Vision measurement of wire EDM parts: fix the recast halo

A wire EDM shop that measures its own blasted parts on a vision system gets numbers that agree with the micrometer. Put an as-received customer part on the same stage and it reads about 0.0005 in small, part after part, and policy forbids touching the surface. The cause is a light-scattering fringe at the recast edge, not a calibration fault, and the fix belongs to the machine: lighting geometry, threshold placement and a telecentric lens. This guide gives the numbers for each, a worked example, and the acceptance test that proves the result.

Why do wire EDM parts measure about 0.0005 in small on a vision system?

A gray fringe of scattered light sits just inside the true edge, and the threshold traces it: 0.0005 in equals 12.7 µm, which is about 6 µm of inward pull on each of the two edges that form the dimension.

The wire-cut flank is not clean metal. It carries a recast layer — material the spark melted and re-deposited, roughly 2–15 µm thick and thickest on roughing cuts and with weak flushing — plus micro-debris fused onto it. Under a wide-angle backlight that fringe scatters light into the shadow side of the edge, so the dark-to-bright transition begins a few microns inside the metal instead of at it. Both edges of a dimension pull inward, so the part reads small, never large. Bead-blasted parts lose the fringe, which is why they agree with gauge blocks.

How much of the error is the recast layer and how much is the lighting?

On an as-received wire-cut part, budget a total bias of 6–20 µm: the recast and debris build a 2–15 µm scattering fringe at the edge, and a wide-angle backlight turns that fringe into a threshold shift.

A diffuse backlight glows from every point of its face, so light reaches the edge from shallow and steep angles alike; some of it grazes past the flank and scatters off the debris. A collimated backlight sends nearly parallel rays, so the shadow boundary stays sharp and the only fringe left is the part's own. When that fringe reaches mid-level brightness, the half-level crossing lands about half a fringe width inside the metal.

Your 12.7 µm sits in the bottom row — older generators and weaker flushing leave more debris, so parts from older machines read worse than parts from new ones. Treat the table as the working budget, then replace it with your own numbers from the gauge-pin test below.

What lighting gives the true edge on a part I am not allowed to clean?

A collimated backlight raised 50–100 mm behind the part, run at the lowest drive that still holds the background near 230 grey levels, cuts the 12.7 µm bias to under 3 µm with no surface contact.

Order backlights in the collimated type for this station. The geometry behind the choice is its own subject, covered in collimated backlight or diffuse backlight for edge gauging. Set exposure by the histogram, not by the picture: background 230 ± 10 grey levels, part body under 30, nothing clipped. Clipping flattens the bright side of the transition and widens the fringe on screen. Shroud the station from room light; ambient glare re-adds the wide-angle component you just removed.

The LS-BKL-0830 is a 120 x 120 mm collimated backlight. That size suits most punch and die details; smaller features simply leave you more light headroom. A high-luminance standard unit such as the LS-BKL-0750, a 180 x 100 mm backlight, stays useful for blasted parts and through-thickness checks. Collimated against high-luminance standard, side by side:

All 2 share: Voltage 24 V; Cable (mm) 500; Recommended working distance (series range) 1~100 mm; Colour temperature of white 6600K; Operating temperature 0 to 50 °C; Operating humidity 20-85% (non-condensing). Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Which lens and camera keep a 12 µm bias visible and correctable?

Work at an object-space pixel of 3–7 µm with a telecentric lens: at 4.8 µm per pixel the 12.7 µm bias is 2.6 pixels — plain to see and chart — while a 20 µm pixel buries it under half a pixel of gradient.

The telecentric lens matters as much as the pixel. Its magnification is fixed over working distance, so part thickness, fixture height and nest tilt do not rescale the edge — the failure mode when you need a telecentric lens for gauging explains in detail. Keep the whole part thickness inside the telecentric depth of field, or the top edge softens and hands the threshold the same kind of fringe you just removed from the light.

Worked example, an 8 mm punch on a 1" sensor:

• Camera: 20 MP, 5472 × 3648 pixels, 2.4 µm pixels, 1" format.
• Sensor width: 5472 × 2.4 µm = 13.1 mm.
• Field of view at 0.5x: 13.1 / 0.5 = 26.3 mm wide — the punch sits centered with margin.
• Object pixel: 2.4 / 0.5 = 4.8 µm.
• Bias: 0.0005 in = 0.5 mil → 0.5 × 25.4 µm = 12.7 µm → 12.7 / 4.8 = 2.6 pixels across the dimension, 1.3 per edge.
• Subpixel locator at 1/10 pixel: 0.5 µm, so the fringe is about 25 times your resolution step.

The LS-TCL-2194 fits the example: 0.5x magnification, 110 mm working distance, 1" format, C-mount, with a built-in coaxial light for a surface view.

The derivation behind the first two bullets is short and worth a bookmark: calculate object space pixel size from magnification. For a different field of view, the resolution calculator runs the same arithmetic:

For planning the measurement itself — which features, which views, which tolerances — machine vision dimensional measurement for small parts is the companion read.

How do I set the threshold so the fringe stops pulling the edge in?

Put the threshold at half the measured span between background and body — (230 + 20) / 2 = 125 on a typical backlit profile — then prove it on a certified pin before any as-received part passes.

Subpixel edge finding without a vision library comes down to this one level, and edge detection without a vision library: threshold basics covers the mechanics. The procedure:

1. Fixture a certified gauge pin or an edge-on gauge block at the same height as the parts.
2. Read background B and body D from a line profile; set the threshold to (B + D) / 2.
3. Measure the pin ten times. The mean must land within 2.5 µm (0.0001 in) of its certificate value before you go on.
4. Swap in the EDM part and measure the same feature type; the gap to your micrometer is the fringe bias.
5. If the gap exceeds 3 µm, switch the locator to the steepest-gradient point of the profile. The gradient peak rides the sharp outer flank and ignores the soft inner fringe, and it usually lands within 1–2 µm of the CMM.
6. Lock exposure and light drive. On a 5-pixel flank a 10% level drift moves the crossing 0.5 px — 2.4 µm at 4.8 µm pixels, your whole acceptance band.

When is it right to clean the part, and how?

Clean only parts you own or have written approval to process — customer surfaces belong to the customer, and the lighting fix above already brings you inside 0.0001 in on most work.

The menu, least invasive first:

• Ultrasonic bath in a mild soap solution: lifts loose debris and oil, attacks no base metal, dries clean. It still needs approval.
• Light oil wipe, blown dry afterwards: loose swarf only, no effect on fused recast.
• Dilute hydrochloric acid dip, 2–3% (shops go to about 10% on stainless and non-ferrous): removes fused flash in a short dip. Rinse, blow dry, and prove it on a scrap setup piece before any customer part.
• Glass bead blast, 300 grit or finer at 20–30 psi: the standard for your own parts. It rounds very sharp edges, which is exactly why it stays off other people's work.

If you cut the parts as well as measure them, skim passes shrink the recast layer at the source — configuring wire EDM for micro parts and paired contours walks through pass strategy and flushing.

How do I prove the fix before measuring customer parts?

Run a ten-part comparison against your micrometer or CMM: with the collimated backlight and the half-level threshold, as-received parts should average within ±2.5 µm of the reference, and the spread across the ten decides whether a fixed correction is safe.

1. Pull ten parts from one machine, one material, one wire run.
2. Measure each on the vision system and on the reference instrument; record both numbers.
3. The mean error is your candidate correction; the spread is max minus min.
4. Spread of 6 µm or less: apply that one correction per machine and material. Wider spread: the fringe varies part to part — fix lighting, not arithmetic.
5. Recheck after any change of wire, flush setting or power schedule, and once a month regardless.

Report both numbers, as-measured and corrected, with the surface photographed under coaxial light; a customer trusts a bias they can see.

What should a no-clean measuring station look like?

Four items do it: a collimated backlight, a telecentric lens between 0.3x and 1x, a 5–12 MP camera on a 1/1.8" to 1" sensor, and a controller holding a constant drive — plus a coaxial front light when reports need surface photos.

For a second station — tall fixtures, or surface documentation:

The LS-TCL-1972 is a 1x telecentric lens for the 2/3" format, C-mount, with a 440 mm working distance for tall setups. The LS-CXL-0092 is a 70 x 70 mm unit from coaxial lights for the as-received surface view.

Final checks before you trust the station:

• Pin or gauge block within 2.5 µm, ten repeats, first part of every shift.
• Background 230 ± 10, body under 30, nothing clipped.
• Gap to the backlight held constant by the fixture, part never on the diffuser.
• One correction factor per machine and material, with its spread recorded.
• Raw and corrected values both on the report, with a coaxial photo of the flank.

Build the station in the system configurator or browse the backlight finder; if a profile still shows a lifted fringe after the changes above, send it to us through the contact page and we will read it with you.