Predictive Quality
Machine learning experiments using upstream manufacturing variables to predict downstream quality outcomes.
Evolution
This work now sits inside a wider area: Industrial AI IOT. Predictive quality on historical data is the foundation — the next phase connects live industrial data, edge signals and connected equipment, so the same models run against streams instead of extracts.
Problem
Quality issues are often detected after value has already been added, when the cost of correction is highest and the causal trail has gone cold.
Hypothesis
Upstream process variables carry enough signal to flag elevated quality risk before the downstream inspection step.
Architecture
- Aligned time-series of upstream process variables.
- Joined downstream quality outcomes as labels.
- Feature engineering around process windows and transitions.
- Model evaluation focused on early-warning value rather than raw accuracy.
Experiment
Whether upstream variables provide a usable early-warning signal for a specific downstream defect mode.
Observations
- Data alignment between systems consumed most of the effort.
- Label quality is the limiting factor — inconsistent defect coding blurs any signal.
- No performance claims yet; the evaluation protocol is still being fixed.
Technologies
- Time-series feature engineering
- Gradient boosting
- Classification models
- Data alignment
Next steps
- Stabilise defect labelling before drawing conclusions.
- Define the operational decision the prediction is meant to support.