摘要
•Unsupervised monitoring enables real-time decisions when fault labels are scarce.•Deseasonalization and autoencoders yield machine-specific 0-100 health scores.•Expert-weighted errors pinpoint abnormal parameters using control charts.•Two-level warning logic supports early caution and urgent intervention triggers.•A fab case study demonstrates actionable health trends for maintenance decisions.
Semiconductor fabs rely on air compressors as critical utilities whose unexpected failures can disrupt production and induce costly tool downtime. Yet air-compressor condition data are typically high-dimensional, non-stationary, and largely unlabeled, making supervised fault classification difficult in practice. This paper presents a deployable prognostics and health management (PHM) framework for real-time predictive maintenance of semiconductor air compressors using unsupervised deep learning. The approach integrates (i) temperature deseasonalization to reduce environmental confounding, (ii) a deep autoencoder trained on historical healthy behavior, and (iii) an expert-informed, weighted reconstruction-error health indicator that is mapped to an interpretable 0-100 health score. A two-level warning logic, combining a moving-average health drop criterion with absolute health thresholds, supports both early cautionary alerts and urgent intervention flags. A real-world case study using multi-year sensor data from seven compressors in a semiconductor plant demonstrates how the proposed health score and visualization dashboard enable continuous monitoring, anomaly localization by contributing parameters, and maintenance decision support. For a representative compressor, the one-week moving-average health score dipped to just above 80 before the 2020 preventive-maintenance date and stayed above 85 before the 2022 date, suggesting that major maintenance could plausibly be deferred when health remains stable. Overall, the proposed framework provides a practical, interpretable, label-free alternative to fault-label-dependent PHM methods for real-time industrial operation where fault labels are scarce.