Predictive quality: how AI spots off-spec batches before they run
How predictive quality uses machine learning to score defect risk before production runs, in any plant, using data you already have.
In most factories, quality is checked after a product or batch is made. If dimensions, strength, colour or composition are out of specification, the material, machine time and energy are already spent. The batch is reworked, downgraded or scrapped.
Predictive quality turns this around. It uses machine learning to estimate the risk that a planned batch will go off-spec before it runs, and suggests the process change that lowers that risk.
How predictive quality works
Every batch or order leaves a trail of data: recipe or bill of materials, raw material lots, line, operator, shift, temperatures, pressures, speeds, throughput, ambient conditions and, finally, the QC result. A machine learning model learns which combinations of these variables tend to come before off-spec results.
- Collect: link each batch's process settings and inputs to its QC outcome.
- Train: a classification model learns the patterns behind pass and fail.
- Score: each planned batch gets a risk score, for example 12% or 85%.
- Explain: the model shows which variables push the risk up, such as a temperature that runs high for a certain product.
- Recommend: it tests small setting changes and suggests the one that lowers risk most.
What results look like
Accuracy is usually reported as AUC (area under the ROC curve). A score of 0.5 is a coin flip and 1.0 is perfect. On our simulated pilot-profile factory, DMIS reaches an AUC of 0.73 on batches it has never seen, and catches 34% of off-spec batches in the riskiest 20%. In one example, lowering one process temperature from 215.4 to 209.9 °C cut a batch's predicted risk from 97% to 39%. Real-plant accuracy is always confirmed in a pilot, because every plant is different.
A model does not need to be perfect to pay back. If quality engineers check the riskiest fifth of batches first, they catch a large share of problems with a fraction of the effort.
What data you need
- Batch records with recipe, line and date
- Key process settings per batch, even if recorded by hand
- QC results linked to each batch
- Ideally, raw material lot numbers
A few months of history is usually enough to start. You do not need new sensors: logbooks and spreadsheets can be imported. See our guide to moving from paper logbooks to digital records.
Common mistakes
- Unlinked data: QC results that cannot be matched to the batch they came from.
- Black-box models: engineers will not act on a score they cannot understand. Explanations matter.
- No feedback loop: models should retrain as new batches are recorded, or accuracy drifts.
- Ignoring the people: pair recommendations with your experts' know-how through playbooks and training.
Getting started
Predictive quality is one of the core modules of DMIS and part of our AI and machine learning solutions. Request a demo to see it on a factory profile like yours.
