02 · Manufacturing
Connect engineering, manufacturing, quality and planning through shared representations.
The factory becomes an AI research environment when product structures, recipes, sensor histories, images, maintenance states, schedules and market demand can be combined without collapsing them into one undifferentiated dataset.
Extreme gradient boosting
Strong baseline for nonlinear relationships in process, ingredient and machine-state data.
Computer vision
Converts each cookie into a measurable quality object.
Digital twins
Represent the physical process as an executable scenario environment.
Optimisation
Turns predictions into feasible production sequences and capacity choices.
Engineering knowledge services
Connect recipes, specifications, deviations and lessons learned.
Human–AI control
Separates prediction, recommendation, approval and automatic execution.
Programmatic pipeline
Example: predictive quality is a data pipeline, not just a model.
Ingesttemperature, speed, moisture, recipe, flour lot
Validatetimestamps, units, missingness, sensor plausibility
Featurerolling means, gradients, batch context
PredictXGBoost quality and defect risk
Explainfeature contribution + uncertainty
Actoperator advice or optimisation request
Feature vector
batch_vector = [
flour_moisture,
mixer_energy_mean,
oven_z1_temp,
oven_z2_temp,
oven_z3_temp,
belt_speed,
ambient_humidity,
previous_batch_changeover
]
Decision separation
| Layer | Role |
|---|---|
| Model | Predict defect risk |
| Policy | Decide whether intervention is allowed |
| Optimiser | Find feasible parameter change |
| Human | Approve high-risk changes |
| MES | Execute authorised setting |