Applied machine learning to large-scale manufacturing time-series data at LG Display — building defect detection models, production forecasting, and an explainability workflow using SHAP that enabled engineers to act on model outputs rather than just receive alerts.
ML models for early defect detection and production forecasting on display panel and glass substrate manufacturing lines. The core contribution wasn't model complexity — tree-based methods worked well — but rather the end-to-end workflow from raw sensor data to actionable root-cause explanations that engineers and process stakeholders could use directly.
Tech stack: Python · LightGBM · Random Forest · SHAP · Pandas · Scikit-learn
Trained supervised models on high-dimensional manufacturing time-series (many sensors, mixed sampling rates, significant missingness). Key challenges were class imbalance (defects are rare) and non-stationarity from equipment aging, maintenance events, and recipe changes.
Used LightGBM and Random Forest on tabularized time-series features with time-aware validation splits to prevent leakage — a common pitfall in manufacturing ML where temporal ordering is critical.
Developed models forecasting production outcomes for display panels and glass substrates, supporting resource planning and operational efficiency decisions.
The most impactful part of this work was making predictions actionable. Applied SHAP (global + local explanations) to:
Positioned explainability not as an afterthought but as the primary handoff mechanism from ML output → engineering action. This was the main driver of stakeholder trust and adoption.