anomaly detection

**Anomaly Detection** in semiconductor manufacturing is the **identification of abnormal process conditions, wafer measurements, or equipment behaviors** — using statistical, model-based, or ML methods to flag observations that deviate significantly from normal operating patterns. **Key Anomaly Detection Approaches** - **Multivariate SPC**: Hotelling T² and Q-statistics detect multivariate outliers. - **Isolation Forest**: Randomly partitions data and measures how quickly observations are isolated. - **Autoencoders**: Neural networks trained to reproduce normal data — anomalies have high reconstruction error. - **One-Class SVM**: Learns the boundary of normal operation and flags points outside it. **Why It Matters** - **Excursion Detection**: Catches process excursions before they produce wafers out of spec. - **Predictive Maintenance**: Detects early equipment degradation signatures before failure. - **Rare Events**: Anomaly detection is more practical than classification for rare failure modes (limited examples). **Anomaly Detection** is **the automatic alarm system** — continuously monitoring process data to flag anything that doesn't look normal.