Knowledge
Calibration and Measurement Stability on a Running Line
A vision measurement station is qualified on the day it is commissioned and then lives in a workshop. Keeping it inside its tolerance is a routine, not a product feature: a short verification against a reference every shift, a full recalibration whenever the optical configuration changes, and a control chart on one reference feature that turns slow drift into a visible trend. Most escapes on a running line are traceable to one of four causes — contamination, temperature, vibration and illumination ageing — and each leaves a different signature.
Part of the POMEAS Machine Vision Optics Technical Guides series — see the Machine Vision Optics Technical Guides index for all 20 articles in this set, grouped by topic.
What each disturbance does to the number
- Contamination lowers contrast and adds veiling glare, which shifts apparent edge position. On backlit silhouettes it mainly destroys edge steepness and repeatability.
- Temperature changes both the part and the machine: steel expands about 11.5 µm per metre per kelvin, aluminium about 23, so a 100 mm steel part moves roughly 1.2 µm per kelvin — 10 K of ambient swing is already about 12 µm.
- Vibration blurs the exposure by roughly v · t: 100 mm/s during a 200 µs exposure moves the image about 20 µm, comparable with the object-side pixel of a large-field station.
- LED ageing reduces output and forces longer exposures or higher gain; exposure value trending upward over weeks is the early warning.
Calibration, verification and the difference between them
Calibration establishes scale factors and any distortion model. Verification checks that they still hold, using a known artefact, without changing anything. Confusing the two hides drift: teams recalibrate when a value looks wrong instead of logging the check and watching it move. The workable routine is a one-button artefact check at a fixed interval with pass/fail limits set to a fraction of the tolerance, typically a tenth, logging the value automatically; and a full recalibration only when the configuration changes — camera, lens, aperture, focus, working distance, mounting — or when the check leaves its band.
Detecting drift before it produces bad parts
Measure one reference feature automatically at a fixed cadence and plot it against time. A step change points to a mechanical or configuration event, a steady ramp to thermal drift or wear, and an increase in scatter to contamination, illumination ageing or vibration. Because the reference is measured by the same chain as the product, it captures the whole system — which is exactly what an audit will ask about. Keep the configuration record, the calibration result with residuals, the verification values with timestamps and temperatures, the maintenance actions and the measured data per batch; six categories that make an accepted batch reconstructible months later.
Turning an artefact into a pixel-to-millimetre scale
Verification assumes a scale already exists. Calibration is what creates it: one measured image of a known artefact, the scale factors that follow from it, and — whenever the field is large or the lens is not telecentric — a distortion model that keeps the scale valid away from the centre. Four things belong in the procedure before any number is written down.
Choose the artefact for the field, not for the drawer
| Artefact | What it establishes | What to watch |
|---|---|---|
| Chrome-on-glass grid or checkerboard | Scale factor plus distortion across the whole field in one image | Must sit in the same plane and attitude as the part; viewing through the glass adds refraction if the plate is tilted |
| Ceramic or anodised dot / line plate | Scale factor where a glass surface would mirror the illumination | Softer edges than chrome; larger dimensional change with temperature than glass |
| Certified gauge block or pin gauge | An end-to-end length check of the whole chain, optics included | Bright edges and chamfers move the apparent edge; best used as a confirmation, not as the only artefact |
| Certified master part | Scale factor in the real fixture, lighting and orientation | Only as good as its own certificate and its own storage — and it wears like a part |
Five steps that must appear in the procedure
- Mount the artefact where the part sits. Same plane, same attitude, same fixture, same working distance, with production lighting on.
- Focus on the artefact and then lock it. Use the production aperture. Every later acquisition uses that same setting.
- Acquire a set of images, not one. Use the same edge tool and threshold that will measure production parts, so the calibration and the measurement share one definition of an edge.
- Fit the scale across the field. Measure the same features at several positions and both axes, and derive object-side pixel size from known length divided by measured pixel length. If the spread across the field is larger than the tolerance budget allows, fit a distortion model instead of a single factor.
- Record what the number depends on. Scale factors per axis, residuals, artefact identity and certificate, ambient temperature, illumination and tool settings, software version. Without these the next calibration cannot be compared with this one.
Judge the result on residuals, not on a single scale factor
| What the residuals show | What it means | What to do |
|---|---|---|
| Small and random, no trend across the field | The model fits the pixel data | Accept it and record it; this is the baseline every later check is compared against |
| Growing towards the edges of the field | Radial distortion is not being modelled | Fit a distortion model, or restrict measurement to the zone where the single factor holds |
| X and Y scales differ consistently | Anisotropy from the sensor or the optics | Keep separate per-axis scale factors rather than averaging them |
| Scale moves after the focus was touched | Magnification is focus-dependent | Lock focus mechanically and re-verify; treat focus as a configuration change, not an adjustment |
| Repeated acquisitions give different scales | The acquisition or the artefact placement is not repeatable | Fix the fixture and the edge tool before changing anything in the model |
Setting a re-verification interval you can defend
There is no universal interval, and a number quoted without context is worthless in an audit. The interval is the outcome of four inputs, and it should be re-derived whenever any of them changes.
| Input | Pushes the interval shorter when… | Pushes it longer when… |
|---|---|---|
| How often the configuration changes | Lens, aperture, focus, working distance, camera, mount or software are touched regularly | The station is mechanically sealed and only the part changes |
| How much of the tolerance the process already uses | Measured values sit close to the specification limits, so little uncertainty is left to spend | The process runs near the middle of a wide tolerance |
| Observed drift rate | Verification values move measurably between checks | Check values have been flat over many cycles |
| Environment and duty | Ambient temperature swings, vibration, dust, or a station that runs continuously | Stable, clean, intermittent operation |
A usable rule of thumb is to set the check limit at a fixed fraction of the tolerance — a tenth is a common working choice — and then size the interval so that the drift expected within it cannot consume that limit: if the reference feature moves by r per week, and the check limit is T/10, the interval is roughly (T/10) ÷ r. That reasoning produces an interval that can be explained to an auditor, unlike one inherited from a template.
Two practical constraints sit on top of the calculation. A short verification and a full recalibration are different jobs, so do not let the interval force a full recalibration that nobody will run: keep the frequent check cheap and automatic, and reserve the full procedure for a configuration change or a failed check. And tie the outcome to something visible — a pass/fail band, an automatic log, and a trend line — because a check that is not recorded cannot shorten or lengthen the interval that justified it.
When a calibration will not converge
Some failures look like a calibration problem and are not. The table separates the ones that belong to the model from the ones that belong to the machine.
| Symptom | Most likely cause | What to check first |
|---|---|---|
| Scale correct in the centre, wrong at the edge | Distortion model missing, or fitted to too few points | How many features were sampled, and how widely they were spread across the field |
| Scale drifts with no optical change | Temperature, or the artefact was not in the part plane | Temperature logged alongside the check value; artefact height against the fixture datum |
| Every calibration gives a slightly different answer | Focus not locked, or the artefact moves between images | Focus lock and fixture repeatability, before touching the model |
| An artefact that passed last month now fails | The artefact, not the station | Clean it, inspect it, and check its certificate before recalibrating the machine |
| Scale is correct but measurements scatter | Repeatability, not calibration | Illumination stability, exposure, vibration and edge-tool settings |
The scale-factor arithmetic itself is worked through in the telecentric FOV and magnification formulas, and the wider selection arithmetic is collected in industrial lens selection calculations and errors. If the residual keeps growing towards the edge of the field on a non-telecentric lens, the distortion question is covered in object-side resolution versus measurement accuracy.
Related reading: building a telecentric measurement station, object-side resolution versus accuracy, and the telecentric lens range. The practical questions on recalibration intervals, records and operator checks are answered below — describe your line and we will suggest a verification interval.
Further reading
Subscribe to Our Newsletter
Simply enter your email to receive the latest news and insights from Pomeas. Stay connected with Pomeas and be the first to discover new innovations in optical excellence.
