Vision system out of calibration: are your previous inspections valid?

A quarterly check on a dimensional vision gauge has just failed, and the number in your head is not the error — it is the count of parts the system passed while it drifted. This explainer is for the quality or manufacturing engineer who owns that call. It walks the containment sequence in the order you need it: bound the window of exposure, divide the calibration error by the tolerance, clear most records with one inequality, re-measure only the parts that can truly be out, and reset the interval from the measured drift rate. A 12 mm gauge with 30 µm of drift carries through every step with the arithmetic shown.

What does out of calibration mean for a vision gauge?

Your gauge is out of calibration the moment a check against a traceable reference shows an error beyond the limit you set for it — for dimensional work, past roughly 10% of the tolerance band.

Run the check with a certified artifact — a gauge block, a certified pin or a golden part — measured at points spanning the range the gauge actually works over, not one value in the middle. The result is a signed error, E = reading − reference, and the sign matters: a gauge reading 30 µm low pushes accepted parts toward a different limit than one reading 30 µm high.

For a camera, calibration covers the pixel scale, the lens distortion model and the plane the part is measured in. Where presentation varies, the pose is part of the calibration too, and calibration reference pose changes that shift measured size treats that failure separately from true scale drift; camera calibration depth range and accuracy covers when a changed geometry forces a recalibration. ISO 9001 then makes the consequence binding: when measuring equipment is found unfit, you determine whether the validity of previous results is impaired, and you act on the answer. The rest of this article is that action.

Which previous inspections become suspect?

Every result since the last verification that passed — worst case the full calibration interval, which in the worked example below is 90 days and 45,000 parts.

The window closes at the last passing check, not at the date you believe the drift began. Without interim checks you assume the full interval, because nothing in the data proves a shorter one. A daily golden-part check shrinks the window to a single day of production — 500 parts here instead of the whole quarter.

ISO/IEC 17025:2017 writes the laboratory version in section 7.10, nonconforming work: evaluate the significance, recall the work where risk exists, notify the customer. An internal plant runs the same sequence through its own containment and customer-notification procedure.

Freeze the results database the hour you find the failure. Quarantine the records, keep the as-found readings, and stop anyone re-inspecting over the top of them — the stored values are the evidence the bound in the re-measurement section stands on.

How do I size the error against the tolerance?

Divide the absolute error by the total tolerance band and apply one threshold: 10% or less closes with paperwork, more than 10% triggers containment and re-measurement.

The 10% figure matches what gauge studies accept as total measurement error against a tolerance in the first place, so a calibration error that consumes the same fraction leaves the station no worse than the acceptance your GRR already justified. A larger error is unknown bias stacked on top of everything else, and unknown bias is what recalls are made of.

Guard banding runs alongside it: subtract the measurement uncertainty from each limit before accepting parts, so shipped product keeps margin for the drift you have not found yet. A calibration error earns its own threshold rather than being absorbed, because bias does not average out.

What does the arithmetic look like on a real gauge?

A gauge reading 11.970 mm on a certified 12.000 mm reference is 30 µm low, 30% of a 0.100 mm band — containment starts, and 1,590 records end up re-measured.

The part is a turned spacer, width 12.00 ± 0.05 mm, gauged on an inline station with the same architecture as an inline dimensional measurement line for brass parts. The LS-TCL-2194 is an object-side telecentric 0.5x lens with a 110 mm working distance and a built-in coaxial light path.

The camera is a MV-CH120-11UC, a 12 MP global-shutter USB3 camera with a 1.1-inch sensor, and C-mount keeps it inside the largest group of area scan cameras. Pixels on this sensor class run about 5.32 µm; divided by the 0.5x magnification, object-side sampling is 5.32 µm ÷ 0.5 = 10.64 µm per pixel. The feature spans about 1,128 pixels and the full 100 µm band spans 9.4 pixels, so tenth-pixel edge interpolation repeats near 1 µm — the optics hold this tolerance, which points the investigation at the set-up rather than the camera.

The quarterly check puts a certified 12.000 mm gauge block in the fixture and reads 11.970 mm: E = |11.970 − 12.000| = 30 µm, a 0.25% scale error, low. The last study gave a measurement uncertainty of 8 µm, so guard-banded acceptance ran between 11.958 and 12.042 mm, and the capability study gave Cpk 1.5 centred on nominal — one standard deviation is 0.050 ÷ (3 × 1.5) = 11.1 µm.

Reading low by 30 µm means true size = reported + 30 µm, so a part escapes only when its reported value exceeds 12.020 mm. That sits 20 ÷ 11.1 = 1.80 sigma above centre, or 3.6% of production — about 0.036 × 45,000 ≈ 1,620 parts expected. The database query returns 1,590 records above 12.020 mm, close enough to the estimate to trust both.

Which parts do I actually have to re-measure?

Only the records the drifting gauge reported within E of a limit — here every part reported above 12.020 mm, 1,590 records from the query.

The inequality does the containment for you: every part reported at or below 12.020 mm is arithmetically inside 12.050 mm whatever the drift did, so 43,410 records clear without being touched. The bound holds because the as-found check delivered both the size and the direction of the error, which is why the window section insisted on freezing the database before anything else.

Re-measure on a traceable reference at least four times better than the tolerance — 25 µm or finer on this band — which a bench micrometer at about 3 µm clears by an order of magnitude. Sort the set by reported value, highest first, because the worst cases settle the disposition fastest. Parts that re-measure inside the limits ship; the rest go to quarantine, and anything already delivered gets a customer notification carrying the re-measured values.

If the station stores only pass or fail, the bound is unavailable and you fall back on stratified sampling: 8 to 12 parts per lot across the whole window, re-measured on the reference, and one out-of-limit part screens its entire lot. Storing the measured value costs one field in the results table and buys this whole section back.

When can I close it out with paperwork only?

When the error sits at or under 10% of the band and the process ran centred — an error of 6 µm here leaves expected escapes near 38 ppm, under a 100 ppm containment trigger.

Run the same arithmetic with a 6 µm error, a 0.05% scale fault: escape needs a reported value above 12.044 mm, which is 44 ÷ 11.1 = 3.96 sigma above centre, about 38 ppm — under two parts across the window. Below the plant trigger, the disposition is no recall with the interval halved, and it stands in an audit because the arithmetic is in the record. Parts later verified on a second, conforming gauge stand on that gauge's own evidence.

When does the 10% threshold not apply?

Safety-critical and regulated parts get recall at any error you cannot bound outright, whatever the percentage says. Destructive tests cannot be re-measured, so the response is record review plus sample evidence. One-sided or asymmetric tolerances compare E against the smaller side, not the band. Attribute checks — presence, count, colour — survive small scale errors but not edges leaving their regions of interest.

The file itself needs seven entries: the as-found error with direction and date; the suspect window and part count; the E ÷ T arithmetic; the re-measurement instrument, its uncertainty and the results; the disposition of every affected lot; the root cause with corrective action; the revised interval. Customer notifications, where product shipped, close it out.

How short should the new calibration interval be?

Short enough that expected drift stays under 10% of the tolerance — this gauge moved 30 µm in 90 days, about 0.33 µm per day, so 10 µm of budget buys 30 days.

Derive it rather than guessing: interval ≤ (0.10 × T) ÷ drift rate = 10 µm ÷ 0.33 µm per day ≈ 30 days. Halving the old interval is the field rule when no rate data exists; three consecutive passing checks at the shorter interval justify stepping back up, re-derived from the rate you have measured by then.

The daily golden part makes the interval survivable: one certified part, measured at shift start, plotted with an alert at 5 µm — half the action threshold — so the suspect clock restarts every 24 hours instead of every quarter.

What makes a vision gauge drift in the first place?

In order of frequency: unlocked focus or iris rings, illumination decay against a fixed edge threshold, mechanical movement of camera or fixture, and thermal growth.

Any turn of a focus or iris ring changes magnification, so lock the rings and recalibrate after every touch — microscope lens calibration: lock magnification, then measure applies to any fixed-magnification gauge. Distance is next: an entocentric lens changes scale with working distance, while telecentric lenses hold magnification through theirs, which is the split entocentric vs telecentric lens selection exists to explain.

Light decay is the quiet one. LED output falls gradually with age and temperature, and a fixed edge threshold translates dimmer light into a one-directional size error. Auto gain and auto exposure chase that decay and make it worse, so camera image brightness that drifts over time starts with turning every auto mode off. A constant-current controller removes the cause: the LS-CTL-0180 drives 4 channels at up to 120 W for 24 V lights, and the light controllers range covers strobed and networked variants. Edge quality closes the loop — collimated backlight versus diffuse for edge gauging keeps the subpixel edge where the calibration found it.

Thermal growth is real but usually the alibi: steel expands about 12 µm per metre per °C, so a 12 mm part measured 10 °C above its calibration temperature is only 1.4 µm larger — too small for this failure. A long camera bracket can still move more than the part does, so check mount torque before blaming physics.

What should the first hour after the failure look like?

Freeze the evidence and size the error before anyone touches the hardware; the checklist below runs the whole sequence in the order that protects both the parts and the record.

1. Freeze the results database and record the as-found error, its direction and the date.
2. Set the suspect window at the last passing verification and count the parts inside it.
3. Compute E ÷ T: at or under 10%, take the paperwork path; above 10%, start containment.
4. Query every record within E of a limit — that set is the whole re-measurement workload.
5. Book a traceable reference at 4:1 or better and re-measure the set, highest reported values first.
6. Lock the mechanics, restore the light output, recalibrate, then prove the station against the reference.
7. Reset the interval from the measured drift rate and start the daily golden part the same shift.
8. Write the seven-entry file and notify customers wherever escaped product shipped.

If the station is due for a rebuild, the system configurator walks camera, lens, light and controller in the order they constrain each other, and the contact page puts the containment arithmetic for your own tolerance and as-found error in front of an applications engineer.