Steel surface inspection is the most demanding AI vision application in manufacturing. Hot strip exits the finishing mill at 850-950°C, traveling at 15-20 m/s, covered in mill scale and cooling water spray that creates steam clouds obscuring the surface. Cold-rolled coils present mirror-finish surfaces where hairline cracks 0.1mm wide must be detected at 30+ m/s line speed under lighting conditions that change with every coil's surface reflectivity. The cameras must survive ambient temperatures exceeding 60°C near hot mills, vibration from 2,000-ton roll stands, and iron oxide dust that coats every surface within weeks. In twenty years of designing vision systems for steel plants, I've watched millions wasted on inspection systems installed as afterthoughts — cameras mounted on structures that vibrate at roll change frequency, housings positioned where water spray blinds the optics within hours, and lighting designed for laboratory conditions that fails completely against mill scale reflectivity. We design AI vision inspection into steel greenfield plants from the ground up — specifying water-cooled camera housings, vibration-isolated mounting structures, synchronized LED strobe arrays, and high-speed linescan cameras positioned at optimal inspection points along the mill layout — so every meter of steel is inspected from the first coil. Book a Demo
Why Steel Vision Is the Hardest AI Vision Challenge
Extreme Temperature
Hot strip at 850-1,200°C emits blackbody radiation that saturates standard cameras. Air shimmer distorts images. Radiant heat melts unprotected optics. Water-cooled, air-purged housings with infrared-blocking filters are mandatory — and must be engineered into the mill structure during construction, not retrofitted onto existing stands where space doesn't exist.
Line Speed: 20+ m/s
At 20 m/s, a 0.1mm defect crosses the camera field of view in 5 microseconds. Exposure times must be under 10μs to freeze motion — requiring LED strobe arrays delivering 100,000+ lux synchronized to the linescan camera. Standard area-scan cameras are useless. Only high-speed linescan cameras (16K+ pixels, 100+ kHz line rate) can achieve the resolution and speed simultaneously.
Mill Scale & Water
Hot-rolled steel is covered in iron oxide scale (Fe₂O₃/Fe₃O₄) that varies in thickness, color, and texture with temperature and chemistry. Cooling water spray creates steam clouds and water rivulets that look like defects. The AI must distinguish real surface defects from scale patterns, water artifacts, and normal surface texture — a classification problem that defeated rule-based vision for decades.
Vibration Near Roll Stands
Roll stands generate 5-50 mm/s vibration during rolling. Camera structures attached to mill housings vibrate at roll gap frequencies. Image blur from 0.1mm of camera displacement at 20 m/s line speed destroys defect detection. Camera mounting structures must be vibration-isolated — separate foundations with damping mounts, not bolted to mill frames.
Planning a new hot strip or cold rolling mill? Book a demo to see how we engineer AI vision systems that operate reliably at 1,000°C and 20+ m/s from the first coil rolled.
Inspection Station Design by Mill Position
| Station | Camera | Resolution | Lighting | Housing | Mounting | Data Rate |
|---|---|---|---|---|---|---|
| IP-01: Roughing Exit | Linescan 4K + thermal LWIR | 0.5 mm/pixel | LED bar 50,000 lux (diffuse, penetrate steam) | Water-cooled SS, IR filter, air purge 6 bar | Isolated gantry, 3m above strip | 400 MB/s |
| IP-02: Finishing Exit | Linescan 8K × 2 (top + bottom) | 0.2 mm/pixel | LED strobe 100,000+ lux, <10μs pulse | Water-cooled SS, double air curtain, wiper | Isolated frame on separate foundation | 1.6 GB/s |
| IP-03: Coiler Entry | Linescan 8K × 2 (top + bottom) | 0.15 mm/pixel | LED bar 80,000 lux, brightfield + darkfield | Air-cooled IP66, anti-condensation heater | Dedicated gantry, vibration-damped | 1.6 GB/s |
| IP-04: Cold Mill Exit | Linescan 16K × 2 (top + bottom) | 0.05 mm/pixel | LED strobe 200,000+ lux, multi-angle (BF+DF+coaxial) | IP54 temperature-controlled enclosure | Precision rail, anti-vibration mounts | 3.2 GB/s |
| IP-05: Inspection Line | Linescan 16K + area-scan for defect review | 0.03 mm/pixel | Multi-technique: BF, DF, structured, coaxial | Standard industrial IP54 | Precision gantry, clean environment | 2.0 GB/s |
Steel Surface Defect Catalog: 50+ Types
Origin: mold, tundish, or solidification. Subsurface defects exposed during rolling. AI correlates with cast number and caster position for upstream root cause.
Origin: descaler, roll surface, guide, looper, coiler. AI identifies repeating patterns (roll circumference period) and maps to specific stand and roll position.
Origin: cold mill stands, emulsion system, tension reel. Mirror-finish detection requires 0.05mm resolution and multi-angle lighting to reveal shallow surface features.
Origin: galvanizing pot, air knife, passivation, skin pass. Coating-specific lighting (UV for zinc flower, IR for thermal patterns) needed alongside visible spectrum.
Camera Housing & Cooling Engineering
Water-Cooled Housing (Hot Mill)
Air-Cooled Housing (Warm/Cold Mill)
Need camera housings engineered for hot mill environments? Book a demo to see our water-cooled housing designs that maintain 25°C camera temperature in 80°C ambient with 1,000°C strip radiation.
High-Speed Data Pipeline
Each camera generates 1.6-3.2 GB/s raw image data. Dual cameras (top + bottom) per station = 3.2-6.4 GB/s per inspection point. CoaXPress or Camera Link HS interface. Frame grabber with onboard FPGA for real-time preprocessing (flat-field correction, strip edge detection, coordinate transformation).
NVIDIA L4 or A2 GPU per inspection station. CNN/Vision Transformer models classify 50+ defect types in real-time. Inference latency <50ms per image strip. Defect location mapped to coil coordinates (length × width). Models trained on 500K+ labeled defect images per defect class — transfer learning from pre-trained steel defect models accelerates deployment.
Full-resolution defect images stored per coil (typical: 200-500 MB per coil for defect crops). Complete coil surface map archived with defect locations, classifications, and severity scores. Time-series database for trending. Storage: 50-100 TB per year per mill. NAS/SAN with automated lifecycle management — hot/warm/cold tiering.
Defect map compared against customer specification limits. Automatic grade assignment: prime, secondary, reject, or conditional release. Disposition recommendation pushed to quality system. Operator review UI with defect gallery for override. Full traceability: every coil, every defect, every decision, every customer spec — audit-ready.
Key Benefits & ROI
Every Meter of Steel. Every Defect. Every Coil.
iFactory designs AI vision inspection for steel greenfield plants — water-cooled cameras, high-speed linescan, synchronized lighting, and edge AI classification — engineered into the mill layout and operational from the first coil.
Frequently Asked Questions
Retrofit Costs $50K-$100K Per Station. Greenfield Costs $10K-$20K.
Water-cooled housings need chilled water piping. Isolated mounts need separate foundations. Strobe lighting needs dedicated power feeds. All trivial during construction. All expensive after commissioning.
