Manufacturing environments implementing AI-based predictive maintenance face critical constraints imposed by limited data and label availability. This study examines two contrasting industrial cases to determine how anomaly detection model selection should align with data conditions. Case A (Direct Loom, Unsupervised): From 636,540 rows of programmable logic controller (PLC) timeseries data collected over nine days without anomaly labels, a six-stage preprocessing pipeline and sliding-window feature extraction (W = 64, stride = 16) produced 160-dimensional vectors, which were fed into an AutoEncoder (160-80-40-80-160). Five-fold time-series cross-validation yielded an anomaly detection rate (ADR) of 1.0% (±0.0%) with a mean false positive rate (FPR) of 0.2%. This FPR was well below the theoretical upper bound of 1.0% implied by the p99 threshold definition, indicating that the model discriminates beyond random chance. The inference time of 0.43 s remained well below the PLC sampling interval of 1.14 s. Case B (Press Induction Motor, Supervised): An XGBoost classifier with class-imbalance compensation, trained on 717 vibration feature records collected over 61 days with anomaly labels, achieved an F1-score of 0.870 and a ROC-AUC of 0.847. These findings support a practical stepwise strategy: begin with an unsupervised AutoEncoder to log anomalies, and then progressively transition to supervised models as labeled data accumulates.
