Machine Vision & Industrial Inspection AI worked example

Model Drift Exposure at 68% current model accuracy on recent inspection data: a worked example in machine vision & industrial inspection ai

This worked example runs the model drift exposure numbers for a tougher week than the baseline: 68% current model accuracy on recent inspection data instead of the typical 94.5%. 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.

The inputs for this scenario

  • Current model accuracy on recent inspection data: 68 % (the input this scenario stresses; the baseline uses 94.5)
  • Retraining trigger threshold (minimum acceptable accuracy): 90 % (held at the documented default)
  • Target accuracy at original training/validation: 98 % (held at the documented default)

Working through the calculation

  • The calculation starts from the formula this tool documents: Accuracy status: model is within acceptable window when current accuracy is above the retraining threshold.
  • Model within operating accuracy window works out to 0 outside at these inputs, and this is the headline figure for the scenario.
  • Accuracy margin to nearest threshold works out to -22 value at these inputs.
  • Retraining trigger threshold works out to 90 value at these inputs.
  • Target accuracy at original training works out to 98 value at these inputs.

How this compares with the baseline

  • Against the tool's baseline example, where current model accuracy on recent inspection data sits at 94.5% and the headline result is 1 inside, this scenario comes in 100% below the baseline at 0 outside.
  • 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. A result at this level usually justifies acting on the stressed input before touching anything else, because every other figure in the table is downstream of it.

Results at a glance

  • Model within operating accuracy window: 0 outside (headline result)
  • Accuracy margin to nearest threshold: -22 value
  • Retraining trigger threshold: 90 value
  • Target accuracy at original training: 98 value

Run it with your numbers

  • To rerun this with your own numbers, open the live Model Drift Exposure calculator, set current model accuracy on recent inspection data to your actual value, and adjust the remaining inputs to match your operation.

Last reviewed 2026-05-12.