Vision systems that cannot fail: design for zero service visits

Most machine vision systems are designed on the quiet assumption that someone can drive to them. Remote lines, offshore platforms, cleanrooms and machines that run unattended break that assumption: the fixes a service visit provides — reseating a cable, swapping a camera, retuning an exposure — do not exist there. Designing vision systems that cannot fail means moving every one of those fixes back to the workbench, before the panel closes. This page gives you the discipline: rank the failure modes, lock the connectors and the exposure, size the thermal and power margins with arithmetic instead of habit, and burn the assembled system in until the weak part shows itself.

What changes when nobody can reseat a connector

A standard design accepts marginal choices because service will clean them up: a latch-tab connector, an auto-exposure loop, a supply running at its rating limit, a fan that clogs over months. Each becomes a defect the first time nobody can open the cabinet. Zero-service design is not the same design with better parts — it is the same failure list, closed at the bench instead of in the field. It is the discipline flight hardware lives by, and industrial versions of it appear wherever the cabinet cannot be opened.

The four fixes a visit normally provides map one-to-one onto design tasks:

• Reseating a connector becomes connector selection and strain relief.
• Swapping a failing part becomes derating and burn-in screening.
• Retuning exposure becomes locked settings sized from optics and vibration.
• Power-cycling a hung system becomes a watchdog with a known start-up state.

Everything below is one of those four, with numbers attached.

Rank the failure modes before picking parts

Rank them in the order a technician meets them. Connectors and cables come first: they are the only parts handled during installation, they carry the machine's vibration, and their failure is instantaneous. Thermal follows, and it is slow — months of drift, then a shutdown on the hottest day of the year. Drift in settings and brightness comes next, then hung acquisition software, then the sensors themselves, which fail far less often than everything wired to them. Spend design effort in that order.

Set the acceptance bar first: worst-case error-rate planning tells you how many missed parts the process can absorb, and that number decides how much of the budget the reliability work below deserves.

Lock every connector and strain-relieve every cable

A plain RJ45 latch tab is a spring, and a spring on a vibrating machine walks out of its socket or fatigue-breaks. Specify metal-to-metal locking: RJ45 with a locking stud, M8 or M12 threaded rings for power and I/O. Then stop the cable from fighting the connector: clamp it within about 100 mm of the plug, leave a service loop so tension never reaches the contacts, and run drag-chain-rated (high-flex) cable anywhere the run moves or sits on vibrating structure.

Locked, flex-rated cable for the three runs every camera needs:

The LS-CBL-0199 is a 3 m GigE camera cable with screw locks and high-flex, drag-chain construction. Its USB3 counterpart is the LS-CBL-0219, also 3 m and drag-chain rated. Power and I/O ride the LS-CBL-0300, an M8 8-pin, 8-core cable of 7 m built for drag-chain use. For a longer GigE run, the LS-CBL-0117 carries RJ45 with a locking stud over 15 m. Pick the remaining lengths and connector styles in the camera cable finder or the cables category.

Global shutter, locked exposure

If the part never moves but the camera's mount does, shutter type still decides image quality. A rolling shutter exposes rows in sequence across the frame time, so vibration during readout shears or tears the image of a stationary part. A global shutter exposes every row at the same instant and removes that failure at the source. On any machine with motors, fans or passing loads, treat global shutter as the default; it is standard across the area scan cameras range.

Auto exposure and auto gain are the other quiet failure. The control loop re-aims whenever a pixel statistic moves, and sensor temperature, LED aging and part shading all move them over months of unattended running — each re-aim silently changes the exposure you validated. Lock exposure, gain and white balance at commissioning, store the values, and treat any later change as an alarm. The brightness-drift diagnosis shows what that drift looks like in data, and locked exposure is one pillar of the stable 2D inspection set-up.

Lock the start-up path too: after a power cut the camera must reload the same stored configuration, never a partially applied one. The known-state restart method is written around a 3D camera, and the discipline transfers directly.

The MV-CH120-11UC is a 12 MP USB3 camera with a global shutter and a 1.1-inch sensor, a workhorse class for a locked-exposure station.

Size the exposure for vibration, not for brightness

Once exposure is locked, choose the value from the vibration budget. For sinusoidal vibration with peak amplitude A at frequency f, the peak velocity of camera or part is v = 2πfA. During an exposure t the image smears by blur = v × t, and the requirement is that blur stays below one pixel.

Worked example. A camera mount sits near a pump running at 50 Hz; a displacement probe reads 0.25 mm peak amplitude, so v = 2π × 50 Hz × 0.25 mm = 78.5 mm/s. The camera covers a 60 mm field of view with 4096 pixels, so one pixel is 60 mm ÷ 4096 = 14.6 µm. The exposure budget is t_max = 14.6 µm ÷ 78.5 mm/s = 0.19 ms. Set the exposure to 100 µs: blur = 78.5 mm/s × 0.1 ms = 7.9 µm, about 0.54 pixels — inside the budget with margin. Now narrow the field of view to 30 mm on the same sensor: the pixel halves to 7.3 µm, t_max drops to 0.09 ms, and 100 µs no longer passes — bring the exposure down to about 50 µs, where blur is 3.9 µm, again 0.54 pixels. Halving the pixel size halves the exposure budget; that is the sensitivity to watch.

Going from 1 ms to 100 µs costs 3.3 stops of light. Recover it with a strobed LED pulse during the exposure window rather than by opening the lens: aperture buys brightness and spends depth of field, which the depth-of-field calculator quantifies. Choose the strobe from the machine vision lights finder and set its pulse width to the locked exposure.

If the disturbance is an impact rather than a steady sinusoid, skip the formula and measure displacement directly during a representative exposure. The pixel size in the budget comes from field of view and sensor resolution — run your own numbers here and shortlist cameras with enough pixels:

Budget power and heat for the hottest ambient

Every watt a vision component dissipates ends up in the cabinet air, and a sealed unattended cabinet has no filter to clean and no fan curve to trust. Work the budget at design time. Say the camera draws 6 W, the controller 15 W, and two 15 W strobe lights pulse at a 50% duty cycle: steady dissipation is 6 + 15 + (2 × 15 × 0.5) = 36 W. At 24 V that is 36 ÷ 24 = 1.5 A. Rate the supply at roughly twice the steady draw — a 3 A floor, 5 A in practice — so inrush, LED pulses and component aging never become a visit. Take the real wattages from each part's data sheet and write the finished budget inside the panel door.

Controllers carry most of that budget, and the vision controllers range is built around it: the MV-VC2140-128G40-1T runs an Intel N-series processor (N97) with 8 GB of memory, 3-4 LAN ports, 2 USB 3 ports and four built-in light-output channels.

Sealing raises its own question — what the enclosure rating means for a housing nobody will open for years; the NEMA 4X versus IP66 ratings guide sorts out which rating a washdown or outdoor cabinet actually needs. And if processing outgrows a vision controller, a fanless industrial mini PC keeps the cabinet closed — the alternative, a filtered fan, is a maintenance schedule you have just promised not to have.

Watchdogs and the alarm path

Software hangs differently from hardware: the stream goes quiet while everything stays powered. Assume the hang will happen. Put a watchdog on the acquisition process — if the frame counter stops advancing for a set interval, stop the stream, close the device and reopen it from the stored configuration. Cap the retries: a camera restart, then a controller restart, then a halt into a safe state with the alarm raised, never an endless reboot loop. Prove the restart during burn-in by pulling a cable and letting the watchdog recover on its own. Multi-camera stations need this per camera; the multi-camera capture design article covers the capture-side structure.

Map health to the PLC as well. A system nobody can fix should at least be visible: run a heartbeat bit, an error code and a frame counter per camera into the alarm map, using the vision-to-PLC diagnostics mapping. The failure may still need a visit; the difference is that the visit starts with the fault already named.

Burn it in before it ships

Burn-in is the last chance to fail on the bench. Assemble exactly as installed — same cable types and lengths, same clamps, same cabinet temperature — and run continuous acquisition for 48 hours at the hottest ambient the enclosure will see, heating the room or box artificially if it cannot reach that temperature naturally. Acceptance is zero dropped or torn frames in the camera's frame counter across the full soak. Force at least 10 power cycles during the run and confirm every cold boot loads the identical locked configuration.

If a fault appears during the soak, isolate it before anything ships: the unit-specific fault isolation method separates a marginal cable from a marginal unit in a few swaps, on the bench where swaps are free.

When the usual answer is wrong

• "Keep a spare on the shelf." The shelf is a service visit in disguise. Where the cabinet cannot be opened, overrate the installed parts instead; genuine redundancy means both sensors installed and both burned in.
• "Auto-exposure will handle the variation." It handles it by moving the exposure you validated. Lock it, and fix the variation in the light.
• "Any Cat6 with a decent latch." Latch stiffness is not locking. Under vibration the latch tab is the first part to go — specify the locking stud or threaded ring.
• "Add a fan with a filter." A filter is a maintenance schedule. Seal the cabinet and spend the budget on low power instead.

Pre-install checklist

• Every connector specified with mechanical locking — threaded ring or locking stud; no bare latch tabs.
• Every cable clamped within about 100 mm of its connector; service loops keep tension off the contacts.
• Motion and vibration paths run drag-chain-rated cable; lengths recorded.
• Exposure, gain and white balance locked, stored, and identical after every cold boot.
• Exposure sized against the vibration velocity budget, not against brightness alone.
• Power budget written down; supply rated at about twice the steady draw at worst-case ambient.
• 48-hour burn-in complete: zero dropped frames, at least 10 proven power-cycle restarts.
• Watchdog restart proven by pulling a cable during the soak.
• Heartbeat, error code and frame counter mapped to the PLC.
• Spare cables pulled to every camera while the conduit is open, labelled at both ends.

From bench to sealed cabinet

Size the camera and pixel budget in the camera finder, take the cable list from the section above, and let the system configurator walk the rest of the chain — lens, light, controller, trigger. The capture-side companions to this page are collected on the systems guides topic page. If the deployment is genuinely unreachable, send the layout through the contact page and go through the failure list with us before the panel door closes.