As smart manufacturing environments become more complex and automation levels advance, conventional rule-based anomaly detection methods struggle to capture multi-stage delay anomalies effectively. To address this gap, we propose an autoencoder-based anomaly detection framework with a dual-decoder architecture that simultaneously reconstructs process-level sojourn times and predicts routing-level lead times. Decoder 1 restores each process’s sojourn time to compute a process anomaly score (PAS), employs a dual control chart using the interquartile range (IQR) technique for visual detection of process-level anomalies, and triggers an early alert as soon as the cumulative PAS exceeds a predefined threshold—providing prompt warning at the routing level. Decoder 2 ingests the full sequence of process times to predict real-time lead times, enabling quantitative assessment of both short- and long-term delay risks. By combining fine-grained, process-level anomaly detection with routing-level lead-time prediction, our framework delivers precise, rapid monitoring of deviations at both levels. Experiments on real-time production data from an operational manufacturing site confirm that the model achieves excellent lead-time prediction accuracy for lots following the standard process flow.
Park et al. (Thu,) studied this question.