Internship Report: Improvement of Anomaly Detection Indicators for Machine Sensors, ST Microelectronics, Rousset, France

This work focuses on enhancing anomaly detection in semiconductor manufacturing equipment within the Fault Detection and Classification (FDC) framework. It addresses limitations in alarm accuracy and anomaly sensitivity by improving reference data selection and optimizing statistical methods across diverse variable types. The project advanced through four phases: (1) a literature review on robust alternatives to traditional z-scores, including MAD, Sn estimators, and percentile-based approaches; (2) an algorithm effectiveness study showing that strict reference dataset selection substantially reduces false positives; (3) a variable characterization analysis establishing KPIs for stability, volatility, maintenance correlations, and regime-switching behaviors; and (4) the development of two complementary solutions—a normalization strategy correcting predictable maintenance-related drift, and a strategic sampling methodology combining hierarchical clustering with cross-variable selection. The results demonstrate that rigorous reference construction and adaptive statistical techniques can significantly improve anomaly detection performance in semiconductor manufacturing.