Automating plant inspection rounds: what vision can digitize

Most plants still run operator rounds on paper: someone walks a route once or twice a shift, reads gauges, writes numbers in a log, and someone else types them into a spreadsheet days later. The readings arrive late, carry transcription errors, and sit too far apart in time to trend. This guide is for the plant or maintenance engineer who wants to digitize that round with machine vision, one checkpoint at a time. You will sort each stop on the route into the question it asks, match it to a fixed camera, smart camera or code reader, size the optics with the arithmetic shown, and route timestamped results into the systems where decisions get made.

The decision, up front

Digitize a checkpoint when the reading is quantitative and the location is fixed; keep it on the walk when it needs a hand, an ear or a flashlight. That one rule splits almost any route:

• Area scan cameras take the gauge faces, level columns and sight glasses — every stop where you would type a number into the log.
• Smart cameras take the binary stops: RUN lamp lit, valve handle in the open position, coupling guard fitted, no part left on the fixture.
• Fixed code readers take identity: the Data Matrix plate on each station that binds every image to an equipment number.
• The operator keeps the sensory stops — bearing noise, hot smells, leaks behind insulation, anything behind a cover that must be opened. The round shrinks; it does not disappear.

If you are still arguing whether a given stop should be automated at all, read when manual visual inspection beats automation for the honest counter-case, then automating manual visual inspection: when one click works for the stops where one camera removes the whole task.

Four questions, three device classes

Every stop on a round asks one of four questions, and the question picks the hardware:

A value stop buys you a trend, so it earns pixels and a lens calculation. A yes/no stop buys you a fact, so it earns an on-device decision wired straight to the PLC — no PC at the skid, no image to store. An identity stop buys you traceability, and a reader answers it without interpreting anything. Sequence the rollout in that order of payback: numbers first, identity second, binary checks third, because a number captured twice a shift is the reading your maintenance system can actually trend.

Worked example: four gauges, one camera

A compressor skid carries four 63 mm dial gauges on a 300 × 200 mm panel behind a rail. The round wants four pressures every shift; the rail sets the working distance at 600 mm.

Pick the sensor for pixels per graduation first:

The LS-CMR-0166 is a 12 MP USB3 area scan camera: 4000 × 3000 pixels on a 1/1.7-inch IMX226 sensor, 30 fps, rolling shutter. A 1/1.7-inch optical format is about 7.4 × 5.5 mm, so 4000 pixels across the width are 1.85 µm apart. The rolling shutter costs nothing here, because the scene is static.

Pick the field of view. The 300 × 200 mm panel needs mounting margin: budget a 340 mm wide scene.

Magnification m = 7.4 mm ÷ 340 mm = 0.0218, so f = 600 mm × 0.0218 ÷ 1.0218 = 12.8 mm. The nearest standard C-mount focal length is 12 mm. Turning that back into coverage: m = 12 ÷ (600 − 12) = 0.0204, so the scene is 7.4 ÷ 0.0204 = 363 mm wide and 5.5 ÷ 0.0204 = 270 mm high. The panel and its margin fit.

Check the resolution. 363 mm ÷ 4000 pixels = 0.09 mm per pixel. The 1 mm graduations on a dial land 11 pixels apart, and the needle tip locates to about a tenth of a millimetre — roughly a fifth of a degree of needle angle on a 63 mm dial, an order of magnitude finer than reading between printed divisions. The gauge itself is now the accuracy limit, not the camera.

Set the aperture. Depth of field at f/4:

With c = 2 pixels = 3.7 µm: DOF ≈ 2 × 4 × 0.0037 × 1.0204 ÷ 0.0204² ≈ 73 mm. The four gauges sit within ±30 mm of the panel plane, so f/4 holds all of them. Mount at 500 mm instead and depth of field falls to about 50 mm; one stop to f/5.6 brings it back near 70 mm.

Set the light and exposure. One small LED area light off the camera axis, a 1 ms pulse inside a 5 ms exposure, gain at minimum. The pulse freezes needle flutter, out-delivers the high-bay lighting, and f/4 keeps the full panel sharp.

Prove it on the skid. Capture one frame and measure the pixel distance between two adjacent graduations: the arithmetic above says 11 pixels, and a healthy set-up lands within a pixel of it. Then run two weeks of shadow rounds — camera records, paper continues — and diff every pair before the paper stops. Converting needle position into a pressure value is its own calibration; reading an analog gauge needle with a camera walks through the angle-to-value math.

Smart cameras take the yes/no stops

A smart camera carries sensor, processor, light and I/O in one housing, so the decision happens at the skid and the PLC reads a discrete bit: guard fitted, valve vertical, RUN lamp lit and green. For a stop visited twice a shift that is the whole point — no PC, no image pipeline, one wire.

The range spans both ends of that job:

The LS-SMC-0063 is a 1.6 MP smart camera with a 16 mm lens and a polarized white light, capturing at 60 fps. The polarized source is what tames brushed stainless and glass, the two surfaces that fake a lit lamp. The MV-SC5060XM-08M-WBN is a 6 MP mono smart camera with an 8 mm M12 lens on GigE — the wider field that suits station-level checks, where presence, position and lamp state land in one frame.

Trigger these from the round clock, never free-run: one frame per stop. Triggering a smart camera over TCP with ASCII commands shows the command a scheduler or PLC sends and how the result comes back on the same socket. When the consuming side is LabVIEW, smart camera LabVIEW integration covers turning those strings into logged records.

Code readers close the traceability loop

A pressure image that cannot be tied to an equipment tag is a picture, not a record. Put a Data Matrix plate on each station and a fixed reader beside the camera on the same bracket; both fire together, and the asset ID plus the reading land in one record. DMC code readers for workpiece tracking works through the camera-versus-reader choice; a round needs only the reader half of it.

Two readers cover the standoff bands a route meets:

The LS-CRD-0047 is a 1.2 MP compact fixed reader with a 60–120 mm reading distance over USB 2.0 — the right band for a plate on the gauge-panel bracket itself. The MV-ID3013PM-12M-WBN is a 1.3 MP compact reader with a 12 mm focal length lens, which carries the read out across an aisle to the next station. Fire it on demand rather than continuously; triggering a fixed code reader from a GPIO pin instead of free-run keeps every read stamped to the round. For symbology, contrast and grading questions on existing tags, how to choose 2D code readers for automotive traceability has the detail.

Route the results where decisions happen

The payback of a digitized round is the trend, not the frame. Three practices hold up:

1. One record per stop. Asset ID, value or pass/fail, timestamp, image path — written over Modbus TCP, OPC UA or a plain TCP string to whatever already collects plant data.
2. Size the image store honestly. A 12 MP mono frame is 12 MB raw; compressed, budget 2 MB. Eight digitized stops at twelve rounds a day is 8 × 12 × 2 MB = 192 MB per day, about 70 GB a year — one disk, not a project. Keep 90 days of frames on a rolling window and archive the extracted numbers long term.
3. Surface it where the shift starts. A live round dashboard at the operator shelter turns the digitized round into a decision tool; industrial panel PCs are built for exactly that mounting.

If the plant already runs operator-rounds or asset-performance software, feed it rather than replace it: those systems take readings by design, and they carry the audit trail and compliance reporting that paper never had.

Where the camera loses to the operator

Leave these on the walk, and design the shortened route around them:

• Sound. Bearing whine and pump cavitation have no optical signature at round frequency.
• Heat and smell. Overheated windings and solvent leaks announce themselves to a person before any fixed camera sees the effect.
• The hidden. A drip forming behind insulation or a seal weeping under a skid is visible only from the angle a walking operator happens to have.
• Everything behind a guard. If seeing the checkpoint means opening an interlocked cover, the camera adds nothing over the operator who already opens it.

Checks before you order

• Measured working distance, taken on the skid with rails and doors in place — the focal length arithmetic above moves with it.
• Smallest feature size at every stop, in millimetres, with 4 pixels or more across it after the field-of-view division.
• Ambient light audit: high-bay duty cycles, windows, forklift headlights — then fix pulse and exposure to match, as in the worked example.
• One trigger authority per station — PLC output, scheduler or GPIO — and one clock. Camera timestamp and PLC timestamp together guarantee a mismatch.
• Network and power at each bracket: one GigE drop serves a camera and a reader through a small switch.
• Vibration: a gauge panel on a running compressor moves; judge image sharpness at the worst hour, not on the bench.
• Retention sized from frames per day: stops × rounds × 2 MB × retention days, plus room for the audit trail.
• A two-week shadow run against the paper log before anyone stops walking the value stops.

From route list to hardware

Per-case decisions, restated:

• A scale with a number on it → an area scan camera, a 12 mm-class C-mount lens at your measured standoff, a pulsed area light: the worked example above.
• A lamp, position or presence → a smart camera with its own light and a discrete output to the PLC.
• An asset tag → a fixed code reader fired on the round clock.
• Sound, heat, smell or anything behind a guard → the operator, on a shortened route.

The camera finder filters by interface, resolution and mount; the system configurator walks camera, lens, light and controller in one pass. More field write-ups sit on the applications topic page. Send the route list to our applications team and we will mark it up with device classes and standoff numbers.