A vision AI model is at its best on the day it goes live. After that the plant keeps changing: a new paint colour, a new supplier, a light that has dimmed, a lens with a film of dust. The model stays the same, so it slowly gets worse, and nobody notices until rejects climb or a defect escapes. This guide sets out a practical retraining cadence for automotive plants: what to check and how often, which drift triggers should start a retrain, how to collect and label the right samples, and how to prove a new model is better before it replaces the old one. To map this onto your own stations, book a cadence review.
Automotive Vision AI Model Retraining Cadence for Plants
Check on a calendar. Retrain on a trigger. Release through a gate. Three habits that keep an inspection model healthy for years, not months.
- Seven drift triggers with starting limits
- A six-step sample collection routine
- An A/B scorecard for releasing a new model
Brightness, focus, reject rate.
Scores, overrides, trends.
Humans re-check a random set.
Full revalidation and sign-off.
Collect, train, A/B, release.
The cadence in one line. Four checks run on a calendar. Retraining itself runs only when a check says it should, or when a planned change is coming.
Why a Good Vision Model Gets Worse
The model does not change. The plant does.
A model learns what good and bad parts looked like in its training images. Every later change in what the camera sees moves the line away from that picture. Engineers call this drift. It is normal, it is gradual, and on a busy line it is invisible until it has already cost something. Our support team can review which of these apply to your stations.
Light and lens
LEDs dim with age. Dust, oil mist and weld fume leave a film on the glass.
New colour or variant
A new paint, trim or model year that the training images never contained.
Supplier or material
The same part from a second source, with a slightly different surface.
Tool and process wear
Worn weld tips and dies change how a good part looks, not only a bad one.
Camera or fixture moved
A replaced camera, a knocked bracket, a different angle after maintenance.
New kinds of defect
A fault the process never produced before, so the model has never seen one.
The loud failure: false rejects
Good parts are rejected. Operators complain, re-inspection queues grow, and people start overruling the system.
Annoying, but visible. You will hear about it within a shift.
The quiet failure: escapes
Bad parts are passed. Nothing happens at the station. The defect turns up at end of line, at the customer, or in the field.
Silent, and far more costly. Only an audit sample or a later finding reveals it.
Drift Triggers: When to Start a Retrain
A calendar tells you when to look. A trigger tells you when to act.
Retraining every month wastes effort when nothing has changed, and is far too slow when something has. So set clear triggers instead. Four of the seven below need no labels at all, which means they can be watched on every shift. To set limits for your own lines, book a trigger workshop.
Starting points, not standards. Tune each limit to the station and to how serious a miss would be.
The rise began in week 7. A weekly review with a clear limit catches it by week 10. Without one, it is found when someone complains.
Check the camera before the model
A large share of what looks like drift is a dirty lens, a failed light or a bracket that has moved.
Clean, check and re-measure first. If the signal returns to normal, no retrain is needed.
Retraining on bad images is worse than doing nothing. It teaches the model that a dirty lens is normal.
Sample Collection: What Goes Into the Next Training Set
A retrain is only as good as the images you feed it, and the labels on them.
More images is not the goal. The right images, correctly labelled, is. A few hundred well-chosen examples of what the model gets wrong will usually help more than fifty thousand it already handles. Ask our vision specialists how collection would work at your stations.
Four places to collect from
Unsure images
Parts where the model's score sat close to the pass or fail line.
Disagreements
Parts an operator overruled, or re-inspection judged differently.
Random audit
A small random sample, to catch what the model is confidently wrong about.
Known cases
Confirmed escapes, and samples of every planned new colour or variant.
Six steps from image to training set
Capture with context
Keep the image, the score, the station, the variant and the shift together.
Remove near-duplicates
Fifty pictures of the same scratch teach the model very little.
Label twice
Two trained people label each image without seeing the other's answer.
Settle disagreements
A senior inspector decides, against the written defect standard.
Split by part, not by image
The same part must never appear in both training and test images.
Protect the golden set
A locked set of reference images, never used for training, only for testing.
About 2% of the week's images. The two labellers disagreed on 63 of them, or 7%.
Why two labellers?
Researchers found at least 3.3% wrong labels, on average, in ten of the most widely used AI test sets. Those were built with care.
A model trained on wrong labels learns the mistakes. A model tested on wrong labels gets a score nobody can trust.
If your labellers often disagree, the defect standard is unclear. Fix the standard before the model.
One Retrain, Start to Finish
A retrain is a short project with a clear end, not an open-ended task. This is a realistic plan for one station.
A/B Evaluation: Proving the New Model Is Better
A new model is only a challenger until it has beaten the champion on the same parts.
The model in production is the champion. The retrained one is the challenger. Never swap them on trust. Run both on identical parts, compare the results, and release only if the challenger wins without getting worse anywhere that matters. To see a scorecard built from your own data, book an evaluation session.
Golden set
Both models score the locked reference images. The challenger must not be worse on any defect type.
Shadow run
Both models see every live part. Only the champion's verdict counts.
Review the differences
People judge only the parts where the two models disagree.
Gate and release
Quality signs off. The old model stays ready as a fallback.
An illustrative scorecard. Every gate is met, so this challenger would be released.
Why review only the disagreements?
In a five-day shadow run the two models agree on almost every part. Those parts tell you nothing about which is better.
The few hundred where they differ tell you everything. In the example, 212 parts were reviewed and the challenger was right on 151.
If the split is close to even, the difference may be chance. Keep the shadow run going.
How many samples is enough?
- The rule of three. If a model misses none of 300 real defects, you can be about 95% sure its miss rate is under 1%.
- For 0.1%, you would need about 3,000 defects with no miss.
- So be honest about what a small test can prove, most of all on safety items.
Divide three by the number of defect samples tested to get the limit.
Who Owns It: Roles, Records and Change Control
A retrain changes a measuring system. Treat it like one.
In an automotive plant a vision station is an inspection gauge, and gauges are controlled. A new model version deserves the same care as a recalibrated gauge: a reason, a test, an approval and a record. Without a named owner, retraining either never happens or happens without anyone checking. Our quality team can share a role and record layout.
Who does what
- Quality engineer. Owns the triggers, the golden set and the release decision.
- Process engineer. Owns camera, lens and lighting health.
- Vision engineer. Trains the challenger and runs the comparison.
- Senior inspectors. Label samples and settle disagreements.
What to keep for every model version
- Why it was retrained, and which trigger fired
- Which images it was trained on
- Golden set and shadow run results
- Who approved it, and when
- Which version to go back to
Retraining on a fixed date whether or not anything changed. Letting the model label its own training images. Testing on images it was trained on. Releasing a new version at the end of a Friday shift. Releasing with no way back. Each one looks like a saving, and each one costs more later.
One set of data, three views
- Line lead. Is my station healthy this shift, and what needs cleaning?
- Quality engineer. Which triggers are close, which retrains are open, what is waiting for sign-off?
- Plant quality head. Model health across every station, and the history behind each version.
iFactory watches the signals and prepares the evidence. People make the release decision.
Turnkey AI: Delivered, Connected and Live in 6–12 Weeks
You do not build this. It arrives ready.
iFactory ships as a pre-configured NVIDIA AI server, racked and ready, with the software pre-loaded. Rack it, plug in power and Ethernet, and the AI is live on your network.
Our team handles cabling, network setup, PLC and SCADA integration, operator training and 24×7 remote monitoring. Training and testing run on the server inside your plant, so part images stay on site. For a scope matched to your stations, request a turnkey quote.
Ship, network and data
Server installed. Cameras, re-inspection results and line data connected. Baselines recorded and a first golden set agreed with quality.
Model training and pilot
Triggers switched on for the pilot stations. One full retrain run end to end, with a shadow run and a scorecard.
Go-live and training
Cadence live across the agreed stations. Line leads, labellers and quality engineers trained. 24×7 remote monitoring begins.
Frequently Asked Questions
How often should an automotive vision model be retrained?
There is no fixed number. Check image health every shift, review drift every week, score an audit sample every month and revalidate every quarter. Retrain when a trigger is met or a planned change is coming. A steady station may need it twice a year; one with frequent variant changes, far more often.
What is model drift?
It is the gap that opens between what the model was trained on and what the camera sees today. The model has not changed, but lighting, parts, suppliers and tooling have. Drift shows up as more false rejects, more operator overrides, or, worst of all, defects that pass unnoticed.
Can the model retrain itself automatically?
Collection, training and testing can all run automatically. Release should not. A new model changes what the plant accepts and rejects, so a person from quality should review the scorecard and approve it. Models that label their own training data tend to repeat their own mistakes.
How many images does a retrain need?
Fewer than most people expect, if they are the right ones. A few hundred correctly labelled examples of what the model gets wrong often do more than tens of thousands of easy cases. A new colour or variant needs enough samples to cover its normal range of appearance.
What is a golden set?
A locked collection of reference images with labels that quality has approved: clear passes, clear defects and borderline cases. It is never used for training, only for testing. Every model version is scored against it, so results can be compared fairly from one version to the next.
Does a retrained model need customer approval?
It depends on your customer's requirements and on whether the station checks a special characteristic. Many plants handle a retrain inside their own change control, with records kept. Where a control plan or customer rule calls for notice, give it before release. Check before you assume either way.
How long does it take to go live?
Six to twelve weeks from delivery. We need a place for the server with power and Ethernet, access to the camera images and re-inspection results, and time with your quality engineers and inspectors. To check your set-up first, contact our team.
Bring One Station. Leave With Its Cadence.
In thirty minutes we take one inspection station and write down its triggers, its sample routine and its release gate. You keep the one-page plan whether or not you go further with iFactory.
- 1Reject and override rates for the last three months
- 2The date the current model went live
- 3Changes since then: colours, variants, suppliers
- 4Your written defect standard for the station
- 5Any escapes traced back to it
