Unsupervised Prognostics and Health Management for Semiconductor Manufacturing Air Compressors: An LSTM Autoencoder Framework with Seasonal Decomposition and Real-Time Anomaly Detection
Main article
Abstract
Semiconductor manufacturing relies on a complex ecosystem of critical auxiliary equipment, among which air compressors play an indispensable role in providing clean, dry compressed air for pneumatic actuation, purging, and process gas blending. Unplanned compressor failures cause cascading production disruptions that can idle entire fabrication lines, incurring costs estimated at USD 1–5 million per hour in advanced technology nodes. This paper presents a comprehensive Prognostics and Health Management (PHM) framework for semiconductor facility air compressors, leveraging unsupervised machine learning to circumvent the chronic shortage of labeled fault data that plagues supervised approaches in industrial settings. The proposed system employs a stacked Long Short-Term Memory Autoencoder (LSTM-AE) that learns normal operating patterns from unlabeled sensor streams spanning eight monitored parameters per compressor unit. A Seasonal-Trend decomposition using Loess (STL) preprocessing stage explicitly removes periodic operational cycles—arising from production shift schedules and environmental temperature fluctuations—before computing reconstruction errors, substantially reducing false alarm rates compared to decomposition-naive baselines. Health scores derived from normalized reconstruction errors trigger a three-tier alert system (normal, warning, critical) with dynamically adjusted thresholds that adapt to gradual baseline drift. The system was validated in collaboration with a major Taiwanese semiconductor manufacturer operating 24 air compressor units across three fabrication buildings. Over an 18-month evaluation period, the PHM framework achieved a 96.3% anomaly detection AUC, a mean time-to-detection of 18.7 hours before confirmed failure, and a 34.2% reduction in unplanned downtime costs. A custom web-based visualization dashboard providing intuitive health score trends and parameter heatmaps was deployed for on-site maintenance personnel, demonstrating practical adoption success in a high-stakes production environment.
