AI Defect Detection: Six Methods, and the One Question None of Them Answer

AI Defect Detection: Six Methods, and the One Question None of Them Answer

5 November 2026

Every AI defect detection system sold into a factory answers the same question, and most of them answer it well: is this unit good. Very few of them answer the question a production manager asks about four seconds later: what did we do to make it bad.

On the line, a wire harness gets routed on the left of the bracket instead of the right. The cover goes on, and the last surface anybody could have looked at closes over it. The unit passes functional test, because the routing only matters at 60 degrees C in a customer’s rack, eleven weeks from now. The inspection camera photographed a closed cover and reported nothing, correctly.

❓What is AI defect detection?

AI defect detection uses machine learning on camera or sensor data to identify defective units during production. Six methods, in practice, account for most of what gets sold under the name: rule-based machine vision, supervised deep learning, anomaly detection, dimensional measurement, sensor-based error proofing, and action recognition. Each one judges a different thing, and each one misses a different thing.

Which of the six belongs on your line depends on what your defects actually are.

Six AI defect detection methods, and what each one is actually good at

Diagram of six AI defect detection methods

Six AI defect detection methods and the specific blind spots each one has.

Rule-based machine vision: fast, brittle, and only as good as the fixture holding the part

At the station: thresholding, edge detection, template matching, blob analysis (a blob is a group of connected pixels the camera treats as one object). The engineer defines the rule, the camera applies it, and the answer arrives in milliseconds with a reason attached. Cheap per station, relatively, and auditable.

It buys that speed with rigidity. The part has to arrive in the same position under the same light, which means a fixture and a controlled enclosure. Change the part revision and somebody reprograms the camera. The system does not learn. It measures what it was told to measure, very quickly, and stops being right the moment the world moves.

Supervised deep learning: accurate on the defects you have already had

A convolutional model trained on labeled images of good and bad units handles what rules cannot: texture, gradient, cosmetic variation. On high-volume electronics with a mature defect catalog this is the strongest of the six for surface and appearance work.

The requirement is labeled defect images, usually hundreds per class, and it lands hardest on the lines that need help least. A station running at, say, 99% yield produces very few examples of the failure you want caught. So teams wait for defects to accumulate, seed them deliberately, or buy synthetic data. None of those is free, and all of them delay the model past the ramp (the weeks when a new product first goes into volume), which is when new defects appear.

Anomaly detection: finds the defect it was never shown, and cannot tell you what it is

Train on good units only, then flag anything that deviates from normal. Autoencoders, one-class models, patch-based feature comparison. It solves the cold-start problem (no defect examples exist yet) and catches the failure mode nobody wrote into the inspection plan.

What comes back, however, is a signal that says “different”, not a verdict that says “wrong”. Somebody adjudicates every flag, and the economics of the method sit in the false positive rate. Set the threshold tight and the flags turn into a second inspection queue. Set it loose and you have a camera that agrees with you.

Dimensional and geometric measurement: the only method that returns a number you can argue with

Laser triangulation, structured light, 3D profilometry (measuring a surface point by point to build its profile), calibrated 2D metrology. Instead of a pass or a fail you get a value, a tolerance and a traceable calibration path. In a customer dispute that matters more than accuracy, because a measurement can be re-run and a judgment cannot.

Its boundary is stated in its name. It measures geometry. Contamination on a sealing surface, a missing adhesive bead, a fastener torqued in the wrong order: all outside scope.

Sensor-based error proofing: catches the wrong part, not the missing action

Light curtains across component bins, pick-to-light racks, RFID on carriers, torque tools that report to the line controller. Deterministic, cheap per point, and the reason classic poka-yoke (mistake-proofing, from the Toyota Production System) works.

Each sensor verifies the one discrete event it was wired to see. The driver reports 2.4 Nm and the station light goes from amber to green; what it does not report is which of the six fasteners received it. Wiping a mating surface, seating a gasket by feel, checking an orientation mark before the lid closes: none of these produces an electrical signal, so this layer is silent about them. Technically a defect prevention system. Also blind to the manual work that leaves no signal.

Action recognition: watches the operation, not the output

Models trained on a station’s own standard procedure judge the work rather than the workpiece: which tool was picked up, whether step 4 preceded step 5, whether the second person at a two-operator station completed their half. The trigger is a human action, not a part arriving in frame.

This is the layer PowerArena HOP (Human Operation Platform) covers, through Digital Station for guidance and per-step recording, and AI Poka-Yoke for the intervention, which holds the unit at the station instead of releasing it. HOP has a real boundary of its own, though: it needs a defined SOP and a usable camera angle. It reports how the unit was built, not whether the die inside it is good.

The cost of an escaped defect is set outside your factory, and it is not small

US-based manufacturers paid a total of $30.37 billion in warranty claims in 2025, at an average claims rate of 1.30% of product sales. The figures come from the SEC filings of more than 1,400 warranty-issuing manufacturers.

(Warranty Week, 23rd Annual Product Warranty Report, April 2026)

That 1.30% is what the defect detection layer of an entire economy failed to catch, priced by the people who received the units.

Stat card showing 30.37 billion dollars in warranty claims paid by US manufacturers in 2025.

Missed defects cost US manufacturers an average of 1.30% in product sales to warranty claims.

The counterweight is what happens when the manual steps are watched rather than sampled.

At a semiconductor reticle box assembly station, first pass rate reached 97.6% with yield holding at 95% after AI vision was applied to the manual assembly steps.

(PowerArena customer, precision semiconductor components, reported 2025)

The question none of them answer: which step caused it

Five of the six methods above judge the artifact, the finished object. They differ in what they need and in how gracefully they fail, but the record they leave behind has the same shape: a unit identifier, a timestamp, a verdict.

Read that record backwards and see how far you get. A scratch on a housing tells you a housing was scratched. It does not tell you at which of the four stations that handled it, by which tool, in which shift, or whether the same thing happened to 30 other units this month and the inspector caught 11. The measurement was correct, the investigation still starts from zero, in a meeting room, with recollection as its primary instrument. Maybe somebody remembers which shift it was, maybe nobody does.

a defect record that identifies a scratch, next to missing data for the station, tool, and shift that actually caused it.

Traditional inspection records show a failure occurred but cannot reveal which production step caused it.

 

This is how a detection loop can run in seconds while the corrective action loop runs in weeks. Detection latency is a technical parameter. Cause latency is an organizational one, and camera resolution does not reduce it. The missing input is a record of the work.

If you run one of these lines, you already know which of your steps has no witness. That is usually the shortlist.

Cause data also changes what an FMEA is built from: observed failure modes with frequencies attached, instead of remembered ones.

97.6% first pass and 0.7 minutes off the cycle, at a reticle box station

At a precision manufacturer with more than 25 years in photomask, wafer and substrate carriers, reticle pod assembly (the sealed carrier that protects a photomask in transit) ran by hand, 10 to 30 steps per operator. Quality moved with the shift roster: new hires had no experience to fall back on, veterans had habits of their own, and nobody could prove an improvement because the data did not exist.

AI vision went onto the manual steps, with continuous recording in place of stopwatch sampling and alerts on incorrect steps and abnormal pauses. Cycle time went from 3.5 minutes to 2.8 minutes, UPH (units per hour) rose 19%, and first pass reached 97.6% with yield stable at 95%. Reported in 2025. The model also forced an SOP rewrite, specifying part placements, tool positions and left and right hand assignments. The process turned out to need that anyway.

Continuous video data at a reticle box station improved first pass yield and boosted UPH by 19%.

Continuous video data at a reticle box station improved first pass yield and boosted UPH by 19%.

100% first pass yield at a station that already had MES, AGVs and IoT-connected tools

By the time AI vision was raised, a global EV manufacturer had already built the automated version of a good line. MES (manufacturing execution system) tracked production, AGVs (automated guided vehicles) moved material, IoT-connected tools logged process parameters. Early field issues kept appearing anyway, from steps that happened inside a chassis that was sealed three stations later.

AI vision was added to the existing architecture as the behavioral confirmation layer, and the AGV was interlocked to it: a unit did not move on until its actions were confirmed, so a missed step turned a moving carrier into a stopped one. First pass yield at the poka-yoke station reached 100%. Reported in 2025. Nothing was replaced. The layer added was the one watching the hands.

AI vision added on top of MES, AGVs, and IoT tools to reach 100% first pass yield on an EV production line.

Adding AI vision to an automated EV line improved the poka-yoke station yield to 100%.

How to decide: matching computer vision quality control to the defect you actually have

If the defect is surface-visible, the part is fixtured and the design is stable, then rule-based machine vision, and do not pay for a model you will have to retrain.

If the defect is visible but varies in appearance, and you have several hundred labeled examples, then supervised deep learning. Count the labeled images before you scope the project.

If you cannot supply labeled defect images because the defect has not happened yet, then anomaly detection, and budget the review labor for false positives as part of the system.

If the specification is a dimension with a tolerance, then dimensional and geometric measurement, because it returns evidence rather than an opinion.

If the failure is the wrong part from the wrong bin, or a fastener left loose, then sensor-based error proofing at the point of use. It is the cheapest reliable answer in this list.

If the defect is a missing, mis-ordered or mis-performed action, then action recognition at the station, with guidance in front of the operator and a hold on release when the sequence is wrong. This is the only one of the six that produces cause data, because it is the only one looking at the cause.

If you cannot say which of the six your defects belong to, then that is the project. Classify last quarter’s escapes by what produced them rather than by what they looked like.

Computer vision for manufacturing QC is six instruments sold under one name. The expensive mistakes come from reaching for an instrument that answers a question nobody was asking.

Every one of the six answers the first question well

A defect detection system tells you that a unit is bad while the unit is standing in front of you. Only one of the six tells you what happened to it thirty seconds earlier, and that is what a corrective action is actually made of.

Where to take this next

Six methods, one blind spot they share, and a short list of if/then rules. These are the places to go one layer deeper on any of them.

Related reading

Case studies

See it on your own line

The quickest way to find out which of the six your escapes belong to is to pick one station and look at what its manual steps are producing.

Book a consultation to see what your manual steps are producing:

Book a Consultation

Table of Contents

    Back to top