Updated July 2026 · 8 min read · Adente Vision Engineering Team
What actually changes at a product changeover?
Four things move when a station switches from one part to another: the geometry the unit expects, the lighting that makes the part readable, the decision threshold that sets sensitivity, and the trigger that fires the capture. A changeover that treats these as a saved recipe is a switch; one that re-derives them by hand every time is a re-commissioning, and that is the difference between minutes and hours.
Geometry changes because a different part sits at a different position and size in the frame. Lighting changes because a matte part and a glossy one need different angle and colour to show their features. The threshold changes because the acceptable variation is not the same across parts. The trigger can change because a longer or shorter part reaches the sensor at a different moment. Treat all four as one named configuration and a changeover becomes select-and-verify.
Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators. It runs four inspection modes, Anomaly, Defect, Counting and Quality, so a changeover may also mean recalling the mode the next part needs, not only its geometry and lighting.
How do saved configurations make a changeover repeatable?
A saved configuration is the recipe that turns a changeover into a repeatable step instead of an expert task. Rather than re-aiming, re-lighting and re-tuning for each part, the station stores the full set of settings for a part under a name and recalls it on demand, so the operator switches parts without owning the commissioning knowledge behind each one. This is the same idea as recipe management in batch control, applied to an inspection station.
The payoff is on lines that change often. A cell that runs three or four parts across a shift, or swaps products by customer order, cannot afford a fresh commissioning at each change, and across the industry the downtime of an unstructured changeover is a recognised drag on line availability. A named configuration per part collapses that to a recall and a short validation, and because the settings are stored on the unit, the next shift inherits the same recipe rather than rebuilding it.
Because lighting on the unit is configurable in colour and angle, the light is part of the recipe, not a separate manual step. That matters because lighting is the setting most likely to be got wrong under time pressure, and folding it into the saved configuration removes it as a source of first-part rejects after a switch.
Changeover steps: what changes, the action, the check
Every changeover follows the same short loop. This table names what moves, what the operator or integrator does, and how to confirm the switch before the line runs at rate.
| What changes | Action at changeover | First-parts check |
|---|---|---|
| Geometry and part position | Recall the saved configuration for the incoming part | Confirm the part is framed and in focus on the preview |
| Lighting (angle and colour) | Apply the light stored in the recipe | Check the reject image looks like the good reference, no new glare |
| Decision threshold | Load the part’s operating point, do not re-guess it | Run known-good and known-bad samples, watch escapes and false rejects |
| Trigger timing | Set the trigger source and offset for the part length | Confirm capture fires with the part centred in frame |
| A genuinely new variant | Capture a fresh set of about 20 good images and retrain | Validate on held-back good and defect parts before full rate |
How do you handle a seasonal lighting or material shift?
Seasonal drift is a slow change in what a good part looks like, and it is handled by re-baselining, not by loosening the check. Ambient light shifts with the seasons, a skylight or a window changes the cell over the year, and a supplier's material can vary by lot or by season in colour, sheen or texture. Each of these moves the definition of normal for a good part, so the reference has to move with it.
The fix is to recapture good parts under the new condition and re-baseline the configuration, keeping the same escape rate. Because the unit trains on good parts only, a seasonal re-baseline is a short capture step rather than a defect-collection project, and because updates load by USB stick the change can be applied on an air-gapped line. Do not reach for the threshold to absorb a seasonal shift, because loosening the check to accept a differently-lit part quietly raises escapes; re-baseline the normal instead.
Build the expected spread into the good set from the start where you can. Capturing good parts that already include the natural variation of material and light, rather than a single pristine example, gives a configuration that tolerates ordinary drift and reduces how often a season forces a re-baseline.
How do you validate the first parts after a switch?
Never run at rate on trust; validate the first parts after every changeover. Recall the configuration, then run a handful of known-good and known-bad parts and watch two numbers move in opposite directions: good parts should pass, and the planted defects should be caught. If good parts reject, the switch has a lighting or presentation problem; if defects pass, the threshold moved too far.
This first-parts check is cheap insurance against a whole batch of scrap or a whole batch of escapes. It is also the moment to read the on-device preview and the dashboard reject images, so a bad switch is caught at the station in minutes rather than found later in a customer complaint. A confident variant recognition is a good sign here: in an automotive bumper cell, variant recognition ran above 0.9 confidence per part, and a clean switch should reproduce that kind of confident, correct call on the first parts, not a string of borderline results.
When does a variant need a fresh 20-image set?
A recall handles a known part; a genuinely new variant needs a fresh reference set. If the incoming part has never been on the station, no saved configuration exists to recall, so you capture about 20 good images of it and train, with training completing under 48 hours. That is the line between a changeover, which is a recall, and an onboarding, which is a short training task.
Judge it by whether the part is new to the model, not by whether it looks similar to an existing one. A near-mirror variant, like a left versus a right part, is still a distinct class the model must learn from its own good images. For the full method behind capturing that set and retraining on the line, see the sibling operator-handover post and the pillar guide on AI visual inspection for how few-shot training works; to see where changeover sits among the tasks the unit runs, browse the real applications.