outlier detection

**Outlier Detection** in semiconductor data analysis is the **identification and handling of data points that are significantly different from the majority** — distinguishing real process excursions (which need investigation) from measurement errors or artifacts (which need removal). **Key Outlier Detection Methods** - **Statistical**: Z-score ($|z| > 3$), IQR method ($< Q_1 - 1.5 cdot IQR$ or $> Q_3 + 1.5 cdot IQR$), Grubbs' test. - **Multivariate**: Mahalanobis distance, PCA residuals (Q-statistic), robust covariance. - **ML-Based**: Isolation forest, Local Outlier Factor (LOF), autoencoders. - **Domain-Specific**: EE box (Equipment Engineering spec limits), out-of-control SPC rules. **Why It Matters** - **Data Quality**: Outliers can corrupt statistical models, virtual metrology, and SPC charts. - **Root Cause**: Some outliers indicate real process issues — automatic removal without investigation risks missing critical signals. - **Balanced Approach**: Industrial practice flags outliers for review rather than automatic deletion. **Outlier Detection** is **separating signal from noise** — identifying abnormal data points that need investigation or removal for reliable analysis.