Anomaly detection
Learn what a good part looks like and flag anything unusual — useful when defects are rare or hard to describe.
Back to all solutions
AUTOMATE INSPECTION · COMPUTER VISION
Every part,
checked.
Inspection is often expensive, inconsistent or too slow, and it relies on people staying sharp through a long shift. Cameras, microphones and machine learning can take on part of the job: finding defects, checking assemblies, reading labels and listening to finished units.
01 / IN PRACTICE
Six examples,
problem to result.
Each example is a demonstration on illustrative data, laid out the way a real project runs: the business problem, the data, the AI, the result and the potential value.
01 / DEFECT DETECTION
Could a camera find this defect automatically?
- BUSINESS PROBLEM
- Moulded housings are checked by eye at the end of the line. Cracks are missed late in a shift, and the occasional bad part reaches a customer.
- DATA
- A few hundred images of good and defective parts from a camera above the line, labelled by your quality team.
- AI
- A computer-vision model compares every part with what good looks like and classifies the type of defect, in around 20 milliseconds.
- RESULT
- Doubtful parts diverted for a person to check, with every image stored against the batch.
POTENTIAL VALUEConsistent inspection at line speed, with inspectors focused on the parts that need them.
Defect detectionDEMONSTRATION · ILLUSTRATIVE DATA
PASS 0.99 CRACK 0.94
INSPECTED TODAY4,812FLAGGED17PER CHECK21 ms
DEFECTS DIVERTED FOR HUMAN REVIEW 0.35%
02 / END-OF-LINE SOUND TESTS
Could every gearbox be listened to before it ships?
- BUSINESS PROBLEM
- Testers listen to each unit on the test stand, but judgement varies between people and shifts, and noisy units still get through.
- DATA
- Recordings of each unit’s run-up test from a microphone on the stand, linked to serial numbers and past returns.
- AI
- A model trained on thousands of good units flags tones, rattles and knocks that do not belong, and names the stage of the test where they appear.
- RESULT
- A pass or fail with the reason, logged against the serial number. Failed units are held for rework.
POTENTIAL VALUEThe same test on every unit, and evidence to hand if a customer complains.
End-of-line sound testsDEMONSTRATION · ILLUSTRATIVE DATA
- Start-up · 0–1 sOverall level within limits
- Run-up · 1–3 sNo rattle or knock
- Full speed · 3–5 s1.9 kHz gear-mesh tone, 6 dB above good units
- Run-down · 5–6 sBearing noise normal
VERDICT · FAIL
ACTIONSGear-mesh whine at full speed, unlike the 2,400 good units it was trained on.
Unit held · sent to rework
Result logged against SN 24-0920
FAILED UNIT HELD BEFORE SHIPPING 6 sec test
03 / ASSEMBLY & KIT CHECKS
Could one photo confirm a kit is complete?
- BUSINESS PROBLEM
- Kits, trays and stores are counted by hand, and a missing component is only found when the assembler reaches for it.
- DATA
- Photos of the tray, rack or pallet from a fixed camera or a phone.
- AI
- Object detection finds and counts each item and compares the result with the bill of materials or stock record.
- RESULT
- Counts synced to the ERP, with gaps and missing parts highlighted on the photo.
POTENTIAL VALUEFewer stoppages for missing parts, with the photo kept as the record.
Assembly & kit checksDEMONSTRATION · ILLUSTRATIVE DATA
gapgap
DETECTEDBRG-6204LUB-EP2PNT-7035Gaps to restock
COUNT SYNCED TO ERP 1.2 sec
04 / CLASSIFYING & SORTING
Could a camera grade and sort material consistently?
- BUSINESS PROBLEM
- Grading depends on who is on shift, so the same material can end up in different bins.
- DATA
- Images of items on the conveyor, with the grades your experienced staff assign.
- AI
- A classification model gives each item a grade or material type from its appearance, with a confidence score.
- RESULT
- Items diverted automatically, or the grade recorded for a person to confirm.
POTENTIAL VALUEConsistent grading on every shift, and a record of what went where.
Classifying & sortingDEMONSTRATION · ILLUSTRATIVE DATA
Copper+1Steel+1Aluminium+1
GRADED + DIVERTED AUTOMATICALLY 0.3 sec / item
05 / READING CERTS & LABELS
Could goods-in read the certificate for you?
- BUSINESS PROBLEM
- Material certificates and delivery notes are checked and typed in by hand, slowing goods-in and risking traceability errors.
- DATA
- Scanned or photographed certificates, delivery notes and labels, and the purchase orders they should match.
- AI
- A vision model reads the fields that matter — supplier, grade, heat number, PO — and checks them against the order.
- RESULT
- Fields filled in automatically, with anything uncertain sent to a person to check.
POTENTIAL VALUETraceability without the typing.
Reading certs & labelsDEMONSTRATION · ILLUSTRATIVE DATA
CERTIFICATEDale Steel Supplies LtdNo. CERT-20418Grade S355J2Heat no. 7H4410PO 5521
- supplier
- Dale Steel Supplies Ltd0.99
- cert_no
- CERT-204180.99
- grade
- S355J20.98
- heat_no
- 7H44100.97
- po_ref
- 5521check
4 FIELDS MATCHED TO PO · 1 SENT FOR REVIEW 0.8 sec
06 / SAFETY MONITORING
Could your existing cameras warn of a hazard as it happens?
- BUSINESS PROBLEM
- Near misses around presses, robot cells and forklift routes are only known about when someone reports them.
- DATA
- Video from cameras you already have, with the exclusion zones and PPE rules for each area.
- AI
- A detection model spots people, vehicles and PPE and checks them against the zone rules — without identifying anyone.
- RESULT
- An alert to the area supervisor within a second, and a log of events for safety reviews.
POTENTIAL VALUEEarlier warnings, designed with privacy in mind.
Safety monitoringDEMONSTRATION · ILLUSTRATIVE DATA
EXCLUSION ZONEforkliftperson · PPE ✓person · PPE ✓IN ZONE
- Person entered exclusion zone — supervisor alerted
- Fire exit B clear
- Hi-vis check · 12 of 12 people
ALERT SENT · NO FACES STORED 0.4 sec
02 / AND MORE
Other ways to
automate inspection.
Measurement & gauging
Measure dimensions, gaps, weld beads or coverage directly from an image and track them against tolerance.
Thermal imaging
Spot hot spots and uneven heating from thermal camera images.
Process monitoring
Watch flames, melt pools, foam levels or conveyor loading and flag when the process starts to drift.
Spare part identification
Photograph a worn part and find the closest matches in your stores catalogue, part numbers attached.
Privacy protection
Blur faces and number plates automatically before footage is stored or shared.
03 / BEFORE WE BUILD
What we look at
before building anything.
Representative examples
Images or recordings from the line itself — its real lighting, speeds and part variety. Often fewer than you might expect to test an idea.
A clear measure of success
Agree what accuracy is useful and what a mistake costs. A missed defect and a false reject rarely matter equally.
The right place to run it
On an industrial PC beside the camera, a server on site or in the cloud — chosen for speed, cost and where data must stay.
People in the loop
Uncertain results go to a person, and their decisions can be used to improve the model over time.
04 / HOW WE'D START
Idea → evidence
→ investment.
You never commit to a large project on a hunch. Each step answers whether the next one is worth taking — and “no” is a useful answer too.
- 01
Free conversation
30 minutes about a process, machine or business problem. No technical sales pitch.
- 02
AI Opportunity Review
A short paid review of the problem, your data and systems, ending in a roadmap and a recommendation.
- 03
Proof of concept
One problem tested on your real data, with a clear Go, Modify or Stop at the end.
- 04
Build & improve
Only if it works: a practical system, deployed, monitored and refined.
05 / AUTOMATE INSPECTION
Inspection slow
or inconsistent?
Tell me what you make and where inspection struggles. In a free 30-minute conversation we can talk through whether computer vision could take on part of the job, and how to test it on your own parts first.
Independent · North Yorkshire · working across the UK