Machine vision and automated QC in hardware manufacturing

Traditional hardware QC is statistical: pull a handful of parts per batch, measure them, and infer the rest. It works when defects are systematic — a worn die drifts, and sampling catches the drift. It fails when defects are random: one missing spring, one dented damper, one rivet formed at half force. Random defects at 0.1% will sail through any sampling plan and land, one per thousand, in your customer's kitchens.

What in-line vision actually does

A vision station is a camera, controlled lighting, and software that answers narrow questions fast: is the spring present? is the rivet head diameter within limits? is the damper seated flush? is the plating free of bare patches? Each part is imaged as it indexes past, judged in tens of milliseconds, and a pneumatic gate diverts failures before packing.

The critical discipline is asking narrow questions. A camera that checks “is this hinge good?” is a research project. A camera that checks “is there a spring in the spring pocket?” is a solved problem that will run for years.

Where vision earns its keep on hardware lines

• Presence/absence — springs, dampers, screws, balls in retainers. Near-perfect reliability, trivial lighting.
• Dimensional gauging — rivet heads, cup depth, hole positions, to a few hundredths of a millimetre with telecentric lenses.
• Surface plating checks — bare steel, blisters and gross scratches show clearly under angled light; subtle cosmetic grading is harder and needs realistic expectations.
• Assembly verification — arm seated in cup, clip engaged, correct component variant used on mixed lines.

Where it disappoints

Vision struggles with judgements humans make holistically: “does this finish look premium?”, faint colour mismatches between batches, oil films that read as defects. It also fails quietly when lighting drifts or a lens collects dust — so a serious deployment includes calibration targets and a daily golden part that must fail, proving the system still rejects.

A reject gate that never fires is not evidence of quality. It is evidence the camera stopped looking.

Deploying without drama

The pattern that works: install the camera in shadow mode first, logging verdicts without diverting, and compare a week of its decisions against your QC team's. Tune thresholds, then switch the gate on. Feed every rejection into a simple Pareto — the point of vision is not just catching defects, it is telling you which upstream station makes them, so you can fix the cause and watch the reject rate fall.

On Klinq lines, vision stations ship pre-integrated at the points where our own production data says defects actually occur — spring insertion, damper seating, final assembly — rather than bolted on where a camera was easiest to mount.