Production ML & Shadow Inference for Manufacturing

Problem

Defect modeling on live equipment required more than training a classifier: raw PLC snapshots had to be aligned with machine modes, temporal context and trustworthy defect labels before any model could be evaluated safely.

What I built

• Prepared time-windowed datasets from ClickHouse telemetry and investigated leakage, defect precursors and atypical operating modes.
• Trained and compared supervised models with LightGBM, XGBoost and CatBoost, alongside rule-based experiments.
• Implemented online inference with historical windows and shadow-mode validation that observed the production process without controlling the machine.

Architecture

PLC history → dataset and labels → feature/window preparation → model training → online inference → shadow validation

Inference outputs and supporting signals were logged so predictions could be reviewed against subsequent machine behavior.

Scale

The largest confirmed experiment processed 643,924 snapshots and 6,299 PLC signals. The best reported experiment reached PR-AUC up to 0.971; hundreds of experimental models were evaluated during the research.

Result

• Produced a reproducible path from industrial telemetry to live model validation.
• Separated offline metrics from evidence collected on real equipment, without claiming autonomous control or prevented defects.