Edge AI in Steel Mills 2026:
Inference Hardware for Surface Inspection, Ladle Tracking & Mill Condition Monitoring

A steel mill is one of the few places where a computer can be destroyed by the product it is measuring. Hot strip leaves the finishing mill at 850–900 °C, the air carries conductive mill scale, and mill stands shake anything bolted to them. None of that is friendly to an office-grade PC.

This guide covers where edge inference earns its place, how line rates set the compute class, which decisions must stay inside the control loop, and what to specify.

Why the Mill Floor Eats Computers

Airborne contamination sets the baseline. ANSI/ISA-71.04-2013 grades process-control environments G1 (mild) through G3 (harsh) to GX (severe). An electric-arc meltshop with its fume and fine metallic dust, a sinter plant, and any area near a pickling line handling hydrochloric acid mist sit in the harshest bands, where unprotected copper and silver corrode by design. A fan-cooled computer pulls that air straight across its boards; a sealed or positive-pressure cabinet is what actually buys service life.

Radiated heat is the second filter. A camera or edge node standing two metres from a hot strip sees a radiant load that lifts its internal temperature far above the ambient reading on the wall, and an EAF or ladle metallurgy station adds short, brutal radiant peaks during a heat. This is why mill enclosures carry vortex or water cooling, and why the compute must be specified at its full-load temperature, not at idle in a test lab.

Vibration and electrical noise are the third. Roughing and finishing stands, shears and coilers inject continuous vibration into anything mounted on the structure, so on-machine hardware must be rated to IEC 60068-2-6 for vibration and IEC 60068-2-27 for shock. Electrically, megawatt variable-frequency drives conduct noise into the same trays that carry camera links and sensor wiring; shielded cabling, isolated digital I/O and a supply that tolerates a dirty 24 VDC bus are the difference between a model that runs and a node that resets mid-coil.

Where Edge Compute Lands Across the Plant

The split below follows where the sensor sits and how fast the process moves. The tighter the timing, the closer the compute belongs to the machine — and the smaller the node that should be doing it.

Sizing a Rolling Line: the Numbers That Set the Class

Line-scan arithmetic decides the platform. Take a galvanising line running a 1,500 mm strip at 3 m/s with a 0.2 mm minimum defect. Cross-web coverage needs 1,500 mm ÷ 0.2 mm = 7,500 pixels, and along the strip 3 m/s ÷ 0.2 mm is 15,000 lines per second — about 340 MB/s at three bytes per pixel, per camera. Three cameras put roughly 1 GB/s of raw image data on the wire.

No edge node classifies that stream, and none needs to. The architecture that works pushes pixel-level pre-filtering into the camera or frame-grabber FPGA and hands the host only defect crops with coordinates, typically a few dozen events per second. The buying rule follows directly: size the host by defect crops per second, not by camera bandwidth. The corollary matters as much: spend the budget on the frame grabber whose FPGA you can actually program, not on the largest GPU in the catalogue.

Typical 2026 street pricing runs about $250–600 for a Jetson Orin Nano Super or Orin NX module, $800–3,500 for a fanless industrial PC, $8,000–20,000 for an AGX Orin 64GB industrial system, and $2,000–9,000 for an L4- or L40S-class edge GPU server — before the mill-specific purged enclosure, vortex cooling and mounting hardware that a casting or hot-rolling location adds.

The Deterministic Half: Decisions That Cannot Leave the Loop

Some mill functions tolerate a 200 ms round trip to a server; some do not. Continuous-casting breakout prediction samples thermocouple arrays in the mould at 10–100 Hz, and the output has to reach strand withdrawal control within tens of milliseconds, because a breakout costs a strand, a tundish and hours of production. Ladle slag detection is similar: the acoustic signature of slag carry-over must be judged in the first seconds of the pour.

Both cases argue for the same architecture: the model runs on an edge node, but the trip or alarm decision stays owned by the PLC, and the edge node must publish a judgement fast and repeatably enough to be trusted. That means deterministic execution — isolated CPU cores or a real-time kernel, a watchdog, and a network path free of camera and telemetry contention.

Condition monitoring is the third case and the least time-critical, but it pays most reliably. Gearbox and bearing faults appear as sidebands around gear-mesh and bearing frequencies, which demands sample rates in the tens of kilohertz and local spectral analysis; shipping raw vibration off the machine is a bandwidth decision nobody can justify.

Five purchasing rules:

1. Design the enclosure before the board. A purged or vortex-cooled enclosure buys more service life in a steel plant than any silicon feature.
2. Match the tier to the task. ARM for sensors and slag acoustics, x86 for control and gauge data paths, GPU only for full-width strip imaging.
3. Keep the trip decision on the PLC. An inference result that has to travel to a data centre and back will never save a strand.
4. Budget storage for continuous writes. Inspection logs and event clips write far more than most mills expect; specify power-loss protection and real endurance.
5. Ask for a full-load thermal figure. Any vendor can quote an idle temperature; only a real one survives a radiant peak at the caster.

QSCompute supplies fanless industrial PCs, ARM edge gateways and GPU edge servers for steel and heavy-metals plants — wide-temperature fanless x86 for casting and gauge data paths, Jetson Orin NX and RK3588 nodes for slag detection and mill condition monitoring, and AGX Orin or L4-class systems for full-width strip inspection, plus 316L/IP69K enclosures, industrial SSD with power-loss protection and wide-input DC supplies. DDP shipping to 85+ countries.