Edge AI for Shipyards & Shipbuilding 2026 — Weld Quality Vision, Block Assembly & Harsh-Environment Compute

Building a ship is two problems in one site. The first is heavy fabrication: steel is cut, formed, and joined by thousands of metres of weld, then assembled into blocks weighing hundreds of tonnes each. The second is a marine environment that attacks electronics year-round with salt, humidity and abrasive blast media. A shipyard is therefore an unusual specification target, because the compute has to survive a salt-air factory as well as run vision and arc-sensing workloads next to equipment that draws thousands of amps and then shuts that current off in microseconds.

Almost everything valuable that can be measured in a yard is already measured by someone: the welding power source logs its arc, the dimensional survey records the block, the paint shop records film thickness. The gap is compute at the point of work, turning those streams into a pass/fail or a rework instruction before the block moves on. This guide covers the shipyard environment, the workloads worth running at the edge, and the specification rules that decide whether a box lasts a decade or fails at the first quayside winter.

A Sea-Air Factory with a Hazard Zone Inside It

No single enclosure suits a shipyard, because the salt load, the dust and the atmosphere change hall by hall. The table below maps the yard into zones, because the hardware posture follows the zone, not the brand of the controller or camera.

Three environmental facts drive the specification. First, shipyard air is a chloride-laden marine atmosphere, which places exposed electronics in the G2–G3 classes of ANSI/ISA-71.04, where bare copper corrodes without a conformal coating or a sealed cabinet. Second, the paint and blast halls are a genuine explosion hazard from solvent vapour and dust, so equipment there falls under ATEX/IECEx and cannot be “a normal IPC with a steel lid.” Third, welding itself is an electromagnetic event: the arc current switches by hundreds of amps in microseconds, which is why I/O isolation and EMC immunity are field-reliability requirements rather than compliance paperwork.

The Workloads and Where the Compute Sits

A shipyard produces high-rate data in a few places and slow, steady data everywhere else, and the two need opposite treatment. Arc welding emits a kilohertz-rate signal that only means something once a local processor reduces it to features; a paint film-thickness reading is a value per panel that any tablet can carry. The rule that keeps the architecture honest is the same across heavy fabrication: reduce at the sensor, transmit the conclusion, and keep the archive for what genuinely needs history.

The clearest case where the edge decision is made by physics rather than by convenience is the weld. A merchant ship hull carries on the order of hundreds of kilometres of weld in total, and each metre is a potential defect that will be re-inspected, re-ground and re-welded at sea-day cost if it is found late. The welding power source already produces a faithful electrical signature of the arc, and a camera watching the molten pool produces a faithful geometric one. Neither the electrical signature nor the frames are useful as a raw archive; the useful output is a per-seam record — was this pass in specification, and if not, where along the joint should a human look. The node reduces the stream and stores the conclusion.

It is equally important to say where this inference must not sit. Weld quality and dimensional fit are minutes-scale decisions, so the fabric node is a data-reduction and anomaly-detection job, not a hard real-time control loop. The genuinely fast and safety-critical functions — crane motions, personnel interlocks and the welding interlock itself — belong in the PLC and drive safety chain, and an edge AI box must never be inserted into their path. The correct relationship is one-way: the AI node consumes data, produces alarms and recommendations, and never closes a safety loop.

Selection Rules for a Shipyard

Sizing the hardware is mostly a matter of refusing to compromise on the environment. The table below lists the specifications worth writing into a purchase order, because each one maps to a failure mode that returns within a year or two if it is skipped.

• Map the yard into environmental zones before choosing a single part number; one enclosure will not serve the welding hall and the quayside.
• Reduce at the source and transmit the conclusion — arc and camera streams in, per-seam pass/fail and rework flags out.
• Keep the weld arc analysis local, because a kilohertz power-source signal only becomes useful as features, not as a stream.
• Keep every crane and personnel-safety function in the PLC; the edge node advises and never closes a critical loop.
• Treat chloride-laden sea air as a design input and specify salt-fog qualification for anything that lives outdoors.
• Timestamp across the yard with PTP so a weld record, a survey deviation and an NDT result can be correlated years later.
• Specify welding-current-grade EMC immunity rather than relying on a datasheet class that was tested in a lab, not next to an arc.
• Version-pin firmware and keep a documented update path, so a decade-old weld node still receives signed fixes.

QSCompute supplies the compute and storage that sit behind the instruments: fanless wide-temperature industrial PCs, salt-fog and vibration-qualified units for quayside and block environments, ATEX-aware edge systems for blast and paint halls, and power-loss-protected industrial storage sized for the yard's trending window. Send us the yard area, the sensor and camera list, and the data rate, and we will size the node that survives the zone it lives in.