Find the loss. Then choose the tool.

02Start with a consistent loss definition and the decision that follows detection.

03A camera that finds anomalies does not diagnose every quality issue.

04A model that predicts a failure is only useful if maintenance can act in time.

05Separate inspection, diagnosis, maintenance planning and machine control.

06The explanation uses globally relevant manufacturing methods and fictional plant arithmetic.

07It makes no Saudi or UAE defect-rate claim.

08A local implementation must account for the line, materials, suppliers, working conditions and available maintenance resources.

02It is a documented benchmark for industrial visual anomaly detection.

03It does not represent a production plant's defect prevalence or the full range of hidden material and process faults.

04Bergmann et al., MVTec AD, CVPR 2019 paper (source 1).

05The dataset owner states that MVTec AD uses a Creative Commons Attribution-NonCommercial-ShareAlike licence.

06A public download does not grant unrestricted commercial reuse; the infographic uses original artwork and reproduces no benchmark images.

07MVTec, dataset page, checked 16 September 2026 (source 2).

08Vorne's OEE educational material defines the product of availability, performance and quality, and explains the underlying time and output measures.

09This is a supplier-authored description of a widely used operating method, not neutral evidence that a particular target is optimal for every plant.

10Vorne, OEE calculation, live page checked 16 September 2026 (source 3).

11Those sources support a distinction between measuring losses and evaluating one detection method.

12They do not establish a universal AI return, a typical downtime reduction or a “world-class” threshold appropriate to every operation.

13Avoid importing such percentages from vendor sales material without a documented denominator and comparable context.

02Let operators record “unknown” rather than force a guess.

03Distinguish a stoppage caused by equipment failure from one caused by missing material or downstream congestion.

04The machine being stopped is not necessarily the machine causing the constraint.

05For quality, distinguish first-pass defects, rework, scrap and customer escapes.

06Count units consistently.

07A part repaired and accepted later should not silently become first-pass good output.

08Link defect classification to product revision, lot and inspection method.

09For downtime, ask what actionable warning exists.

10A bearing temperature signal may support condition monitoring; a material shortage may require inventory discipline; a changeover delay may need better standard work.

11These interventions have different data and ownership requirements.

12A language model summarising maintenance notes is a different project from predictive maintenance based on sensor data.

13Map the response to every alert: inspect, adjust, stop, replace, schedule or escalate.

14State who decides and how long the response takes.

15An alert that arrives five minutes before failure is not useful if spare parts require three days, although it may still support a safe shutdown where properly engineered.

16AI assistance must remain within the approved operating boundary.

17This explanation does not recommend autonomous machine-control changes.

18Safety and process control require qualified engineering review and the plant's applicable procedures.

02The line stops for 60 minutes and runs for 420.

03It produces 700 units at an agreed ideal cycle time of 0.5 minutes each, with 665 first-pass good units.

04These numbers are illustrative, not an observed factory.

05Availability is 420/480 = 87.5%.

06Performance is (0.5 × 700)/420 = 83.3%.

07Quality is 665/700 = 95%.

08OEE is their product, approximately 69.3%.

09The equivalent direct calculation is (0.5 × 665)/480 = 69.3%.

10The 60 stopped minutes are visible, but the running-period speed loss also matters: the 700 units account for only 350 ideal minutes out of 420 actual run minutes.

11That 70-minute gap needs investigation.

12Installing a defect camera addresses a different loss from either stopped time or slow running.

13If a future test prevents ten defective units at the same total production count, good output rises from 665 to 675.

14Quality becomes 96.4% and OEE approximately 70.3%, a gain of about one percentage point.

15This is conditional arithmetic, not a forecast.

16It does not establish additional sales, lower labour expense or the best investment.

02A detector catches 90 defects and misses ten.

03If it also flags 1% of the 9,900 good items, it produces 99 false alerts.

04Of 189 total alerts, only 90 are genuine defects, giving precision of about 47.6% despite 90% defect recall.

05This demonstrates why a headline “90% detection” is incomplete.

06The quality team must handle 189 alerts, and ten defects still escape detection.

07Changing the alert threshold trades one type of error against another.

08Select thresholds using the cost and consequence of errors, not just a visually impressive score.

09The assumed 1% defect prevalence and false-positive rate are fictional.

10Real prevalence can shift by product, lot and season.

11A balanced benchmark test set can substantially overstate production precision if its positive share is much higher than the plant's.

02Include new lighting, camera movement, dirt, packaging and legitimate product changes.

03Record rejected or uninspectable images.

04A model should not quietly count unreadable items as good.

05For maintenance, evaluate warnings at an action-relevant horizon and count false alarms per asset-time.

06Use chronological validation to avoid information from after a failure leaking into predictions.

07Record planned interventions that prevented a failure, because the observed outcome is affected by maintenance action.

08Compare performance against a practical baseline: current inspection, a simple threshold or existing preventive maintenance.

09Measure inspection effort, alert response, missed faults and line disruption.

10An offline score is evidence about the dataset, not proof of operating benefit.

11Financial evaluation should distinguish avoided scrap, recovered throughput and reduced downtime.

12If recovered output cannot be sold or used, do not value it automatically at full sales price.

13Avoid counting the same recovered unit in both throughput and scrap benefits.

02Each source supports only the scope stated beside the claim.

02Identify the loss and its owner before choosing an inspection, maintenance or workflow intervention.

02Better classifiers do not establish root cause.

03A model may detect a stain while missing an internal material fault.

04A plant with inconsistent downtime coding may benefit more from a reliable log and daily review than a prediction model.

05OEE can also be gamed by changing planned time or ideal cycle assumptions.

06Product mixes and asset roles differ.

07Use consistent definitions within a relevant process before comparing plants or countries.

08The target should follow customer and operating needs, not a generic league table.