Before anyone signs a purchase order for an AOI inspection system, a feasibility study has to show four things. The defect classes are written down with optical criteria that someone has signed off. Nuisance variation (supplier tint, gloss, humidity, lighting) separates from those defects. The sensing modality was chosen on that evidence. The bench trial is large enough to support a go/no-go call. Take the question of whether a vision system can tell supplier-batch tint drift from a genuine insulation defect. Doing so needs white balance and greyscale calibration before every inspection and comparisons made under the same lighting.

Each of these lines brings its own confound. Varnish tint drifts between suppliers on wire, discoloration sits next to delamination on coated coil, visible stitching faults share a part with hidden foam-density faults in trim, and tyre lines run through wet and dry seasons. There is no single lighting setup, vendor demo or standard that is a reliable shortcut across all of these. Each one gets its own protocol and its own reference samples. TechnoLynx’s work on AI engineering for manufacturing and automotive lines follows the same order: the defect taxonomy and an optics and lighting trial come before the imaging and model approach is chosen. How these questions split across a wider programme, and why they rarely share one answer, is covered in scoping an AOI programme’s feasibility and defect taxonomy questions.

The AOI feasibility sequence, in the order it has to run

1. Define each defect class and its accept/reject criteria with the quality team, and get sign-off before any modality or model is chosen.
2. Lock down imaging: white balance and greyscale calibration before every inspection, and comparisons made within the same acquisition under the same lighting.
3. Assemble OK and NG samples and run real imaging and judgement tests on them to see whether imaging captures the defect features.
4. Test whether defects separate from nuisance variation in every supplier, season and lighting condition, and pick the modality on that result.
5. Define the inspection objective, then size the bench trial and document its AQL and LTPD to show it meets that objective.
6. Requalify against the same reference standards whenever a supplier, batch, coating or lighting component changes.

Can a vision system tell supplier-batch tint drift apart from a real insulation defect?

Yes, under conditions. Separating varnish tint drift from insulation flaws depends first on the optics and lighting design and on the reference panel it is validated against, and only second on the classifier. According to NITR’s colour-difference inspection guidance, colour difference in coated and dyed materials comes mainly from batch variation in the coating, inconsistent base colour of the material and uneven illumination. Only the first two are about the material, because the guidance classes illumination and uncalibrated white balance as systematic errors of the inspection, not a material problem.

The same guidance says light source colour temperature drift, camera white balance changes and gloss differences all make one colour read as different values, so calibration has to come before every inspection. A 2024 survey of machine-learning defect detection says industrial environments frequently present challenges such as inconsistent lighting, noise and complex backgrounds. Where shop-floor lighting is inconsistent, defect visibility can vary before tint variation even enters the picture.

Standard lighting plus software, or UV imaging?

Whether standard lighting plus software is enough depends on what the panel shows. Product preference does not come into it. NITR’s guidance judges colour in a perceptually uniform space such as Lab and checks the total ΔE separately from its lightness, red-green and yellow-blue components, because a directional shift can hide inside a small total difference. Suppose tint and defect separate cleanly on a golden-sample panel across several supplier batches under controlled visible-spectrum lighting. Then visible-spectrum lighting is enough to proceed with. If the two overlap, that is the point to test other modalities.

Why a model trained on one supplier’s varnish can fail on another’s

A 2025 analysis of industrial anomaly detection found current detectors prone to changes in data distribution, because they are designed to flag minimal deviations from their training data. On the AeBAD aero-engine blade dataset, whose shift comes mainly from illumination and view, shifted normal samples raised the anomaly score of normal regions for every method tested, producing large numbers of false positives. A new supplier’s base tint is the same kind of shift. The loud failure is a wave of false rejects. The quiet one happens when the variation is filtered away and some of the defect signal goes with it.

Requalification therefore belongs in the protocol. MIL-HDBK-1823A (2009) holds that the calibration standards used for verification should be substantially the same as those used in the original demonstration, so that the demonstrated detection performance still applies. For varnish, that means a golden-sample panel with a known tint range and known defect types. The panel is rerun whenever a supplier or batch is introduced and refreshed when the coating or lighting hardware changes.

Writing a tighter tint tolerance into supplier contracts helps by narrowing what the system must tolerate. It still depends on supplier compliance and does not replace separation. NITR’s point that the quality department must set the allowed colour difference quantitatively cuts both ways: set it too strictly and the system over-rejects, too loosely and defects get through.

Tighter supplier colour tolerance or a smarter vision model?

A vision system can treat supplier tint drift as a nuisance variable and still catch insulation defects only when the imaging modality separates the two signatures and the defect taxonomy is fixed before the trial. Tightening supplier colour tolerance is a lever to use alongside that, not a precondition.

TechnoLynx runs this work through its Computer Vision R&D engagement, which trials optics and lighting to separate tint drift from insulation defect signatures and validates false accept and reject rates across supplier batches before deployment; the first step is agreeing with your team what counts as a genuine defect and what counts as tint drift.

The FibreEUse evaluation describes hyperspectral imaging as capturing spectral data across a wide range alongside spatial information, which lets it detect small differences in moisture and chemical composition. It also comes with costs a production line feels. The large number of spectral variables demands many samples and invites classifier overfitting, and illumination varies across scan lines enough that converting radiance to reflectance through calibration is essential. Line-rate constraints for hyperspectral or multispectral equipment have to be validated in the trial. A datasheet will not settle them.

The trade-off in filtering variation out

Filtering variation out of the model’s evidence is the obvious fix, and it has a trade-off that needs stating plainly. The ShiftSplit-AD study tested whether domain shift and defects split cleanly into low-rank and sparse components on aero-engine blade images, and its results rejected that split. Real defects also carry correlated structure, so aggressive filtering suppressed nuisance activity but threw away defect information as well. On one defect class, breakdown, the decomposition suppressed residual structure that the raw detector had been relying on. The authors keep their conclusion deliberately conditional: filtering helps when domain variation contaminates the raw evidence, and its value depends on whether what it removes is nuisance or defect. Applied to varnish, the risk is that a model tuned to ignore tint can fail silently, because it can stop seeing a discoloured insulation flaw that looks like tint.

Asking suppliers for tighter colour tolerance may just move the variance somewhere else instead of removing it. The decision follows from the trial panel. If one modality separates tint from defect cleanly on the trial panel, build around it and leave supplier tolerances as they are. If none does, tighten tolerances, or pair a modest tightening with a less demanding robustness target. Separation found on sampled batches is still only a sample, so validation has to cover supplier variation over time.

Discoloration or delamination on coated coil: write the classes before choosing the camera

Discoloration and delamination need separate class definitions because they are physically different defects: one is a change in surface colour or reflectance, the other a loss of adhesion between layers. Enamelled-wire practice shows the split. LP Winding Wire’s eight-type defect taxonomy defines peeling as loss of adhesion between coating and conductor, which shows up as delamination and lifting, and lists scorching and discoloration as a class of its own.

Coated-steel practice already offers classes to align with. The American Galvanizers Association’s inspection guide judges surface conditions by their effect on long-term corrosion performance. Appearance variation such as bright, spangled or matte grey finishes has no bearing on protection, while heavy wet storage stain must be removed or the part rejected. Guidelines on downgraded steel products define annealing halo as an oxidation discoloration at the edges of cold-rolled products, ranging from straw yellow to dark blue.

Where an industry taxonomy has no class separating discoloration from early delamination, the extension should be explicit and recorded as a deviation. MIL-HDBK-1823A asks for the defect feature to be defined so that it can be measured unambiguously and similar targets produce similar outputs. The VDI/VDE/VDMA 2632 Part 3 standard on classifying machine vision systems introduces characteristics for assessing their classification performance at acceptance, which gives sign-off a shared vocabulary between user and supplier.

Modality follows the physical signature. Visible light is enough if colour and reflectance separate the classes. Multispectral helps if subtle delamination needs extra bands. Thermal earns its place only if the air gap leaves a detectable thermal signature. Thermal is not a free option: the FibreEUse authors note that it needs sensitive, expensive instrumentation and skilled operators, and that conduction and convection can make localisation harder. The taxonomy needs a rule for cases between discoloration and early delamination and a reference method to settle them, and the modality choice needs validating on the reader’s own coatings and severities.

Stitching versus foam density: which defects can a camera actually see?

Camera-based inspection, black-light and RGB included, suits visible stitching defects. Foam-density defects sit below the surface and should be presumed to need a different sensing route, such as X-ray or indentation force testing, until a feasibility test shows otherwise. The 2024 defect-detection survey separates surface defects (localised irregularities such as cracks, scratches and colour variations) from structural defects in the overall structure and shape of the material. Steel guidance on physio-chemical faults such as inclusions and piping says they generally cannot be seen with the naked eye and are found by ultrasonic, mechanical or chemical tests.

A hidden defect can still be detected by other means. Laboratory X-ray scattering detected local fibre shifts in glass-fabric thermoplastic composites, and a shift in the hidden middle layer was as detectable as one in the top layer. That study was on glass-fabric thermoplastic composites, not foam, so it shows what kind of method can work. It is not a result for seat foam.

The taxonomy follows from that split. Stitching-type defects go to the camera system with agreed severity thresholds. Foam-density defects go to a separate inspection stage, and the edge cases near the surface that could fall to either modality get an explicit rule. If a stitching fault is visible only under the lighting a human inspector uses, the camera setup has to reproduce it or the class has to be redefined.

Who signs off that the taxonomy matches the inspectors?

The evidence for sign-off is that the taxonomy reproduces what the current black-light inspectors flag, within a tolerance agreed in advance. MIL-HDBK-1823A treats the human inspector as often the most significant variable in an inspection process and asks for several randomly chosen operators in the test plan. With two inspectors, the first number to measure is how often they disagree with each other, because that sets the ceiling the automated system is judged against. Suppliers and the quality team can agree thresholds jointly, but supplier agreement does not guarantee inspector agreement, so a reconciliation step has to come before manual inspection is replaced.

How do you get wet-season and dry-season variation into a bench trial without waiting a year?

A defensible bench trial samples both suppliers and both seasonal production regimes and treats each supplier-season combination as a stratum with its own minimum count. Archived or conditioned samples can stand in for the missing season, but only where they can be shown to reproduce the defect that real production does. Humidity is a documented driver of coating defects. LP Winding Wire lists ambient humidity above 70% RH among the causes of enamel pinholes, because moisture mixes into the coating. Escobar and Meeker’s review of accelerated test models notes that humidity matters for the service life of paints and coatings, and that it is usually combined with temperature when used as an accelerating variable.

The University of Maryland’s CALCE centre describes humidity cycling as accelerating damage from repeated moisture absorption and desorption. Under HAST conditions above 100 °C, it puts the acceleration of moisture-driven corrosion in integrated circuits at up to a factor of 50. Acceleration does not make a sample equivalent to the real thing. The review is explicit that every accelerated test involves extrapolation, that the test must produce the same failure mode seen in the field, and that high stress levels can create failure modes that never occur in use. For a vision trial, a conditioned defect has to look like a seasonal one to the camera. Sharing a cause is not enough. The review also recommends comparing against field data, since field data usually carries extra variability and understanding the gap is what makes laboratory results usable. If conditioned samples cannot be shown equivalent to archived real wet-season samples, the go/no-go waits for real seasonal data.

Sizing the trial for a go/no-go call

Taylor Enterprises’ guidance on statistically valid sampling plans is to define what the inspection must achieve first, then document the plan’s AQL and LTPD to show it meets that objective. MIL-HDBK-1823A asks for at least 60 targeted sites for a hit/miss system and at least 40 for a system with a quantitative response, plus at least three times as many unflawed sites as flawed ones so that false positives can be estimated. False-positive precision has its own floor in the handbook: showing a rate below 1% needs at least 100 false-positive opportunities, and below 0.1% needs at least 1,000. The handbook also notes that repeated inspections of the same specimen tell you about the inspection, not the specimens, so imaging one defective tyre ten times does not count as ten samples.

Worked example: two suppliers, two seasons

Assume a classifier with a quantitative score, two suppliers and two seasonal regimes. Assume too that the handbook’s minimums apply per stratum, which is a judgement call: pooling the strata would hide a weak one. If the target is a false-positive rate below 1% in each stratum, each stratum also needs at least 100 false-positive opportunities, and the 40 targeted sites per stratum, times three, give 120 unflawed sites, which covers that. For contrast, the X-ray scattering study showed a detectable signal with 9 samples (3 controls and 3 for each defect location). That is enough to show a signal exists and far short of a capability claim. ShiftSplit-AD’s protocol used 521 normal training images and 1,639 test images across four defect classes.

Defect samples are the bottleneck. The 2024 survey notes that tight yield control makes defective samples scarce and that collecting and annotating them is costly. If one supplier’s defect rate is too low to fill its stratum in practical time, that is a finding to report. Synthetic images should not paper over it.

The go/no-go rule follows: move to a pilot cell only when detection meets the agreed threshold in every supplier-season stratum at the computed sample size. The workflow needs fault tolerance and a willingness to have humans review uncertain parts. The costs are not symmetric either. The 2025 anomaly detection analysis points out that misclassifying a defective part costs far more than a false reject, and AUROC alone does not reflect that. Bench results do not carry over to the line automatically. MIL-HDBK-1823A asks that demonstrations follow production procedures and are not tuned to score well. The pilot cell is where transfer to the line gets proven, and repeatability under line conditions is covered in validating feasibility and repeatability of vision-based defect detection before production.

Bench data fell short: is it the camera, the model or the test set?

Three different limits produce the same disappointing number, and what tells them apart is where the defect signal disappears. If the signal is missing from the raw capture, the limit is hardware or optics. If it is in the capture but the model misses it, the limit is the algorithm. If the system works on what it was tested on but the test did not cover the variation or count enough sites, the limit is coverage. MIL-HDBK-1823A defines the inspection threshold as the signal level below which the response cannot be told apart from noise. The handbook’s components-of-variance approach assigns the observed variance to sources such as probe, operator and underlying variance, so the cause of poor performance can be identified.

Symptom on the bench	Likely limit	Check that confirms it
A trained inspector cannot see the defect in raw images, even at best focus and lighting	Hardware: resolution, optics or illumination geometry	Re-image the reference defects at higher resolution or another geometry; if the defect appears, the original setup was the limit
Repeated captures of the same defect give scores that straddle the threshold	Acquisition noise	Repeat inspections of the same specimens; the spread describes the inspection, not the parts
The defect is visible in raw images, but the model misses it or labels it as a nuisance class	Algorithm	Try other features or models on the same captured images; if separation improves, the hardware was adequate
Good results overall, failure on one supplier, season or lighting condition	Coverage of variation	Score each stratum separately; the test set has to span the full range of the confounding variable
Point estimates look fine, but confidence bounds are too wide to decide	Coverage of sample count	Compare site counts with the 60 hit/miss or 40 quantitative minimums and the three-to-one unflawed ratio

Coverage limits can be disguised. MIL-HDBK-1823A warns that when specimen surface texture differs from the production part’s, the false-positive rate on real parts can come out higher. A limit can also be specific to one defect class. In the FibreEUse hyperspectral study, pointed-impact damage was detected at 97.23% precision and 94.66% recall, but wedge-shaped impact damage reached only 87.56% precision and 41.24% recall. The authors put this down to impact energy too low to leave visible surface damage and to complex surface texture hiding what damage there was. A better model fixes neither cause: the first is a physics limit and the second is a confound, so the right response is a different sensor or a narrower claim.

Two feasibility tracks to settle before writing a camera spec

Two questions need separate feasibility studies before any camera or lighting specification exists: whether delamination and discoloration on galvanised coil separate optically, and whether foam-density defects can be detected optically at all. TechnoLynx runs the coil question as a feasibility-first study: a small labelled pilot dataset imaged under several illumination geometries shows whether the two classes can be told apart before any optics or lighting are bought. Whether a separation found in that pilot holds across surface variation and line conditions is the subject of validating feasibility and repeatability before production.

Is foam density visible to light-based imaging at all?

Foam density is a gating decision, not a tuning one. If density variation leaves no signature on the surface, no lighting geometry will recover it. The 2024 survey notes that combining thermal, ultrasonic and optical sensing addresses the limits of single-modal methods, and cites phased-array ultrasonics fused with pulsed thermography for internal defects such as delamination in composites. Near-infrared and ultrasonic thickness or density sensing are candidates to test against the foam. If stitching and foam density turn out to need different sensors, they become two tracks with separate budgets and separate go/no-go decisions.

Who can run an AOI feasibility study before we commit capital?

TechnoLynx runs AOI feasibility studies for manufacturers who need to know whether real defects separate from their own nuisance variation before they buy inspection hardware. When choosing a partner, ask to see the validation protocol: per-stratum sample counts, false accept and reject rates by supplier, a requalification plan and a willingness to return a no-go.

Off-the-shelf systems can work in controlled settings, and the case for engineering work grows with uncontrolled variation. For foam density, the sub-surface track may belong with an NDT specialist, and our part there is the engineering, not certification of the test method.

FAQ

What is defect detection rate?

In the FibreEUse study, detection was reported as pixel-level precision (the share of flagged pixels that are real damage) and recall (the share of real damage found), computed against manually defined ground truth. NDE reliability work reports probability of detection curves together with false-positive estimates.