Why good machined parts fail a vision check: oil, chips and fixtures

What is actually failing?

Save an exception image with the part family, recipe revision, result and cycle reference. Then inspect the physical component independently. The part may be acceptable but imaged badly; it may be genuinely defective; or the camera may have produced the correct result and the reject mechanism removed the wrong part. These causes require different corrections. A dashboard with only one reject total hides the distinction.

How should oil and chips be tested?

Use paired observations of the same parts before and after the permitted cleaning operation, without losing their identities. Include known defective parts as well as clean good parts. Oil can change reflection, and a loose chip can create an apparent raised edge. Decide with the process owner whether the station must inspect in that condition, route uncertain parts to cleaning, or operate after a controlled clean-and-dry step. A cleaning recommendation is a process change that must fit the production sequence, not a software workaround.

What should the fixture repeat?

The fixture must establish the agreed datum while leaving the inspection region exposed. Examine seating against chips, worn contacts, clamping variation and mirrored or rotated loading. Test allowed movement at the limits of the locating arrangement. Do not clamp over the edge you need to inspect or teach the system to find a fixture feature instead of the part feature. The challenge set should include incomplete seating and an empty fixture, because both can otherwise resemble a missing feature.

When is a lighting change useful?

A backlight can isolate an exposed outline while revealing little about the top surface. Directional lighting can reveal relief but can also make harmless marks conspicuous. Diffuse light may help read a reflective printed region while hiding shallow surface relief. Test the critical fault and normal variation under each candidate arrangement. Record light position, shielding, exposure and product pose so the result can be reproduced after maintenance.

How do you accept the correction?

Change one documented factor at a time where practical, then run a held-out challenge set with acceptable, faulty and borderline samples. Report false rejects and missed defects separately by defect class and part family. Include reloading, production contamination and the permitted operating range. Restore the previous recipe if the revised one has not met its acceptance conditions. Keep a versioned record of the reason, evidence and authorised change.

Burr inspection scope · Reject diagnosis

How contamination changes the feature the camera measures

An oil film changes reflection at the surface. Droplets can create bright highlights or dark boundaries that resemble edges, while a loose chip can obscure the actual outline or hold the component away from its locating face. The same part can therefore produce a different image and a different apparent dimension without any change in its machined geometry. Conversely, wiping the part can remove loose contamination while leaving a genuine burr that the original image partly hid.

Distinguish part contamination from optical contamination. A mark attached to the component normally moves with it; a dirty protective window tends to stay in image coordinates. A useful diagnostic is to move or reload the identified part within the allowed fixture range and compare the suspect feature. This is not an automatic classification rule: some illumination artefacts also change with pose, so inspect the physical part before assigning a reference judgement.

HIKROBOT’s lighting glossary explains how directional and diffuse illumination change surface appearance. (HIKROBOT Light Source Catalog, 2025 Q3, glossary) A diffuse arrangement can suppress glare but also reduce contrast from shallow relief. A backlight may stabilise an exposed silhouette without showing the top-surface burr. Evaluate the critical defect and acceptable tooling marks together when changing the light.

Run a paired cleaning and reloading trial

Hypothetical example: an acceptable turned part fails an edge check while oily. After the approved cleaning step it passes. A known burr sample also changes appearance after cleaning but must remain rejected under the same quality requirement. If both now pass, the apparent improvement in production yield is not evidence of a correct fix. Preserve both before-and-after image pairs and the independent physical judgements.

Next, reload the cleaned acceptable part repeatedly using the normal handling procedure. If the result changes with seating, examine contact faces, chips under the datum and clamp repeatability before adjusting the edge threshold. NIST’s gauge R&R framework treats measurement-system variability as something to characterise, including the conditions that vary in use. (NIST e-Handbook, 2.4 Gauge R&R studies) The hypothetical trial does not establish a universal number of repeats or a station accuracy.

Decide whether contamination is tolerated or controlled upstream

The scope must state the part condition at inspection. If the validated method assumes clean, dry parts, that condition must be produced and maintained by the process. If the plant needs inspection before cleaning, test the permitted oil and chip conditions as part of the application rather than accepting a clean bench demonstration. An unresolved image can be routed to cleaning and reinspection, but the rework identity must remain linked to the original hold.

For dimensional decisions, examine bias as well as spread. A stable oil boundary may move the detected edge consistently, producing repeatable but wrong readings. NIST separates analysis of bias from analysis of variability. (NIST e-Handbook, 2.4.5 Analysis of bias) JCGM 100 sections 4–6 provide the framework for evaluating uncertainty contributions; a pixel size or a narrow repeatability spread is not a complete uncertainty statement. (JCGM 100:2008, Guide to the expression of uncertainty in measurement)

South African production teams can use the national measurement framework when specifying traceable reference measurements. The Measurement Units and Measurement Standards Act establishes that context; it does not set the permissible burr size or the customer drawing tolerance. (South Africa: Measurement Units and Measurement Standards Act 18 of 2006) Agree the actual datum, feature definition and decision boundary with the product owner. A fixture that makes the image easy but measures from the wrong datum has not solved the inspection problem.

Preserve the correction as a controlled configuration

Record the cleaning method, fixture contact condition, lighting position, exposure, region and edge settings that produced the accepted result. Photograph the mechanical setup for maintenance records where permitted. Add a known reference check after fixture cleaning, lens-window replacement or clamp adjustment. A clean station can still be misaligned, and a replacement lamp can change contrast even when its model name matches.

Keep the limits of the correction visible. If an attached burr and a loose chip cannot be distinguished reliably in the chosen view, do not describe the algorithm as determining attachment. Use the agreed containment or secondary check. If a fixture feature sometimes substitutes for a missing part feature, introduce an explicit part-presence and location validity check before measurement.

The burr-inspection solution covers the possible station scope; the false-reject guide covers wider diagnosis. This article provides a sequence for separating physical contamination, optical change, seating variation and incorrect removal. The intended outcome is a reproducible method backed by reference evidence, not merely a smaller reject counter.