Machine vision maintenance: a daily-to-annual checklist that catches drift before your customer does.

Machine vision maintenance is a short, scheduled routine, and it isn't complicated. Run golden good and bad samples every shift and log the scores. Clean the protective window with the right materials. Check light output, mounts, focus locks and cables monthly. Back up every job, put changes under control, and recalibrate and re-run your test set once a year.

None of that's hard. The trouble is that a vision station fails soft, so nobody notices until the reject bin fills up or a customer finds what the camera missed. This checklist is the routine we hand over with the stations we build, written so a maintenance tech can run it without calling anyone. If your station is already rejecting good parts, start with our guide to why your vision system rejects good parts, then come back here to keep it fixed.

What does machine vision maintenance involve?

It covers four things. Keeping the image the same (optics, lighting, mounting), keeping the decision the same (jobs, thresholds, models), keeping the data safe (backups, image archives) and proving all of that with logged checks. Most of it takes minutes a week, and it's cheap. The budget side is bigger, since spares, recipe changes and retraining add up, and our post on machine vision system cost suggests planning 5 to 15 percent of installed cost a year.

A vision station is different from most machines on your floor. A worn bearing gets noisy. A vision station with a dirty window just gets slightly darker, and the inspection keeps producing confident answers, slightly worse each day. So vision system preventive maintenance is mostly about measuring the image, not inspecting the hardware. If the image is the same as it was at acceptance, the station is fine. If it isn't, you want to know before the threshold does.

What should be on a vision system maintenance checklist?

A good vision system maintenance checklist isn't long. It has tasks by interval, a pass criterion for each one and a named owner. Daily covers golden samples and a look at the live image. Weekly covers cleaning and light checks. Monthly covers mounts, cables, backups and archives. Annually covers recalibration, firmware review and a full test-set re-run.

Here's the version we start from. Adjust intervals to your environment. A sealed station in clean assembly can stretch the cleaning interval. A station near a grinder or in coolant mist may need cleaning every shift.

Fig. 01Machine vision maintenance by interval, on one station

How do you run a golden-sample check on a vision system?

Keep a signed good part and a signed bad part at the station. At the start of every shift, run both through the normal inspection and write down the scores. Don't just note pass or fail. Good must pass with margin, bad must fail. A score drifting toward the threshold is your early warning.

The scores are the point. Pass and fail tell you the station worked today; the score tells you whether it'll still work next week. A golden good part that scored 92 at acceptance and scores 84 now is passing, and it's also telling you that something in the image has changed. That's exactly the kind of thing a shift log catches and a pass/fail light doesn't.

A few practical details matter. Store the golden parts in a labeled case so they don't get scratched or swapped with production parts. Keep a second set locked away, because the station set gets handled every day and wears. And if your station inspects several part numbers, rotate which one you run, or keep a golden pair for the highest-risk one. For deep learning stations, log the confidence score the model reports, since that's the number that drifts first. Our deep learning vs rule-based vision guide explains why the two kinds of job drift differently.

How do you clean a machine vision lens without damaging it?

Blow loose grit off first, then wipe once with lens tissue wetted with a suitable solvent, working from the center outward, with a fresh tissue for each pass. Don't wipe dry, don't use a shop rag, and never put acetone on a plastic window. Then check the live image and the golden-sample score.

Edmund Optics' cleaning guidance covers the basics. Remove dust with compressed air or a blower before any wiping, because dragging grit across a coating scratches it. Use lens tissue with reagent-grade isopropyl alcohol or lens cleaner, and work from the center out. Clean fingerprints off coated optics quickly, because skin oils can stain the coating permanently. And don't use acetone on plastic optics.3 That last one matters more than it sounds, since a lot of camera enclosure windows are polycarbonate or acrylic. Check yours before someone grabs the wrong bottle.

Two floor habits cause most of the trouble we see. The first is the shop rag on second shift, which leaves an oily film that softens edges and scatters light. The second is cleaning the lens itself when there's a protective window in front of it. Every touch risks turning the focus or aperture ring, a far bigger problem than a little dust, so the procedure should say to clean the window and leave the lens alone. Skip unfiltered shop air too, since it can carry oil and water.

Fig. 02Cleaning a machine vision lens window, step by step

How fast do machine vision LED lights lose brightness?

Slowly, but faster than your inspection can tolerate. Lighting makers typically rate life as hours until output falls to 70% of new, often tens of thousands of hours. Most inspections need the image to stay within a few percent of acceptance, so plan around your threshold's tolerance, not the datasheet's L70 number.

LUMIMAX, for example, rates its visible and infrared lights for at least 80,000 hours at full load before intensity falls more than 30%, and its ultraviolet lights for at least 55,000 hours.1 That 70% point is the industry's habit for LED life, and it comes from the lighting world, where the IES LM-80 test measures lumen depreciation for at least 6,000 hours and the TM-21 method projects it further out. Researchers at Pacific Northwest National Laboratory have pointed out that LED life depends heavily on operating temperature, which is why that testing happens at several temperatures.2

Now do the arithmetic for a vision station. Run 24/7, 80,000 hours is about 9.1 years (80,000 ÷ 8,760). That sounds great. But if your inspection holds only until the image is about 10% darker, an exponential decay fitted to the 70% point gets there at roughly 23,600 hours, or about 2.7 years. Run the light hotter and it gets there sooner. Fig. 03 shows both curves.

Fig. 03LED output vs operating hours · illustrative model

You can slow the aging down. LUMIMAX recommends switching the light on only while the image is taken, because the lower temperature in switched operation reduces brightness loss, and it notes that lower brightness settings age the LEDs less.1 Strobing from the camera trigger, good airflow around the light's fins and a mount on a thermally conductive bracket all help.

Watch for one trap. Many cameras and jobs use auto-exposure or auto-gain, which quietly compensate as the light dims. The image looks fine, but it isn't. Noise creeps up and contrast drops, and the station loses margin without anyone seeing it. Lock exposure and gain after acceptance, or at least log them, so dimming shows up as a number rather than hiding in the settings.

Why do mounts, focus locks and cables need regular checks?

Because a camera that moves a millimeter or a lens that turns a few degrees changes the image as much as a dirty window. Vibration loosens brackets, changeovers bump fixtures, and cables on moving axes fatigue. A monthly check with witness marks and a reference overlay catches most of it, and it's quick.

Paint a witness mark across every adjustable joint at acceptance, on the bracket clamps, the lens focus and aperture rings and the light mounts. A broken mark tells you something moved without needing to measure anything. Vision Doctor's maintenance notes make the same point about vibration loosening camera and light mounts, and suggest thread-locker where mounts keep backing off.4 Then compare the live image against a saved reference image with an overlay. If the part edges don't line up, the camera moved, even if the inspection still passes. On dimensional gauging and vision-guided robotics stations, a moved camera also means a stale calibration, so recheck it against the target before you trust another measurement.

Cables are the other quiet failure. The usual story is a connector that isn't fully locked, or a cable flexing in a drag chain past its bend rating, and the symptom is intermittent image drops that are miserable to chase on a running line. Check that M12 and GigE connectors are locked, that strain relief is in place, and that nothing is rubbing.

How should you back up vision jobs and control changes?

Back up every job, recipe, calibration file, model and camera setting monthly and after every change, to a controlled location with a version number. Only named people should edit jobs, every edit gets a reason and a record (that's the part that slips), and firmware updates get tested on a spare before they touch production.

Backups are where we've seen the most avoidable downtime. A camera dies, the spare comes off the shelf, and nobody can find the job file, or they find three versions and don't know which one was running. Vision Doctor's maintenance notes recommend backing up inspection programs, calibration data and camera settings, and warn that important configuration files often live in more places than you'd expect.4 Our rule is simple. If you can't rebuild the station from the backup and a spare camera in an afternoon, the backup isn't done. Test the restore at least once a year.

Change control is the other half. In regulated plants it's not optional. FDA's drug GMP rules in 21 CFR 211.68 require automatic and electronic equipment to be routinely checked under a written program, require controls so that only authorized personnel make changes, and require backups that are exact and complete.5 Device makers now work under FDA's Quality Management System Regulation, which took effect on February 2, 2026 and incorporates ISO 13485:2016 by reference.6 And for computerized systems in GxP settings, ISPE's GAMP 5 guide sets out the risk-based approach most validation teams follow.7

Even if none of that applies to you, borrow the habit. The threshold someone loosened at 2 a.m. to keep the line running is the most common unrecorded change we find, and it's usually still loosened months later. A one-line entry in a change log (what, why, who, the before and after values) fixes that. When a change does touch the inspection logic, our post on vision system validation covers when it needs revalidation.

How do you spot vision system drift early?

Log the same few numbers every shift, the golden scores, the mean gray level, exposure and gain, and the reject rate, and plot them on a simple control chart. A run of points on one side of the average flags drift days before any point crosses a limit, and long before a good part fails. It's the cheapest early warning you've got.

Here's what we'd put in the log. It fits on a clipboard or a three-column spreadsheet, and the station's HMI can often record most of it automatically.

Then chart the golden good score. You don't need special software. NIST's engineering statistics handbook describes the classic run rules for Shewhart control charts, including one point beyond three standard deviations and eight consecutive points on one side of the center line.8 The eight-in-a-row rule is the useful one for vision, because drift is slow. The handbook also notes the cost. Adding run rules raises the false-alarm rate, so treat a run as a reason to look, not to stop the line.

Fig. 04Golden-sample control chart catching drift · illustrative

When the chart flags, work through the cheap checks first. Clean the window and rerun the golden part, compare the live image to the reference overlay, then check exposure, gain and the light's gray level. Image archive housekeeping belongs here too. A full disk can stop a PC-based station from saving images or even running, so set a retention rule that matches your traceability needs and an alarm well before the drive fills.

Which spare parts should you keep for a vision system?

Keep a spare of anything that would stop the line and has a long lead time. For most stations that's the light, the camera cable and the power supply. Critical lines should add a spare camera with the job preloaded and a lens preset and locked to the same focus and aperture.

The preset lens is the one people skip, and it's the one that saves the most time. Swapping a camera takes minutes. Getting a new lens to match the old focus and aperture by eye can take an hour and still leave the image slightly different. PC-based stations add a drive image to the list, one of the trade-offs we cover in smart camera vs PC-based vision. Here's how we'd stock a typical single-view station.

Who owns vision system preventive maintenance?

Operators own the shift checks, maintenance techs own cleaning and the mechanical checks, and one named vision owner, usually a controls or manufacturing engineer, owns backups, changes and the log. Quality owns the golden samples and decides when drift has become a quality problem, since that's their call.

The vision owner matters most. On a lot of lines we visit, the vision job belongs to whoever installed it, and that person has moved on, so nobody wants to touch it and nobody's watching the log. Name an owner and a backup, put the tasks in your CMMS like any other preventive maintenance, and review the log monthly. That review is ten minutes of looking at a chart, and it's where drift gets caught. If you're buying support rather than doing it in-house, our machine vision integrators guide lists what to ask about service.

If you're building a new station, put this checklist into the project from the start. It's much easier to design in witness marks, a reference target in the field of view and a locked exposure than to add them later. Our guide to starting a machine vision project covers handover, the machine vision lighting guide covers choosing lights that age gracefully, and the machine vision library covers individual applications like surface defect inspection, where drift bites hardest. We build vision into robotic cells with these routines included, as described on our machine vision capability page.

Frequently asked questions

How often should you clean a machine vision camera lens?

Clean when the shift check shows film, droplets or a falling golden score, and on a fixed schedule as a backstop. Weekly suits most indoor assembly stations. Stations near grinding, coolant mist, washdown or welding may need it every shift. Clean the protective window rather than the lens wherever there is one, and log each cleaning so you can see its effect on the score.

Can you use glass cleaner or a shop rag on a vision lens?

No. Household glass cleaners can leave residue and may attack coatings or plastic windows, and shop rags carry oil and grit that film or scratch the surface. Use lens tissue with reagent-grade isopropyl alcohol or a lens cleaning solution, after blowing off loose dust. Keep a labeled cleaning kit at the station so nobody improvises on second shift.

How often does a vision system need recalibration?

Recalibrate on a schedule, often yearly, and after any event that could move the camera, the lens or the part position, such as a crash, a bracket change, a lens swap or a fixture rebuild. Gauging stations and robot guidance stations need it most, because their output is a measurement or a position rather than a pass or fail. Verify with a calibration target and keep the results.

Should you install firmware updates on vision cameras?

Not automatically. Read the release notes and install only updates that fix a problem you have or a security issue you need closed. Test the update on a spare camera with the locked test set, compare results to the running version, back up the old firmware and job, and record the change before updating production.

How long should you keep vision inspection images?

Keep images as long as your traceability, customer and regulatory requirements say, and no longer than your storage can handle. Many plants keep every failed image and a sample of passes for a set period, such as 90 days, and keep the images tied to any quality escape permanently. Write the retention rule down and automate the purge.

Does a deep learning vision system need different maintenance?

It needs everything a rule-based station needs plus model care. Log the model's confidence scores, keep the training images and the exact model version under backup and change control, and collect new images when parts, suppliers or lighting change. Retraining is a change like any other, so re-run the locked test set before the new model goes live.