the square of magnification. The lighting budget therefore has to be re-solved at the new magnification before any colour decision is made.

The brightness budget at higher magnification

Object-side irradiance falls approximately as 1/m², so moving from 0.25X to 0.54X costs about 4.6× in light before any aperture change. If the original recipe used a 2 ms exposure at 0.25X, the same image at 0.54X needs either a longer exposure, a wider aperture, more gain, or more light — and the first three all cost something else: motion blur, depth of field, or noise.

This is where a controller with defined timing matters. A four-channel constant-current controller such as the PMS-DC5V3W-4T gives 256 dimming steps across 0–1000 mA per channel, with external trigger up to 1 kHz and a 200 µs response. That is enough to run short, bright pulses synchronised to the camera instead of raising gain, and enough to keep the coaxial, low-angle and colour channels independently calibrated — which matters because a “level 128” on blue and a “level 128” on white are not the same amount of light.

A validation sequence that survives production

The lighting recipe is only as good as the test that produced it. A sequence that has held up on AOI lines looks like this:

• Fix the geometry first, then the wavelength, then the brightness. Changing two at once makes the result unattributable.
• Photograph a golden board and a known-bad board under each candidate; keep the exposure fixed across candidates so the comparison is about contrast, not about light.
• Measure, do not judge: record the grey-level difference between feature and background, and the background uniformity at the field corners.
• Repeat with the board rotated 90° and shifted within the fixture; a recipe that only works in one pose is not a recipe.
• Re-run after a mask or solder-paste batch change. Wavelength choices that depend on mask chemistry are the first to drift.

If the same cell must handle several board types, the honest answer is often two recipes with a fixed mechanical datum, not one compromised recipe. Recording ring height, module part number and controller channel levels in the work instruction is what makes that possible.

Where the lighting decision ends and the lens decision begins

Lighting cannot rescue a sampling limit. If the smallest feature is below two or three pixels on the object side, no illumination geometry will make it measurable, and colour cycling only adds frames. In that case the conversation moves to magnification: a higher zoom step, a different telecentric lens, or a larger sensor. The zoom lens range and the PCB AOI optical selection guide cover that side of the trade-off; lighting’s job is to make sure the feature is visible once the optics can resolve it.

Frequently asked questions

Can one ring light handle pads, solder and silkscreen on the same board?

Not at one height. A high-angle ring saturates the pads and flattens the joints; a low-angle ring reveals the joints but leaves the pads uneven. The workable compromise is a dual-height or dual-zone arrangement with the height recorded in the recipe, plus a separate coaxial channel for pads.

Is RGB lighting better than white for AOI?

Only when the feature and background differ in reflectance at some wavelength. If the difference is geometric — a scratch, a fillet, a burr — colour adds frames without adding contrast. Test the same geometry with two wavelengths before adding channels.

Why do my pads saturate even at low brightness?

Because the reflection is specular: the light returns almost unchanged, so the sensor clips long before the rest of the image is properly exposed. Fix the geometry (coaxial or polarised) rather than lowering the source, which will under-expose everything else.

Does polarisation help on solder mask?

It helps where the problem is a specular reflection from a smooth dielectric surface. A crossed polariser pair suppresses that reflection but costs most of the light — published figures put a single polariser near 40 % transmission and a crossed pair below 3 % at extinction, requiring roughly 2.5 f-stops of compensation.

How do I keep the recipe stable after a board revision?

Record geometry, wavelength, channel levels and exposure as part of the recipe, and re-validate on a golden board after any laminate, mask or paste change. Lighting that depends on mask chemistry is the most common source of silent drift.

What to send for a lighting review

A useful review needs five things: a photograph of the board as it sits in the fixture, the field of view and the smallest feature that must be judged, the lens and camera in use (or planned), the stand-off available above the board, and one known-good plus one known-bad sample. With those, the lighting geometry and the lens magnification can be checked against each other instead of being argued separately — which is usually where the real constraint turns out to be.