Predictive Quality for Injection Molding
Machine learning on line-level IoT telemetry that predicts defect probability before parts reach inspection — and surfaces the parameter drift driving each call.
Results
The Problem
Defects in high-volume molding surface at inspection — after the material, the cycle time, and the machine hours are already spent. By the time a scrap trend is visible on a control chart, a shift's worth of parts may already be compromised. Process engineers know the causes are in the parameters, but the relationships are nonlinear and interact across dozens of variables.
What We Build
Machine-learning models trained on line-level IoT telemetry — barrel and mold temperatures, injection pressure and velocity profiles, hold time, cooling rate, cycle time, material lot — to predict defect probability at the shot or batch level, before parts reach inspection.
Critically, the models are built to be explainable: rather than a black-box score, engineers get the specific parameter drift driving the elevated risk, so the response is a machine adjustment rather than a guess.
Outcome
- ◆Defect prediction ahead of inspection, shifting quality control from detection to prevention
- ◆Root-cause attribution surfaced per prediction, not just a flag
Techniques
- ◆Gradient-boosted classifiers
- ◆Anomaly detection
- ◆SHAP-based attribution
- ◆Time-series feature engineering on sensor streams