PlatformsFaststream SiliconFaststream RadioFaststream VisionConnected EdgeFaststream SecureMobility & Rail
ProductsSemiconductor IPWireless & RANEdge & GatewaysTracking & IdentificationSoftware & FrameworksConnected Systems
Technology5G protocol stackWireless and RF architectureBaseband and low PHYForward error correctionControl and data planeHigh-speed interfacesFirmware and bootSilicon root of trustSoftware-defined vehicleAutomotive OTAFunctional safety
AIAI Engineering ServicesEdge AI & Embedded MLComputer Vision EngineeringSensor Fusion & PerceptionAI Silicon & AccelerationMLOps for DevicesAI Visual InspectionPredictive MaintenanceDriver MonitoringVideo Analytics & Safety
SolutionsSemiconductorIndustrial AIConnected ProductsAsset TrackingBluetooth AoA RTLSWearable TrackingAutomotive & MobilitySmart InfrastructureSecure IdentityWireless & SatellitePrivate 5GSmart WashroomsFuel ManagementSmart BuildingsWorker SafetyEnergy MonitoringSmart AgricultureSmart CityAutonomous PlatformsAssembly AutomationLiDAR Rail SafetyHardware Wallet
IndustriesSemiconductorTelecommunicationsIndustrial & ManufacturingAutomotive & MobilityTransportation & RailAerospace & DefenceHealthcare & MedicalEnergy & UtilitiesOil & GasRetailConsumer ElectronicsMedia & EntertainmentSmart Infrastructure & IoT
ServicesSystem Integration overviewASIC & SoC DesignRTL to GDSIIVerification methodologyDFT and silicon testLow-power designMixed-signal integrationDesign enablementFPGA DesignFPGA-to-ASIC ConversionAnalog, Mixed-Signal & RFHardware & High-Speed PCBEmbedded SoftwareCloud, OTA & Device ManagementManufacturing TransitionHow we engage
CompanyAbout FaststreamEngineering ExcellenceLeadership & OrganisationHow We EngageQuality & ComplianceStandards & EcosystemPartners & EcosystemTrust CentreLocations & DeliveryNewsroom & MediaCareersCase StudiesKnowledge CenterWhite PapersGlossaryNewsletterResources & Support
ContactStart a projectHow we engage
Talk to us
ENGINEERING INSIGHT

Catching the defect you have barely seen

The defects that matter most are usually the ones you have the fewest examples of. On a mature line the dangerous fault is rare, and a supervised classifier needs examples it does not have. Anomaly detection flips the problem: learn what normal looks like from the plentiful good parts, and flag anything that deviates.

ShareLinkedInXEmail
Train on what you have plenty of, not what you barely see.
SUPERVISED CLASSIFIERANOMALY DETECTIONWhat it learnsEach defect from examplesWhat a good part looks likeData it needsMany examples per defectMostly good partsThe rare defectToo few to train onCaught as a deviationA never-seen faultMissed entirelyFlagged as abnormalBest whenDefects are common, well-sampledDefects are rare or unknownThe riskBlind to the unseenTuning normal vs false alarmsYou cannot train on a defect you have barely seen. Model normal instead, and the rare fault stands out.
THE CHOICE

Which framing fits the defect.

Classification versus anomaly detection
SituationBetter framingWhy
Common, well-sampled defectsSupervised classificationEnough examples to learn each class
Rare, safety-critical defectsAnomaly detectionToo few examples to classify reliably
Unknown or novel faultsAnomaly detectionNo examples exist to train on
Many defect types, imbalancedBoth, layeredClassify the common, flag the rest
High cost of an escapeAnomaly detectionCatches the fault you never saw
Legitimate part variationCareful normal modellingStops variation reading as a fault
THE DISTINCTION

Model normal, and the fault reveals itself.

A supervised classifier learns each defect from examples, which works beautifully when defects are common and well sampled. The trouble on a mature production line is that the defects that matter most — contamination, a coating void, a safety-critical flaw — are rare, sometimes a fraction of a percent of output, and are precisely the ones that must never pass. There are simply too few examples to train a reliable classifier for them, and a fault the line has never produced before has no examples at all. Optimising a classifier for the frequent, benign defects while missing the rare, dangerous one is the classic failure.

Anomaly detection inverts the problem. Instead of learning what each defect looks like, it learns what a good part looks like — and good parts are abundant. Anything that deviates from that learned normal is flagged, whether or not the system has ever seen that particular defect. This is what makes it possible to catch the rare, the novel and the safety-critical: the model does not need an example of the fault, only a thorough understanding of normal. On a line where an escape is expensive, that is exactly the property you want.

The discipline moves to modelling normal well. A real part has legitimate variation — sheen, texture, print, tolerance — and a model that has learned too narrow a slice of normal flags that variation as a fault, driving false alarms. So the work is training on enough good-part variation that ordinary differences are understood as normal, and setting the sensitivity so genuine deviations stand out without burying the line in false calls. In practice the two framings are layered: classify the common defects where examples exist, and let anomaly detection catch everything else — including the fault no one has seen yet.

IN PRACTICE

Using anomaly detection well.

COMMON QUESTIONS

What engineers ask before they call.

01

Why not just train a classifier on the defects?

Because on a mature line the defects that matter most are rare — sometimes a fraction of a percent of output — and a supervised classifier needs many examples of each defect to learn it reliably. Those critical faults are exactly the ones you have the fewest examples of, and a fault the line has never produced before has none at all. Framing the problem as anomaly detection, trained on the plentiful good parts, lets the system flag deviations it has never seen a labelled example of.

02

How does anomaly detection catch a defect it has never seen?

By learning what a good part looks like rather than what each defect looks like. Good parts are abundant, so the model can build a thorough picture of normal, and anything that deviates from that normal is flagged — regardless of whether the system has ever encountered that particular fault. This is what makes it suited to rare, novel and safety-critical defects, where the whole difficulty is that examples of the fault are scarce or non-existent.

03

How do you stop it flagging good parts as anomalies?

By modelling normal thoroughly and setting the sensitivity carefully. A real part has legitimate variation in sheen, texture and tolerance, and a model that has learned too narrow a version of normal reads that variation as a fault. Training on enough good-part variation that ordinary differences are understood as normal, and tuning the deviation threshold so only genuine faults stand out, keeps the false-alarm rate low. Borderline parts are routed to human review rather than force-failed.

FOUND THIS USEFUL?

Pass it on.

Written for engineers. Share it with one.

ShareLinkedInXEmail
KEEP READING

Related work.

BUILD WITH FASTSTREAM

Bring us the difficult part.

Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.