Model drift and retraining: keeping an inspection model accurate over time.

What is model drift on an inspection line?

Model drift is the gradual gap that opens between the parts a model was trained on and the parts it sees today. The model has not changed; the line has. As the incoming parts move away from the training set, the model's picture of what counts as normal stops matching the line's real normal, and accuracy erodes without anyone touching the software.

For an inspection model the practical worry is data drift: the distribution of good-part images shifts, so the same anomaly threshold that fit last quarter now sits in the wrong place. The model still runs at the same speed and gives a confidence score for every part, but that score is being measured against a picture of normal that has quietly gone stale.

Drift is a line reality, not a defect in the model. Any inspector trained on a fixed set of examples will drift if the process underneath it changes, which on a real production line it always does over time. The question is not whether drift happens but how fast you notice it and how quickly you can retrain.

What actually causes an inspection model to drift?

On a real line, drift comes from physical change in the process, not from the algorithm. A handful of sources account for most of it, and each one shifts the appearance of a good part rather than introducing a new defect.

The common triggers are material and supplier changes (a new resin batch, a different coating, a fresh coil with a slightly different finish), tool and mold wear (surface texture, flash and burr on a good part all evolve as tooling ages), and lighting decay (LED output falls over thousands of hours and seasonal ambient light changes what the sensor records). Camera or fixture nudges from cleaning and maintenance add framing shifts on top.

None of these are failures you can write a rule against in advance, which is exactly why they erode a learned model. As an industry pattern, drift from wear and lighting tends to build slowly over weeks or months, while a lot or supplier change can shift the picture of normal overnight. Treat the timescale as process-specific: the only reliable way to know your line's drift rate is to watch the numbers it produces.

How do you know an inspection model is drifting?

The first visible symptom of drift is usually a rising false-reject rate: the line starts flagging good parts it used to pass. Because the model's idea of normal has narrowed or shifted, more genuinely good parts now fall outside it, and operators feel the pain as a climbing reject bin and more manual re-checks.

A second signal sits in the confidence scores. When a good part that used to clear the threshold comfortably now scores near the borderline, the distribution of scores is sliding, and that slide is an early warning before the reject rate itself becomes obvious. Watching the trend of scores and reject rate over time, not a single shift, is what separates real drift from normal noise.

The reason to care is asymmetric. A rising false-reject rate is annoying and costly, but the more dangerous form of drift lets defects start scoring as normal, which raises the escape rate instead. That second failure mode is the one to design your monitoring around, because a missed defect leaves the plant while a false reject only fills a bin.

What does the retraining loop look like?

The retraining loop is deliberately short so drift never gets far ahead of the line. It has four steps: detect the drift, recapture current good parts, retrain the model, and redeploy it to the unit. Kept as a routine, the loop turns drift from a slow accuracy leak into scheduled maintenance.

Recapture is the step that keeps the loop cheap. Because the model trains on good parts only, you do not collect or stage defects to retrain; you photograph about 20 good parts that represent the line's current normal, including the new lot or the current tool condition. 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, and the retraining loop runs on the same unit that does the inspection, so recapture uses the production camera and lighting rather than a separate rig.

Training then completes in under 48 hours, and the updated model is validated on sample parts and redeployed to the unit. One useful detail: the unit runs a hybrid of classical computer vision and AI inference. The deterministic measurement side (a fixed dimension in millimetres, a barcode) does not drift the way a learned anomaly model does, so retraining is usually about refreshing the learned picture of normal, not re-engineering the whole check.

Can you retrain an inspection model on an air-gapped line?

Yes. Because the unit runs and trains at the edge, retraining does not require a cloud connection or shipping images off the line. Model updates move by USB stick, so a factory that keeps its inspection cell air-gapped can still refresh a drifting model without opening a network path to the outside.

That matters for two reasons. First, raw part imagery often is sensitive process data, and keeping capture, training and redeployment on the line means the images never leave the plant. Second, an air-gapped update path removes the excuse to let drift accumulate: you do not have to schedule a connectivity window or a vendor visit to retrain, because the loop is a local operation your integrator or a trained operator can run. For the full case on why edge deployment keeps data and uptime on your side, see the sibling post on edge versus cloud visual inspection; for what makes the recapture images themselves good enough to retrain on, see the companion post on training-data quality versus quantity.

This post is a spoke of the pillar guide on AI visual inspection; to see the on-device unit that runs the recapture and retrain loop, take the system overview.