In a Shah Alam injection-molding line, manual visual inspection runs 250–400 parts/hour per inspector at RM0.05–0.09 per unit after shift allowances, while a trained Cognex In-Sight D900 or Keyence CV-X500F inspects 4,000–7,000 parts/hour with under-kill below 0.1% — provided the training dataset, lighting booth, and PLC reject loop are specified before purchase, not after.
Manual Inspection Physics in Malaysian Plants
Manual visual inspection is scheduled fatigue management, not quality control. A QA inspector at a Sungai Buloh injection plant holds stable attention for 35–45 minutes, then degrades. Over an 8-hour shift, effective crisp-focus time is 4.5–5.5 hours. On a hand-held connector housing, inspectors process 2,500–3,200 units/day at 6–10 seconds per unit (rotating for gate scars, ejector pin marks, weld lines, flash). A secondary inspector re-screens 5–10% of doubtful units, adding 10–15% labor overhead.
Most mid-sized plants run IEC 2859-1 AQL 0.65 sampling on bulk parts, meaning a statistically accepted lot still contains 1.0–1.5% defective units. That is precisely the failure mode that triggers debit notes and line-stop penalties from Tier-1 MNC customers in automotive and medical devices.
The Cost-Per-Unit Math: RM25,000 vs RM400,000
Klang Valley QA employers pay RM2,100–2,800 per inspector per month including EPF, SOCSO, and shift allowance. A 3-shift visual inspection station needs 6–9 heads (including relief and re-screening), which is RM150k–300k per year in payroll. Per-unit direct cost lands at RM0.05–0.09, plus another RM0.02–0.05 in over-kill scrap when 1–5% of good parts are mistakenly scrapped.
An AI station — industrial PC with NVIDIA GPU, 6MP camera, telecentric lens, light booth, encoder, frame, and reject cylinder — costs RM150k–400k installed in Selangor. With one process engineer at RM7k/month and RM15k/year maintenance, the annual cost is roughly RM55k in year one and RM35k thereafter. At 4,000 parts/hour and 60% line availability over 16 hours, that is 38,000 parts/day — 12x a single inspector’s output, at about one-tenth the unit cost.
Actual Detection Performance and False-Reject Rates
Field figures match what integrators admit privately. Sustained human defect detection on cosmetic surfaces is 85–92%, with misses concentrated on low-contrast features and defects smaller than 0.8 mm. False rejects run 1–5% of inspected good parts. On a 10,000-pack blister line, experienced operators miss 6–14 packs per 10,000 when tablet color closely matches the foil backing.
A supervised deep-learning classifier trained on 1,200 images per defect class reaches 99.0–99.5% detection, with under-kill of 0.05–0.1% and false-reject of 0.2–0.4% on the same parts. The trap is lighting sensitivity: a camera operated without an enclosure under factory fluorescent that shifts at 3 p.m. will drift. Fixed-intensity LED ring-light booths are mandatory, and golden-sample validation must run at every shift change.
Integration With PLCs, MES, and Operator Stations
The camera is the easy 20%; the station is the rest. A photoelectric sensor or encoder triggers the camera. The pass/fail decision travels via PROFINET or EtherNet/IP to a Siemens S7-1200 or Mitsubishi FX5U PLC, which fires a 2-position divert cylinder or air-blast nozzle. Data output is JSON or OPC UA into the plant MES — AVEVA, Wonderware, or SAP Digital Manufacturing — for traceability and Pareto reporting by defect code.
Operator involvement is a daily reality: cleaning optics every 2–4 hours, reshooting false-reject images as “good” in the training loop, and reviewing the shift Pareto on a 21-inch HMI. Malaysian plants should budget 4–6 weeks of image collection and labeling before go-live — 500–2,000 images per defect class — and insist the vendor runs a wet-run on the actual line for five consecutive shift days before acceptance.
System Landscape: Cognex, Keyence, Omron, Hikrobot
Cognex In-Sight D900 — uses the ViDi deep-learning toolkit with strong SMT/electronics libraries. The reference choice for PCB and connector inspection, with official presence in Petaling Jaya and Penang.
Keyence CV-X500F with AI unit — aggressive direct sales from the KL office, free demos on your actual parts, but per-node licensing adds RM10k–20k for each additional camera set.
Omron FH-Series — multi-camera synchronization and built-in traceability, strong in food/pharma/automotive packaging lines with the mature FH controller ecosystem.
Hikrobot Vision Master with deep-learning SDK — priced 30–40% below the top three, but engineering support is thinner and typically routed through Penang-based automation houses. Acceptable when your in-house controls team is strong.
| Option | Key Feature | Best For |
|---|---|---|
| Manual visual inspection (baseline) | 250–400 parts/hour, RM0.05–0.09/unit, 85–92% detection | Low-volume, high-mix runs; AQL 0.65 acceptance sampling |
| Cognex In-Sight D900 | Deep learning via ViDi; 99%+ detection; native PROFINET/EtherNet/IP | SMT/PCB, connectors, Penang and Shah Alam electronics lines |
| Keyence CV-X500F | AI unit, free in-plant demo, per-node licensing | Food/pharma blister packs, cosmetic surface inspection |
| Omron FH-Series | 16-camera sync, built-in traceability, mature FH controller | Automotive Tier-1/2, multi-angle multi-camera stations |
| Hikrobot Vision Master | Deep-learning SDK, 30–40% lower price, thinner support | In-house controls teams willing to self-integrate |
Ready to Accelerate Your Digital Growth Strategy?
Partner with an industry-leading digital agency to upscale your infrastructure today.
