Machine Vision & Industrial Inspection AI calculator

AI Model Drift Exposure Calculator

Model Drift Exposure tells you whether a deployed machine-vision inspection model is still performing inside its acceptable accuracy window, and how much headroom you have before it crosses the retraining trigger. Quality engineers and ML-ops teams running automated optical inspection (AOI) use it to decide when to schedule a retrain instead of waiting for a customer escape. Vision models degrade silently as lighting drifts, new defect modes appear, or the part mix shifts, so a single accuracy number means little without the threshold it sits above. This calculator turns three numbers into a clear in-window / out-of-window status plus the exact accuracy margin you have left.

What this calculator does

  • Check whether an AI inspection model's current accuracy is within its acceptable operating window by comparing current performance against the minimum acceptable accuracy threshold and the original trained accuracy.
  • Use it when reviewing AI inspection model performance during production and you need a quick check of whether the current accuracy is within operating spec or has drifted enough to trigger retraining.
  • It computes whether current model accuracy sits above the retraining trigger threshold and the percentage-point margin between them.

Formula used

  • Accuracy status: model is within acceptable window when current accuracy is above the retraining threshold
  • Margin to retraining trigger = current accuracy - retraining threshold

Inputs explained

  • Current model accuracy on recent inspection data:
  • Retraining trigger threshold (minimum acceptable accuracy):
  • Target accuracy at original training/validation:

How to use the result

  • Use it at every periodic model-health review or after any line change (new fixturing, lens, lighting, or part variant) that could shift the image distribution.
  • Aggregate accuracy can stay above threshold while a single critical defect class collapses, so pair this with per-class metrics before clearing a model.

Common questions

  • How do you calculate model drift exposure? Subtract the retraining threshold from current accuracy to get the margin. With 94.5% current accuracy and a 90% trigger, the margin is 3.5 points, and because current accuracy is above the trigger the model is inside its operating window.
  • What is a good accuracy margin before retraining? There is no universal number, but most AOI teams want at least 2-4 points of margin above the trigger so normal day-to-day variation does not flip the model out of window. A 3.5-point margin like the example is comfortable but worth watching.
  • What causes machine-vision model drift? Lighting and ambient changes, new or evolving defect signatures, camera or lens aging, fixturing and part-presentation shifts, and changes in upstream process that alter surface finish. Any of these can pull accuracy down toward the trigger even with no code change.
  • Model drift vs data drift, what's the difference? Data drift is a change in the input images (new lighting, new part variant); model drift is the resulting drop in prediction accuracy. Data drift is the cause, the accuracy loss this tool tracks is the effect.
  • What should the retraining trigger threshold be set to? Set it at the lowest accuracy your quality plan can tolerate given downstream containment, then add margin. If 90% accuracy is the minimum that keeps escapes within your PPM target, 90 is your trigger, as in the default.

Last reviewed 2026-05-12.